model : add support for ERNIE 4.5 0.3B model (#14408)

Add Day-0 support for Baidu ERNIE 4.5 0.3B model.

Signed-off-by: Weizhao Ouyang <weizhao.ouyang@arm.com>

.github/workflows/build-cmake-pkg.yml

@@ -0,0 +1,51 @@
+name: Build relocatable cmake package
+on:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  linux:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y build-essential tcl
+
+      - name: Build
+        run: |
+          PREFIX="$(pwd)"/inst
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX" \
+                -DLLAMA_CURL=OFF -DLLAMA_BUILD_TESTS=OFF -DLLAMA_BUILD_TOOLS=OFF \
+                -DLLAMA_BUILD_EXAMPLES=OFF -DCMAKE_BUILD_TYPE=Release
+          cmake --build build --config Release
+          cmake --install build --prefix "$PREFIX" --config Release
+
+          export LLAMA_CONFIG="$PREFIX"/lib/cmake/llama/llama-config.cmake
+          tclsh <<'EOF'
+          set build(commit)  [string trim [exec git rev-parse --short HEAD]]
+          set build(number)  [string trim [exec git rev-list  --count HEAD]]
+          set build(version) "0.0.$build(number)"
+
+          set llamaconfig [read [open "$env(LLAMA_CONFIG)" r]]
+          set checks [list "set\\(LLAMA_VERSION     \\s+$build(version)\\)" \
+                           "set\\(LLAMA_BUILD_COMMIT\\s+$build(commit)\\)" \
+                           "set\\(LLAMA_BUILD_NUMBER\\s+$build(number)\\)"]
+
+          puts -nonewline "Checking llama-config.cmake version... "
+          foreach check $checks {
+              if {![regexp -expanded -- $check $llamaconfig]} {
+                  puts "\"$check\" failed!"
+                  exit 1
+              }
+          }
+          puts "success."
+          EOF
+
+          cd examples/simple-cmake-pkg
+          cmake -S . -B build -DCMAKE_PREFIX_PATH="$PREFIX"/lib/cmake
+          cmake --build build

.github/workflows/build.yml

@@ -5,10 +5,43 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]
+
   pull_request:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/build.yml', '.github/workflows/build-linux-cross.yml', '**/CMakeLists.txt', '**/.cmake', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal', '**/*.comp']
+    paths: [
+      '.github/workflows/build.yml',
+      '.github/workflows/build-linux-cross.yml',
+      '.github/workflows/build-cmake-pkg.yml',
+      '**/CMakeLists.txt',
+      '**/.cmake',
+      '**/*.h',
+      '**/*.hpp',
+      '**/*.c',
+      '**/*.cpp',
+      '**/*.cu',
+      '**/*.cuh',
+      '**/*.swift',
+      '**/*.m',
+      '**/*.metal',
+      '**/*.comp'
+    ]
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
@@ -478,6 +511,9 @@ jobs:
   build-linux-cross:
     uses: ./.github/workflows/build-linux-cross.yml
 
+  build-cmake-pkg:
+    uses: ./.github/workflows/build-cmake-pkg.yml
+
   macOS-latest-cmake-ios:
     runs-on: macos-latest
 
@@ -628,7 +664,7 @@ jobs:
           ./build-xcframework.sh
 
   windows-msys2:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       fail-fast: false
@@ -678,27 +714,33 @@ jobs:
             cmake --build build --config ${{ matrix.build }} -j $(nproc)
 
   windows-latest-cmake:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
       SDE_VERSION: 9.33.0-2024-01-07
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
         include:
           - build: 'cpu-x64 (static)'
+            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
+            arch: 'x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
+            arch: 'x64'
             defines: '-DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_VULKAN=ON'
           - build: 'llvm-arm64'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON'
           - build: 'llvm-arm64-opencl-adreno'
+            arch: 'arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/opencl-arm64-release" -DGGML_OPENCL=ON -DGGML_OPENCL_USE_ADRENO_KERNELS=ON'
          # - build: 'kompute-x64'
+         #   arch: 'x64'
          #   defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON'
 
     steps:
@@ -736,7 +778,7 @@ jobs:
         id: get_vulkan
         if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -769,6 +811,8 @@ jobs:
       - name: libCURL
         id: get_libcurl
         uses: ./.github/actions/windows-setup-curl
+        with:
+          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
 
       - name: Build
         id: cmake_build
@@ -789,7 +833,7 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'llvm-arm64' && matrix.build != 'llvm-arm64-opencl-adreno' }}
+        if: ${{ matrix.arch == 'x64' }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -894,7 +938,7 @@ jobs:
           cmake --build build --config Release
 
   windows-latest-cmake-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -928,7 +972,7 @@ jobs:
 
   windows-latest-cmake-hip:
     if: ${{ github.event.inputs.create_release != 'true' }}
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     steps:
       - name: Clone

.github/workflows/release.yml

@@ -235,7 +235,7 @@ jobs:
           name: llama-bin-ubuntu-vulkan-x64.zip
 
   windows-cpu:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     strategy:
       matrix:
@@ -271,7 +271,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch }}
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
             -DGGML_NATIVE=OFF ^
@@ -288,7 +288,7 @@ jobs:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
-          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.42.34433\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
+          Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
       - name: Upload artifacts
@@ -298,11 +298,11 @@ jobs:
           name: llama-bin-win-cpu-${{ matrix.arch }}.zip
 
   windows:
-    runs-on: windows-latest
+    runs-on: windows-2025
 
     env:
       OPENBLAS_VERSION: 0.3.23
-      VULKAN_VERSION: 1.4.309.0
+      VULKAN_VERSION: 1.4.313.2
 
     strategy:
       matrix:
@@ -332,7 +332,7 @@ jobs:
         id: get_vulkan
         if: ${{ matrix.backend == 'vulkan' }}
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -448,7 +448,7 @@ jobs:
           name: cudart-llama-bin-win-cuda-${{ matrix.cuda }}-x64.zip
 
   windows-sycl:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     defaults:
       run:
@@ -520,7 +520,7 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-hip:
-    runs-on: windows-latest
+    runs-on: windows-2022
 
     strategy:
       matrix:

CMakeLists.txt

@@ -95,7 +95,7 @@ endif()
 if (NOT DEFINED LLAMA_BUILD_COMMIT)
     set(LLAMA_BUILD_COMMIT        ${BUILD_COMMIT})
 endif()
-set(LLAMA_INSTALL_VERSION 0.0.${BUILD_NUMBER})
+set(LLAMA_INSTALL_VERSION 0.0.${LLAMA_BUILD_NUMBER})
 
 # override ggml options
 set(GGML_ALL_WARNINGS   ${LLAMA_ALL_WARNINGS})

ci/run.sh

@@ -779,7 +779,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?\thi\nwhat is panda?\tit's a bear\nwhat is panda?\tThe giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

common/arg.cpp

@@ -2706,6 +2706,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.embd_sep = value;
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
+    add_opt(common_arg(
+        {"--cls-separator"}, "STRING",
+        "separator of classification sequences (default \\t) for example \"<#seq#>\"",
+        [](common_params & params, const std::string & value) {
+            params.cls_sep = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--host"}, "HOST",
         string_format("ip address to listen, or bind to an UNIX socket if the address ends with .sock (default: %s)", params.hostname.c_str()),
@@ -3210,6 +3217,32 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.speculative.model.path = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_MODEL_DRAFT"));
+    add_opt(common_arg(
+        {"-ctkd", "--cache-type-k-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for K for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_k)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_k = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_K_DRAFT"));
+    add_opt(common_arg(
+        {"-ctvd", "--cache-type-v-draft"}, "TYPE",
+        string_format(
+            "KV cache data type for V for the draft model\n"
+            "allowed values: %s\n"
+            "(default: %s)",
+            get_all_kv_cache_types().c_str(),
+            ggml_type_name(params.speculative.cache_type_v)
+        ),
+        [](common_params & params, const std::string & value) {
+            params.speculative.cache_type_v = kv_cache_type_from_str(value);
+        }
+    ).set_env("LLAMA_ARG_CACHE_TYPE_V_DRAFT"));
 
     add_opt(common_arg(
         {"-mv", "--model-vocoder"}, "FNAME",

common/common.cpp

@@ -706,11 +706,17 @@ bool fs_validate_filename(const std::string & filename) {
         // disable C++17 deprecation warning for std::codecvt_utf8
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
+
         std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
         filename_utf32 = converter.from_bytes(filename);
@@ -1284,6 +1290,9 @@ std::vector<llama_token> common_tokenize(
     int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
     n_tokens = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
+    if (n_tokens == std::numeric_limits<int32_t>::min()) {
+        throw std::runtime_error("Tokenization failed: input text too large, tokenization result exceeds int32_t limit");
+    }
     if (n_tokens < 0) {
         result.resize(-n_tokens);
         int check = llama_tokenize(vocab, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);

common/common.h

@@ -199,6 +199,9 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
 
+    ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
+    ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
+
     struct cpu_params cpuparams;
     struct cpu_params cpuparams_batch;
 
@@ -355,6 +358,7 @@ struct common_params {
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embeddings
+    std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
     int32_t port           = 8080;         // server listens on this network port

common/json-schema-to-grammar.cpp

@@ -41,49 +41,6 @@ static std::string build_repetition(const std::string & item_rule, int min_items
     return result;
 }
 
-/* Minimalistic replacement for std::string_view, which is only available from C++17 onwards */
-class string_view {
-    const std::string & _str;
-    const size_t _start;
-    const size_t _end;
-public:
-    string_view(const std::string & str, size_t start = 0, size_t end  = std::string::npos) : _str(str), _start(start), _end(end == std::string::npos ? str.length() : end) {}
-
-    size_t size() const {
-        return _end - _start;
-    }
-
-    size_t length() const {
-        return size();
-    }
-
-    operator std::string() const {
-        return str();
-    }
-
-    std::string str() const {
-        return _str.substr(_start, _end - _start);
-    }
-
-    string_view substr(size_t pos, size_t len = std::string::npos) const {
-        return string_view(_str, _start + pos, len == std::string::npos ? _end : _start + pos + len);
-    }
-
-    char operator[](size_t pos) const {
-        auto index = _start + pos;
-        if (index >= _end) {
-            throw std::out_of_range("string_view index out of range");
-        }
-        return _str[_start + pos];
-    }
-
-    bool operator==(const string_view & other) const {
-        std::string this_str = *this;
-        std::string other_str = other;
-        return this_str == other_str;
-    }
-};
-
 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();
@@ -112,14 +69,14 @@ static void _build_min_max_int(int min_value, int max_value, std::stringstream &
         }
         out << "}";
     };
-    std::function<void(const string_view &, const string_view &)> uniform_range =
-        [&](const string_view & from, const string_view & to) {
+    std::function<void(const std::string_view &, const std::string_view &)> uniform_range =
+        [&](const std::string_view & from, const std::string_view & to) {
             size_t i = 0;
             while (i < from.length() && i < to.length() && from[i] == to[i]) {
                 i++;
             }
             if (i > 0) {
-                out << "\"" << from.substr(0, i).str() << "\"";
+                out << "\"" << from.substr(0, i) << "\"";
             }
             if (i < from.length() && i < to.length()) {
                 if (i > 0) {

convert_hf_to_gguf.py

@@ -310,6 +310,8 @@ class ModelBase:
                             gguf.MODEL_TENSOR.POSNET_NORM2,
                             gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
                             gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
+                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
+                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
                         )
                     )
                     or not new_name.endswith(".weight")
@@ -320,7 +322,11 @@ class ModelBase:
                     self.match_model_tensor_name(new_name, key, bid)
                     for key in (
                         gguf.MODEL_TENSOR.TOKEN_EMBD,
+                        gguf.MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
                         gguf.MODEL_TENSOR.OUTPUT,
+                        gguf.MODEL_TENSOR.ALTUP_ROUTER,
+                        gguf.MODEL_TENSOR.LAUREL_L,
+                        gguf.MODEL_TENSOR.LAUREL_R,
                     )
                 ):
                     if self.ftype in (
@@ -921,13 +927,20 @@ class TextModel(ModelBase):
         tokenizer = SentencePieceProcessor()
         tokenizer.LoadFromFile(str(tokenizer_path))
 
-        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+        vocab_size = self.find_hparam([
+            "vocab_size_per_layer_input", # gemma3n
+            "vocab_size",
+        ], optional=True) or tokenizer.vocab_size()
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
         toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
         for token_id in range(tokenizer.vocab_size()):
+            if token_id >= vocab_size:
+                logger.warning(f'ignore tokens from {token_id}: id is out of range, max={vocab_size - 1}')
+                break
+
             piece = tokenizer.IdToPiece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.GetScore(token_id)
@@ -2145,7 +2158,6 @@ class Llama4Model(LlamaModel):
 
     def set_vocab(self):
         self._set_vocab_gpt2()
-        self.gguf_writer.add_add_bos_token(True)
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
@@ -2194,7 +2206,7 @@ class Llama4VisionModel(MmprojModel):
                 name += ".weight"
             if "multi_modal_projector.linear_1" in name:
                 # despite the name with number postfix, this is a single fully connected layer
-                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)]
+                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC] + '.weight', data_torch)]
             return [(self.map_tensor_name(name), data_torch)]
         return []
 
@@ -2731,6 +2743,52 @@ class Qwen2Model(TextModel):
         yield from super().modify_tensors(data_torch, name, bid)
 
 
+@ModelBase.register("Ernie4_5_ForCausalLM")
+class Ernie4_5Model(TextModel):
+    model_arch = gguf.MODEL_ARCH.ERNIE4_5
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        num_heads = self.hparams["num_attention_heads"]
+        num_kv_heads = self.hparams["num_key_value_heads"]
+        head_dim = self.hparams["head_dim"]
+
+        if "ernie." in name:
+            name = name.replace("ernie.", "model.")
+        # split the qkv weights
+        # qkv_proj shape: [(num_heads + 2 * num_kv_heads) * head_dim, hidden_size]
+        if "qkv_proj" in name:
+            name_q = name.replace("qkv_proj.weight", "q_proj.weight")
+            name_k = name.replace("qkv_proj.weight", "k_proj.weight")
+            name_v = name.replace("qkv_proj.weight", "v_proj.weight")
+            total_q_dim = num_heads * head_dim
+            total_k_dim = num_kv_heads * head_dim
+            total_v_dim = num_kv_heads * head_dim
+            q_proj_weight, k_proj_weight, v_proj_weight = data_torch.split([total_q_dim, total_k_dim, total_v_dim], dim=0)
+            return [
+                (self.map_tensor_name(name_q), q_proj_weight),
+                (self.map_tensor_name(name_k), k_proj_weight),
+                (self.map_tensor_name(name_v), v_proj_weight)
+            ]
+        # split the up_gate_proj into gate and up
+        # up_gate_proj shape: [2 * intermediate_size, hidden_size]
+        if "up_gate_proj" in name:
+            name_up = name.replace("up_gate_proj.weight", "up_proj.weight")
+            name_gate = name.replace("up_gate_proj.weight", "gate_proj.weight")
+            dim_half = data_torch.shape[0] // 2
+            gate_proj_weight, up_proj_weight = data_torch.split(dim_half, dim=0)
+            return [
+                (self.map_tensor_name(name_gate), gate_proj_weight),
+                (self.map_tensor_name(name_up), up_proj_weight)
+            ]
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register(
     "Qwen2VLModel",
     "Qwen2VLForConditionalGeneration",
@@ -3918,9 +3976,6 @@ class BertModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(True)
-
 
 @ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
 class DistilBertModel(BertModel):
@@ -3962,8 +4017,6 @@ class RobertaModel(BertModel):
         bpe_tok_path = self.dir_model / "tokenizer.json"
         if bpe_tok_path.exists():
             self._set_vocab_gpt2()
-            self.gguf_writer.add_add_bos_token(True)
-            self.gguf_writer.add_add_eos_token(True)
 
             # we need this to validate the size of the token_type embeddings
             # though currently we are passing all zeros to the token_type embeddings
@@ -4223,6 +4276,7 @@ class Gemma2Model(TextModel):
 @ModelBase.register("Gemma3ForCausalLM", "Gemma3ForConditionalGeneration")
 class Gemma3Model(TextModel):
     model_arch = gguf.MODEL_ARCH.GEMMA3
+    norm_shift = 1.0  # Gemma3RMSNorm adds 1.0 to the norm value
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -4244,9 +4298,8 @@ class Gemma3Model(TextModel):
         self.gguf_writer.add_value_length(hparams.get("head_dim", 256))
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1_000_000.0)) # for global layers
-        # both attn_logit_softcapping and final_logit_softcapping are removed in Gemma3
+        # attn_logit_softcapping is removed in Gemma3
         assert hparams.get("attn_logit_softcapping") is None
-        assert hparams.get("final_logit_softcapping") is None
         self.gguf_writer.add_sliding_window(hparams["sliding_window"])
         self.gguf_writer.add_head_count_kv(hparams.get("num_key_value_heads", 4))
         if hparams.get("rope_scaling") is not None:
@@ -4258,7 +4311,7 @@ class Gemma3Model(TextModel):
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         del bid  # unused
 
-        if name.startswith("language_model."):
+        if "language_model." in name:
             name = name.replace("language_model.", "")
 
         elif name.startswith("multi_modal_projector.") or name.startswith("vision_tower.") \
@@ -4273,8 +4326,9 @@ class Gemma3Model(TextModel):
 
         # ref code in Gemma3RMSNorm
         # output = output * (1.0 + self.weight.float())
+        # note: this is not the case on gemma3n
         if name.endswith("norm.weight"):
-            data_torch = data_torch + 1
+            data_torch = data_torch + self.norm_shift
 
         return [(self.map_tensor_name(name), data_torch)]
 
@@ -4331,6 +4385,104 @@ class Gemma3VisionModel(MmprojModel):
         return [] # skip other tensors
 
 
+@ModelBase.register("Gemma3nForConditionalGeneration")
+class Gemma3NModel(Gemma3Model):
+    model_arch = gguf.MODEL_ARCH.GEMMA3N
+    norm_shift = 0.0 # same value with Gemma3p5RMSNorm scale_shift on python code
+
+    _altup_proj: list[Tensor] = []
+    _altup_unembd: list[Tensor] = []
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams["altup_num_inputs"] == 4, "Current conversion only supports 4 altup inputs"
+        self._altup_proj = [
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+        ]
+        self._altup_unembd = [
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+            torch.Tensor(), # to be replaced
+        ]
+
+    def set_vocab(self):
+        with open(self.dir_model / "chat_template.jinja") as f:
+            # quick hack to make sure chat template is added
+            self.gguf_writer.add_chat_template(f.read())
+        super().set_vocab()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_altup_active_idx(self.hparams["altup_active_idx"])
+        self.gguf_writer.add_altup_num_inputs(self.hparams["altup_num_inputs"])
+        self.gguf_writer.add_embedding_length_per_layer_input(self.hparams["hidden_size_per_layer_input"])
+        self.gguf_writer.add_shared_kv_layers(self.hparams["num_kv_shared_layers"])
+
+        activation_sparsity_scale = []
+        for s in self.hparams["activation_sparsity_pattern"]:
+            normal_dist = torch.distributions.normal.Normal(0, 1)
+            std_multiplier = normal_dist.icdf(torch.tensor(s, dtype=torch.float32))
+            activation_sparsity_scale.append(std_multiplier.item())
+        self.gguf_writer.add_activation_sparsity_scale(activation_sparsity_scale)
+
+        sliding_window_pattern = []
+        for t in self.hparams["layer_types"]:
+            sliding_window_pattern.append(t == "sliding_attention")
+        self.gguf_writer.add_sliding_window_pattern(sliding_window_pattern)
+
+    def _stack_matrices(self, matrices: list[Tensor]) -> Tensor | None:
+        has_all = all(m.numel() > 0 for m in matrices)
+        if not has_all:
+            return None
+        else:
+            return torch.stack(matrices, dim=0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.endswith("_scale"):
+            name = name + ".weight"
+
+        # TODO: implement self.prediction_coefs.weight.clamp_(...)
+
+        if "language_model." not in name:
+            return [] # skip non-language model tensors
+
+        if "altup_unembed_projections" in name:
+            data_torch = data_torch.to(device="cpu")
+            if ".0." in name:
+                self._altup_unembd[0] = data_torch
+            elif ".1." in name:
+                self._altup_unembd[1] = data_torch
+            elif ".2." in name:
+                self._altup_unembd[2] = data_torch
+            else:
+                raise ValueError(f"Unknown name: {name}")
+            out = self._stack_matrices(self._altup_unembd)
+            if out is not None:
+                return [(self.map_tensor_name("model.altup_unembed_projections.weight"), out)]
+            else:
+                return []
+
+        if "altup_projections" in name:
+            data_torch = data_torch.to(device="cpu")
+            if ".0." in name:
+                self._altup_proj[0] = data_torch
+            elif ".1." in name:
+                self._altup_proj[1] = data_torch
+            elif ".2." in name:
+                self._altup_proj[2] = data_torch
+            else:
+                raise ValueError(f"Unknown name: {name}")
+            out = self._stack_matrices(self._altup_proj)
+            if out is not None:
+                return [(self.map_tensor_name("model.altup_projections.weight"), out)]
+            else:
+                return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
 @ModelBase.register("Starcoder2ForCausalLM")
 class StarCoder2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.STARCODER2
@@ -4848,8 +5000,6 @@ class JinaBertV2Model(BertModel):
             self.gguf_writer.add_token_type_count(2)
         else:
             raise NotImplementedError(f'Tokenizer {tokenizer_class} is not supported for JinaBertModel')
-        self.gguf_writer.add_add_bos_token(True)
-        self.gguf_writer.add_add_eos_token(True)
 
 
 @ModelBase.register("OpenELMForCausalLM")
@@ -5451,9 +5601,6 @@ class T5Model(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(False)
-        self.gguf_writer.add_add_eos_token(True)
-
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -5591,9 +5738,6 @@ class T5EncoderModel(TextModel):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-        self.gguf_writer.add_add_bos_token(False)
-        self.gguf_writer.add_add_eos_token(True)
-
     def set_gguf_parameters(self):
         if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
             logger.warning("Couldn't find context length in config.json, assuming default value of 512")
@@ -6389,8 +6533,8 @@ def parse_args() -> argparse.Namespace:
         help="model is executed on big endian machine",
     )
     parser.add_argument(
-        "model", type=Path,
-        help="directory containing model file",
+        "model", type=str,
+        help="directory containing model file or huggingface repository ID (if --remote)",
         nargs="?",
     )
     parser.add_argument(
@@ -6493,18 +6637,20 @@ def main() -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
-    dir_model = args.model
-
     if args.remote:
+        hf_repo_id = args.model
         from huggingface_hub import snapshot_download
         local_dir = snapshot_download(
-            repo_id=str(dir_model),
+            repo_id=hf_repo_id,
             allow_patterns=["LICENSE", "*.json", "*.md", "*.txt", "tokenizer.model"])
         dir_model = Path(local_dir)
         logger.info(f"Downloaded config and tokenizer to {local_dir}")
+    else:
+        hf_repo_id = None
+        dir_model = Path(args.model)
 
     if not dir_model.is_dir():
-        logger.error(f'Error: {args.model} is not a directory')
+        logger.error(f'Error: {dir_model} is not a directory')
         sys.exit(1)
 
     ftype_map: dict[str, gguf.LlamaFileType] = {
@@ -6524,9 +6670,9 @@ def main() -> None:
 
     if args.outfile is not None:
         fname_out = args.outfile
-    elif args.remote:
+    elif hf_repo_id:
         # if remote, use the model ID as the output file name
-        fname_out = Path("./" + str(args.model).replace("/", "-") + "-{ftype}.gguf")
+        fname_out = Path("./" + hf_repo_id.replace("/", "-") + "-{ftype}.gguf")
     else:
         fname_out = dir_model
 
@@ -6555,7 +6701,7 @@ def main() -> None:
                                      split_max_tensors=args.split_max_tensors,
                                      split_max_size=split_str_to_n_bytes(args.split_max_size), dry_run=args.dry_run,
                                      small_first_shard=args.no_tensor_first_split,
-                                     remote_hf_model_id=str(args.model) if args.remote else None)
+                                     remote_hf_model_id=hf_repo_id)
 
         if args.vocab_only:
             logger.info("Exporting model vocab...")

docs/backend/SYCL.md

@@ -757,7 +757,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | Name              | Value            | Function                                                                                                                  |
 |-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
 | GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
-| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features based on Intel GPU type, to compare the performance increase |
+| GGML_SYCL_DISABLE_OPT | 0 (default) or 1 | Disable optimize features for Intel GPUs. (Recommended to 1 for intel devices older than Gen 10) |
 | GGML_SYCL_DISABLE_GRAPH | 0 or 1 (default) | Disable running computations through SYCL Graphs feature. Disabled by default because graph performance isn't yet better than non-graph performance. |
 | GGML_SYCL_DISABLE_DNN | 0 (default) or 1 | Disable running computations through oneDNN and always use oneMKL. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |

docs/build-s390x.md

@@ -0,0 +1,246 @@
+> [!IMPORTANT]
+> This build documentation is specific only to IBM Z & LinuxONE mainframes (s390x). You can find the build documentation for other architectures: [build.md](build.md).
+
+# Build llama.cpp locally (for s390x)
+
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
+**To get the code:**
+
+```bash
+git clone https://github.com/ggml-org/llama.cpp
+cd llama.cpp
+```
+
+## CPU Build with BLAS
+
+Building llama.cpp with BLAS support is highly recommended as it has shown to provide performance improvements. Make sure to have OpenBLAS installed in your environment.
+
+```bash
+cmake -S . -B build             \
+    -DCMAKE_BUILD_TYPE=Release  \
+    -DGGML_BLAS=ON              \
+    -DGGML_BLAS_VENDOR=OpenBLAS
+
+cmake --build build --config Release -j $(nproc)
+```
+
+**Notes**:
+
+-   For faster repeated compilation, install [ccache](https://ccache.dev/)
+-   By default, VXE/VXE2 is enabled. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_VXE=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   By default, NNPA is enabled when available. To disable it (not recommended):
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DGGML_NNPA=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+-   For debug builds:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Debug    \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS
+    cmake --build build --config Debug -j $(nproc)
+    ```
+
+-   For static builds, add `-DBUILD_SHARED_LIBS=OFF`:
+
+    ```bash
+    cmake -S . -B build             \
+        -DCMAKE_BUILD_TYPE=Release  \
+        -DGGML_BLAS=ON              \
+        -DGGML_BLAS_VENDOR=OpenBLAS \
+        -DBUILD_SHARED_LIBS=OFF
+
+    cmake --build build --config Release -j $(nproc)
+    ```
+
+## Getting GGUF Models
+
+All models need to be converted to Big-Endian. You can achieve this in three cases:
+
+1. **Use pre-converted models verified for use on IBM Z & LinuxONE (easiest)**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    You can find popular models pre-converted and verified at [s390x Ready Models](https://huggingface.co/collections/taronaeo/s390x-ready-models-672765393af438d0ccb72a08).
+
+    These models have already been converted from `safetensors` to `GGUF Big-Endian` and their respective tokenizers verified to run correctly on IBM z15 and later system.
+
+2. **Convert safetensors model to GGUF Big-Endian directly (recommended)**
+
+    ![File Type - safetensors](https://img.shields.io/badge/File_Type-safetensors-da1e28)
+
+    The model you are trying to convert must be in `safetensors` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct)). Make sure you have downloaded the model repository for this case.
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile model-name-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        model-directory/
+    ```
+
+    For example,
+
+    ```bash
+    python3 convert_hf_to_gguf.py \
+        --outfile granite-3.3-2b-instruct-be.f16.gguf \
+        --outtype f16 \
+        --bigendian \
+        granite-3.3-2b-instruct/
+    ```
+
+3. **Convert existing GGUF Little-Endian model to Big-Endian**
+
+    ![File Type - gguf](https://img.shields.io/badge/File_Type-gguf-fff)
+
+    The model you are trying to convert must be in `gguf` file format (for example [IBM Granite 3.3 2B](https://huggingface.co/ibm-granite/granite-3.3-2b-instruct-GGUF)). Make sure you have downloaded the model file for this case.
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py model-name.f16.gguf BIG
+    ```
+
+    For example,
+
+    ```bash
+    python3 gguf-py/gguf/scripts/gguf_convert_endian.py granite-3.3-2b-instruct-le.f16.gguf BIG
+    mv granite-3.3-2b-instruct-le.f16.gguf granite-3.3-2b-instruct-be.f16.gguf
+    ```
+
+    **Notes:**
+
+    - The GGUF endian conversion script may not support all data types at the moment and may fail for some models/quantizations. When that happens, please try manually converting the safetensors model to GGUF Big-Endian via Step 2.
+
+## IBM Accelerators
+
+### 1. SIMD Acceleration
+
+Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 2. NNPA Vector Intrinsics Acceleration
+
+Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned on when available) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+
+### 3. zDNN Accelerator
+
+_Only available in IBM z16 or later system. No direction at the moment._
+
+### 4. Spyre Accelerator
+
+_No direction at the moment._
+
+## Performance Tuning
+
+### 1. Virtualization Setup
+
+It is strongly recommended to use only LPAR (Type-1) virtualization to get the most performance.
+
+Note: Type-2 virtualization is not supported at the moment, while you can get it running, the performance will not be the best.
+
+### 2. IFL (Core) Count
+
+It is recommended to allocate a minimum of 8 shared IFLs assigned to the LPAR. Increasing the IFL count past 8 shared IFLs will only improve Prompt Processing performance but not Token Generation.
+
+Note: IFL count does not equate to vCPU count.
+
+### 3. SMT vs NOSMT (Simultaneous Multithreading)
+
+It is strongly recommended to disable SMT via the kernel boot parameters as it negatively affects performance. Please refer to your Linux distribution's guide on disabling SMT via kernel boot parameters.
+
+### 4. BLAS vs NOBLAS
+
+IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongly recommended to use BLAS.
+
+## Frequently Asked Questions (FAQ)
+
+1. I'm getting the following error message while trying to load a model: `gguf_init_from_file_impl: failed to load model: this GGUF file version 50331648 is extremely large, is there a mismatch between the host and model endianness?`
+
+    Answer: Please ensure that the model you have downloaded/converted is GGUFv3 Big-Endian. These models are usually denoted with the `-be` suffix, i.e., `granite-3.3-2b-instruct-be.F16.gguf`.
+
+    You may refer to the [Getting GGUF Models](#getting-gguf-models) section to manually convert a `safetensors` model to `GGUF` Big Endian.
+
+2. I'm getting extremely poor performance when running inference on a model
+
+    Answer: Please refer to the [Appendix B: SIMD Support Matrix](#appendix-b-simd-support-matrix) to check if your model quantization is supported by SIMD acceleration.
+
+3. I'm building on IBM z17 and getting the following error messages: `invalid switch -march=z17`
+
+    Answer: Please ensure that your GCC compiler is of minimum GCC 15.1.0 version, and have `binutils` updated to the latest version. If this does not fix the problem, kindly open an issue.
+
+## Getting Help on IBM Z & LinuxONE
+
+1. **Bugs, Feature Requests**
+
+    Please file an issue in llama.cpp and ensure that the title contains "s390x".
+
+2. **Other Questions**
+
+    Please reach out directly to [aionz@us.ibm.com](mailto:aionz@us.ibm.com).
+
+## Appendix A: Hardware Support Matrix
+
+|         | Support | Minimum Compiler Version |
+| ------- | ------- | ------------------------ |
+| IBM z15 | ✅      |                          |
+| IBM z16 | ✅      |                          |
+| IBM z17 | ✅      | GCC 15.1.0               |
+
+-   ✅ - supported and verified to run as intended
+-   🚫 - unsupported, we are unlikely able to provide support
+
+## Appendix B: SIMD Support Matrix
+
+|            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
+| ---------- | ----------- | ---- | ---- | ----- |
+| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP16       | ✅          | ✅   | ❓   | ❓    |
+| BF16       | 🚫          | 🚫   | ❓   | ❓    |
+| Q4_0       | ✅          | ✅   | ❓   | ❓    |
+| Q4_1       | ✅          | ✅   | ❓   | ❓    |
+| Q5_0       | 🚫          | 🚫   | ❓   | ❓    |
+| Q5_1       | 🚫          | 🚫   | ❓   | ❓    |
+| Q8_0       | ✅          | ✅   | ❓   | ❓    |
+| Q2_K       | 🚫          | 🚫   | ❓   | ❓    |
+| Q3_K       | ✅          | ✅   | ❓   | ❓    |
+| Q4_K       | ✅          | ✅   | ❓   | ❓    |
+| Q5_K       | ✅          | ✅   | ❓   | ❓    |
+| Q6_K       | ✅          | ✅   | ❓   | ❓    |
+| TQ1_0      | 🚫          | 🚫   | ❓   | ❓    |
+| TQ2_0      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_XS     | 🚫          | 🚫   | ❓   | ❓    |
+| IQ2_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_XXS    | 🚫          | 🚫   | ❓   | ❓    |
+| IQ3_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_S      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ1_M      | 🚫          | 🚫   | ❓   | ❓    |
+| IQ4_NL     | ✅          | ✅   | ❓   | ❓    |
+| IQ4_XS     | ✅          | ✅   | ❓   | ❓    |
+| FP32->FP16 | 🚫          | ✅   | ❓   | ❓    |
+| FP16->FP32 | 🚫          | ✅   | ❓   | ❓    |
+
+-   ✅ - acceleration available
+-   🚫 - acceleration unavailable, will still run using scalar implementation
+-   ❓ - acceleration unknown, please contribute if you can test it yourself

docs/build.md

@@ -1,6 +1,6 @@
 # Build llama.cpp locally
 
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](../include/llama.h).
 
 The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
 
@@ -557,6 +557,10 @@ ninja
 
 To read documentation for how to build on Android, [click here](./android.md)
 
+## IBM Z & LinuxONE
+
+To read documentation for how to build on IBM Z & LinuxONE, [click here](./build-s390x.md)
+
 ## Notes about GPU-accelerated backends
 
 The GPU may still be used to accelerate some parts of the computation even when using the `-ngl 0` option. You can fully disable GPU acceleration by using `--device none`.

examples/embedding/embedding.cpp

@@ -133,10 +133,36 @@ int main(int argc, char ** argv) {
     // max batch size
     const uint64_t n_batch = params.n_batch;
 
+    // get added sep and eos token, if any
+    const std::string added_sep_token = llama_vocab_get_add_sep(vocab) ? llama_vocab_get_text(vocab, llama_vocab_sep(vocab)) : "";
+    const std::string added_eos_token = llama_vocab_get_add_eos(vocab) ? llama_vocab_get_text(vocab, llama_vocab_eos(vocab)) : "";
+
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = common_tokenize(ctx, prompt, true, true);
+        std::vector<llama_token> inp;
+
+        // split classification pairs and insert expected separator tokens
+        if (pooling_type == LLAMA_POOLING_TYPE_RANK && prompt.find(params.cls_sep) != std::string::npos) {
+            std::vector<std::string> pairs = split_lines(prompt, params.cls_sep);
+            std::string final_prompt;
+
+            for (size_t i = 0; i < pairs.size(); i++) {
+                final_prompt += pairs[i];
+                if (i != pairs.size() - 1) {
+                    if (!added_eos_token.empty()) {
+                        final_prompt += added_eos_token;
+                    }
+                    if (!added_sep_token.empty()) {
+                        final_prompt += added_sep_token;
+                    }
+                }
+            }
+
+            inp = common_tokenize(ctx, final_prompt, true, true);
+        } else {
+            inp = common_tokenize(ctx, prompt, true, true);
+        }
         if (inp.size() > n_batch) {
             LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
@@ -145,11 +171,11 @@ int main(int argc, char ** argv) {
         inputs.push_back(inp);
     }
 
-    // check if the last token is SEP
+    // check if the last token is SEP/EOS
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_vocab_sep(vocab)) {
-            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
+        if (inp.empty() || (inp.back() != llama_vocab_sep(vocab) && inp.back() != llama_vocab_eos(vocab))) {
+            LOG_WRN("%s: last token in the prompt is not SEP or EOS\n", __func__);
             LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }

examples/eval-callback/eval-callback.cpp

@@ -55,6 +55,8 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                         v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                     } else if (type == GGML_TYPE_F32) {
                         v = *(float *) &data[i];
+                    } else if (type == GGML_TYPE_I64) {
+                        v = (float) *(int64_t *) &data[i];
                     } else if (type == GGML_TYPE_I32) {
                         v = (float) *(int32_t *) &data[i];
                     } else if (type == GGML_TYPE_I16) {

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == -1;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);

ggml/CMakeLists.txt

@@ -131,6 +131,7 @@ option(GGML_RVV              "ggml: enable rvv"              ON)
 option(GGML_RV_ZFH           "ggml: enable riscv zfh"        OFF)
 option(GGML_XTHEADVECTOR     "ggml: enable xtheadvector"     OFF)
 option(GGML_VXE              "ggml: enable vxe"              ON)
+option(GGML_NNPA             "ggml: enable nnpa"             ON)
 
 option(GGML_CPU_ALL_VARIANTS "ggml: build all variants of the CPU backend (requires GGML_BACKEND_DL)" OFF)
 set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")

ggml/include/ggml-cpu.h

@@ -101,6 +101,7 @@ extern "C" {
     GGML_BACKEND_API int ggml_cpu_has_riscv_v    (void);
     GGML_BACKEND_API int ggml_cpu_has_vsx        (void);
     GGML_BACKEND_API int ggml_cpu_has_vxe        (void);
+    GGML_BACKEND_API int ggml_cpu_has_nnpa       (void);
     GGML_BACKEND_API int ggml_cpu_has_wasm_simd  (void);
     GGML_BACKEND_API int ggml_cpu_has_llamafile  (void);
 
@@ -133,6 +134,7 @@ extern "C" {
 
     GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
 
+    GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *,       float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
     GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

ggml/include/ggml.h

@@ -470,6 +470,7 @@ extern "C" {
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
         GGML_OP_GET_ROWS_BACK,
+        GGML_OP_SET_ROWS,
         GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
         GGML_OP_DIAG_MASK_ZERO,
@@ -489,6 +490,7 @@ extern "C" {
         GGML_OP_UPSCALE, // nearest interpolate
         GGML_OP_PAD,
         GGML_OP_PAD_REFLECT_1D,
+        GGML_OP_ROLL,
         GGML_OP_ARANGE,
         GGML_OP_TIMESTEP_EMBEDDING,
         GGML_OP_ARGSORT,
@@ -686,6 +688,9 @@ extern "C" {
     // true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
     GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
 
+    // true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
+    GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
+
     GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
     GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
 
@@ -1374,6 +1379,23 @@ extern "C" {
             struct ggml_tensor  * b,  // row indices
             struct ggml_tensor  * c); // data for ggml_get_rows, only used for its shape
 
+    // a TD  [n_embd, ne1,    ne2,    ne3]
+    // b TS  [n_embd, n_rows, ne02,   ne03] | ne02 == ne2, ne03 == ne3
+    // c I64 [n_rows, ne11,   ne12,   1]    | c[i] in [0, ne1)
+    //
+    // undefined behavior if destination rows overlap
+    //
+    // broadcast:
+    //   ne2 % ne11 == 0
+    //   ne3 % ne12 == 0
+    //
+    // return view(a)
+    GGML_API struct ggml_tensor * ggml_set_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,  // destination
+            struct ggml_tensor  * b,  // source
+            struct ggml_tensor  * c); // row indices
+
     GGML_API struct ggml_tensor * ggml_diag(
         struct ggml_context     * ctx,
         struct ggml_tensor      * a);
@@ -1801,6 +1823,17 @@ extern "C" {
             int                   p0,
             int                   p1);
 
+    // Move tensor elements by an offset given for each dimension. Elements that
+    // are shifted beyond the last position are wrapped around to the beginning.
+    GGML_API struct ggml_tensor * ggml_roll(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   shift0,
+            int                   shift1,
+            int                   shift2,
+            int                   shift3);
+
+
     // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
     // timesteps: [N,]
     // return: [N, dim]

ggml/src/CMakeLists.txt

@@ -286,6 +286,10 @@ function(ggml_add_cpu_backend_variant tag_name)
         foreach (feat ${ARGN})
             set(GGML_INTERNAL_${feat} ON)
         endforeach()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        foreach (feat ${ARGN})
+            set(GGML_INTERNAL_${feat} ON)
+        endforeach()
     endif()
 
     ggml_add_cpu_backend_variant_impl(${tag_name})
@@ -330,9 +334,26 @@ if (GGML_CPU_ALL_VARIANTS)
             ggml_add_cpu_backend_variant(android_armv8.2_1    DOTPROD)
             ggml_add_cpu_backend_variant(android_armv8.2_2    DOTPROD FP16_VECTOR_ARITHMETIC)
             ggml_add_cpu_backend_variant(android_armv8.6_1    DOTPROD FP16_VECTOR_ARITHMETIC MATMUL_INT8)
+        elseif (APPLE)
+            ggml_add_cpu_backend_variant(apple_m1             DOTPROD)
+            ggml_add_cpu_backend_variant(apple_m2_m3          DOTPROD MATMUL_INT8)
+            ggml_add_cpu_backend_variant(apple_m4             DOTPROD MATMUL_INT8 NOSVE SME)
         else()
             message(FATAL_ERROR "Unsupported ARM target OS: ${CMAKE_SYSTEM_NAME}")
         endif()
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
+        if (CMAKE_SYSTEM_NAME MATCHES "Linux")
+            ggml_add_cpu_backend_variant(power0)
+            ggml_add_cpu_backend_variant(power7_1       POWER7)
+            ggml_add_cpu_backend_variant(power7_2       POWER7  VSX)
+            ggml_add_cpu_backend_variant(power8_1       POWER8)
+            ggml_add_cpu_backend_variant(power8_2       POWER8  VSX)
+            ggml_add_cpu_backend_variant(power9         POWER9  VSX)
+            ggml_add_cpu_backend_variant(power10        POWER10 VSX)
+            ggml_add_cpu_backend_variant(power11        POWER11 VSX)
+        else()
+            message(FATAL_ERROR "Unsupported PowerPC target OS: ${CMAKE_SYSTEM_NAME}")
+        endif()
     else()
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported with ${GGML_SYSTEM_ARCH} on ${CMAKE_SYSTEM_NAME}")
     endif()

ggml/src/ggml-backend-reg.cpp

@@ -69,6 +69,9 @@
 #if defined(__clang__)
 #    pragma clang diagnostic push
 #    pragma clang diagnostic ignored "-Wdeprecated-declarations"
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic push
+#    pragma GCC diagnostic ignored "-Wdeprecated-declarations"
 #endif
 
 namespace fs = std::filesystem;
@@ -91,6 +94,8 @@ static std::string path_str(const fs::path & path) {
 
 #if defined(__clang__)
 #    pragma clang diagnostic pop
+#elif defined(__GNUC__)
+#    pragma GCC diagnostic pop
 #endif
 
 #ifdef _WIN32

ggml/src/ggml-cann/common.h

@@ -359,7 +359,7 @@ struct ggml_backend_cann_context {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
 
-        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
+        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/src/ggml-cpu/CMakeLists.txt

@@ -190,6 +190,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                         set(ARCH_TAGS "${ARCH_TAGS}+sve2")
                         list(APPEND ARCH_DEFINITIONS GGML_USE_SVE2)
                     endif()
+                    if (GGML_INTERNAL_NOSVE)
+                        set(ARCH_TAGS "${ARCH_TAGS}+nosve")
+                    endif()
                     if (GGML_INTERNAL_SME)
                         set(ARM_MCPU "armv9.2-a")
                         set(ARCH_TAGS "${ARCH_TAGS}+sme")
@@ -385,6 +388,27 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             else()
                 list(APPEND ARCH_FLAGS -mcpu=native -mtune=native -mpowerpc64)
             endif()
+        elseif(GGML_CPU_ALL_VARIANTS)
+            # Begin with the lowest baseline
+            set(ARCH_DEFINITIONS "")
+
+            # When a feature is selected, bump the MCPU to the first
+            # version that supported it
+            foreach(PVER RANGE 7 11)
+                if(DEFINED GGML_INTERNAL_POWER${PVER})
+                    set(POWERPC_MCPU "power${PVER}")
+                    list(APPEND ARCH_DEFINITIONS GGML_USE_POWER${PVER})
+                endif()
+            endforeach()
+            if (GGML_INTERNAL_VSX)
+                list(APPEND ARCH_DEFINITIONS GGML_USE_VSX)
+                list(APPEND ARCH_FLAGS -mvsx)
+            endif()
+
+            if (DEFINED POWERPC_MCPU)
+                list(APPEND ARCH_FLAGS -mcpu=${POWERPC_MCPU})
+            endif()
+            ggml_add_cpu_backend_features(${GGML_CPU_NAME} powerpc ${ARCH_DEFINITIONS})
         else()
             if (GGML_CPU_POWERPC_CPUTYPE)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
@@ -424,6 +448,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # TODO: Separation to determine activation of VX/VXE/VXE2
         if (${S390X_M} MATCHES "8561|8562")
+            set(GGML_NNPA OFF)
             message(STATUS "z15 target")
             list(APPEND ARCH_FLAGS -march=z15)
         elseif (${S390X_M} MATCHES "3931")
@@ -440,7 +465,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         endif()
 
         if (GGML_VXE)
+            message(STATUS "VX/VXE/VXE2 enabled")
             list(APPEND ARCH_FLAGS -mvx -mzvector)
+            list(APPEND ARCH_DEFINITIONS GGML_VXE)
+        endif()
+
+        if (GGML_NNPA)
+            message(STATUS "NNPA enabled")
+            list(APPEND ARCH_DEFINITIONS GGML_NNPA)
         endif()
     elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
         message(STATUS "Wasm detected")
@@ -462,9 +494,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
 
         # Fetch KleidiAI sources:
         include(FetchContent)
-        set(KLEIDIAI_COMMIT_TAG "v1.6.0")
+        set(KLEIDIAI_COMMIT_TAG "v1.9.0")
         set(KLEIDIAI_DOWNLOAD_URL "https://github.com/ARM-software/kleidiai/archive/refs/tags/${KLEIDIAI_COMMIT_TAG}.tar.gz")
-        set(KLEIDIAI_ARCHIVE_MD5  "75b4ad68f25ab673dcc01065e5a0b05f")
+        set(KLEIDIAI_ARCHIVE_MD5  "2a8e1bb55d201557553545536489a017")
 
         if (POLICY CMP0135)
             cmake_policy(SET CMP0135 NEW)

ggml/src/ggml-cpu/amx/mmq.cpp

@@ -8,6 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
@@ -453,7 +454,7 @@ void quantize_row_q8_K_vnni(const float * RESTRICT x, void * RESTRICT vy, int64_
 
         // Quantize these floats
         const float iscale = 127.f / amax;
-        y[i].d = GGML_FP32_TO_FP16(1 / iscale);
+        y[i].d = GGML_CPU_FP32_TO_FP16(1 / iscale);
         const float id = ( amax != 0.0f ) ? iscale : 0.f;
         const __m512 vscale = _mm512_set1_ps(id);
 
@@ -1090,7 +1091,7 @@ struct acc_C<block_q8_0, block_q4_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1113,8 +1114,8 @@ struct acc_C<block_q8_1, block_q4_1, is_acc> {
         const __m512 vm0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset + TILE_N * sizeof(ggml_half))));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
-            const __m512 vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].s));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].s));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1137,7 +1138,7 @@ struct acc_C<block_q8_0, block_q8_0, is_acc> {
         const __m512 vd0 = _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)((const char *)packed_B + offset)));
 
         for (int m = 0; m < nr; ++m) {
-            const __m512 vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[m * lda].d));
+            const __m512 vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[m * lda].d));
             const __m512 vtile = _mm512_cvtepi32_ps(_mm512_loadu_si512(tile + m * TILE_N));
 
             __m512 vsum;
@@ -1437,7 +1438,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q4_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     vcomp = _mm512_dpbusd_epi32(vcomp, off, va[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b
@@ -1498,8 +1499,8 @@ struct tinygemm_kernel_vnni<block_q8_1, block_q4_1, float, 1, BLOCK_N, BLOCK_K>
                 for (int k = 0; k < 8; ++k) {
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
-                vs1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].s));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
+                vs1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].s));
             }
 
             // load b
@@ -1571,7 +1572,7 @@ struct tinygemm_kernel_vnni<block_q8_0, block_q8_0, float, BLOCK_M, BLOCK_N, BLO
                     va[k] = _mm512_set1_epi32(a_ptr[k]);
                     va[k] = _mm512_add_epi8(va[k], off);
                 }
-                vd1 = _mm512_set1_ps(GGML_FP16_TO_FP32(A[0 * KB + i].d));
+                vd1 = _mm512_set1_ps(GGML_CPU_FP16_TO_FP32(A[0 * KB + i].d));
             }
 
             // load b

ggml/src/ggml-cpu/arch/arm/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -62,7 +63,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const float32x4_t v  = vmulq_n_f32(srcv[j], id);
@@ -104,7 +105,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         int32x4_t accv = vdupq_n_s32(0);
 
@@ -120,7 +121,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             accv = vaddq_s32(accv, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vaddvq_s32(accv));
     }
 #else
     GGML_UNUSED(nb);
@@ -194,10 +195,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -274,10 +275,10 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     // dot product
                     sumv0 = svmla_n_f32_x(ph4, sumv0, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                     svdot_s32(svdup_n_s32(0), qx0ls, qy0l),
-                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0hs, qy0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(ph4, sumv1, svcvt_f32_s32_x(ph4, svadd_x(ph4,
                                     svdot_s32(svdup_n_s32(0), qx1ls, qy1l),
-                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1hs, qy1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -313,9 +314,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     // dot product
                     sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -354,9 +355,9 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     // dot product
                     sumv0 = svmla_n_f32_x(ph32, sumv0, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0s, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(ph32, sumv1, svcvt_f32_s32_x(ph32,
-                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1s, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(ph32, svadd_f32_x(ph32, sumv0, sumv1));
@@ -404,8 +405,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0l), v0_0hs, v1_0h);
         const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1l), v0_1hs, v1_1h);
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -423,7 +424,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -464,10 +465,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const block_q8_1 * GGML_RESTRICT b_y1 = &vy1[i];
 
             float32_t summs_t[4] = {
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
-                GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
-                GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y0->s),
+                GGML_CPU_FP16_TO_FP32(b_x0->m) * GGML_CPU_FP16_TO_FP32(b_y1->s),
+                GGML_CPU_FP16_TO_FP32(b_x1->m) * GGML_CPU_FP16_TO_FP32(b_y1->s)
             };
             summs0 = vaddq_f32(summs0, vld1q_f32(summs_t));
 
@@ -490,10 +491,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             // mmla into int32x4_t
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -539,7 +540,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const block_q8_1 * GGML_RESTRICT y0 = &y[ib + 0];
         const block_q8_1 * GGML_RESTRICT y1 = &y[ib + 1];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s) + GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
@@ -562,8 +563,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
         const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
 
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@@ -582,7 +583,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -666,10 +667,10 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -694,7 +695,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -739,8 +740,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         const uint8x16_t m4b = vdupq_n_u8(0x0F);
 
-        summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
-        summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
+        summs0 += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
+        summs1 += GGML_CPU_FP16_TO_FP32(x1->m) * GGML_CPU_FP16_TO_FP32(y1->s);
 
         // extract the 5th bit via lookup table ((b) << 4)
         memcpy(&qh0, x0->qh, sizeof(qh0));
@@ -784,10 +785,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
-                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@@ -812,7 +813,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -864,10 +865,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8x16_t y1_h = vld1q_s8(b_y1->qs + 16);
 
             float32_t _scale[4] = {
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x0->d)*GGML_FP16_TO_FP32(b_y1->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y0->d),
-                GGML_FP16_TO_FP32(b_x1->d)*GGML_FP16_TO_FP32(b_y1->d)
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x0->d)*GGML_CPU_FP16_TO_FP32(b_y1->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y0->d),
+                GGML_CPU_FP16_TO_FP32(b_x1->d)*GGML_CPU_FP16_TO_FP32(b_y1->d)
             };
             float32x4_t scale = vld1q_f32(_scale);
 
@@ -934,10 +935,10 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
                     sumv0 = svmla_n_f32_x(pl16, sumv0, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                     svdot_s32(svdup_n_s32(0), qx0_0, qy0_0),
-                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                    svdot_s32(svdup_n_s32(0), qx0_1, qy0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(pl16, sumv1, svcvt_f32_s32_x(pl16, svadd_x(pl16,
                                     svdot_s32(svdup_n_s32(0), qx1_0, qy1_0),
-                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                    svdot_s32(svdup_n_s32(0), qx1_1, qy1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(pl16, svadd_f32_x(pl16, sumv0, sumv1));
@@ -960,9 +961,9 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     const svint8_t qy1 = svld1_s8(svptrue_b8(), y1->qs);
 
                     sumv0 = svmla_n_f32_x(svptrue_b32(), sumv0, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                                svdot_s32(svdup_n_s32(0), qx0, qy0)), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
                     sumv1 = svmla_n_f32_x(svptrue_b32(), sumv1, svcvt_f32_s32_x(svptrue_b32(),
-                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                                svdot_s32(svdup_n_s32(0), qx1, qy1)), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
                 }
 
                 sumf = svaddv_f32(svptrue_b32(), svadd_f32_x(svptrue_b32(), sumv0, sumv1));
@@ -1002,8 +1003,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                     qy_64 = svadd_s8_x(svptrue_b8(), qy_32, qy_64);
 
                     // scale creation
-                    const float32_t deq1 = GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d);
-                    const float32_t deq2 = GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d);
+                    const float32_t deq1 = GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d);
+                    const float32_t deq2 = GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d);
 
                     // duplicate deq1 in first half of vector and deq2 in second half of vector
                     const svfloat32_t temp = svdup_f32_m(svdup_f32_z(ph8, deq1), pl8, deq2);
@@ -1043,11 +1044,11 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), x0_0, y0_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x0_1, y0_1))), GGML_CPU_FP16_TO_FP32(x0->d)*GGML_CPU_FP16_TO_FP32(y0->d));
 
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
                         ggml_vdotq_s32(vdupq_n_s32(0), x1_0, y1_0),
-                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
+                        ggml_vdotq_s32(vdupq_n_s32(0), x1_1, y1_1))), GGML_CPU_FP16_TO_FP32(x1->d)*GGML_CPU_FP16_TO_FP32(y1->d));
     }
 
     sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
@@ -1059,7 +1060,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1217,7 +1218,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
         const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
 #if defined(__ARM_FEATURE_DOTPROD)
         sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -1362,7 +1363,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         const int16x8_t ysum0 = vld1q_s16(y[i].bsums);
         const int16x8_t ysum1 = vld1q_s16(y[i].bsums + 8);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
 #if defined(__ARM_FEATURE_DOTPROD)
         sumi0 = vaddq_s32(sumi0, sumi1);
@@ -1393,7 +1394,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -1425,9 +1426,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     switch (vector_length) {
         case 128:
             for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                 svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                 svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
 
                 const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1570,9 +1571,9 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         case 256:
         case 512:
             for (int i = 0; i < nb; ++i) {
-                const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+                const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
                 svfloat32_t d_broad = svdup_n_f32((float32_t)d);
-                const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+                const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
                 svfloat32_t dmin_broad = svdup_n_f32((float32_t)dmin);
 
                 const uint8_t * GGML_RESTRICT q2 = x[i].qs;
@@ -1671,8 +1672,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sum = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1742,8 +1743,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1805,7 +1806,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3_sv = x[i].qs;
         const uint8_t * GGML_RESTRICT qh_sv = x[i].hmask;
@@ -1981,7 +1982,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].hmask;
@@ -2112,7 +2113,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2258,18 +2259,18 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
                                                vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
                 const float32x4_t dmins = {
-                    GGML_FP16_TO_FP32(x0->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x0->dmin) * y1->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y0->d,
-                    GGML_FP16_TO_FP32(x1->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->dmin) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->dmin) * y1->d,
                 };
                 vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
 
                 const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                 };
                 vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
             }
@@ -2289,8 +2290,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2377,8 +2378,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2478,9 +2479,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2520,8 +2521,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t q8sums = vpaddq_s16(vld1q_s16(y[i].bsums), vld1q_s16(y[i].bsums + 8));
 
@@ -2630,9 +2631,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2827,10 +2828,10 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 const int32x4_t vibias = vmulq_n_s32(vld1q_s32(bias), 32);
 
                 const float32x4_t superblock_scale = {
-                    GGML_FP16_TO_FP32(x0->d) * y0->d,
-                    GGML_FP16_TO_FP32(x0->d) * y1->d,
-                    GGML_FP16_TO_FP32(x1->d) * y0->d,
-                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_CPU_FP16_TO_FP32(x1->d) * y1->d,
                 };
 
                 visum = vsubq_s32(visum, vibias);
@@ -2858,7 +2859,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     svuint8_t q6h_1, q6h_2, q6h_3, q6h_4;
 
     for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q6 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3011,7 +3012,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q6 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3128,7 +3129,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -3199,7 +3200,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         float sumf1 = 0, sumf2 = 0;
@@ -3234,7 +3235,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -3284,7 +3285,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         const uint8x8_t scales8 = vld1_u8(x[i].scales);
@@ -3329,7 +3330,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3398,7 +3399,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -3458,7 +3459,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -3521,7 +3522,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3557,7 +3558,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3630,7 +3631,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3691,7 +3692,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3786,7 +3787,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
         }
 
-        sumf += y[i].d * GGML_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d) * (sumi1 + sumi2 + IQ1S_DELTA * sumi3);
     }
 
     *s = sumf;
@@ -3817,7 +3818,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -3905,7 +3906,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
         }
 
-        sumf += y[i].d * GGML_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
+        sumf += y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16) * (vaddvq_s32(sumi1) + IQ1M_DELTA * vaddvq_s32(sumi2));
     }
 
     *s = sumf;
@@ -3952,7 +3953,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qh += 2;
         }
 
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -4003,13 +4004,13 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         prod_2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q4b.val[2], q8b.val[2]), q4b.val[3], q8b.val[3]);
 
         sumf +=
-            GGML_FP16_TO_FP32(x[ib+0].d) * GGML_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
-            GGML_FP16_TO_FP32(x[ib+1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
+            GGML_CPU_FP16_TO_FP32(x[ib+0].d) * GGML_CPU_FP16_TO_FP32(y[ib + 0].d) * vaddvq_s32(prod_1) +
+            GGML_CPU_FP16_TO_FP32(x[ib+1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) * vaddvq_s32(prod_2);
     }
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4071,7 +4072,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
         }
 
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
     }
 
     *s = sumf;
@@ -4079,7 +4080,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/loongarch/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -474,7 +475,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = (__m256)__lasx_xvreplfr2vr_s( id );
 
@@ -548,7 +549,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = __lasx_xvreplfr2vr_s( id );
 
@@ -576,7 +577,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = __lsx_vadd_w(__lsx_vadd_w(ni0, ni1), __lsx_vadd_w(ni2, ni3));
         const __m128i s1 = __lsx_vadd_w(__lsx_vadd_w(ni4, ni5), __lsx_vadd_w(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(__lsx_vadd_w(s0, s1)));
 
         // Convert int32 to int16
         ni0 = lsx_packs_w( ni0, ni1 );
@@ -667,7 +668,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = __lasx_xvreplfr2vr_s( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
@@ -699,7 +700,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     for (; ib + 1 < nb; ib += 2) {
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = __lsx_vld((const __m128i *)x[ib].qs, 0);
 
@@ -717,7 +718,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         //_mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = (__m128)__lsx_vreplgr2vr_w( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = __lsx_vld((const __m128i *)x[ib + 1].qs, 0);
 
@@ -766,7 +767,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -797,10 +798,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = __lasx_xvreplfr2vr_s( d0 );
         const __m256 d1v = __lasx_xvreplfr2vr_s( d1 );
@@ -834,7 +835,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -865,7 +866,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)); //FIXME
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)); //FIXME
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -902,7 +903,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -934,16 +935,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = __lasx_xvand_v(bxhi, __lasx_xvreplgr2vr_b(0x10));
         qx = __lasx_xvor_v(qx, bxhi);
 
-        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -973,7 +974,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -1003,7 +1004,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = __lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = __lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
         __m256i qx = __lasx_xvld((const __m256i *)x[ib].qs, 0);
         __m256i qy = __lasx_xvld((const __m256i *)y[ib].qs, 0);
 
@@ -1023,7 +1024,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1047,8 +1048,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1116,8 +1117,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1170,7 +1171,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
         // Set up scales
@@ -1294,7 +1295,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1330,8 +1331,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1438,9 +1439,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1593,9 +1594,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1624,7 +1625,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1713,7 +1714,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1780,7 +1781,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = __lasx_xvldi(0);
@@ -1820,7 +1821,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -1895,7 +1896,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -1980,7 +1981,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,7 +2050,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -2108,7 +2109,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -2168,7 +2169,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2213,7 +2214,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2279,7 +2280,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = (__m256)__lasx_xvldi(0);
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -2340,7 +2341,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2451,7 +2452,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(d), __lasx_xvffint_s_w(sumi), accum);
         accum1 += d * sumi1;
     }
@@ -2484,7 +2485,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -2530,9 +2531,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = lasx_madd_h(p16_1, mone);
         const __m256i p_2 = lasx_madd_h(p16_2, mone);
-        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                 __lasx_xvffint_s_w(p_1), accum1);
-        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                 __lasx_xvffint_s_w(p_2), accum2);
     }
 
@@ -2540,7 +2541,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2595,7 +2596,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi1 = __lasx_xvadd_w(p_1, sumi1);
             sumi2 = __lasx_xvadd_w(p_2, sumi2);
         }
-        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = __lasx_xvfmadd_s(__lasx_xvreplfr2vr_s(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 __lasx_xvffint_s_w(__lasx_xvadd_w(sumi1, sumi2)), accum);
     }
 
@@ -2604,7 +2605,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

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

@@ -0,0 +1,82 @@
+# include "ggml-backend-impl.h"
+
+#if defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)
+
+#if defined(__linux__)
+#include <sys/auxv.h>
+#endif
+
+#include <string>
+
+struct powerpc_features {
+    std::string platform = "";
+    int power_version    = -1;
+
+    bool has_vsx         = false;
+
+    powerpc_features() {
+#if defined(__linux__)
+        unsigned long auxval = getauxval(AT_PLATFORM);
+        if (auxval) {
+            platform = std::string(reinterpret_cast<const char*>(auxval));
+            // TBD: Do systems exist that return this in uppercase?
+            if (platform.substr(0, 5) == "power") {
+                // Extractt a numeric suffix, if one exists
+                int vpos = -1;
+                for (int i = platform.length() - 1; i >= 0; i--) {
+                    if (std::isdigit(platform[i])) {
+                        vpos = i;
+                    } else {
+                        break;
+                    }
+                }
+                if (vpos > -1) {
+                    power_version = std::stoi(platform.substr(vpos));
+                }
+            }
+        }
+#endif
+        if (power_version >= 9) {
+            has_vsx = true;
+        }
+    }
+};
+
+static int ggml_backend_cpu_powerpc_score() {
+    int score = 1;
+    powerpc_features pf;
+
+// Platform scores
+#if defined(GGML_USE_POWER7)
+    if (pf.power_version < 7) { return 0; }
+    score += 1<<1;
+#endif
+#if defined(GGML_USE_POWER8)
+    if (pf.power_version < 8) { return 0; }
+    score += 1<<2;
+#endif
+#if defined(GGML_USE_POWER9)
+    if (pf.power_version < 9) { return 0; }
+    score += 1<<3;
+#endif
+#if defined(GGML_USE_POWER10)
+    if (pf.power_version < 10) { return 0; }
+    score += 1<<4;
+#endif
+#if defined(GGML_USE_POWER11)
+    if (pf.power_version < 11) { return 0; }
+    score += 1<<5;
+#endif
+
+// Feature scores
+#if defined(GGML_USE_VSX)
+    if (!pf.has_vsx) { return 0; }
+    score += 1<<6;
+#endif
+
+    return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_powerpc_score)
+
+#endif // defined(__powerpc64__) || defined(__ppc64__) || defined(__PPC64__)

ggml/src/ggml-cpu/arch/powerpc/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -67,7 +68,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float id = d ? 1.0f/d : 0.0f;
         const vector float vid = vec_splats(id);
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const vector float v  = vec_round(vec_mul(srcv[j], vid));
@@ -112,7 +113,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float id = d ? 1.0f/d : 0.0f;
         const vector float vid = vec_splats(id);
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vector int accv = vec_splats(0);
 
@@ -127,7 +128,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         accv = vec_add(accv, vec_sld(accv, accv, 4));
         accv = vec_add(accv, vec_sld(accv, accv, 8));
-        y[i].s = GGML_FP32_TO_FP16(d * vec_extract(accv, 0));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * vec_extract(accv, 0));
     }
 
 #else
@@ -170,8 +171,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -214,7 +215,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -249,12 +250,12 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.0f, 0.0f, 0.0f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -291,7 +292,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -326,8 +327,8 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed long long aux64x2_0 = {(uint64_t)(table_b2b_1[x[ib].qh[0]]), (uint64_t)(table_b2b_1[x[ib].qh[1]])};
@@ -379,7 +380,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -415,12 +416,12 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[ib].m));
-        vector float vys = {GGML_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].m));
+        vector float vys = {GGML_CPU_FP16_TO_FP32(y[ib].s), 0.f, 0.f, 0.f};
         vsumf0 = vec_madd(vxmin, vys, vsumf0);
 
         vector unsigned long long aux64x2_0 = {(uint64_t)(table_b2b_0[x[ib].qh[0]]), (uint64_t)(table_b2b_0[x[ib].qh[1]])};
@@ -470,7 +471,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -502,8 +503,8 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char q8x0 = vec_xl( 0, x[ib].qs);
@@ -542,7 +543,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -574,11 +575,11 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -708,8 +709,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -770,7 +771,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -962,7 +963,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1005,11 +1006,11 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         vector signed short q8ysums0 = vec_xl( 0, y[i].bsums);
@@ -1177,9 +1178,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1222,11 +1223,11 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
-        vector float vxmin = vec_splats(GGML_FP16_TO_FP32(x[i].dmin));
+        vector float vxmin = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].dmin));
         vector float vdmin = vec_mul(vxmin, vyd);
 
         UNUSED(kmask1);
@@ -1394,9 +1395,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1432,7 +1433,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1591,7 +1592,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1659,7 +1660,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1742,7 +1743,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -1790,7 +1791,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
     const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -1871,7 +1872,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -1939,7 +1940,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2033,7 +2034,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -2096,7 +2097,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2176,7 +2177,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2236,7 +2237,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     const vector signed char mask2 = (vector signed char)vec_xl( 0, k_mask2);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2329,7 +2330,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -2394,7 +2395,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     vector float vsumf3 = vec_splats(0.0f);
 
     for (int i = 0; i < nb; ++i) {
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[i].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[i].d));
         vector float vyd = vec_splats(y[i].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2505,7 +2506,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -2546,8 +2547,8 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         __builtin_prefetch(y[ib].qs, 0, 1);
 
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ib].d));
-        vector float vyd = vec_splats(GGML_FP16_TO_FP32(y[ib].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        vector float vyd = vec_splats(GGML_CPU_FP16_TO_FP32(y[ib].d));
         vector float vd = vec_mul(vxd, vyd);
 
         vector signed char qxs = (vector signed char)vec_xl( 0, x[ib].qs);
@@ -2582,7 +2583,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -2620,7 +2621,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     for (int ibl = 0; ibl < nb; ++ibl) {
 
-        vector float vxd = vec_splats(GGML_FP16_TO_FP32(x[ibl].d));
+        vector float vxd = vec_splats(GGML_CPU_FP16_TO_FP32(x[ibl].d));
         vector float vyd = vec_splats(y[ibl].d);
         vector float vd = vec_mul(vxd, vyd);
 
@@ -2697,7 +2698,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/riscv/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -45,7 +46,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
 
@@ -85,7 +86,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d  = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         vfloat32m8_t x0 = __riscv_vfmul_vf_f32m8(v_x, id, vl);
 
@@ -102,7 +103,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // set y[i].s
         int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
-        y[i].s = GGML_FP32_TO_FP16(sum*d);
+        y[i].s = GGML_CPU_FP32_TO_FP16(sum*d);
     }
 
 #else
@@ -160,7 +161,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
 #endif
@@ -177,7 +178,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -225,7 +226,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
 #endif
@@ -242,7 +243,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -293,7 +294,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
         int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
 #endif
@@ -316,7 +317,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -366,7 +367,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         vint32m1_t sum = __riscv_vwredsum_vs_i16m4_i32m1(mul, zero, vl);
         int32_t sumi = __riscv_vmv_x_s_i32m1_i32(sum);
 
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
 #endif
@@ -389,7 +390,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -427,7 +428,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         int sumi = __riscv_vmv_x_s_i32m1_i32(v_sum);
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
 #endif
@@ -438,7 +439,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -465,8 +466,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * q2 = x[i].qs;
         const  int8_t * q8 = y[i].qs;
         const uint8_t * sc = x[i].scales;
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
         uint8_t *patmp = atmp;
         int vsums;
         int tmp;
@@ -569,8 +570,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const int8_t *  q8 = y[i].qs;
             const uint8_t * sc = x[i].scales;
 
-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             size_t vl = 16;
 
@@ -644,8 +645,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             const uint8_t * q2 = x[i].qs;
             const  int8_t * q8 = y[i].qs;
             const uint8_t * sc = x[i].scales;
-            const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
             uint8_t *patmp = atmp;
             int vsums;
             int tmp;
@@ -750,8 +751,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -916,7 +917,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             q3 += 32;    q8 += 128;   scale += 8;
         }
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         sumf += d * isum;
     }
 
@@ -1017,7 +1018,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             }
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             sumf += d*sum_t;
 
@@ -1134,7 +1135,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
                 q3 += 32;    q8 += 128;   scale += 8;
             }
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
             sumf += d * isum;
         }
         break;
@@ -1202,7 +1203,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1239,8 +1240,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int tmp, tmp2, sumi;
         __asm__ __volatile__(
@@ -1361,8 +1362,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
             size_t vl = 8;
 
-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             vint16mf2_t q8sums_0 = __riscv_vlse16_v_i16mf2(y[i].bsums, 4, vl);
             vint16mf2_t q8sums_1 = __riscv_vlse16_v_i16mf2(y[i].bsums+1, 4, vl);
@@ -1422,8 +1423,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         break;
     case 128:
         for (int i = 0; i < nb; ++i) {
-            const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-            const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+            const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+            const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
             int tmp, tmp2, sumi;
             __asm__ __volatile__(
@@ -1580,9 +1581,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1627,8 +1628,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT hm = x[i].qh;
         const  int8_t * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
 
         vint16m1_t q8sums_0 = __riscv_vlse16_v_i16m1(y[i].bsums, 4, vl);
         vint16m1_t q8sums_1 = __riscv_vlse16_v_i16m1(y[i].bsums+1, 4, vl);
@@ -1749,9 +1750,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1778,7 +1779,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
         const uint8_t * restrict q6 = x[i].ql;
         const uint8_t * restrict qh = x[i].qh;
@@ -1862,7 +1863,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 256:
         for (int i = 0; i < nb; ++i) {
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * GGML_RESTRICT q6 = x[i].ql;
             const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1943,7 +1944,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     case 128:
         for (int i = 0; i < nb; ++i) {
 
-            const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+            const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 
             const uint8_t * restrict q6 = x[i].ql;
             const uint8_t * restrict qh = x[i].qh;
@@ -2058,7 +2059,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];

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

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include <cmath>
@@ -90,16 +91,16 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
 
                 // vector version needs Zvfhmin extension
-                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float a_scale = GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 const float b_scales[8] = {
-                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                 };
                 const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
                 const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
@@ -129,7 +130,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -181,20 +182,20 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
                     // vector version needs Zvfhmin extension
                     const float a_scales[4] = {
-                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(a_ptr[l].d[3])
                     };
                     const float b_scales[8] = {
-                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
-                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_CPU_FP16_TO_FP32(b_ptr[l].d[7])
                     };
                     const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
 
@@ -382,7 +383,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }

ggml/src/ggml-cpu/arch/s390/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -49,7 +50,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f / d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const __vector float v = vec_mul(srcv[j], vec_splats(id));
@@ -94,7 +95,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f / d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         __vector int32_t acc = vec_splats(0);
 
@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             acc = vec_add(acc, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * (acc[0] + acc[1] + acc[2] + acc[3]));
     }
 #else
     GGML_UNUSED(nb);
@@ -164,7 +165,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __vector int16_t v_xy_ = v_xylso + v_xylse + v_xyhso + v_xyhse; v_xy_ += vec_reve(v_xy_);
 
         const __vector float v_xy = vec_float(vec_unpackh(v_xy_));
-        const __vector float v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __vector float v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -185,7 +186,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -219,7 +220,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         __builtin_prefetch(x[ib].qs, 0, 1);
         __builtin_prefetch(y[ib].qs, 0, 1);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const uint8x16_t v_x = vec_xl(0, x[ib].qs);
         const int8x16_t v_xl = (const int8x16_t)(v_x & v_m);
@@ -231,7 +232,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
         const float32x4_t v_xy = vec_float(v_xy_);
 
-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -252,7 +253,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
         const int32x4_t v_xy_ = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
         const float32x4_t v_xy = vec_float(v_xy_);
-        const float32x4_t v_d = vec_splats(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         acc = vec_madd(v_xy, v_d, acc);
     }
@@ -305,7 +306,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -348,7 +349,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sum = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * restrict x0l = x[i].qs;
         const uint8_t * restrict x0h = x[i].hmask;
@@ -497,7 +498,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -537,8 +538,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
         const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -647,9 +648,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -698,8 +699,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const int16x8_t v_ysumsl = vec_xl(0 , y[i].bsums);
         const int16x8_t v_ysumsh = vec_xl(16, y[i].bsums);
@@ -819,9 +820,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -859,7 +860,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     int8x16_t  v_y[4];
 
     for (int i = 0; i < nb; ++i) {
-        const float d_all = GGML_FP16_TO_FP32(x[i].d);
+        const float d_all = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT x0l = x[i].ql;
         const uint8_t * GGML_RESTRICT x0h = x[i].qh;
@@ -1004,7 +1005,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1071,7 +1072,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 //    float sumf = 0;
 
 //    for (int i = 0; i < nb; ++i) {
-//        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //        const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //        const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -1121,7 +1122,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
 //     float sumf = 0.f;
 //     for (int i = 0; i < nb; ++i) {
-//         const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+//         const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
 //         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
 //         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 //         int32_t bsum = 0;
@@ -1182,12 +1183,12 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const int8x16_t v_yh = vec_xl(QK8_0/2, y0->qs);
         const int32x4_t v_xy = ggml_vec_dot(ggml_vec_dot(vec_splats(0), v_xl, v_yl), v_xh, v_yh);
 
-        sumf += GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
+        sumf += GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d) * (v_xy[0] + v_xy[1] + v_xy[2] + v_xy[3]);
     }
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1257,7 +1258,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi2 += (vsumi1[0] + vsumi1[1] + vsumi1[2] + vsumi1[3]) * ls2;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d * (sumi1 + sumi2);
     }
 
     *s = sumf;
@@ -1265,7 +1266,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/wasm/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -65,7 +66,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         for (int j = 0; j < 8; j++) {
             const v128_t v  = wasm_f32x4_mul(srcv[j], wasm_f32x4_splat(id));
@@ -110,7 +111,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         const float d = amax / ((1 << 7) - 1);
         const float id = d ? 1.0f/d : 0.0f;
 
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
 
         v128_t accv = wasm_i32x4_splat(0);
 
@@ -126,7 +127,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
             accv = wasm_i32x4_add(accv, vi);
         }
 
-        y[i].s = GGML_FP32_TO_FP16(
+        y[i].s = GGML_CPU_FP32_TO_FP16(
                 d * (wasm_i32x4_extract_lane(accv, 0) +
                      wasm_i32x4_extract_lane(accv, 1) +
                      wasm_i32x4_extract_lane(accv, 2) +
@@ -324,8 +325,8 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         );
 
         // Accumulate results with scaling
-        float scale0 = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
-        float scale1 = GGML_FP16_TO_FP32(x1->d) * GGML_FP16_TO_FP32(y1->d);
+        float scale0 = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
+        float scale1 = GGML_CPU_FP16_TO_FP32(x1->d) * GGML_CPU_FP16_TO_FP32(y1->d);
 
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp0), wasm_f32x4_splat(scale0)));
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(dp1), wasm_f32x4_splat(scale1)));
@@ -348,7 +349,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -428,7 +429,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                             wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                            wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
     }
 
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -454,7 +455,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -491,7 +492,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const block_q5_1 * GGML_RESTRICT x0 = &x[ib];
         const block_q8_1 * GGML_RESTRICT y0 = &y[ib];
 
-        summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
+        summs += GGML_CPU_FP16_TO_FP32(x0->m) * GGML_CPU_FP16_TO_FP32(y0->s);
 
         const v128_t m4b = wasm_i8x16_splat(0x0F);
 
@@ -538,7 +539,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
                                            wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
                             wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
                                            wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
-                    wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
+                    wasm_f32x4_splat(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d))));
     }
 
     sumf = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@@ -564,7 +565,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -620,7 +621,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const v128_t sum_dots = wasm_i32x4_add(wasm_i32x4_add(dx0_0, dx0_1), wasm_i32x4_add(dx1_0, dx1_1));
 
         // Convert to float and accumulate
-        const float scale = GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d);
+        const float scale = GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d);
         sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(sum_dots), wasm_f32x4_splat(scale)));
     }
 
@@ -635,7 +636,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -746,8 +747,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             isum += wasm_i32x4_extract_lane(isum_vec, 0);
         }
 
-        const float dall = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dall = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf += dall * isum - dmin * summs;
     }
 
@@ -768,8 +769,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -880,7 +881,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         // Accumulate results
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const v128_t v_d = wasm_f32x4_splat(d);
         v128_t v_sum = wasm_f32x4_add(
             wasm_f32x4_mul(wasm_f32x4_convert_i32x4(v_acc0), v_d),
@@ -957,7 +958,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -991,8 +992,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Corrected sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Corrected sign
 
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1136,9 +1137,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1170,8 +1171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
     float sumf = 0;
 
     for (int i = 0; i < nb; ++i) {
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin); // Fixed sign
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); // Fixed sign
 
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -1331,9 +1332,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1420,7 +1421,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         wasm_v128_store(&aux32[0], acc0);
         wasm_v128_store(&aux32[4], acc1);
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) {
             sums[l] += d * aux32[l];
         }
@@ -1470,7 +1471,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];

ggml/src/ggml-cpu/arch/x86/quants.c

@@ -3,6 +3,7 @@
 #include "ggml-quants.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
+#include "simd-mappings.h"
 
 #include "../../quants.h"
 #include "../../ggml-cpu-impl.h"
@@ -256,9 +257,9 @@ static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_
 
 // quad fp16 delta calculation
 static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
-    // GGML_FP16_TO_FP32 is faster than Intel F16C
-    return _mm256_set_m128(_mm_set1_ps(GGML_FP16_TO_FP32(x1) * GGML_FP16_TO_FP32(y1)),
-                           _mm_set1_ps(GGML_FP16_TO_FP32(x0) * GGML_FP16_TO_FP32(y0)));
+    // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
+    return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+                           _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
 }
 #endif
 #elif defined(__SSSE3__)
@@ -305,7 +306,7 @@ void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = maxScalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -401,7 +402,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
         // Quantize these floats
         const float d = max_scalar / 127.f;
-        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].d = GGML_CPU_FP32_TO_FP16(d);
         const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
         const __m256 mul = _mm256_set1_ps( id );
 
@@ -425,7 +426,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
 
 #if defined(__AVX2__)
         // Compute the sum of the quants and set y[i].s
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
 
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
@@ -455,7 +456,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, i
         // Compute the sum of the quants and set y[i].s
         const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
         const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
-        y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+        y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
 
         // Convert int32 to int16
         ni0 = _mm_packs_epi32( ni0, ni1 );
@@ -552,7 +553,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
 
@@ -613,7 +614,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 0 and 1
-        const __m128 d_0_1 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d) );
+        const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
 
         const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
 
@@ -631,7 +632,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
 
         // Compute combined scale for the block 2 and 3
-        const __m128 d_2_3 = _mm_set1_ps( GGML_FP16_TO_FP32(x[ib + 1].d) * GGML_FP16_TO_FP32(y[ib + 1].d) );
+        const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
 
         const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
 
@@ -680,7 +681,7 @@ void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -711,10 +712,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const float d0 = GGML_FP16_TO_FP32(x[ib].d);
-        const float d1 = GGML_FP16_TO_FP32(y[ib].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         const __m256 d0v = _mm256_set1_ps( d0 );
         const __m256 d1v = _mm256_set1_ps( d1 );
@@ -752,7 +753,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -783,7 +784,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -807,7 +808,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         /* Compute combined scale for the block */
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
 
         __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
         const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -851,7 +852,7 @@ void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -883,16 +884,16 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i qx = bytes_from_nibbles_32(x[ib].qs);
         __m256i bxhi = bytes_from_bits_32(x[ib].qh);
         bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
         qx = _mm256_or_si256(qx, bxhi);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@@ -910,9 +911,9 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     // Main loop
     for (; ib < nb; ++ib) {
-        const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d));
+        const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
 
-        summs += GGML_FP16_TO_FP32(x[ib].m) * GGML_FP16_TO_FP32(y[ib].s);
+        summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
 
         __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
         const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
@@ -926,7 +927,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         bxh = _mm_or_si128(bxh, bxhih);
         bx_0 = MM256_SET_M128I(bxh, bxl);
 
-        const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
         const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
         const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@@ -956,7 +957,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -986,7 +987,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     // Main loop
     for (; ib < nb; ++ib) {
         // Compute combined scale for the block
-        const __m256 d = _mm256_set1_ps(GGML_FP16_TO_FP32(x[ib].d) * GGML_FP16_TO_FP32(y[ib].d));
+        const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
         __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
         __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
 
@@ -1025,7 +1026,7 @@ void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -1144,7 +1145,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         }
 
         const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
 
         sumi0 = _mm256_sub_epi16(sumi0, ysum);
         sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
@@ -1190,7 +1191,7 @@ void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -1244,7 +1245,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
         }
 
         const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(x[i].d));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
 
         sumi0 = _mm256_add_epi16(sumi0, sumi1);
         sumi0 = _mm256_sub_epi16(sumi0, ysum);
@@ -1269,7 +1270,7 @@ void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             }
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -1299,8 +1300,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1366,8 +1367,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1477,8 +1478,8 @@ void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -1533,7 +1534,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1638,7 +1639,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -1824,7 +1825,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -1862,8 +1863,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -1928,8 +1929,8 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
    for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2049,9 +2050,9 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2092,8 +2093,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         memcpy(utmp, x[i].scales, 12);
         utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
@@ -2170,8 +2171,8 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = -y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         const uint8_t * GGML_RESTRICT q5 = x[i].qs;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -2311,9 +2312,9 @@ void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2344,7 +2345,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2422,7 +2423,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 
     for (int i = 0; i < nb; ++i) {
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         const uint8_t * GGML_RESTRICT q4 = x[i].ql;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
@@ -2555,7 +2556,7 @@ void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -2622,7 +2623,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m256i sumi1 = _mm256_setzero_si256();
@@ -2663,7 +2664,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         __m128i sumi1_0 = _mm_setzero_si128();
@@ -2717,7 +2718,7 @@ void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -2792,7 +2793,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -2913,7 +2914,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
 
@@ -3035,7 +3036,7 @@ void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -3104,7 +3105,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3177,7 +3178,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
@@ -3253,7 +3254,7 @@ void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -3313,7 +3314,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3358,7 +3359,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3414,7 +3415,7 @@ void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -3480,7 +3481,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3565,7 +3566,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     __m256 accumf = _mm256_setzero_ps();
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
@@ -3648,7 +3649,7 @@ void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -3753,7 +3754,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
         accum1 += d * sumi1;
 
@@ -3801,7 +3802,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                    + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
         accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
         accum1 += d * sumi1;
 
@@ -3835,7 +3836,7 @@ void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -3947,7 +3948,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 8; qh += 4;
         }
 
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
 
         accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
         accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
@@ -4033,7 +4034,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qs += 8; qh += 4;
         }
 
-        const __m256 d = _mm256_set1_ps(y[i].d * GGML_FP16_TO_FP32(scale.f16));
+        const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
 
         accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
         accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
@@ -4083,7 +4084,7 @@ void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
             qh += 2;
         }
 
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -4129,9 +4130,9 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
         const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
         const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
         const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
-        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 0].d)*GGML_FP16_TO_FP32(x[ib + 0].d)),
+        accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
                 _mm256_cvtepi32_ps(p_1), accum1);
-        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(y[ib + 1].d)*GGML_FP16_TO_FP32(x[ib + 1].d)),
+        accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
                 _mm256_cvtepi32_ps(p_2), accum2);
     }
 
@@ -4164,7 +4165,7 @@ void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const v
 
 #endif
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -4219,7 +4220,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
             sumi1 = _mm256_add_epi32(p_1, sumi1);
             sumi2 = _mm256_add_epi32(p_2, sumi2);
         }
-        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
     }
 
@@ -4267,7 +4268,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
         }
         __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
         __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
-        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+        accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
                 _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
     }
 
@@ -4276,7 +4277,7 @@ void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const v
 #else
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/arch/x86/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include <cmath>
@@ -39,11 +40,11 @@ static inline __m512 __avx512_f32cx8x2_load(ggml_fp16_t *x, ggml_fp16_t *y) {
     float tmp[16];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     for (int i = 0; i < 8; i++) {
-        tmp[i + 8] = GGML_FP16_TO_FP32(y[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(y[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -54,10 +55,10 @@ static inline __m512 __avx512_repeat_f32cx16_load(__m128i x) {
     _mm_storeu_si128((__m128i*)tmphalf, x);
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 8] = GGML_FP16_TO_FP32(tmphalf[i]);
-        tmp[i + 12] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 8] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
+        tmp[i + 12] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm512_loadu_ps(tmp);
@@ -67,7 +68,7 @@ static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -76,8 +77,8 @@ static inline __m256 __avx_repeat_f32cx8_load(ggml_fp16_t *x) {
     float tmp[8];
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
-        tmp[i + 4] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+        tmp[i + 4] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -88,7 +89,7 @@ static inline __m256 __avx_rearranged_f32cx8_load(ggml_fp16_t *x, __m128i arrang
 
     _mm_storeu_si128((__m128i*)tmphalf, _mm_shuffle_epi8(_mm_loadu_si128((const __m128i *) x), arrangeMask));
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(tmphalf[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(tmphalf[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -211,7 +212,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             id[row_iter] = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f; //d ? 1.0f / d : 0.0f;
 
             // Store the scale for the individual block
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
 
             // Store the values in blocks of eight values - Aim is to use these later for block interleaving
             srcv[row_iter][0] = v0;
@@ -297,7 +298,7 @@ void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTR
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -647,7 +648,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                 const __m256 col_scale_f32 = GGML_F32Cx8_REARRANGE_LOAD(b_ptr[b].d, changemask);
 
                 // Load and convert to FP32 scale from block_q8_0
-                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_FP16_TO_FP32(a_ptr[b].d));
+                const __m256 row_scale_f32 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(a_ptr[b].d));
 
                 // Load the block values in block_q8_0 in batches of 16 bytes and replicate the same across 256 bit vector
                 __m256i lhs_vec_0 = _mm256_castsi128_si256(_mm_loadu_si128((const __m128i *)a_ptr[b].qs));
@@ -706,7 +707,7 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -972,13 +973,13 @@ void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                 }
             }
             for (int sb = 0; sb < 8; sb++) {
                 uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -1755,7 +1756,7 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -3259,7 +3260,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                                 sumi2 = sumi2 * scales_1[j];
                                 sumi += sumi1 + sumi2;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -3268,7 +3269,7 @@ void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const vo
                     for(int m = 0; m < 4; m++) {
                         const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                         for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                         }
                     }
                 }

ggml/src/ggml-cpu/common.h

@@ -4,6 +4,7 @@
 #include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
+#include "simd-mappings.h"
 
 #ifdef __cplusplus
 
@@ -12,11 +13,11 @@
 // convenience functions/macros for use in template calls
 // note: these won't be required after the 'traits' lookup table is used.
 static inline ggml_fp16_t f32_to_f16(float x) {
-    return GGML_FP32_TO_FP16(x);
+    return GGML_CPU_FP32_TO_FP16(x);
 }
 
 static inline float f16_to_f32(ggml_fp16_t x) {
-    return GGML_FP16_TO_FP32(x);
+    return GGML_CPU_FP16_TO_FP32(x);
 }
 
 static inline ggml_bf16_t f32_to_bf16(float x) {

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

@@ -62,11 +62,17 @@ struct ggml_compute_params {
 #if defined(__s390x__) && defined(__VEC__)
 #ifndef __VXE__
 #define __VXE__
-#endif
+#endif  // __VXE__
 #ifndef __VXE2__
 #define __VXE2__
-#endif
-#endif
+#endif  // __VXE2__
+#endif  // __s390x__ && __VEC__
+
+#if defined(__s390x__) && defined(GGML_NNPA)
+#ifndef __NNPA__
+#define __NNPA__
+#endif  // __NNPA__
+#endif  // __s390x__ && GGML_NNPA
 
 #if defined(__ARM_FEATURE_SVE)
 #include <sys/prctl.h>
@@ -371,7 +377,7 @@ inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b)
 #define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
 #endif
 
-typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef signed   char char8x16_t  __attribute__((vector_size(16)));
 typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
 
 typedef int8_t  int8x16_t __attribute__((vector_size(16)));
@@ -382,10 +388,10 @@ typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
 typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
 typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
 
-typedef float float32x4_t __attribute__((vector_size(16)));
-typedef double double64x2_t __attribute((vector_size(16)));
+typedef float  float32x4_t  __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));
 
-typedef signed long long long64x2_t __attribute((vector_size(16)));
+typedef signed   long long long64x2_t  __attribute__((vector_size(16)));
 typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
 
 typedef struct ggml_uint8x16x2_t {

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

@@ -72,15 +72,13 @@
 #define UNUSED GGML_UNUSED
 #define SWAP(x, y, T) do { T SWAP = x; (x) = y; (y) = SWAP; } while (0)
 
+// precomputed f32 table for f16 (256 KB) (simd-mappings.h)
+float ggml_table_f32_f16[1 << 16];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
-    int has_neon;
-    int has_dotprod;
-    int has_i8mm;
-    int has_sve;
     int sve_cnt;
-    int has_sme;
-} ggml_arm_arch_features = {-1, -1, -1, -1, 0, -1};
+} ggml_arm_arch_features = { 0 };
 #endif
 
 
@@ -197,6 +195,7 @@ typedef pthread_t ggml_thread_t;
 
 static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32] = {
+        .from_float               = (ggml_from_float_t) ggml_cpu_fp32_to_fp32,
         .vec_dot                  = (ggml_vec_dot_t) ggml_vec_dot_f32,
         .vec_dot_type             = GGML_TYPE_F32,
         .nrows                    = 1,
@@ -678,87 +677,15 @@ bool ggml_is_numa(void) {
 
 #if defined(__linux__) && defined(__aarch64__)
 #include <sys/auxv.h>
-#elif defined(__APPLE__)
-#include <sys/sysctl.h>
-#endif
-
-#if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM (1 << 13)
-#endif
-
-#if !defined(HWCAP2_SME)
-#define HWCAP2_SME (1 << 23)
 #endif
 
 static void ggml_init_arm_arch_features(void) {
-#if defined(__linux__) && defined(__aarch64__)
-    uint32_t hwcap = getauxval(AT_HWCAP);
-    uint32_t hwcap2 = getauxval(AT_HWCAP2);
-
-    ggml_arm_arch_features.has_neon    = !!(hwcap & HWCAP_ASIMD);
-    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
-    ggml_arm_arch_features.has_i8mm    = !!(hwcap2 & HWCAP2_I8MM);
-    ggml_arm_arch_features.has_sve     = !!(hwcap & HWCAP_SVE);
-    ggml_arm_arch_features.has_sme     = !!(hwcap2 & HWCAP2_SME);
-
-#if defined(__ARM_FEATURE_SVE)
+#if defined(__linux__) && defined(__aarch64__) && defined(__ARM_FEATURE_SVE)
     ggml_arm_arch_features.sve_cnt = PR_SVE_VL_LEN_MASK & prctl(PR_SVE_GET_VL);
 #endif
-#elif defined(__APPLE__)
-    int oldp = 0;
-    size_t size = sizeof(oldp);
-    if (sysctlbyname("hw.optional.AdvSIMD", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_neon = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_dotprod = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_i8mm = oldp;
-
-    if (sysctlbyname("hw.optional.arm.FEAT_SME", &oldp, &size, NULL, 0) != 0) {
-        oldp = 0;
-    }
-    ggml_arm_arch_features.has_sme = oldp;
-
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#else
-// Run-time CPU feature detection not implemented for this platform, fallback to compile time
-#if defined(__ARM_NEON)
-    ggml_arm_arch_features.has_neon = 1;
-#else
-    ggml_arm_arch_features.has_neon = 0;
-#endif
-
-#if defined(__ARM_FEATURE_MATMUL_INT8)
-    ggml_arm_arch_features.has_i8mm = 1;
-#else
-    ggml_arm_arch_features.has_i8mm = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SVE)
-    ggml_arm_arch_features.has_sve = 1;
-    ggml_arm_arch_features.sve_cnt = 16;
-#else
-    ggml_arm_arch_features.has_sve = 0;
-    ggml_arm_arch_features.sve_cnt = 0;
-#endif
-
-#if defined(__ARM_FEATURE_SME) || defined(__ARM_FEATURE_SME2)
-    ggml_arm_arch_features.has_sme = 1;
-#else
-    ggml_arm_arch_features.has_sme = 0;
-#endif
-#endif
 }
-#endif
+
+#endif // __ARM_ARCH
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value) {
     GGML_ASSERT(!ggml_get_no_alloc(ctx));
@@ -813,7 +740,7 @@ struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -872,7 +799,7 @@ struct ggml_tensor * ggml_set_f32(struct ggml_tensor * tensor, float value) {
             {
                 assert(tensor->nb[0] == sizeof(ggml_fp16_t));
                 for (int i = 0; i < n; i++) {
-                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_FP32_TO_FP16(value));
+                    ggml_vec_set_f16(nc, (ggml_fp16_t *)(data + i*n1), GGML_CPU_FP32_TO_FP16(value));
                 }
             } break;
         case GGML_TYPE_BF16:
@@ -923,7 +850,7 @@ int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -968,7 +895,7 @@ void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value) {
         case GGML_TYPE_F16:
             {
                 GGML_ASSERT(tensor->nb[0] == sizeof(ggml_fp16_t));
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -997,7 +924,7 @@ int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1024,7 +951,7 @@ void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1062,7 +989,7 @@ float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i) {
             }
         case GGML_TYPE_F16:
             {
-                return GGML_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
+                return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *)(tensor->data))[i]);
             }
         case GGML_TYPE_BF16:
             {
@@ -1101,7 +1028,7 @@ void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value) {
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(tensor->data))[i] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(tensor->data))[i] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1128,7 +1055,7 @@ float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
         case GGML_TYPE_I32:
             return ((int32_t *) data)[0];
         case GGML_TYPE_F16:
-            return GGML_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
+            return GGML_CPU_FP16_TO_FP32(((ggml_fp16_t *) data)[0]);
         case GGML_TYPE_BF16:
             return GGML_BF16_TO_FP32(((ggml_bf16_t *) data)[0]);
         case GGML_TYPE_F32:
@@ -1155,7 +1082,7 @@ void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2,
             } break;
         case GGML_TYPE_F16:
             {
-                ((ggml_fp16_t *)(data))[0] = GGML_FP32_TO_FP16(value);
+                ((ggml_fp16_t *)(data))[0] = GGML_CPU_FP32_TO_FP16(value);
             } break;
         case GGML_TYPE_BF16:
             {
@@ -1891,6 +1818,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_get_rows_back(params, tensor);
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                ggml_compute_forward_set_rows(params, tensor);
+            } break;
         case GGML_OP_DIAG:
             {
                 ggml_compute_forward_diag(params, tensor);
@@ -1967,6 +1898,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad_reflect_1d(params, tensor);
             } break;
+        case GGML_OP_ROLL:
+            {
+                ggml_compute_forward_roll(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -2240,6 +2175,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
                 n_tasks = n_threads;
             } break;
         case GGML_OP_GET_ROWS:
+        case GGML_OP_SET_ROWS:
             {
                 // FIXME: get_rows can use additional threads, but the cost of launching additional threads
                 // decreases performance with GPU offloading
@@ -2291,6 +2227,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
         case GGML_OP_PAD_REFLECT_1D:
+        case GGML_OP_ROLL:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -3193,6 +3130,10 @@ enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct g
     return ggml_graph_compute(cgraph, &cplan);
 }
 
+void ggml_cpu_fp32_to_fp32(const float * x, float * y, int64_t n) {
+    memcpy(y, x, n * sizeof(float));
+}
+
 void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
     int64_t i = 0;
 #if defined(__F16C__)
@@ -3213,9 +3154,24 @@ void ggml_cpu_fp32_to_fp16(const float * x, ggml_fp16_t * y, int64_t n) {
         __m128i y_vec = _mm_cvtps_ph(x_vec, _MM_FROUND_TO_NEAREST_INT);
         _mm_storel_epi64((__m128i *)(y + i), y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        float32x4_t v_xh = vec_xl(0, (const float *)(x + i + 0));
+        float32x4_t v_xl = vec_xl(0, (const float *)(x + i + 4));
+        uint16x8_t v_yd = vec_round_from_fp32(v_xh, v_xl, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
+    for (; i + 3 < n; i += 4) {
+        float32x4_t v_x = vec_xl(0, (const float *)(x + i));
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_yd = vec_round_from_fp32(v_x, v_zero, 0);
+        uint16x8_t v_y = vec_convert_to_fp16(v_yd, 0);
+        vec_xst(v_y, 0, (ggml_fp16_t *)(y + i));
+    }
 #endif
     for (; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(x[i]);
     }
 }
 
@@ -3239,9 +3195,25 @@ void ggml_cpu_fp16_to_fp32(const ggml_fp16_t * x, float * y, int64_t n) {
         __m128 y_vec = _mm_cvtph_ps(x_vec);
         _mm_storeu_ps(y + i, y_vec);
     }
+#elif defined(__NNPA__)
+    for (; i + 7 < n; i += 8) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        float32x4_t v_yl = vec_extend_to_fp32_lo(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i + 0));
+        vec_xst(v_yl, 0, (float *)(y + i + 4));
+    }
+    for (; i + 3 < n; i += 4) {
+        uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)(x + i));
+        uint16x8_t v_yd = vec_convert_from_fp16(v_x, 0);
+        float32x4_t v_yh = vec_extend_to_fp32_hi(v_yd, 0);
+        vec_xst(v_yh, 0, (float *)(y + i));
+    }
 #endif
+
     for (; i < n; ++i) {
-        y[i] = GGML_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 }
 
@@ -3441,9 +3413,17 @@ int ggml_cpu_has_vxe(void) {
 #endif
 }
 
+int ggml_cpu_has_nnpa(void) {
+#if defined(GGML_NNPA)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_neon(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_NEON)
-    return ggml_arm_arch_features.has_neon;
+    return 1;
 #else
     return 0;
 #endif
@@ -3451,7 +3431,7 @@ int ggml_cpu_has_neon(void) {
 
 int ggml_cpu_has_dotprod(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
-    return ggml_arm_arch_features.has_dotprod;
+    return 1;
 #else
     return 0;
 #endif
@@ -3459,7 +3439,7 @@ int ggml_cpu_has_dotprod(void) {
 
 int ggml_cpu_has_sve(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
-    return ggml_arm_arch_features.has_sve;
+    return 1;
 #else
     return 0;
 #endif
@@ -3467,7 +3447,7 @@ int ggml_cpu_has_sve(void) {
 
 int ggml_cpu_has_matmul_int8(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
-    return ggml_arm_arch_features.has_i8mm;
+    return 1;
 #else
     return 0;
 #endif
@@ -3483,14 +3463,14 @@ int ggml_cpu_get_sve_cnt(void) {
 
 int ggml_cpu_has_sme(void) {
 #if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SME)
-    return ggml_arm_arch_features.has_sme;
+    return 1;
 #else
     return 0;
 #endif
 }
 
 void ggml_cpu_init(void) {
-    // needed to initialize f16 tables
+    // needed to initialize ggml_time
     {
         struct ggml_init_params params = { 0, NULL, false };
         struct ggml_context * ctx = ggml_init(params);
@@ -3511,9 +3491,10 @@ void ggml_cpu_init(void) {
                     uint16_t u16;
                     ggml_fp16_t fp16;
                 } u = {i};
-                float f = GGML_FP16_TO_FP32(u.fp16);
-                ggml_table_gelu_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_f32(f));
-                ggml_table_gelu_quick_f16[i] = GGML_FP32_TO_FP16(ggml_gelu_quick_f32(f));
+                float f = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
+                ggml_table_f32_f16[i] = f;
+                ggml_table_gelu_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_f32(f));
+                ggml_table_gelu_quick_f16[i] = GGML_CPU_FP32_TO_FP16(ggml_gelu_quick_f32(f));
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);

ggml/src/ggml-cpu/ggml-cpu.cpp

@@ -416,6 +416,7 @@ static bool ggml_backend_cpu_device_supports_op(ggml_backend_dev_t dev, const st
 
     switch (op->op) {
         case GGML_OP_CPY:
+        case GGML_OP_SET_ROWS:
             return
                 op->type != GGML_TYPE_IQ3_XXS &&
                 op->type != GGML_TYPE_IQ3_S   &&
@@ -578,6 +579,9 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
         if (ggml_cpu_has_vxe()) {
             features.push_back({ "VXE", "1" });
         }
+        if (ggml_cpu_has_nnpa()) {
+            features.push_back({ "NNPA", "1" });
+        }
         if (ggml_cpu_has_wasm_simd()) {
             features.push_back({ "WASM_SIMD", "1" });
         }

ggml/src/ggml-cpu/llamafile/sgemm.cpp

@@ -52,6 +52,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-quants.h"
+#include "simd-mappings.h"
 
 #include <array>
 #include <type_traits>
@@ -62,7 +63,7 @@
 #define NOINLINE __attribute__((__noinline__))
 #endif
 
-#if defined(__ARM_NEON) || defined(__AVX512F__)
+#if defined(__ARM_NEON) || defined(__AVX512F__) || defined(__VXE__) || defined(__VXE2__)
 #define VECTOR_REGISTERS 32
 #else
 #define VECTOR_REGISTERS 16
@@ -73,7 +74,7 @@
 namespace {
 
 inline float unhalf(ggml_fp16_t d) {
-    return GGML_FP16_TO_FP32(d);
+    return GGML_CPU_FP16_TO_FP32(d);
 }
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
@@ -109,6 +110,12 @@ inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
 inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
+#if defined(__VXE__) || defined(__VXE2__)
+inline float32x4_t add(float32x4_t x, float32x4_t y) { return vec_add(x, y); }
+inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vec_sub(x, y); }
+inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vec_mul(x, y); }
+#endif
+
 #if defined(__MMA__)
 typedef vector unsigned char vec_t;
 typedef __vector_quad acc_t;
@@ -162,6 +169,13 @@ inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
 #endif
 #endif
 
+#if defined(__VXE__) || defined(__VXE2__)
+template <>
+inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
+    return vec_madd(a, b, c);
+}
+#endif
+
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // VECTORIZED HORIZONTAL SUM
 
@@ -178,6 +192,13 @@ inline float hsum(float16x8_t x) {
 }
 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
 
+#if defined(__VXE__) || defined(__VXE2__)
+inline float hsum(float32x4_t x) {
+    float32x4_t tmp = x + vec_reve(x);
+    return tmp[0] + tmp[1];
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 inline float hsum(__m128 x) {
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
@@ -227,6 +248,21 @@ template <> inline float32x4_t load(const ggml_fp16_t *p) {
 #endif // _MSC_VER
 #endif // __ARM_NEON
 
+#if defined(__VXE__) || defined(__VXE2__)
+template <> inline float32x4_t load(const ggml_fp16_t * p) {
+    float tmp[4];
+
+    for (int i = 0; i < 4; i++) {
+        tmp[i] = GGML_CPU_FP16_TO_FP32(p[i]);
+    }
+
+    return vec_xl(0, (const float *)(tmp));
+}
+template <> inline float32x4_t load(const float * p) {
+    return vec_xl(0, p);
+}
+#endif
+
 #if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 template <> inline __m128 load(const float *p) {
     return _mm_loadu_ps(p);
@@ -3319,6 +3355,14 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
             (const float *)B, ldb,
             (float *)C, ldc};
         return tb.matmul(m, n);
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
+            k, (const float *)A, lda,
+            (const float *)B, ldb,
+            (float *)C, ldc};
+        return tb.matmul(m, n);
 #elif defined(__MMA__)
         if (k % 8)
             return false;
@@ -3410,6 +3454,16 @@ bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64
                 (float *)C, ldc};
             return tb.matmul(m, n);
         }
+#elif defined(__VXE__) || defined(__VXE2__)
+        if (n < 4)
+            return false;
+        if (Btype == GGML_TYPE_F16) {
+            tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
+                k, (const ggml_fp16_t *)A, lda,
+                (const ggml_fp16_t *)B, ldb,
+                (float *)C, ldc};
+            return tb.matmul(m, n);
+        }
 #endif
         return false;
     }

ggml/src/ggml-cpu/llamafile/sgemm.h

@@ -1,6 +1,11 @@
 #pragma once
 #include <stdint.h>
 #include <stdbool.h>
+
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+#endif
+
 #ifdef __cplusplus
 extern "C" {
 #endif

ggml/src/ggml-cpu/ops.cpp

@@ -108,7 +108,7 @@ static void ggml_compute_forward_dup_f16(
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
                             for (int i00 = 0; i00 < ne00; i00++) {
-                                dst_ptr[id] = GGML_FP16_TO_FP32(src0_ptr[i00]);
+                                dst_ptr[id] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
                                 id++;
                             }
                         }
@@ -130,7 +130,7 @@ static void ggml_compute_forward_dup_f16(
                             const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
 
                             for (int i00 = 0; i00 < ne00; i00++) {
-                                src0_f32[i00] = GGML_FP16_TO_FP32(src0_ptr[i00]);
+                                src0_f32[i00] = GGML_CPU_FP16_TO_FP32(src0_ptr[i00]);
                             }
 
                             quantize_row_q(src0_f32, dst_ptr + id, ne00);
@@ -156,7 +156,7 @@ static void ggml_compute_forward_dup_f16(
                             for (int i00 = 0; i00 < ne00; i00++) {
                                 const ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
-                                dst_ptr[id] = GGML_FP16_TO_FP32(*src0_ptr);
+                                dst_ptr[id] = GGML_CPU_FP16_TO_FP32(*src0_ptr);
                                 id++;
                             }
                         }
@@ -267,7 +267,7 @@ static void ggml_compute_forward_dup_f16(
                         const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
                               char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                        *(float *) dst_ptr = GGML_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr);
+                        *(float *) dst_ptr = GGML_CPU_FP16_TO_FP32(*(const ggml_fp16_t *) src0_ptr);
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -372,7 +372,7 @@ static void ggml_compute_forward_dup_bf16(
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
                             for (int i00 = 0; i00 < ne00; i00++) {
-                                dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00]));
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(src0_ptr[i00]));
                                 id++;
                             }
                         }
@@ -473,7 +473,7 @@ static void ggml_compute_forward_dup_bf16(
                             for (int i00 = 0; i00 < ne00; i00++) {
                                 const ggml_bf16_t * src0_ptr = (ggml_bf16_t *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
-                                dst_ptr[id] = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr));
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*src0_ptr));
                                 id++;
                             }
                         }
@@ -566,7 +566,7 @@ static void ggml_compute_forward_dup_bf16(
                         const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
                               char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                        *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr));
+                        *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(GGML_BF16_TO_FP32(*(const ggml_bf16_t *) src0_ptr));
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -696,24 +696,8 @@ static void ggml_compute_forward_dup_f32(
     if (ggml_is_contiguous(dst)) {
         // TODO: simplify
         if (nb00 == sizeof(float)) {
-            if (dst->type == GGML_TYPE_F32) {
-                size_t id = 0;
-                const size_t rs = ne00 * nb00;
-                char * dst_ptr = (char *) dst->data;
-
-                for (int i03 = 0; i03 < ne03; i03++) {
-                    for (int i02 = 0; i02 < ne02; i02++) {
-                        id += rs * ir0;
-                        for (int i01 = ir0; i01 < ir1; i01++) {
-                            const char * src0_ptr = (char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03;
-                            memcpy(dst_ptr + id, src0_ptr, rs);
-                            id += rs;
-                        }
-                        id += rs * (ne01 - ir1);
-                    }
-                }
-            } else if (ggml_get_type_traits_cpu(dst->type)->from_float) {
-                ggml_from_float_t const quantize_row_q = ggml_get_type_traits_cpu(dst->type)->from_float;
+            if (ggml_get_type_traits_cpu(dst->type)->from_float) {
+                ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
 
                 size_t id = 0;
                 size_t rs = nb0 * (ne00 / ggml_blck_size(dst->type));
@@ -724,7 +708,7 @@ static void ggml_compute_forward_dup_f32(
                         id += rs * ir0;
                         for (int i01 = ir0; i01 < ir1; i01++) {
                             const float * src0_ptr = (float *) ((char *) src0->data + i01*nb01 + i02*nb02 + i03*nb03);
-                            quantize_row_q(src0_ptr, dst_ptr + id, ne00);
+                            from_float(src0_ptr, dst_ptr + id, ne00);
                             id += rs;
                         }
                         id += rs * (ne01 - ir1);
@@ -765,7 +749,7 @@ static void ggml_compute_forward_dup_f32(
                             for (int i00 = 0; i00 < ne00; i00++) {
                                 const float * src0_ptr = (float *) ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
 
-                                dst_ptr[id] = GGML_FP32_TO_FP16(*src0_ptr);
+                                dst_ptr[id] = GGML_CPU_FP32_TO_FP16(*src0_ptr);
                                 id++;
                             }
                         }
@@ -878,7 +862,7 @@ static void ggml_compute_forward_dup_f32(
                         const char * src0_ptr = ((char *) src0->data + i00*nb00 + i01*nb01 + i02*nb02 + i03*nb03);
                               char * dst_ptr  = ((char *)  dst->data + i10*nb0  + i11*nb1  + i12*nb2  + i13*nb3);
 
-                        *(ggml_fp16_t *) dst_ptr = GGML_FP32_TO_FP16(*(const float *) src0_ptr);
+                        *(ggml_fp16_t *) dst_ptr = GGML_CPU_FP32_TO_FP16(*(const float *) src0_ptr);
 
                         if (++i10 == ne0) {
                             i10 = 0;
@@ -1419,7 +1403,7 @@ static void ggml_compute_forward_add1_f16_f32(
         ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
         ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
         for (int i = 0; i < ne0; i++) {
-            dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
         }
     }
 }
@@ -1435,7 +1419,7 @@ static void ggml_compute_forward_add1_f16_f16(
     GGML_ASSERT(ggml_is_scalar(src1));
 
     // scalar to add
-    const float v = GGML_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
+    const float v = GGML_CPU_FP16_TO_FP32(*(ggml_fp16_t *) src1->data);
 
     const int ith = params->ith;
     const int nth = params->nth;
@@ -1467,7 +1451,7 @@ static void ggml_compute_forward_add1_f16_f16(
         ggml_fp16_t * dst_ptr  = (ggml_fp16_t *) ((char *) dst->data  + i3*nb3  + i2*nb2  + i1*nb1 );
         ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01);
         for (int i = 0; i < ne0; i++) {
-            dst_ptr[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(src0_ptr[i]) + v);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(src0_ptr[i]) + v);
         }
     }
 }
@@ -1889,7 +1873,7 @@ static void ggml_compute_forward_sum_f16(
             }
         }
     }
-    ((ggml_fp16_t *) dst->data)[0] = GGML_FP32_TO_FP16(sum);
+    ((ggml_fp16_t *) dst->data)[0] = GGML_CPU_FP32_TO_FP16(sum);
 }
 
 static void ggml_compute_forward_sum_bf16(
@@ -2300,6 +2284,12 @@ void ggml_compute_forward_repeat(
             {
                 ggml_compute_forward_repeat_f32(params, dst);
             } break;
+        // TODO: templateify the implemenation and support for I64
+        //       ref https://github.com/ggml-org/llama.cpp/pull/14274#discussion_r2169492225
+        //case GGML_TYPE_I64:
+        //    {
+        //        ggml_compute_forward_repeat_i64(params, dst);
+        //    } break;
         default:
             {
                 GGML_ABORT("fatal error");
@@ -2660,7 +2650,7 @@ static void ggml_compute_forward_gelu_f16(
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
-            const float v = GGML_FP16_TO_FP32(x);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2763,7 +2753,7 @@ static void ggml_compute_forward_gelu_erf_f16(
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
-            const float v = GGML_FP16_TO_FP32(x);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2866,7 +2856,7 @@ static void ggml_compute_forward_gelu_quick_f16(
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
-            const float v = GGML_FP16_TO_FP32(x);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -2969,7 +2959,7 @@ static void ggml_compute_forward_silu_f16(
 #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const ggml_fp16_t x = ((ggml_fp16_t *) ((char *) dst->data + i1*(dst->nb[1])))[k];
-            const float v = GGML_FP16_TO_FP32(x);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -3163,7 +3153,7 @@ static void ggml_compute_forward_silu_back_f16(
     #ifndef NDEBUG
         for (int k = 0; k < nc; k++) {
             const float x = ((ggml_fp16_t *) ((char *) dst->data + i1*( dst->nb[1])))[k];
-            const float v = GGML_FP16_TO_FP32(x);
+            const float v = GGML_CPU_FP16_TO_FP32(x);
             GGML_UNUSED(v);
             assert(!isnan(v));
             assert(!isinf(v));
@@ -4470,6 +4460,74 @@ void ggml_compute_forward_get_rows(
     //}
 }
 
+static void ggml_compute_forward_set_rows_f32(
+        const ggml_compute_params * params,
+              ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const ggml_tensor * src1 = dst->src[1];
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    const int64_t nc = ne00;
+    const int64_t nr = ne01;
+
+    assert(ne0  == nc);
+    assert(ne2  == ne02);
+    assert(ne3  == ne03);
+    assert(src0->type == GGML_TYPE_F32);
+    assert(ne02 % ne11 == 0);
+    assert(ne03 % ne12 == 0);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    // rows per thread
+    const int64_t dr = (nr + nth - 1)/nth;
+
+    // row range for this thread
+    const int64_t ir0 = dr*ith;
+    const int64_t ir1 = std::min(ir0 + dr, nr);
+
+    ggml_from_float_t const from_float = ggml_get_type_traits_cpu(dst->type)->from_float;
+
+    for (int64_t i03 = 0; i03 < ne03; ++i03) {
+        for (int64_t i02 = 0; i02 < ne02; ++i02) {
+            for (int64_t i = ir0; i < ir1; ++i) {
+                const int64_t i12 = i03%ne12;
+                const int64_t i11 = i02%ne11;
+                const int64_t i10 = i;
+
+                const int64_t i1 = *(int64_t *) ((char *) src1->data + i10*nb10 + i11*nb11 + i12*nb12);
+
+                GGML_ASSERT(i1 >= 0 && i1 < ne1);
+
+                from_float(
+                        (const float *) ((char *) src0->data +  i*nb01 + i02*nb02 + i03*nb03),
+                                        ((char *)  dst->data + i1*nb1  + i02*nb2  + i03*nb3), nc);
+            }
+        }
+    }
+}
+
+void ggml_compute_forward_set_rows(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_set_rows_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("src0->type = %d (%s) not supported", src0->type, ggml_type_name(src0->type));
+            }
+    }
+}
+
 // ggml_compute_forward_get_rows_back
 
 static void ggml_compute_forward_get_rows_back_f32_f16(
@@ -4500,7 +4558,7 @@ static void ggml_compute_forward_get_rows_back_f32_f16(
 
         for (int j = 0; j < nc; ++j) {
             ggml_fp16_t v = ((ggml_fp16_t *) ((char *) src0->data + i*src0->nb[1]))[j];
-            ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_FP16_TO_FP32(v);
+            ((float *) ((char *) dst->data + r*dst->nb[1]))[j] += GGML_CPU_FP16_TO_FP32(v);
         }
     }
 }
@@ -4792,7 +4850,7 @@ static void ggml_compute_forward_soft_max_f32(
         if (mp_f32) {
             if (use_f16) {
                 for (int i = 0; i < nc; ++i) {
-                    wp[i] += slope*GGML_FP16_TO_FP32(mp_f16[i]);
+                    wp[i] += slope*GGML_CPU_FP16_TO_FP32(mp_f16[i]);
                 }
             } else {
                 for (int i = 0; i < nc; ++i) {
@@ -5018,8 +5076,8 @@ static void ggml_compute_forward_clamp_f16(
         ggml_fp16_t * src0_ptr = (ggml_fp16_t *) ((char *) src0->data + j*nb01);
 
         for (int i = 0; i < nc; i++) {
-            float v = GGML_FP16_TO_FP32(src0_ptr[i]);
-            dst_ptr[i] = GGML_FP32_TO_FP16(MAX(MIN(v, max), min));
+            float v = GGML_CPU_FP16_TO_FP32(src0_ptr[i]);
+            dst_ptr[i] = GGML_CPU_FP32_TO_FP16(MAX(MIN(v, max), min));
         }
     }
 }
@@ -5476,11 +5534,11 @@ static void ggml_compute_forward_rope_f16(
                             const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
                             ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
-                            const float x0 = GGML_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
+                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
 
-                            dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                            dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     } else {
                         for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
@@ -5492,11 +5550,11 @@ static void ggml_compute_forward_rope_f16(
                             const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
                             ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
-                            const float x0 = GGML_FP16_TO_FP32(src[0]);
-                            const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+                            const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims/2]);
 
-                            dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                            dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                            dst_data[0]        = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims/2] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                         }
                     }
                 } else {
@@ -5507,11 +5565,11 @@ static void ggml_compute_forward_rope_f16(
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
                               ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
 
-                        const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[1]);
+                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_CPU_FP16_TO_FP32(src[1]);
 
-                        dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        dst_data[0] = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[1] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 }
 
@@ -5525,11 +5583,11 @@ static void ggml_compute_forward_rope_f16(
                         const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
                         ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
 
-                        const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
+                        const float x0 = GGML_CPU_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_CPU_FP16_TO_FP32(src[n_dims]);
 
-                        dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        dst_data[0]      = GGML_CPU_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims] = GGML_CPU_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
                 } else {
                     for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
@@ -5640,7 +5698,7 @@ static void ggml_compute_forward_conv_transpose_1d_f16_f32(
             for (int64_t i11 = 0; i11 < ne11; i11++) {
                 const float * const src = (float *)((char *) src1->data + i11*nb11);
                 for (int64_t i10 = 0; i10 < ne10; i10++) {
-                    dst_data[i10*ne11 + i11] = GGML_FP32_TO_FP16(src[i10]);
+                    dst_data[i10*ne11 + i11] = GGML_CPU_FP32_TO_FP16(src[i10]);
                 }
             }
         }
@@ -5933,7 +5991,7 @@ static void ggml_compute_forward_im2col_f16(
                                 if (iih < 0 || iih >= IH || iiw < 0 || iiw >= IW) {
                                     dst_data[iic*(KH*KW) + ikh*KW + ikw] = 0;
                                 } else {
-                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_FP32_TO_FP16(src_data[iih*IW + iiw]);
+                                    dst_data[iic*(KH*KW) + ikh*KW + ikw] = GGML_CPU_FP32_TO_FP16(src_data[iih*IW + iiw]);
                                 }
                             }
                         }
@@ -6109,7 +6167,7 @@ void ggml_compute_forward_conv_transpose_2d(
                     const float * const src = (float *)((char *) src1->data + i12*nb12 + i11*nb11);
                     ggml_fp16_t * dst_data = wdata + i11*ne10*ne12;
                     for (int i10 = 0; i10 < ne10; i10++) {
-                        dst_data[i10*ne12 + i12] = GGML_FP32_TO_FP16(src[i10]);
+                        dst_data[i10*ne12 + i12] = GGML_CPU_FP32_TO_FP16(src[i10]);
                     }
                 }
             }
@@ -6358,7 +6416,7 @@ static void ggml_compute_forward_pool_1d_sk_p0(
                 case GGML_OP_POOL_COUNT: GGML_ABORT("fatal error");
             }
             for (int ki = 0; ki < k; ++ki) {
-                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
+                const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                 switch (op) {
                     case GGML_OP_POOL_AVG:                         drow[i] += srow_j; break;
                     case GGML_OP_POOL_MAX:   if (srow_j > drow[i]) drow[i]  = srow_j; break;
@@ -6450,7 +6508,7 @@ void ggml_compute_forward_pool_2d(
                     for (int kx = 0; kx < k0; ++kx) {
                         int j = ix + kx;
                         if (j < 0 || j >= src->ne[0]) continue;
-                        const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
+                        const float srow_j = (src->type == GGML_TYPE_F32) ? ((const float*)srow)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t*)srow)[j]);
                         switch (op) {
                             case GGML_OP_POOL_AVG:                     *out += srow_j; break;
                             case GGML_OP_POOL_MAX: if (srow_j > *out)  *out  = srow_j; break;
@@ -6538,7 +6596,7 @@ void ggml_compute_forward_pool_2d_back(
                             }
 
                             const float val = dst->type == GGML_TYPE_F32 ?
-                                ((const float *) drowf)[j] : GGML_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
+                                ((const float *) drowf)[j] : GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drowf)[j]);
                             if (val <= maxval) {
                                 continue;
                             }
@@ -6558,7 +6616,7 @@ void ggml_compute_forward_pool_2d_back(
                     if (dst->type == GGML_TYPE_F32) {
                         ((float *) drow)[j] += grad0;
                     } else {
-                        ((ggml_fp16_t *) drow)[j] = GGML_FP32_TO_FP16(grad0 + GGML_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j]));
+                        ((ggml_fp16_t *) drow)[j] = GGML_CPU_FP32_TO_FP16(grad0 + GGML_CPU_FP16_TO_FP32(((const ggml_fp16_t *) drow)[j]));
                     }
                 } else if (op == GGML_OP_POOL_AVG) {
                     const float grad = grad0 / ka;
@@ -6577,7 +6635,7 @@ void ggml_compute_forward_pool_2d_back(
                             if (dst->type == GGML_TYPE_F32) {
                                 ((float *) drow)[j] += grad;
                             } else {
-                                ((ggml_fp16_t *) drow)[j] += GGML_FP32_TO_FP16(grad);
+                                ((ggml_fp16_t *) drow)[j] += GGML_CPU_FP32_TO_FP16(grad);
                             }
                         }
                     }
@@ -6793,6 +6851,73 @@ void ggml_compute_forward_pad_reflect_1d(
     }
 }
 
+// ggml_compute_forward_roll
+
+static int64_t ggml_wrap_index(int64_t i, int64_t ne) {
+    if (i < 0) {
+        return i + ne;
+    } else if (i >= ne) {
+        return i - ne;
+    }
+    return i;
+}
+
+static void ggml_compute_forward_roll_f32(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+    const float * src_data = (const float *) src0->data;
+    float * dst_data = (float *) dst->data;
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+
+    const int s0 = ggml_get_op_params_i32(dst, 0);
+    const int s1 = ggml_get_op_params_i32(dst, 1);
+    const int s2 = ggml_get_op_params_i32(dst, 2);
+    const int s3 = ggml_get_op_params_i32(dst, 3);
+
+    const int64_t total = ne1 * ne2 * ne3;
+    const int64_t per_thread = (total + params->nth) / params->nth;
+    const int64_t start = params->ith * per_thread;
+    const int64_t end   = std::min(start + per_thread, total);
+
+    for (int64_t i = start; i < end; ++i) {
+        const int64_t i1 = i % ne1;
+        const int64_t i2 = (i / ne1) % ne2;
+        const int64_t i3 = i / (ne2 * ne1);
+        float * dst_row = dst_data + (i3*nb3 + i2*nb2 + i1*nb1) / sizeof(float);
+
+        const int64_t i01 = ggml_wrap_index(i1 - s1, ne01);
+        const int64_t i02 = ggml_wrap_index(i2 - s2, ne02);
+        const int64_t i03 = ggml_wrap_index(i3 - s3, ne03);
+        const float * src_row = src_data + (i03*nb03 + i02*nb02 + i01*nb01) / sizeof(float);
+
+        const int64_t s = ggml_wrap_index(-s0, ne00);
+        const int64_t n = ne00 - s;
+        ggml_vec_cpy_f32(n, dst_row,     src_row + s);
+        ggml_vec_cpy_f32(s, dst_row + n, src_row);
+    }
+}
+
+void ggml_compute_forward_roll(
+        const ggml_compute_params * params,
+        ggml_tensor * dst) {
+
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_roll_f32(params, dst);
+            } break;
+        default:
+            {
+                GGML_ABORT("fatal error");
+            }
+    }
+}
+
 // ggml_compute_forward_arange
 
 static void ggml_compute_forward_arange_f32(
@@ -7075,7 +7200,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
         // loop over n_kv and n_head_kv
         // ref: https://arxiv.org/pdf/2112.05682.pdf
         for (int64_t ic = 0; ic < nek1; ++ic) {
-            const float mv = mp ? slope*GGML_FP16_TO_FP32(mp[ic]) : 0.0f;
+            const float mv = mp ? slope*GGML_CPU_FP16_TO_FP32(mp[ic]) : 0.0f;
             if (mv == -INFINITY) {
                 continue;
             }
@@ -7143,7 +7268,7 @@ static void ggml_compute_forward_flash_attn_ext_f16(
 
         if (v->type == GGML_TYPE_F16) {
             for (int64_t d = 0; d < DV; ++d) {
-                VKQ32[d] = GGML_FP16_TO_FP32(VKQ16[d]);
+                VKQ32[d] = GGML_CPU_FP16_TO_FP32(VKQ16[d]);
             }
         }
 

ggml/src/ggml-cpu/ops.h

@@ -53,6 +53,7 @@ void ggml_compute_forward_permute(const struct ggml_compute_params * params, str
 void ggml_compute_forward_transpose(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_get_rows_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_set_rows(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_inf(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_diag_mask_zero(const struct ggml_compute_params * params, struct ggml_tensor * dst);
@@ -72,6 +73,7 @@ void ggml_compute_forward_pool_2d_back(const struct ggml_compute_params * params
 void ggml_compute_forward_upscale(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_pad_reflect_1d(const struct ggml_compute_params * params, struct ggml_tensor * dst);
+void ggml_compute_forward_roll(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_arange(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_timestep_embedding(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_argsort(const struct ggml_compute_params * params, struct ggml_tensor * dst);

ggml/src/ggml-cpu/quants.c

@@ -2,6 +2,7 @@
 #include "ggml-common.h"
 
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "ggml-quants.h"
 #include "quants.h"
 
@@ -137,7 +138,7 @@ void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += sumi*GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d);
+        sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
     }
 
     *s = sumf;
@@ -174,7 +175,7 @@ void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -217,7 +218,7 @@ void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d)) * sumi;
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
     }
 
     *s = sumf;
@@ -260,7 +261,7 @@ void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, c
         }
 
         int sumi = sumi0 + sumi1;
-        sumf += (GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d))*sumi + GGML_FP16_TO_FP32(x[ib].m)*GGML_FP16_TO_FP32(y[ib].s);
+        sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
     }
 
     *s = sumf;
@@ -290,7 +291,7 @@ void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             sumi += x[ib].qs[j]*y[ib].qs[j];
         }
 
-        sumf += sumi*(GGML_FP16_TO_FP32(x[ib].d)*GGML_FP16_TO_FP32(y[ib].d));
+        sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
     }
 
     *s = sumf;
@@ -342,7 +343,7 @@ void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
             }
         }
 
-        sumf += (float) sum * (GGML_FP16_TO_FP32(x[i].d) * y[i].d);
+        sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
     }
 
     *s = sumf;
@@ -372,7 +373,7 @@ void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
             }
         }
 
-        const float d = y[i].d * GGML_FP16_TO_FP32(x[i].d);
+        const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
 
         sumf += (float) sumi * d;
     }
@@ -405,8 +406,8 @@ void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             summs += y[i].bsums[j] * (sc[j] >> 4);
         }
 
-        const float dall = y[i].d * GGML_FP16_TO_FP32(x[i].d);
-        const float dmin = y[i].d * GGML_FP16_TO_FP32(x[i].dmin);
+        const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+        const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
 
         int isum = 0;
         int is = 0;
@@ -504,7 +505,7 @@ void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -577,9 +578,9 @@ void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -657,9 +658,9 @@ void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
-        const float dmin = GGML_FP16_TO_FP32(x[i].dmin) * y[i].d;
+        const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
         sumf -= dmin * sumi;
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -714,7 +715,7 @@ void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, c
             for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
             q8 += 8; a += 8;
         }
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
     }
     for (int l = 0; l < 8; ++l) sumf += sums[l];
@@ -739,7 +740,7 @@ void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
         int32_t bsum = 0;
@@ -778,7 +779,7 @@ void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint16_t * GGML_RESTRICT q2 = x[i].qs;
         const uint8_t  * GGML_RESTRICT sc = x[i].scales;
         const int8_t   * GGML_RESTRICT q8 = y[i].qs;
@@ -829,7 +830,7 @@ void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
     float sumf = 0;
     for (int i = 0; i < nb; i++) {
 
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const int8_t  * q8 = y[i].qs;
         const uint8_t * qs = x[i].qs;
         const uint8_t * qh = x[i].qh;
@@ -882,7 +883,7 @@ void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT q3 = x[i].qs;
         const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
         const int8_t  * GGML_RESTRICT q8 = y[i].qs;
@@ -924,7 +925,7 @@ void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     float sumf = 0.f;
     for (int i = 0; i < nb; ++i) {
-        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
         const uint8_t * GGML_RESTRICT qs = x[i].qs;
         const uint8_t * GGML_RESTRICT qh = x[i].qh;
         const uint8_t * GGML_RESTRICT signs = x[i].signs;
@@ -1002,7 +1003,7 @@ void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
             qs += 4;
         }
 
-        sumf += GGML_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+        sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
     }
 
     *s = sumf;
@@ -1063,7 +1064,7 @@ void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
             qh += 2;
         }
 
-        sumf += GGML_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+        sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
     }
 
     *s = sumf;
@@ -1087,7 +1088,7 @@ void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     float sumf = 0;
 
     for (; ib < nb; ++ib) {
-        const float d = GGML_FP16_TO_FP32(y[ib].d)*GGML_FP16_TO_FP32(x[ib].d);
+        const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
         int sumi1 = 0, sumi2 = 0;
         for (int j = 0; j < QK4_NL/2; ++j) {
             sumi1 += y[ib].qs[j+       0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
@@ -1113,7 +1114,7 @@ void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
 
     float sumf = 0;
     for (int ibl = 0; ibl < nb; ++ibl) {
-        const float d4d8 = GGML_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+        const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
         uint16_t h = x[ibl].scales_h;
         const uint8_t * qs = x[ibl].qs;
         const int8_t  * q8 = y[ibl].qs;

ggml/src/ggml-cpu/repack.cpp

@@ -6,6 +6,7 @@
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
 #include "traits.h"
 
 #include "arch-fallback.h"
@@ -72,7 +73,7 @@ void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GG
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -110,7 +111,7 @@ void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GG
             const float d = amax / ((1 << 7) - 1);
             id[row_iter] = d ? 1.0f / d : 0.0f;
 
-            y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
+            y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d);
         }
 
         for (int j = 0; j < QK8_0 * 4; j++) {
@@ -236,7 +237,7 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -280,7 +281,7 @@ void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                         const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                         sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
         }
@@ -325,7 +326,7 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                             const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -396,13 +397,13 @@ void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                         sumi2 = sumi2 * scales_1[j];
                         sumi += sumi1 + sumi2;
                     }
-                    sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d;
                 }
             }
             for (int sb = 0; sb < 8; sb++) {
                 uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16;
                 for (int j = 0; j < ncols_interleaved; j++) {
-                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
+                    sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d;
                 }
             }
         }
@@ -449,7 +450,7 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                             const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
                             sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                         }
-                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                     }
                 }
             }
@@ -500,7 +501,7 @@ void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                                 }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }
@@ -555,7 +556,7 @@ void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -609,7 +610,7 @@ void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                 sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                          (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                         }
                     }
                 }
@@ -688,7 +689,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                                 sumi2 = sumi2 * scales_1[j];
                                 sumi += sumi1 + sumi2;
                             }
-                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -697,7 +698,7 @@ void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs,
                     for(int m = 0; m < 4; m++) {
                         const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6);
                         for(int j = 0; j < ncols_interleaved; j++) {
-                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
+                            sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j]) * a_ptr[l].d[m];
                         }
                     }
                 }
@@ -753,7 +754,7 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
                                     sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
                                             (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
                                 }
-                                sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                                sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
                             }
                         }
                     }
@@ -1163,13 +1164,24 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
         // not realy a GGML_TYPE_Q8_0 but same size.
         switch (op->op) {
             case GGML_OP_MUL_MAT:
-                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                return true;
+                {
+                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                    return true;
+                }
             case GGML_OP_MUL_MAT_ID:
-                size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
-                size = GGML_PAD(size, sizeof(int64_t));  // + padding for next bloc.
-                size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
-                return true;
+                {
+                    size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
+                    size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
+
+                    const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert
+                    const int64_t ne12 = op->src[1]->ne[2]; // n_tokens
+
+                    const size_t sizeof_mmid_row_mapping = sizeof(int64_t);
+
+                    size += sizeof_mmid_row_mapping*ne02*(ne12 + 1);
+
+                    return true;
+                }
             default:
                 // GGML_ABORT("fatal error");
                 break;
@@ -1305,14 +1317,17 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
             int32_t i2;
         };
 
-        GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
-                                      n_as * ne12 * sizeof(mmid_row_mapping)));
+        GGML_ASSERT(params->wsize >=
+                (GGML_PAD(nbw3, sizeof(int64_t)) +
+                 n_as*(ne12 + 1)*sizeof(mmid_row_mapping))
+                );
 
-        auto * wdata             = (char *)     params->wdata;
-        auto * wdata_src1_end    = (char *)     wdata + GGML_PAD(nbw3, sizeof(int64_t));
-        auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
+        auto * wdata          = (char *)params->wdata;
+        auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t));
 
-        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as);  // [n_as][ne12]
+        // total of [n_as][ne12 + 1] elemets of type mmid_row_mapping (2*int32_t = int64_t)
+        auto * matrix_row_counts = (int64_t *) (wdata_src1_end);                                        // [n_as]
+        struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
 
         // src1: float32 => param type
         for (int64_t i12 = 0; i12 < ne12; ++i12) {
@@ -1397,44 +1412,45 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
     }
 };
 
-// instance for Q4
-static const tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
-static const tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
-static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
-static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
-
-// instance for IQ4
-static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
-
 }  // namespace ggml::cpu::repack
 
 static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) {
+
+    // instance for Q4
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
+
+    // instance for IQ4
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
             if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::repack::q4_0_8x8_q8_0;
+                return &q4_0_8x8_q8_0;
             }
         }
         if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
             if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::q4_0_4x8_q8_0;
+                return &q4_0_4x8_q8_0;
             }
         }
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::q4_0_4x4_q8_0;
+                return &q4_0_4x4_q8_0;
             }
         }
     } else if (cur->type == GGML_TYPE_Q4_K) {
         if (ggml_cpu_has_avx2()) {
             if (cur->ne[1] % 8 == 0) {
-                return &ggml::cpu::repack::q4_K_8x8_q8_K;
+                return &q4_K_8x8_q8_K;
             }
         }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
-                return &ggml::cpu::repack::iq4_nl_4x4_q8_0;
+                return &iq4_nl_4x4_q8_0;
             }
         }
     }

ggml/src/ggml-cpu/simd-mappings.h

@@ -2,10 +2,167 @@
 
 #include "ggml-cpu-impl.h"
 
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+//   $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__F16C__)
+#include <immintrin.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
 //
 // simd mappings
 //
 
+// FP16 to FP32 conversion
+
+// 16-bit float
+// on Arm, we use __fp16
+// on x86, we use uint16_t
+//
+// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
+// for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
+//
+#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+
+    static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
+        __fp16 tmp;
+        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+        return (float)tmp;
+    }
+
+    static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __fp16 tmp = f;
+        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+        return res;
+    }
+#elif defined(__F16C__)
+    #ifdef _MSC_VER
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+    #else
+        #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+        #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+    #endif
+#elif defined(__POWER9_VECTOR__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x)
+    /* the inline asm below is about 12% faster than the lookup method */
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        double d;
+        __asm__(
+            "mtfprd %0,%2\n"
+            "xscvhpdp %0,%0\n"
+            "frsp %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=f"(f):
+            /* in */   "r"(h));
+        return f;
+    }
+
+    static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) {
+        double d;
+        ggml_fp16_t r;
+        __asm__( /* xscvdphp can work on double or single precision */
+            "xscvdphp %0,%2\n"
+            "mffprd %1,%0\n" :
+            /* temp */ "=d"(d),
+            /* out */  "=r"(r):
+            /* in */   "f"(f));
+        return r;
+    }
+#elif defined(__riscv) && defined(__riscv_zfhmin)
+    static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
+        float f;
+        __asm__(
+            "fmv.h.x %[f], %[h]\n\t"
+            "fcvt.s.h %[f], %[f]"
+            : [f] "=&f" (f)
+            : [h] "r" (h)
+        );
+        return f;
+    }
+
+    static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
+        ggml_fp16_t res;
+        __asm__(
+            "fcvt.h.s %[f], %[f]\n\t"
+            "fmv.x.h %[h], %[f]"
+            : [h] "=&r" (res)
+            : [f] "f" (f)
+        );
+        return res;
+    }
+
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+#elif defined(__NNPA__)
+    #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) nnpa_compute_fp16_to_fp32(x)
+    #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) nnpa_compute_fp32_to_fp16(x)
+
+    #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+    #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+    static inline float nnpa_compute_fp16_to_fp32(ggml_fp16_t h) {
+        uint16x8_t v_h = vec_splats(h);
+        uint16x8_t v_hd = vec_convert_from_fp16(v_h, 0);
+        return vec_extend_to_fp32_hi(v_hd, 0)[0];
+    }
+
+    static inline ggml_fp16_t nnpa_compute_fp32_to_fp16(float f) {
+        float32x4_t v_f = vec_splats(f);
+        float32x4_t v_zero = vec_splats(0.0f);
+        uint16x8_t v_hd = vec_round_from_fp32(v_f, v_zero, 0);
+        uint16x8_t v_h = vec_convert_to_fp16(v_hd, 0);
+        return vec_extract(v_h, 0);
+    }
+#endif
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_CPU_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+    uint16_t s;
+    memcpy(&s, &f, sizeof(uint16_t));
+    return ggml_table_f32_f16[s];
+}
+
+#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_CPU_FP32_TO_FP16)
+#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+
 // we define a common set of C macros which map to specific intrinsics based on the current architecture
 // we then implement the fundamental computation operations below using only these macros
 // adding support for new architectures requires to define the corresponding SIMD macros
@@ -415,7 +572,7 @@ static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
     float tmp[8];
 
     for (int i = 0; i < 8; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     return _mm256_loadu_ps(tmp);
@@ -426,7 +583,7 @@ static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
     _mm256_storeu_ps(arr, y);
 
     for (int i = 0; i < 8; i++)
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
 }
 #define GGML_F32Cx8_LOAD(x)     __avx_f32cx8_load(x)
 #define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
@@ -574,10 +731,10 @@ static inline unsigned char ggml_endian_byte(int i) {
 inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) {
     float tmp[4];
 
-    tmp[0] = GGML_FP16_TO_FP32(p[0]);
-    tmp[1] = GGML_FP16_TO_FP32(p[1]);
-    tmp[2] = GGML_FP16_TO_FP32(p[2]);
-    tmp[3] = GGML_FP16_TO_FP32(p[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]);
 
     return wasm_v128_load(tmp);
 }
@@ -587,10 +744,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 
     wasm_v128_store(tmp, x);
 
-    p[0] = GGML_FP32_TO_FP16(tmp[0]);
-    p[1] = GGML_FP32_TO_FP16(tmp[1]);
-    p[2] = GGML_FP32_TO_FP16(tmp[2]);
-    p[3] = GGML_FP32_TO_FP16(tmp[3]);
+    p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]);
+    p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]);
+    p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]);
+    p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]);
 }
 
 #define GGML_F16x4             v128_t
@@ -690,10 +847,10 @@ inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
 static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
     float tmp[4];
 
-    tmp[0] = GGML_FP16_TO_FP32(x[0]);
-    tmp[1] = GGML_FP16_TO_FP32(x[1]);
-    tmp[2] = GGML_FP16_TO_FP32(x[2]);
-    tmp[3] = GGML_FP16_TO_FP32(x[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
 
     return _mm_loadu_ps(tmp);
 }
@@ -703,10 +860,10 @@ static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
     _mm_storeu_ps(arr, y);
 
-    x[0] = GGML_FP32_TO_FP16(arr[0]);
-    x[1] = GGML_FP32_TO_FP16(arr[1]);
-    x[2] = GGML_FP32_TO_FP16(arr[2]);
-    x[3] = GGML_FP32_TO_FP16(arr[3]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }
 
 #define GGML_F32Cx4             __m128
@@ -828,7 +985,7 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F32x4_ZERO    __lsx_vldi(0)
 #define GGML_F32x4_SET1(x) __lsx_vinsgr2vr_w(__lsx_vldi(0),(x), 0)
 #define GGML_F32x4_LOAD(x) __lsx_vld((x), 0)
-#define GGML_F32x4_STORE((x),(y))   __lsx_vst((y), (x), 0)
+#define GGML_F32x4_STORE(x, y)   __lsx_vst(y, x, 0)
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
@@ -874,10 +1031,10 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
     float tmp[4];
 
-    tmp[0] = GGML_FP16_TO_FP32(x[0]);
-    tmp[1] = GGML_FP16_TO_FP32(x[1]);
-    tmp[2] = GGML_FP16_TO_FP32(x[2]);
-    tmp[3] = GGML_FP16_TO_FP32(x[3]);
+    tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+    tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+    tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+    tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
 
     return __lsx_vld(tmp, 0);
 }
@@ -887,10 +1044,10 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 
     __lsx_vst(y, arr, 0);
 
-    x[0] = GGML_FP32_TO_FP16(arr[0]);
-    x[1] = GGML_FP32_TO_FP16(arr[1]);
-    x[2] = GGML_FP32_TO_FP16(arr[2]);
-    x[3] = GGML_FP32_TO_FP16(arr[3]);
+    x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+    x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+    x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+    x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
 }
 
 #define GGML_F32Cx4             __m128
@@ -922,7 +1079,7 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F32_STEP 32
 #define GGML_F32_EPR  4
 
-#define GGML_F32x4              __vector float
+#define GGML_F32x4              float32x4_t
 #define GGML_F32x4_ZERO         vec_splats(0.0f)
 #define GGML_F32x4_SET1         vec_splats
 #define GGML_F32x4_LOAD(p)      vec_xl(0, p)
@@ -944,10 +1101,8 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
     for (int i = 0; i < offset; ++i) {              \
         x[i] = vec_add(x[i], x[offset + i]);        \
     }                                               \
-    res = vec_extract(x[0], 0) +                    \
-          vec_extract(x[0], 1) +                    \
-          vec_extract(x[0], 2) +                    \
-          vec_extract(x[0], 3);                     \
+    float32x4_t tmp = x[0] + vec_reve(x[0]);        \
+    res = tmp[0] + tmp[1];                          \
 }
 
 #define GGML_F32_VEC        GGML_F32x4
@@ -964,28 +1119,45 @@ static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
 #define GGML_F16_STEP GGML_F32_STEP
 #define GGML_F16_EPR  GGML_F32_EPR
 
-static inline __vector float __lzs_f16cx4_load(const ggml_fp16_t * x) {
+static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
+#if defined(__NNPA__)
+    uint16x8_t v_x = vec_xl(0, (const ggml_fp16_t *)x);
+    uint16x8_t v_xd = vec_convert_from_fp16(v_x, 0);
+    return vec_extend_to_fp32_hi(v_xd, 0);
+#else
     float tmp[4];
 
     for (int i = 0; i < 4; i++) {
-        tmp[i] = GGML_FP16_TO_FP32(x[i]);
+        tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
     }
 
     // note: keep type-cast here to prevent compiler bugs
     // see: https://github.com/ggml-org/llama.cpp/issues/12846
     return vec_xl(0, (const float *)(tmp));
+#endif
 }
 
-static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
+#if defined(__NNPA__)
+    float32x4_t v_zero = vec_splats(0.0f);
+    uint16x8_t v_xd = vec_round_from_fp32(v_y, v_zero, 0);
+    uint16x8_t v_x = vec_convert_to_fp16(v_xd, 0);
+
+    x[0] = vec_extract(v_x, 0);
+    x[1] = vec_extract(v_x, 1);
+    x[2] = vec_extract(v_x, 2);
+    x[3] = vec_extract(v_x, 3);
+#else
     float arr[4];
 
     // note: keep type-cast here to prevent compiler bugs
     // see: https://github.com/ggml-org/llama.cpp/issues/12846
-    vec_xst(y, 0, (float *)(arr));
+    vec_xst(v_y, 0, (float *)(arr));
 
     for (int i = 0; i < 4; i++) {
-        x[i] = GGML_FP32_TO_FP16(arr[i]);
+        x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
     }
+#endif
 }
 
 #define GGML_F16_VEC                GGML_F32x4
@@ -1006,3 +1178,7 @@ static inline void __lzs_f16cx4_store(ggml_fp16_t * x, __vector float y) {
 #define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
 #define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
 #endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/vec.cpp

@@ -219,11 +219,11 @@ void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * G
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
 #else
     for (int i = 0; i < n; ++i) {
-        sumf += (ggml_float)(GGML_FP16_TO_FP32(x[i])*GGML_FP16_TO_FP32(y[i]));
+        sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
     }
 #endif
 

ggml/src/ggml-cpu/vec.h

@@ -58,7 +58,7 @@ inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf
 inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
 inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
     for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) + GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
     }
 }
 inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
@@ -67,7 +67,7 @@ inline static void ggml_vec_acc1_f32(const int n, float * y, const float   v)
 inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] - y[i]; }
 inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
     for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) - GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i]));
     }
 }
 inline static void ggml_vec_set_f32 (const int n, float * x, const float   v)                  { for (int i = 0; i < n; ++i) x[i]  = v;           }
@@ -75,20 +75,20 @@ inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x)
 inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i]  = -x[i];       }
 inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(-GGML_FP16_TO_FP32(x[i]));
+        y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_FP16_TO_FP32(x[i]));
     }
 }
 
 inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]*y[i];   }
 inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
     for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) * GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i]));
     }
 }
 inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i]/y[i];   }
 inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
     for (int i = 0; i < n; ++i) {
-        z[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(x[i]) / GGML_FP16_TO_FP32(y[i]));
+        z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i]));
     }
 }
 
@@ -131,13 +131,13 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
     // leftovers
     for (int i = np; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
 #else
     for (int i = 0; i < n; ++i) {
         for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
-            sumf[j] += (ggml_float)(GGML_FP16_TO_FP32(x[j][i])*GGML_FP16_TO_FP32(y[i]));
+            sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
         }
     }
 #endif
@@ -280,12 +280,12 @@ inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y,
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i]) + GGML_FP16_TO_FP32(x[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
     }
 #endif
 }
@@ -430,12 +430,12 @@ inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float
 
     // leftovers
     for (int i = np; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(y[i])*v);
+        y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
     }
 #endif
 }
@@ -444,103 +444,103 @@ inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) {
 inline static void ggml_vec_sqr_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i];   }
 inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v*v);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v*v);
     }
 }
 inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
 inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(sqrtf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_log_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]);  }
 inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(logf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_sin_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]);  }
 inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(sinf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_cos_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]);  }
 inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(cosf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_abs_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
 inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(fabsf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_sgn_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
 inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
     }
 }
 inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
 inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16((GGML_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
+        y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
     }
 }
 inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]);  }
 inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(tanhf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_elu_f32  (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
 inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(expm1f(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expm1f(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
 inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16((v > 0.f) ? v : 0.f);
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : 0.f);
     }
 }
 inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
 inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
     }
 }
 inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
 inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(1.f / (1.f + expf(-GGML_FP16_TO_FP32(x[i]))));
+        y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_FP16_TO_FP32(x[i]))));
     }
 }
 // TODO: optimize performance
 inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
     }
 }
 inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
 inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
+        y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
     }
 }
 inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
 inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        y[i] = GGML_FP32_TO_FP16(expf(GGML_FP16_TO_FP32(x[i])));
+        y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i])));
     }
 }
 
@@ -562,9 +562,9 @@ inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp
 
 inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float xi = GGML_FP16_TO_FP32(x[i]);
+        float xi = GGML_CPU_FP16_TO_FP32(x[i]);
         float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
-        y[i] = GGML_FP32_TO_FP16(res);
+        y[i] = GGML_CPU_FP32_TO_FP16(res);
     }
 }
 
@@ -577,9 +577,9 @@ inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
         } else if (x[i] >= 10.0f) {
             y[i] = x[i];
         } else {
-            ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+            ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
             memcpy(&t, &fp16, sizeof(uint16_t));
-            y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+            y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]);
         }
     }
 }
@@ -613,9 +613,9 @@ inline static float ggml_gelu_quick_f32(float x) {
 inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
     uint16_t t;
     for (int i = 0; i < n; ++i) {
-        ggml_fp16_t fp16 = GGML_FP32_TO_FP16(x[i]);
+        ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
         memcpy(&t, &fp16, sizeof(uint16_t));
-        y[i] = GGML_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
+        y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
     }
 }
 #else
@@ -628,8 +628,8 @@ inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float *
 
 inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
     for (int i = 0; i < n; ++i) {
-        float v = GGML_FP16_TO_FP32(x[i]);
-        y[i] = GGML_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
+        float v = GGML_CPU_FP16_TO_FP32(x[i]);
+        y[i] = GGML_CPU_FP32_TO_FP16(v*(1.0f/(1.0f+expf(GELU_QUICK_COEF*v))));
     }
 }
 
@@ -638,8 +638,8 @@ inline static float ggml_silu_f32(float x) {
     return x/(1.0f + expf(-x));
 }
 inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
-    float v = GGML_FP16_TO_FP32(x);
-    return GGML_FP32_TO_FP16(v/(1.0f + expf(-v)));
+    float v = GGML_CPU_FP16_TO_FP32(x);
+    return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v)));
 }
 
 #if __FINITE_MATH_ONLY__
@@ -888,9 +888,9 @@ inline static float ggml_silu_backward_f32(float x, float dy) {
 }
 
 inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
-    const float v = GGML_FP16_TO_FP32(x);
+    const float v = GGML_CPU_FP16_TO_FP32(x);
     const float s = 1.0f/(1.0f + expf(-v));
-    return GGML_FP32_TO_FP16(GGML_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
+    return GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
 }
 
 inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
@@ -928,7 +928,7 @@ inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float
 inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) {
     float sum = 0.0f;
     for (int i = 0; i < n; ++i) {
-        sum += GGML_FP16_TO_FP32(x[i]);
+        sum += GGML_CPU_FP16_TO_FP32(x[i]);
     }
     *s = sum;
 }

ggml/src/ggml-cuda/common.cuh

@@ -19,10 +19,10 @@
 #endif
 #include "ggml-common.h"
 
-#include <cstdio>
 #include <array>
 #include <cassert>
 #include <cfloat>
+#include <cstdio>
 #include <string>
 #include <vector>
 
@@ -76,11 +76,9 @@
 #define GGML_CUDA_CC_IS_CDNA(cc)  (cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1)
 
 // Moore Threads
-#define GGML_CUDA_MUSA_ARCH_IS_QY1 (__MUSA_ARCH__ <= 210)
-
-#define GGML_CUDA_CC_QY1  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
-#define GGML_CUDA_CC_QY2  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
-#define GGML_CUDA_CC_NG   (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
+#define GGML_CUDA_CC_QY1 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x210) // MTT S80, MTT S3000
+#define GGML_CUDA_CC_QY2 (GGML_CUDA_CC_OFFSET_MTHREADS + 0x220) // MTT S4000
+#define GGML_CUDA_CC_NG  (GGML_CUDA_CC_OFFSET_MTHREADS + 0x310) // TBD
 
 #define GGML_CUDA_CC_IS_MTHREADS(cc) (cc >= GGML_CUDA_CC_OFFSET_MTHREADS && cc < GGML_CUDA_CC_OFFSET_AMD)
 #define GGML_CUDA_CC_IS_QY1(cc)      (cc >= GGML_CUDA_CC_QY1 && cc < GGML_CUDA_CC_QY2)
@@ -203,9 +201,9 @@ typedef float2 dfloat2;
 #define FAST_FP16_AVAILABLE
 #endif // defined(FP16_AVAILABLE) && __CUDA_ARCH__ != 610
 
-#if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#if (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 #define FP16_MMA_AVAILABLE
-#endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // (!defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA) || defined(GGML_USE_MUSA)
 
 #if defined(GGML_HIP_ROCWMMA_FATTN) && (defined(CDNA) || defined(RDNA3) || (defined(GGML_HIP_ROCWMMA_FATTN_GFX12) && defined(RDNA4)))
 #define FP16_MMA_AVAILABLE
@@ -219,9 +217,9 @@ typedef float2 dfloat2;
 #define CP_ASYNC_AVAILABLE
 #endif // !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 
-#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#if !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 #define FLASH_ATTN_AVAILABLE
-#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && GGML_CUDA_MUSA_ARCH_IS_QY1)
+#endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL;
@@ -233,7 +231,8 @@ static bool fast_fp16_available(const int cc) {
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fast_fp16_hardware_available(const int cc) {
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc);
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_PASCAL && cc != 610) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }
 
 // Any FP16 tensor core instructions are available for ggml code.
@@ -241,15 +240,35 @@ static bool fp16_mma_available(const int cc) {
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
     return false;
 #else
-    return (GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
+    if ((GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) ||
+        GGML_CUDA_CC_IS_MTHREADS(cc)) {
+        return true;
+    } else if (GGML_CUDA_CC_IS_RDNA4(cc)) {
+#if defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+        return true;
+#else
+        return false;
+#endif // defined(GGML_HIP_ROCWMMA_FATTN) && defined(GGML_HIP_ROCWMMA_FATTN_GFX12)
+    } else {
+        return false;
+    }
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
     return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+}
+
+static bool bf16_mma_hardware_available(const int cc) {
+    return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_AMPERE) || GGML_CUDA_CC_IS_CDNA(cc) || cc >= GGML_CUDA_CC_RDNA3;
+}
+
+static bool fp32_mma_hardware_available(const int cc) {
+    return GGML_CUDA_CC_IS_CDNA(cc);
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.
@@ -362,6 +381,26 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template<bool norm>
+static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float sum = 0.0f;
+    for (int i = col; i < ncols; i += blockDim.x) {
+        sum += x[row * ncols + i];
+    }
+
+    sum = warp_reduce_sum(sum);
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}
+
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
@@ -767,21 +806,7 @@ struct ggml_backend_cuda_context {
         name(GGML_CUDA_NAME + std::to_string(device)) {
     }
 
-    ~ggml_backend_cuda_context() {
-        if (copy_event != nullptr) {
-            CUDA_CHECK(cudaEventDestroy(copy_event));
-        }
-        for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
-            for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
-                if (streams[i][j] != nullptr) {
-                    CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
-                }
-            }
-            if (cublas_handles[i] != nullptr) {
-                CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
-            }
-        }
-    }
+    ~ggml_backend_cuda_context();
 
     cudaStream_t stream(int device, int stream) {
         if (streams[device][stream] == nullptr) {

ggml/src/ggml-cuda/conv2d-dw.cu

@@ -0,0 +1,161 @@
+#include "conv2d-dw.cuh"
+
+struct conv_params {
+    int in_w, in_h;
+    int out_w, out_h;
+    int kernel_w, kernel_h;
+    int stride_x, stride_y;
+    int padding_x, padding_y;
+    int dilation_x, dilation_y;
+    int channels, batches;
+};
+
+struct kernel_bounds {
+    int y_min, y_max;
+    int x_min, x_max;
+};
+
+__device__ __forceinline__ kernel_bounds calculate_kernel_bounds(int out_x, int out_y, const conv_params & params) {
+    kernel_bounds bounds;
+    bounds.y_min = max(0, (params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.y_max =
+        min(params.kernel_h,
+            (params.in_h + params.padding_y - out_y * params.stride_y + params.dilation_y - 1) / params.dilation_y);
+    bounds.x_min = max(0, (params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    bounds.x_max =
+        min(params.kernel_w,
+            (params.in_w + params.padding_x - out_x * params.stride_x + params.dilation_x - 1) / params.dilation_x);
+    return bounds;
+}
+
+__device__ __forceinline__ int calculate_input_coord(int out_coord, int kern_coord, int stride, int dilation, int padding) {
+    return out_coord * stride + kern_coord * dilation - padding;
+}
+
+struct whcn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + c * params.in_w * params.in_h + y * params.in_w + x;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return c * params.kernel_h * params.kernel_w + ky * params.kernel_w + kx;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + c * params.out_w * params.out_h +
+               y * params.out_w + x;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        out_x = global_idx % params.out_w;
+        out_y = (global_idx / params.out_w) % params.out_h;
+        c     = (global_idx / (params.out_w * params.out_h)) % params.channels;
+        n     = global_idx / (params.out_w * params.out_h * params.channels);
+    }
+};
+
+struct cwhn_layout {
+    __device__ static int input_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.in_w * params.in_h) + (y * params.in_w + x) * params.channels + c;
+    }
+
+    __device__ static int kernel_index(int c, int ky, int kx, const conv_params & params) {
+        return (ky * params.kernel_w + kx) * params.channels + c;
+    }
+
+    __device__ static int output_index(int n, int c, int y, int x, const conv_params & params) {
+        return n * (params.channels * params.out_w * params.out_h) + y * (params.out_w * params.channels) +
+               x * params.channels + c;
+    }
+
+    __device__ static void unpack_indices(int global_idx, const conv_params & params, int & n, int & c, int & out_y,
+                                          int & out_x) {
+        c     = global_idx % params.channels;
+        out_x = (global_idx / params.channels) % params.out_w;
+        out_y = (global_idx / (params.channels * params.out_w)) % params.out_h;
+        n     = global_idx / (params.channels * params.out_w * params.out_h);
+    }
+};
+
+template <typename T, typename Layout>
+__global__ void conv2d_dw_kernel(const T * __restrict__ input, const T * __restrict__ kernel, T * __restrict__ output,
+                                 const int in_w, const int in_h, const int out_w, const int out_h,
+                                 const int kernel_w, const int kernel_h, const int stride_x, const int stride_y,
+                                 const int padding_x, const int padding_y, const int dilation_x, const int dilation_y,
+                                 const int channels, const int batches) {
+    const int global_idx     = blockIdx.x * blockDim.x + threadIdx.x;
+    const int total_elements = batches * channels * out_h * out_w;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    conv_params params = { in_w,     in_h,      out_w,     out_h,      kernel_w,   kernel_h, stride_x,
+                           stride_y, padding_x, padding_y, dilation_x, dilation_y, channels, batches };
+
+    int batch_idx, channel_idx, out_y_idx, out_x_idx;
+    Layout::unpack_indices(global_idx, params, batch_idx, channel_idx, out_y_idx, out_x_idx);
+
+    T accumulator = 0;
+    kernel_bounds bounds = calculate_kernel_bounds(out_x_idx, out_y_idx, params);
+
+    for (int kern_y = bounds.y_min; kern_y < bounds.y_max; ++kern_y) {
+        int in_y_idx = calculate_input_coord(out_y_idx, kern_y, params.stride_y, params.dilation_y, params.padding_y);
+
+        for (int kern_x = bounds.x_min; kern_x < bounds.x_max; ++kern_x) {
+            int in_x_idx = calculate_input_coord(out_x_idx, kern_x, params.stride_x, params.dilation_x, params.padding_x);
+
+            const T input_val  = input[Layout::input_index(batch_idx, channel_idx, in_y_idx, in_x_idx, params)];
+            const T kernel_val = kernel[Layout::kernel_index(channel_idx, kern_y, kern_x, params)];
+
+            accumulator += input_val * kernel_val;
+        }
+    }
+
+    output[Layout::output_index(batch_idx, channel_idx, out_y_idx, out_x_idx, params)] = accumulator;
+}
+
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F32 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    const float * w_d = (const float *) kernel->data;
+    const float * x_d = (const float *) input->data;
+    float *       y_d = (float *) dst->data;
+
+    const int32_t * p          = (const int32_t *) dst->op_params;
+    const int       stride_x   = p[0];
+    const int       stride_y   = p[1];
+    const int       padding_x  = p[2];
+    const int       padding_y  = p[3];
+    const int       dilation_x = p[4];
+    const int       dilation_y = p[5];
+
+    const int in_w     = input->ne[0];
+    const int in_h     = input->ne[1];
+    const int kernel_w = kernel->ne[0];
+    const int kernel_h = kernel->ne[1];
+    const int out_w    = dst->ne[0];
+    const int out_h    = dst->ne[1];
+    const int channels = dst->ne[2];
+    const int batches  = dst->ne[3];
+
+    cudaStream_t st = ctx.stream();
+
+    const int total  = batches * channels * out_h * out_w;
+    const int blocks = (total + CUDA_CONV2D_DW_BLOCK_SIZE - 1) / CUDA_CONV2D_DW_BLOCK_SIZE;
+
+    if (ggml_is_contiguous(input)) {
+        conv2d_dw_kernel<float, whcn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else if (ggml_is_contiguous_channels(input)) {
+        conv2d_dw_kernel<float, cwhn_layout><<<blocks, CUDA_CONV2D_DW_BLOCK_SIZE, 0, st>>>(
+            x_d, w_d, y_d, in_w, in_h, out_w, out_h, kernel_w, kernel_h, stride_x, stride_y, padding_x, padding_y,
+            dilation_x, dilation_y, channels, batches);
+    } else {
+        GGML_ABORT("Unsupported memory layout for conv_2d_dw");
+    }
+}

ggml/src/ggml-cuda/conv2d-dw.cuh

@@ -0,0 +1,5 @@
+#pragma once
+#include "common.cuh"
+
+#define CUDA_CONV2D_DW_BLOCK_SIZE 256
+void ggml_cuda_op_conv2d_dw(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/conv2d-transpose.cu

@@ -0,0 +1,91 @@
+#include <algorithm>
+
+#include "conv2d-transpose.cuh"
+#include "ggml.h"
+
+__global__ void conv2d_transpose_kernel(const float * __restrict__ input, const half * __restrict__ kernel,
+                                        float * __restrict__ output, const int in_w, const int in_h, const int out_w,
+                                        const int out_h, const int kernel_w, const int kernel_h, const int stride,
+                                        const int c_in, const int c_out, const int batches) {
+    const int global_idx = blockIdx.x * blockDim.x + threadIdx.x;
+
+    const int total_elements = out_w * out_h * c_out * batches;
+
+    if (global_idx >= total_elements) {
+        return;
+    }
+
+    const int out_x_idx = global_idx % out_w;
+    const int out_y_idx = (global_idx / out_w) % out_h;
+    const int c_idx     = (global_idx / (out_w * out_h)) % c_out;
+    const int n_idx     = global_idx / (out_w * out_h * c_out);
+
+    float accumulator = 0;
+    // For each output idx, find the inputs that contribute to it by checking stride alignment and bounds
+
+    for (int c_in_idx = 0; c_in_idx < c_in; c_in_idx++) {
+        for (int kh = 0; kh < kernel_h; ++kh) {
+            int in_y = out_y_idx - kh;
+            if (in_y < 0 || in_y % stride) continue;
+            in_y /= stride;
+            if (in_y >= in_h) continue;
+
+            for (int kw = 0; kw < kernel_w; ++kw) {
+                int in_x = out_x_idx - kw;
+                if (in_x < 0 || in_x % stride) continue;
+                in_x /= stride;
+                if (in_x >= in_w) continue;
+
+                const int input_idx = (in_w * in_h * c_in) * n_idx + (in_w * in_h) * c_in_idx + (in_w) *in_y + in_x;
+                const int kernel_idx =
+                    (kernel_h * kernel_w * c_out) * c_in_idx + (kernel_h * kernel_w) * c_idx + (kernel_w) *kh + kw;
+
+                float input_val = input[input_idx];
+                half  kern_val  = kernel[kernel_idx];
+
+                accumulator += input_val * (float) kern_val;
+            }
+        }
+    }
+
+    output[(out_w * out_h * c_out) * n_idx + (out_w * out_h) * c_idx + (out_w) *out_y_idx + out_x_idx] = accumulator;
+}
+
+//input is (W, H, C_in, N), Kernel is (W, H, C_out, C_in)
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * kernel = dst->src[0];
+    const ggml_tensor * input  = dst->src[1];
+
+    GGML_ASSERT(kernel->type == GGML_TYPE_F16 && input->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+
+    const float * input_data  = (const float *) input->data;
+    float *       output_data = (float *) dst->data;
+    const half * kernel_data = (const half *) kernel->data;
+
+    const int input_w      = input->ne[0];
+    const int input_h      = input->ne[1];
+    const int output_w     = dst->ne[0];
+    const int output_h     = dst->ne[1];
+    const int channels_in  = input->ne[2];
+    const int channels_out = kernel->ne[2];
+    const int kernel_w     = kernel->ne[0];
+    const int kernel_h     = kernel->ne[1];
+    const int stride       = dst->op_params[0];
+    const int batches      = input->ne[3];
+
+    GGML_ASSERT(channels_in == kernel->ne[3]);
+    GGML_ASSERT(stride > 0);
+
+    cudaStream_t st = ctx.stream();
+
+    GGML_ASSERT(ggml_is_contiguous(input));
+    GGML_ASSERT(ggml_is_contiguous(kernel));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int total  = (output_w * output_h * channels_out * batches);
+    const int blocks = (total + CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE - 1) / CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE;
+
+    conv2d_transpose_kernel<<<blocks, CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE, 0, st>>>(
+        input_data, kernel_data, output_data, input_w, input_h, output_w, output_h, kernel_w, kernel_h, stride,
+        channels_in, channels_out, batches);
+}

ggml/src/ggml-cuda/conv2d-transpose.cuh

@@ -0,0 +1,4 @@
+#include "common.cuh"
+
+#define CUDA_CONV2D_TRANSPOSE_BLOCK_SIZE 256
+void ggml_cuda_conv_2d_transpose_p0(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-cuda/fattn-wmma-f16.cu

@@ -9,7 +9,11 @@
 #ifdef FP16_MMA_AVAILABLE
 #if !(defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__))
 #include <mma.h>
+#ifdef GGML_USE_MUSA
+namespace wmma = mtmusa::wmma;
+#else // GGML_USE_MUSA
 namespace wmma = nvcuda::wmma;
+#endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
 #undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>

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

@@ -11,6 +11,8 @@
 #include "ggml-cuda/clamp.cuh"
 #include "ggml-cuda/concat.cuh"
 #include "ggml-cuda/conv-transpose-1d.cuh"
+#include "ggml-cuda/conv2d-dw.cuh"
+#include "ggml-cuda/conv2d-transpose.cuh"
 #include "ggml-cuda/convert.cuh"
 #include "ggml-cuda/count-equal.cuh"
 #include "ggml-cuda/cpy.cuh"
@@ -35,6 +37,7 @@
 #include "ggml-cuda/ssm-scan.cuh"
 #include "ggml-cuda/sum.cuh"
 #include "ggml-cuda/sumrows.cuh"
+#include "ggml-cuda/mean.cuh"
 #include "ggml-cuda/tsembd.cuh"
 #include "ggml-cuda/unary.cuh"
 #include "ggml-cuda/upscale.cuh"
@@ -47,6 +50,7 @@
 #include <atomic>
 #include <charconv>
 #include <cinttypes>
+#include <condition_variable>
 #include <cstddef>
 #include <cstdint>
 #include <float.h>
@@ -54,9 +58,8 @@
 #include <map>
 #include <memory>
 #include <mutex>
-#include <stdint.h>
-#include <stdio.h>
 #include <stdarg.h>
+#include <stdio.h>
 #include <stdlib.h>
 #include <string>
 #include <vector>
@@ -97,8 +100,7 @@ int ggml_cuda_get_device() {
 static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device) {
     ggml_cuda_set_device(device);
     cudaError_t err;
-    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr)
-    {
+    if (getenv("GGML_CUDA_ENABLE_UNIFIED_MEMORY") != nullptr) {
         err = cudaMallocManaged(ptr, size);
 #if defined(GGML_USE_HIP)
         if (err == hipSuccess) {
@@ -116,9 +118,7 @@ static cudaError_t ggml_cuda_device_malloc(void ** ptr, size_t size, int device)
             err = cudaMalloc(ptr, size);
         }
 #endif // defined(GGML_USE_HIP)
-    }
-    else
-    {
+    } else {
         err = cudaMalloc(ptr, size);
     }
     return err;
@@ -514,6 +514,33 @@ std::unique_ptr<ggml_cuda_pool> ggml_backend_cuda_context::new_pool_for_device(i
     return std::unique_ptr<ggml_cuda_pool>(new ggml_cuda_pool_leg(device));
 }
 
+// destroying a cuBLAS handle while a graph is being captured in a different thread can result in a CUDA error
+// this lock is used to ensure that no cuBLAS handle is destroyed while a graph is being captured
+
+static std::mutex ggml_cuda_lock;
+static std::condition_variable ggml_cuda_lock_cv;
+static std::atomic<int> ggml_cuda_lock_counter;
+
+ggml_backend_cuda_context::~ggml_backend_cuda_context() {
+    std::unique_lock<std::mutex> lock(ggml_cuda_lock);
+    ggml_cuda_lock_cv.wait(lock, []{ return ggml_cuda_lock_counter.load(std::memory_order_relaxed) == 0; });
+
+    if (copy_event != nullptr) {
+        CUDA_CHECK(cudaEventDestroy(copy_event));
+    }
+    for (int i = 0; i < GGML_CUDA_MAX_DEVICES; ++i) {
+        for (int j = 0; j < GGML_CUDA_MAX_STREAMS; ++j) {
+            if (streams[i][j] != nullptr) {
+                CUDA_CHECK(cudaStreamDestroy(streams[i][j]));
+            }
+        }
+        if (cublas_handles[i] != nullptr) {
+            CUBLAS_CHECK(cublasDestroy(cublas_handles[i]));
+        }
+    }
+}
+
+
 // cuda buffer
 
 struct ggml_backend_cuda_buffer_context {
@@ -1200,9 +1227,12 @@ static void ggml_cuda_op_mul_mat_cublas(
 
     const int cc = ggml_cuda_info().devices[id].cc;
 
+    const bool supports_bf16 = GGML_CUDA_CC_IS_NVIDIA(cc) || GGML_CUDA_CC_IS_AMD(cc) ||
+        (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
+
     const bool use_fp16 = (src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) && ggml_is_contiguous(src0) && row_diff == src0->ne[1] && dst->op_params[0] == GGML_PREC_DEFAULT;
 
-    if (src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
+    if (supports_bf16 && src0->type == GGML_TYPE_BF16 && ggml_is_contiguous(src0) && row_diff == src0->ne[1]) {
         ggml_cuda_pool_alloc<nv_bfloat16> src1_as_bf16(ctx.pool(id));
         if (src1->type != GGML_TYPE_BF16) {
             const to_bf16_cuda_t to_bf16_cuda = ggml_get_to_bf16_cuda(src1->type);
@@ -1230,7 +1260,7 @@ static void ggml_cuda_op_mul_mat_cublas(
 
         const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
         to_fp32_cuda(dst_bf16.get(), dst_dd_i, row_diff*src1_ncols, stream);
-    } else if (((GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) || GGML_CUDA_CC_IS_AMD(cc)) && use_fp16) {
+    } else if (fast_fp16_hardware_available(cc) && use_fp16) {
         // convert src0 and src1 to fp16, multiply as fp16, convert dst to fp32
         ggml_cuda_pool_alloc<half> src0_as_f16(ctx.pool(id));
         if (src0->type != GGML_TYPE_F16) {
@@ -1916,16 +1946,14 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
         && ggml_nbytes(src0) != ggml_backend_buffer_get_alloc_size(src0->buffer, src0) && src0->view_src;
 
     bool use_mul_mat_vec   = (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
-        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
     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;
     bool use_mul_mat_q     = ggml_is_quantized(src0->type) && !bad_padding_clear
         && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
 
-    bool any_gpus_with_slow_fp16   = false;
-    bool any_gpus_without_fp16_mma = false;
+    bool any_gpus_with_slow_fp16 = false;
 
     if (split) {
         ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) src0->buffer->buft->context;
@@ -1936,16 +1964,16 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
                 continue;
             }
 
-            const int cc              = ggml_cuda_info().devices[id].cc;
-            use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-            any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-            any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+            const int cc            = ggml_cuda_info().devices[id].cc;
+            use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+            use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+            any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
         }
     } else {
-        const int cc              = ggml_cuda_info().devices[ctx.device].cc;
-        use_mul_mat_q             = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
-        any_gpus_with_slow_fp16   = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
-        any_gpus_without_fp16_mma = any_gpus_without_fp16_mma || !fp16_mma_hardware_available(cc);
+        const int cc            = ggml_cuda_info().devices[ctx.device].cc;
+        use_mul_mat_q           = use_mul_mat_q             && ggml_cuda_should_use_mmq(src0->type, cc, src1->ne[1]);
+        use_mul_mat_vec         = use_mul_mat_vec           && ggml_cuda_should_use_mmv(src0->type, cc, src0->ne, src1->ne[1]);
+        any_gpus_with_slow_fp16 = any_gpus_with_slow_fp16   || !fast_fp16_hardware_available(cc);
     }
 
     // debug helpers
@@ -1956,7 +1984,7 @@ static void ggml_cuda_mul_mat(ggml_backend_cuda_context & ctx, const ggml_tensor
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && use_mul_mat_vec && (src0->ne[1] <= MMV_MAX_ROWS || any_gpus_without_fp16_mma)) {
+    if (!split && use_mul_mat_vec) {
         // the custom F16 vector kernel can be used over batched cuBLAS GEMM
         // but this is only faster for GPUs without tensor cores or with a thin src0 matrix (particularly KQV in attention)
         ggml_cuda_mul_mat_vec(ctx, src0, src1, nullptr, dst);
@@ -2310,6 +2338,12 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_IM2COL:
             ggml_cuda_op_im2col(ctx, dst);
             break;
+        case GGML_OP_CONV_2D_DW:
+            ggml_cuda_op_conv2d_dw(ctx, dst);
+            break;
+        case GGML_OP_CONV_TRANSPOSE_2D:
+            ggml_cuda_conv_2d_transpose_p0(ctx, dst);
+            break;
         case GGML_OP_CONV_TRANSPOSE_1D:
             ggml_cuda_op_conv_transpose_1d(ctx,dst);
             break;
@@ -2322,6 +2356,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_SUM_ROWS:
             ggml_cuda_op_sum_rows(ctx, dst);
             break;
+        case GGML_OP_MEAN:
+            ggml_cuda_op_mean(ctx, dst);
+            break;
         case GGML_OP_SSM_CONV:
             ggml_cuda_op_ssm_conv(ctx, dst);
             break;
@@ -2685,6 +2722,11 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
 
             CUDA_CHECK(cudaStreamEndCapture(cuda_ctx->stream(), &cuda_ctx->cuda_graph->graph));
             graph_evaluated_or_captured = true; // CUDA graph has been captured
+
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            if (ggml_cuda_lock_counter.fetch_sub(1, std::memory_order_relaxed) == 1) {
+                ggml_cuda_lock_cv.notify_all();
+            }
         } else {
             graph_evaluated_or_captured = true; // ggml graph has been directly evaluated
         }
@@ -2760,7 +2802,13 @@ static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t backend,
         }
     }
 
-    if (use_cuda_graph && cuda_graph_update_required) { // Start CUDA graph capture
+    if (use_cuda_graph && cuda_graph_update_required) {
+        // Start CUDA graph capture
+        {
+            std::lock_guard<std::mutex> lock(ggml_cuda_lock);
+            ggml_cuda_lock_counter.fetch_add(1, std::memory_order_relaxed);
+        }
+
         CUDA_CHECK(cudaStreamBeginCapture(cuda_ctx->stream(), cudaStreamCaptureModeRelaxed));
     }
 
@@ -3016,9 +3064,16 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     return false;
                 }
 #ifdef GGML_USE_MUSA
-                if (b->type == GGML_TYPE_F16 && b->ne[2]*b->ne[3] > 1 &&
-                    !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
-                    return false;
+                const int cc = ggml_cuda_info().devices[dev_ctx->device].cc;
+                if (b->ne[2]*b->ne[3] > 1 && !ggml_is_transposed(a) && !ggml_is_transposed(b)) {
+                    if (GGML_CUDA_CC_IS_QY1(cc) && op->op == GGML_OP_MUL_MAT &&
+                            a->type == GGML_TYPE_F16 && b->type == GGML_TYPE_F16) {
+                        return false;
+                    }
+                    if (GGML_CUDA_CC_IS_QY2(cc) && op->op == GGML_OP_MUL_MAT_ID &&
+                            a->type == GGML_TYPE_Q2_K && b->type == GGML_TYPE_F32) {
+                        return false;
+                    }
                 }
 #endif // GGML_USE_MUSA
                 switch (a->type) {
@@ -3045,11 +3100,6 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_IQ4_NL:
                     case GGML_TYPE_IQ4_XS:
                     case GGML_TYPE_BF16:
-#ifdef GGML_USE_MUSA
-                        if (a->type == GGML_TYPE_Q3_K) {
-                            return false;
-                        }
-#endif // GGML_USE_MUSA
                         return true;
                     default:
                         return false;
@@ -3209,9 +3259,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             return op->src[0]->nb[0] == ggml_type_size(op->src[0]->type) && ggml_is_contiguous_2(op->src[0]);
         }
         case GGML_OP_IM2COL:
+        case GGML_OP_CONV_2D_DW:
+        case GGML_OP_CONV_TRANSPOSE_2D:
         case GGML_OP_POOL_2D:
         case GGML_OP_SUM:
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_ARGSORT:
         case GGML_OP_ACC:
             return true;

ggml/src/ggml-cuda/mean.cu

@@ -0,0 +1,19 @@
+#include "mean.cuh"
+
+void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0   = dst->src[0];
+    const float *       src0_d = (const float *) src0->data;
+    float *             dst_d  = (float *) dst->data;
+    cudaStream_t        stream = ctx.stream();
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+
+    const int64_t ncols = src0->ne[0];
+    const int64_t nrows = ggml_nrows(src0);
+
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+}

ggml/src/ggml-cuda/mean.cuh

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

ggml/src/ggml-cuda/mmv.cu

@@ -2,25 +2,26 @@
 #include "common.cuh"
 #include "mmv.cuh"
 
-template <typename T, typename type_acc, int block_size>
+template <typename T, typename type_acc, int ncols_dst, int block_size>
 static __global__ void mul_mat_vec(
         const T * __restrict__ x, const float * __restrict__ y, const int32_t * __restrict__ ids, float * __restrict__ dst,
-        const int64_t ncols2, const int64_t nchannels_y, const int64_t stride_row,
-        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 int64_t row         = blockIdx.x;
-    const int64_t channel_dst = blockIdx.y;
-    const int64_t channel_x   = ids ? ids[channel_dst]          : channel_dst / channel_ratio;
-    const int64_t channel_y   = ids ? channel_dst % nchannels_y : channel_dst;
-    const int64_t sample_dst  = blockIdx.z;
-    const int64_t sample_x    = sample_dst / sample_ratio;
-    const int64_t sample_y    = sample_dst;
-    const int     tid         = threadIdx.x;
+        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 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_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_y    = sample_dst;
+    const int tid         = threadIdx.x;
+
     constexpr int warp_size   = ggml_cuda_get_physical_warp_size();
 
-    x   += sample_x  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
-    y   += sample_y  *stride_sample_y   + channel_y  *stride_channel_y;
-    dst += sample_dst*stride_sample_dst + channel_dst*stride_channel_dst;
+    x   += int64_t(sample_x)  *stride_sample_x   + channel_x  *stride_channel_x   + row*stride_row;
+    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;
 
     const float2 * y2 = (const float2 *) y;
 
@@ -34,81 +35,108 @@ static __global__ void mul_mat_vec(
         __syncthreads();
     }
 
-    float sumf = 0.0f;
+    float sumf[ncols_dst] = {0.0f};
 
     if constexpr (std::is_same<T, float>::value) {
         const float2 * x2 = (const float2 *) x;
 
-        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
             const float2 tmpx = x2[col2];
-            const float2 tmpy = y2[col2];
-            sumf += tmpx.x*tmpy.x;
-            sumf += tmpx.y*tmpy.y;
+
+#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;
+            }
         }
     } else if constexpr (std::is_same<T, half>::value) {
         const half2 * x2 = (const half2 *) x;
 
         if (std::is_same<type_acc, float>::value) {
-            for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
                 const float2 tmpx = __half22float2(x2[col2]);
-                const float2 tmpy = y2[col2];
-                sumf += tmpx.x * tmpy.x;
-                sumf += tmpx.y * tmpy.y;
+
+#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;
+                }
             }
         } else {
 #ifdef FP16_AVAILABLE
-            half2 sumh2 = make_half2(0.0f, 0.0f);
+            half2 sumh2[ncols_dst] = {{0.0f, 0.0f}};
 
-            for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
-                const float2 tmp = y2[col2];
-                sumh2 += x2[col2] * make_half2(tmp.x, tmp.y);
+            for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+                const half2 tmpx = 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);
+                }
             }
 
-            sumf = __low2float(sumh2) + __high2float(sumh2);
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                sumf[j] = __low2float(sumh2[j]) + __high2float(sumh2[j]);
+            }
 #else
             NO_DEVICE_CODE;
 #endif // FP16_AVAILABLE
         }
     } else if constexpr (std::is_same<T, nv_bfloat16>::value) {
         const int * x2 = (const int *) x;
-        for (int64_t col2 = tid; col2 < ncols2; col2 += block_size) {
-            const int    tmpx = x2[col2];
-            const float2 tmpy = y2[col2];
-            sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-            sumf += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+        for (int col2 = tid; col2 < ncols2; col2 += block_size) {
+            const int tmpx = x2[col2];
+#pragma unroll
+            for (int j = 0; j < ncols_dst; ++j) {
+                const float2 tmpy = y2[j*stride_col_y2 + col2];
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+            }
         }
     } else {
         static_assert(std::is_same<T, void>::value, "unsupported type");
     }
 
-    sumf = warp_reduce_sum<warp_size>(sumf);
+#pragma unroll
+    for (int j = 0; j < ncols_dst; ++j) {
+        sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
 
-    if (block_size > warp_size) {
-        buf_iw[tid/warp_size] = sumf;
-        __syncthreads();
-        if (tid >= warp_size) {
-            return;
+        if (block_size > warp_size) {
+            buf_iw[tid/warp_size] = sumf[j];
+            __syncthreads();
+            if (tid < warp_size) {
+                sumf[j] = buf_iw[tid];
+                sumf[j] = warp_reduce_sum<warp_size>(sumf[j]);
+            }
+            if (j < ncols_dst) {
+                __syncthreads();
+            }
         }
-        sumf = buf_iw[tid];
-        sumf = warp_reduce_sum<warp_size>(sumf);
     }
 
-    if (tid != 0) {
+    if (tid >= ncols_dst) {
         return;
     }
 
-    dst[row] = sumf;
+    dst[tid*stride_col_dst + row] = sumf[tid];
 }
 
-template <typename T, typename type_acc>
+template <typename T, typename type_acc, int ncols_dst>
 static void launch_mul_mat_vec_cuda(
         const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_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,
         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) {
-    GGML_ASSERT(ncols      % 2 == 0);
-    GGML_ASSERT(stride_row % 2 == 0);
+    GGML_ASSERT(ncols        % 2 == 0);
+    GGML_ASSERT(stride_row   % 2 == 0);
+    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;
@@ -138,44 +166,52 @@ static void launch_mul_mat_vec_cuda(
     const dim3 block_dims(block_size_best, 1, 1);
     switch (block_size_best) {
         case   32: {
-            mul_mat_vec<T, type_acc,  32><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst,  32><<<block_nums, block_dims, smem, 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);
         } break;
         case   64: {
-            mul_mat_vec<T, type_acc,  64><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst,  64><<<block_nums, block_dims, smem, 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);
         } break;
         case   96: {
-            mul_mat_vec<T, type_acc,  96><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst,  96><<<block_nums, block_dims, smem, 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);
         } break;
         case  128: {
-            mul_mat_vec<T, type_acc, 128><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst, 128><<<block_nums, block_dims, smem, 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);
         } break;
         case  160: {
-            mul_mat_vec<T, type_acc, 160><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst, 160><<<block_nums, block_dims, smem, 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);
         } break;
         case  192: {
-            mul_mat_vec<T, type_acc, 192><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst, 192><<<block_nums, block_dims, smem, 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);
         } break;
         case  224: {
-            mul_mat_vec<T, type_acc, 224><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst, 224><<<block_nums, block_dims, smem, 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);
         } break;
         case  256: {
-            mul_mat_vec<T, type_acc, 256><<<block_nums, block_dims, smem, stream>>>
-                (x, y, ids, dst, ncols/2, nchannels_y, stride_row, 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<T, type_acc, ncols_dst, 256><<<block_nums, block_dims, smem, 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);
         } break;
         default: {
             GGML_ABORT("fatal error");
@@ -183,23 +219,91 @@ static void launch_mul_mat_vec_cuda(
     }
 }
 
+template <typename T, typename type_acc>
+static void mul_mat_vec_cuda_switch_ncols_dst(
+        const T * x, const float * y, const int32_t * ids, 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,
+        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) {
+    switch (ncols_dst) {
+        case 1:
+            launch_mul_mat_vec_cuda<T, type_acc, 1>
+                (x, y, ids, 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_cuda<T, type_acc, 2>
+                (x, y, ids, 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_cuda<T, type_acc, 3>
+                (x, y, ids, 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_cuda<T, type_acc, 4>
+                (x, y, ids, 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_cuda<T, type_acc, 5>
+                (x, y, ids, 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_cuda<T, type_acc, 6>
+                (x, y, ids, 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_cuda<T, type_acc, 7>
+                (x, y, ids, 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_cuda<T, type_acc, 8>
+                (x, y, ids, 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;
+        default:
+            GGML_ABORT("fatal error");
+            break;
+    }
+}
+
 template<typename T>
 static void mul_mat_vec_cuda(
         const T * x, const float * y, const int32_t * ids, float * dst,
-        const int64_t ncols, const int64_t nrows, const int64_t stride_row, const int64_t nchannels_x, const int64_t nchannels_y, const int64_t nchannels_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<T, half>::value) {
         if (prec == GGML_PREC_DEFAULT) {
-            launch_mul_mat_vec_cuda<T, half>
-                (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+            mul_mat_vec_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);
             return;
         }
     }
-    launch_mul_mat_vec_cuda<T, float>
-        (x, y, ids, dst, ncols, nrows, stride_row, nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y,
+    mul_mat_vec_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);
 }
 
@@ -246,24 +350,24 @@ void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor *
     const int64_t stride_channel_dst = ids ? s1   : s2;
     const int64_t stride_channel_y   = ids ? s11  : s12;
 
-    GGML_ASSERT(ncols_dst == 1);
+    GGML_ASSERT(!ids || ncols_dst == 1);
 
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0->data;
-            mul_mat_vec_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, 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_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, 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_cuda(src0_d, src1_d, ids_d, dst_d, ne00, ne01, s01,
+            mul_mat_vec_cuda(src0_d, src1_d, ids_d, 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;
@@ -282,16 +386,19 @@ void ggml_cuda_op_mul_mat_vec(
     GGML_ASSERT(dst->type  == GGML_TYPE_F32);
 
     const int64_t ne00 = src0->ne[0];
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne0  =  dst->ne[0];
     const int64_t row_diff = row_high - row_low;
 
-    GGML_ASSERT(src1_ncols == 1);
-
-    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
+    const int id = ggml_cuda_get_device();
+    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;
+    const int64_t stride_col_dst     = id == ctx.device ? ne0 : row_diff; // main device has larger memory buffer
     const int64_t nchannels_x        = 1;
     const int64_t nchannels_y        = 1;
     const int64_t nchannels_dst      = 1;
@@ -307,19 +414,19 @@ void ggml_cuda_op_mul_mat_vec(
     switch (src0->type) {
         case GGML_TYPE_F32: {
             const float * src0_d = (const float *) src0_dd_i;
-            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, 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_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, 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_cuda(src0_d, src1_ddf_i, nullptr, dst_dd_i, ne00, row_diff, stride_row,
+            mul_mat_vec_cuda(src0_d, src1_ddf_i, nullptr, 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;
@@ -334,3 +441,66 @@ void ggml_cuda_op_mul_mat_vec(
     GGML_UNUSED(src1_ncols);
     GGML_UNUSED(src1_padded_row_size);
 }
+
+bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11) {
+    if (src0_ne[0] % 2 != 0) {
+        return false;
+    }
+    switch (type) {
+        case GGML_TYPE_F32:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return ne11 <= 8;
+                }
+                if (cc >= GGML_CUDA_CC_TURING) {
+                    return ne11 <= 4;
+                }
+                return ne11 <= 3;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp32_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_F16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (fp16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (fp16_mma_hardware_available(cc)) {
+                    if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+                        return ne11 <= 5;
+                    }
+                    return ne11 <= 2;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        case GGML_TYPE_BF16:
+            if (GGML_CUDA_CC_IS_NVIDIA(cc)) {
+                const bool src0_small = (src0_ne[1] <= 512 || src0_ne[2]*src0_ne[3] == 1);
+                if (cc >= GGML_CUDA_CC_ADA_LOVELACE) {
+                    return src0_small && ne11 <= 4;
+                }
+                if (bf16_mma_hardware_available(cc)) {
+                    return src0_small && ne11 <= 3;
+                }
+                return ne11 <= 8;
+            } else if (GGML_CUDA_CC_IS_AMD(cc)) {
+                if (bf16_mma_hardware_available(cc)) {
+                    return ne11 <= 3;
+                }
+                return ne11 <= 8;
+            }
+            return ne11 <= 8;
+        default:
+            return false;
+    }
+}

ggml/src/ggml-cuda/mmv.cuh

@@ -1,8 +1,5 @@
 #include "common.cuh"
 
-// maximum number of src0 rows with which to use mul_mat_vec over cuBLAS if FP16 tensor cores are available
-#define MMV_MAX_ROWS 512
-
 void ggml_cuda_mul_mat_vec(ggml_backend_cuda_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * ids, ggml_tensor * dst);
 
 void ggml_cuda_op_mul_mat_vec(
@@ -10,3 +7,5 @@ void ggml_cuda_op_mul_mat_vec(
     const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, const char * src0_dd_i, const float * src1_ddf_i,
     const char * src1_ddq_i, float * dst_dd_i, const int64_t row_low, const int64_t row_high, const int64_t src1_ncols,
     const int64_t src1_padded_row_size, cudaStream_t stream);
+
+bool ggml_cuda_should_use_mmv(enum ggml_type type, int cc, const int64_t * src0_ne, int64_t ne11);

ggml/src/ggml-cuda/sumrows.cu

@@ -1,25 +1,9 @@
 #include "sumrows.cuh"
 
-static __global__ void k_sum_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col == 0) {
-        dst[row] = sum;
-    }
-}
-
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
     const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
-    k_sum_rows_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -35,5 +19,8 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    sum_rows_f32_cuda(src0_d, dst_d, ncols, nrows, stream);
+    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const dim3 block_nums(nrows, 1, 1);
+
+    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
 }

ggml/src/ggml-cuda/sumrows.cuh

@@ -1,5 +1,4 @@
 #include "common.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream);
-
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst);

ggml/src/ggml-impl.h

@@ -317,203 +317,81 @@ struct ggml_cgraph ggml_graph_view(struct ggml_cgraph * cgraph, int i0, int i1);
 GGML_API void * ggml_aligned_malloc(size_t size);
 GGML_API void ggml_aligned_free(void * ptr, size_t size);
 
-// FP16 to FP32 conversion
+// FP16 <-> FP32
+// ref: https://github.com/Maratyszcza/FP16
 
-// 16-bit float
-// on Arm, we use __fp16
-// on x86, we use uint16_t
-//
-// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
-// for     MUSA compilers        , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
-//
-#if defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-    #define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        __fp16 tmp;
-        memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-        return (float)tmp;
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-        ggml_fp16_t res;
-        __fp16 tmp = f;
-        memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-        return res;
-    }
-
-#elif defined(__F16C__)
-
-    #ifdef _MSC_VER
-        #define GGML_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
-        #define GGML_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
-    #else
-        #define GGML_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
-        #define GGML_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
-    #endif
-
-#elif defined(__POWER9_VECTOR__)
-
-    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-    /* the inline asm below is about 12% faster than the lookup method */
-    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        float f;
-        double d;
-        __asm__(
-            "mtfprd %0,%2\n"
-            "xscvhpdp %0,%0\n"
-            "frsp %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=f"(f):
-            /* in */   "r"(h));
-        return f;
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-        double d;
-        ggml_fp16_t r;
-        __asm__( /* xscvdphp can work on double or single precision */
-            "xscvdphp %0,%2\n"
-            "mffprd %1,%0\n" :
-            /* temp */ "=d"(d),
-            /* out */  "=r"(r):
-            /* in */   "f"(f));
-        return r;
-    }
-
-#elif defined(__riscv) && defined(__riscv_zfhmin)
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        float f;
-        __asm__(
-            "fmv.h.x %[f], %[h]\n\t"
-            "fcvt.s.h %[f], %[f]"
-            : [f] "=&f" (f)
-            : [h] "r" (h)
-        );
-        return f;
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-        ggml_fp16_t res;
-        __asm__(
-            "fcvt.h.s %[f], %[f]\n\t"
-            "fmv.x.h %[h], %[f]"
-            : [h] "=&r" (res)
-            : [f] "f" (f)
-        );
-        return res;
-    }
-
-    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-    #define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
-    #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
-
-#else
-
-    // FP16 <-> FP32
-    // ref: https://github.com/Maratyszcza/FP16
-
-    static inline float fp32_from_bits(uint32_t w) {
-        union {
-            uint32_t as_bits;
-            float as_value;
-        } fp32;
-        fp32.as_bits = w;
-        return fp32.as_value;
-    }
-
-    static inline uint32_t fp32_to_bits(float f) {
-        union {
-            float as_value;
-            uint32_t as_bits;
-        } fp32;
-        fp32.as_value = f;
-        return fp32.as_bits;
-    }
-
-    static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-        const uint32_t w = (uint32_t) h << 16;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        const uint32_t two_w = w + w;
-
-        const uint32_t exp_offset = UINT32_C(0xE0) << 23;
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float exp_scale = 0x1.0p-112f;
-    #else
-        const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
-    #endif
-        const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
-
-        const uint32_t magic_mask = UINT32_C(126) << 23;
-        const float magic_bias = 0.5f;
-        const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
-
-        const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
-        const uint32_t result = sign |
-            (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
-        return fp32_from_bits(result);
-    }
-
-    static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    #if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
-        const float scale_to_inf = 0x1.0p+112f;
-        const float scale_to_zero = 0x1.0p-110f;
-    #else
-        const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
-        const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
-    #endif
-        float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
-
-        const uint32_t w = fp32_to_bits(f);
-        const uint32_t shl1_w = w + w;
-        const uint32_t sign = w & UINT32_C(0x80000000);
-        uint32_t bias = shl1_w & UINT32_C(0xFF000000);
-        if (bias < UINT32_C(0x71000000)) {
-            bias = UINT32_C(0x71000000);
-        }
-
-        base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
-        const uint32_t bits = fp32_to_bits(base);
-        const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
-        const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
-        const uint32_t nonsign = exp_bits + mantissa_bits;
-        return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
-    }
-
-    #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-    #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
-
-#endif // defined(__ARM_NEON) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
-
-// precomputed f32 table for f16 (256 KB)
-// defined in ggml.c, initialized in ggml_init()
-GGML_API float ggml_table_f32_f16[1 << 16];
-
-// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
-// so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
-// This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
-inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
-    uint16_t s;
-    memcpy(&s, &f, sizeof(uint16_t));
-    return ggml_table_f32_f16[s];
+static inline float fp32_from_bits(uint32_t w) {
+    union {
+        uint32_t as_bits;
+        float as_value;
+    } fp32;
+    fp32.as_bits = w;
+    return fp32.as_value;
 }
 
-#define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
+static inline uint32_t fp32_to_bits(float f) {
+    union {
+        float as_value;
+        uint32_t as_bits;
+    } fp32;
+    fp32.as_value = f;
+    return fp32.as_bits;
+}
 
-#if !defined(GGML_FP32_TO_FP16)
-#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
+    const uint32_t w = (uint32_t) h << 16;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    const uint32_t two_w = w + w;
+
+    const uint32_t exp_offset = UINT32_C(0xE0) << 23;
+#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+    const float exp_scale = 0x1.0p-112f;
+#else
+    const float exp_scale = fp32_from_bits(UINT32_C(0x7800000));
 #endif
+    const float normalized_value = fp32_from_bits((two_w >> 4) + exp_offset) * exp_scale;
+
+    const uint32_t magic_mask = UINT32_C(126) << 23;
+    const float magic_bias = 0.5f;
+    const float denormalized_value = fp32_from_bits((two_w >> 17) | magic_mask) - magic_bias;
+
+    const uint32_t denormalized_cutoff = UINT32_C(1) << 27;
+    const uint32_t result = sign |
+        (two_w < denormalized_cutoff ? fp32_to_bits(denormalized_value) : fp32_to_bits(normalized_value));
+    return fp32_from_bits(result);
+}
+
+static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
+#if (defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 199901L) || defined(__GNUC__) && !defined(__STRICT_ANSI__)) && (!defined(__cplusplus) || __cplusplus >= 201703L)
+    const float scale_to_inf = 0x1.0p+112f;
+    const float scale_to_zero = 0x1.0p-110f;
+#else
+    const float scale_to_inf = fp32_from_bits(UINT32_C(0x77800000));
+    const float scale_to_zero = fp32_from_bits(UINT32_C(0x08800000));
+#endif
+    float base = (fabsf(f) * scale_to_inf) * scale_to_zero;
+
+    const uint32_t w = fp32_to_bits(f);
+    const uint32_t shl1_w = w + w;
+    const uint32_t sign = w & UINT32_C(0x80000000);
+    uint32_t bias = shl1_w & UINT32_C(0xFF000000);
+    if (bias < UINT32_C(0x71000000)) {
+        bias = UINT32_C(0x71000000);
+    }
+
+    base = fp32_from_bits((bias >> 1) + UINT32_C(0x07800000)) + base;
+    const uint32_t bits = fp32_to_bits(base);
+    const uint32_t exp_bits = (bits >> 13) & UINT32_C(0x00007C00);
+    const uint32_t mantissa_bits = bits & UINT32_C(0x00000FFF);
+    const uint32_t nonsign = exp_bits + mantissa_bits;
+    return (sign >> 16) | (shl1_w > UINT32_C(0xFF000000) ? UINT16_C(0x7E00) : nonsign);
+}
+
+#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
+#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+
+#define GGML_FP16_TO_FP32(x) GGML_COMPUTE_FP16_TO_FP32(x)
+#define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
 
 /**
  * Converts brain16 to float32.

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

@@ -521,6 +521,22 @@ typedef struct {
     uint64_t nb2;
 } ggml_metal_kargs_get_rows;
 
+typedef struct {
+    int32_t  nk0;
+    int32_t  ne01;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int32_t  ne11;
+    int32_t  ne12;
+    uint64_t nb10;
+    uint64_t nb11;
+    uint64_t nb12;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+} ggml_metal_kargs_set_rows;
+
 typedef struct {
     int64_t  ne00;
     int64_t  ne01;

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

@@ -48,22 +48,28 @@ static struct ggml_backend_metal_device_context {
     int            mtl_device_ref_count;
     id<MTLLibrary> mtl_library;
 
+    NSLock * mtl_lock;
+
     bool has_simdgroup_reduction;
     bool has_simdgroup_mm;
     bool has_residency_sets;
     bool has_bfloat;
     bool use_bfloat;
 
+    size_t max_size;
+
     char name[128];
 } g_ggml_ctx_dev_main = {
     /*.mtl_device              =*/ nil,
     /*.mtl_device_ref_count    =*/ 0,
     /*.mtl_library             =*/ nil,
+    /*.mtl_lock                =*/ nil,
     /*.has_simdgroup_reduction =*/ false,
     /*.has_simdgroup_mm        =*/ false,
     /*.has_residency_sets      =*/ false,
     /*.has_bfloat              =*/ false,
     /*.use_bfloat              =*/ false,
+    /*.max_size                =*/ 0,
     /*.name                    =*/ "",
 };
 
@@ -71,6 +77,10 @@ static struct ggml_backend_metal_device_context {
 static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_device_context * ctx) {
     assert(ctx != NULL);
 
+    if (ctx->mtl_lock == nil) {
+        ctx->mtl_lock = [[NSLock alloc] init];
+    }
+
     if (ctx->mtl_device == nil) {
         ctx->mtl_device = MTLCreateSystemDefaultDevice();
     }
@@ -94,6 +104,8 @@ static id<MTLDevice> ggml_backend_metal_device_acq(struct ggml_backend_metal_dev
         ctx->use_bfloat = false;
 #endif
 
+        ctx->max_size = ctx->mtl_device.maxBufferLength;
+
         strncpy(ctx->name, [[ctx->mtl_device name] UTF8String], sizeof(ctx->name) - 1);
     }
 
@@ -110,6 +122,11 @@ static void ggml_backend_metal_device_rel(struct ggml_backend_metal_device_conte
     ctx->mtl_device_ref_count--;
 
     if (ctx->mtl_device_ref_count == 0) {
+        if (ctx->mtl_lock) {
+            [ctx->mtl_lock release];
+            ctx->mtl_lock = nil;
+        }
+
         if (ctx->mtl_library) {
             [ctx->mtl_library release];
             ctx->mtl_library = nil;
@@ -185,6 +202,15 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_F32,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_F16,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,
+    GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_L2_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -194,11 +220,14 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,
     GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32,
+    GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4,
     GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16,
@@ -498,6 +527,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_COS,
     GGML_METAL_KERNEL_TYPE_NEG,
     GGML_METAL_KERNEL_TYPE_SUM_ROWS,
+    GGML_METAL_KERNEL_TYPE_MEAN,
     GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,
     GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,
     GGML_METAL_KERNEL_TYPE_ARGMAX,
@@ -976,7 +1006,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     struct ggml_backend_metal_context * ctx = calloc(1, sizeof(struct ggml_backend_metal_context));
     struct ggml_backend_metal_device_context * ctx_dev = dev->context;
 
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     GGML_LOG_INFO("%s: picking default device: %s\n", __func__, [[device name] UTF8String]);
 
@@ -990,9 +1020,16 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
     ctx->d_queue = dispatch_queue_create("ggml-metal", DISPATCH_QUEUE_CONCURRENT);
 
     // load library
-    if (ctx_dev->mtl_library == nil) {
-        ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat);
+    {
+        [ctx_dev->mtl_lock lock];
+
+        if (ctx_dev->mtl_library == nil) {
+            ctx_dev->mtl_library = ggml_metal_load_library(device, ctx_dev->use_bfloat);
+        }
+
+        [ctx_dev->mtl_lock unlock];
     }
+
     id<MTLLibrary> metal_library = ctx_dev->mtl_library;
     if (metal_library == nil) {
         GGML_LOG_ERROR("%s: error: metal library is nil\n", __func__);
@@ -1141,6 +1178,15 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,                 get_rows_iq4_nl,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,                 get_rows_iq4_xs,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,                    get_rows_i32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_F32,                    set_rows_f32,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_F16,                    set_rows_f16,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16,                   set_rows_bf16,                   use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0,                   set_rows_q8_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0,                   set_rows_q4_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1,                   set_rows_q4_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0,                   set_rows_q5_0,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1,                   set_rows_q5_1,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL,                 set_rows_iq4_nl,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                        rms_norm,                        has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_L2_NORM,                         l2_norm,                         has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                      group_norm,                      has_simdgroup_reduction);
@@ -1150,11 +1196,14 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV6_F32,                   rwkv_wkv6_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RWKV_WKV7_F32,                   rwkv_wkv7_f32,                   true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,                  mul_mv_f32_f32,                  has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4,               mul_mv_f32_f32_c4,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32,                 mul_mv_bf16_f32,                 has_simdgroup_reduction && use_bfloat);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4,              mul_mv_bf16_f32_c4,              use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW,            mul_mv_bf16_f32_1row,            has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4,              mul_mv_bf16_f32_l4,              has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_BF16,                mul_mv_bf16_bf16,                has_simdgroup_reduction && use_bfloat);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,                  mul_mv_f16_f32,                  has_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4,               mul_mv_f16_f32_c4,               true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,             mul_mv_f16_f32_1row,             has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,               mul_mv_f16_f32_l4,               has_simdgroup_reduction);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,                  mul_mv_f16_f16,                  has_simdgroup_reduction);
@@ -1454,6 +1503,7 @@ static struct ggml_backend_metal_context * ggml_metal_init(ggml_backend_dev_t de
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_COS,                             cos,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NEG,                             neg,                             true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                        sum_rows,                        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MEAN,                            mean,                            true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGMAX,                          argmax,                          true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_AVG_F32,                 pool_2d_avg_f32,                 true);
         GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_POOL_2D_MAX_F32,                 pool_2d_max_f32,                 true);
@@ -1603,6 +1653,10 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
     const bool use_bfloat              = ctx_dev->use_bfloat;
 
     if (!use_bfloat) {
+        if (op->type == GGML_TYPE_BF16) {
+            return false;
+        }
+
         for (size_t i = 0, n = 3; i < n; ++i) {
             if (op->src[i] != NULL && op->src[i]->type == GGML_TYPE_BF16) {
                 return false;
@@ -1653,6 +1707,7 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
         case GGML_OP_LOG:
             return false; // TODO: implement
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_GROUP_NORM:
             return has_simdgroup_reduction && ggml_is_contiguous(op->src[0]);
@@ -1771,6 +1826,27 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_device_contex
             {
                 return op->ne[3] == 1;
             }
+        case GGML_OP_SET_ROWS:
+            {
+                if (op->src[0]->type != GGML_TYPE_F32) {
+                    return false;
+                }
+
+                switch (op->type) {
+                    case GGML_TYPE_F32:
+                    case GGML_TYPE_F16:
+                    case GGML_TYPE_BF16:
+                    case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_Q4_0:
+                    case GGML_TYPE_Q4_1:
+                    case GGML_TYPE_Q5_0:
+                    case GGML_TYPE_Q5_1:
+                    case GGML_TYPE_IQ4_NL:
+                        return true;
+                    default:
+                        return false;
+                };
+            }
         default:
             return false;
     }
@@ -2400,11 +2476,31 @@ static bool ggml_metal_encode_node(
                 [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
             } break;
         case GGML_OP_SUM_ROWS:
+        case GGML_OP_MEAN:
             {
                 GGML_ASSERT(src0->nb[0] == ggml_type_size(src0->type));
 
-                id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                id<MTLComputePipelineState> pipeline = nil;
 
+                switch (dst->op) {
+                    case GGML_OP_SUM_ROWS:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SUM_ROWS].pipeline;
+                        break;
+                    case GGML_OP_MEAN:
+                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MEAN].pipeline;
+                        break;
+                    default:
+                        GGML_ABORT("fatal error");
+                }
+
+                int nth = 32; // SIMD width
+
+                while (nth < ne00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
+                nth = MIN(nth, ne00);
 
                 ggml_metal_kargs_sum_rows args = {
                    /*.ne00 =*/ ne00,
@@ -2434,11 +2530,12 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                [encoder setBytes:&args length:sizeof(args) atIndex:2];
+                [encoder setBytes:&args length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
+                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                [encoder setThreadgroupMemoryLength:32*sizeof(float) atIndex:0];
 
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
             } break;
         case GGML_OP_SOFT_MAX:
             {
@@ -3063,14 +3160,23 @@ static bool ggml_metal_encode_node(
                                 nsg = 1;
                                 nr0 = 1;
                                 nr1 = 4;
-                                pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                if (ne00 == 4) {
+                                    nr0 = 32;
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32_C4].pipeline;
+                                } else {
+                                    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32].pipeline;
+                                }
                             } break;
                         case GGML_TYPE_F16:
                             {
                                 nsg = 1;
                                 nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
-                                    if (ne11 * ne12 < 4) {
+                                    if (ne00 == 4) {
+                                        nr0 = 32;
+                                        nr1 = 4;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_C4].pipeline;
+                                    } else if (ne11 * ne12 < 4) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW].pipeline;
                                     } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4].pipeline;
@@ -3089,7 +3195,11 @@ static bool ggml_metal_encode_node(
                                 nsg = 1;
                                 nr0 = 1;
                                 if (src1t == GGML_TYPE_F32) {
-                                    if (ne11 * ne12 < 4) {
+                                    if (ne00 == 4) {
+                                        nr0 = 32;
+                                        nr1 = 4;
+                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_C4].pipeline;
+                                    } else if (ne11 * ne12 < 4) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_1ROW].pipeline;
                                     } else if (ne00 >= 128 && ne01 >= 8 && ne00%4 == 0) {
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_BF16_F32_L4].pipeline;
@@ -3710,13 +3820,74 @@ static bool ggml_metal_encode_node(
                 };
 
                 [encoder setComputePipelineState:pipeline];
-                [encoder setBuffer:id_src0     offset:offs_src0 atIndex:0];
-                [encoder setBuffer:id_src1     offset:offs_src1 atIndex:1];
-                [encoder setBuffer:id_dst      offset:offs_dst  atIndex:2];
-                [encoder setBytes:&args length:sizeof(args) atIndex:3];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
 
                 [encoder dispatchThreadgroups:MTLSizeMake(ne10, ne11, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
             } break;
+        case GGML_OP_SET_ROWS:
+            {
+                id<MTLComputePipelineState> pipeline = nil;
+
+                switch (dst->type) {
+                    case GGML_TYPE_F32:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_F32   ].pipeline; break;
+                    case GGML_TYPE_F16:    pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_F16   ].pipeline; break;
+                    case GGML_TYPE_BF16:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_BF16  ].pipeline; break;
+                    case GGML_TYPE_Q8_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q8_0  ].pipeline; break;
+                    case GGML_TYPE_Q4_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_0  ].pipeline; break;
+                    case GGML_TYPE_Q4_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q4_1  ].pipeline; break;
+                    case GGML_TYPE_Q5_0:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_0  ].pipeline; break;
+                    case GGML_TYPE_Q5_1:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_Q5_1  ].pipeline; break;
+                    case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SET_ROWS_IQ4_NL].pipeline; break;
+                    default: GGML_ABORT("not implemented");
+                }
+
+                const int32_t nk0 = ne0/ggml_blck_size(dst->type);
+
+                int nth = 32; // SIMD width
+
+                while (nth < nk0 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                int nrptg = 1;
+                if (nth > nk0) {
+                    nrptg = (nth + nk0 - 1)/nk0;
+                    nth   = nk0;
+
+                    if (nrptg*nth > (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                        nrptg--;
+                    }
+                }
+
+                nth = MIN(nth, nk0);
+
+                ggml_metal_kargs_set_rows args = {
+                    /*.nk0  =*/ nk0,
+                    /*.ne01 =*/ ne01,
+                    /*.nb01 =*/ nb01,
+                    /*.nb02 =*/ nb02,
+                    /*.nb03 =*/ nb03,
+                    /*.ne11 =*/ ne11,
+                    /*.ne12 =*/ ne12,
+                    /*.nb10 =*/ nb10,
+                    /*.nb11 =*/ nb11,
+                    /*.nb12 =*/ nb12,
+                    /*.nb1  =*/ nb1,
+                    /*.nb2  =*/ nb2,
+                    /*.nb3  =*/ nb3,
+                };
+
+                [encoder setComputePipelineState:pipeline];
+                [encoder setBytes:&args    length:sizeof(args) atIndex:0];
+                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:1];
+                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:2];
+                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:3];
+
+                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nrptg - 1)/nrptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, nrptg, 1)];
+            } break;
         case GGML_OP_RMS_NORM:
             {
                 GGML_ASSERT(ne00 % 4 == 0);
@@ -3733,6 +3904,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_rms_norm args = {
@@ -3769,6 +3941,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_l2_norm args = {
@@ -3841,6 +4014,7 @@ static bool ggml_metal_encode_node(
                     nth *= 2;
                 }
 
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
                 nth = MIN(nth, ne00/4);
 
                 ggml_metal_kargs_norm args = {
@@ -4927,8 +5101,39 @@ static bool ggml_metal_encode_node(
                     default: GGML_ABORT("not implemented");
                 }
 
+                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+                // TODO: support
+                //const int32_t nk00 = ne00/ggml_blck_size(dst->type);
+                const int32_t nk00 = ne00;
+
+                int nth = 32; // SIMD width
+
+                while (nth < nk00 && nth < (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                    nth *= 2;
+                }
+
+                nth = MIN(nth, (int) pipeline.maxTotalThreadsPerThreadgroup);
+
+                // when rows are small, we can batch them together in a single threadgroup
+                int nrptg = 1;
+
+                // TODO: relax this constraint in the future
+                if (ggml_blck_size(src0->type) == 1 && ggml_blck_size(dst->type) == 1) {
+                    if (nth > nk00) {
+                        nrptg = (nth + nk00 - 1)/nk00;
+                        nth   = nk00;
+
+                        if (nrptg*nth > (int) pipeline.maxTotalThreadsPerThreadgroup) {
+                            nrptg--;
+                        }
+                    }
+                }
+
+                nth = MIN(nth, nk00);
+
                 ggml_metal_kargs_cpy args = {
-                    /*.ne00 =*/ ne00,
+                    /*.ne00 =*/ nk00,
                     /*.ne01 =*/ ne01,
                     /*.ne02 =*/ ne02,
                     /*.ne03 =*/ ne03,
@@ -4951,11 +5156,7 @@ static bool ggml_metal_encode_node(
                 [encoder setBuffer:id_src0 offset:offs_src0 atIndex:1];
                 [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
 
-                GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
-                int nth = MIN(1024, ne00/ggml_blck_size(src0->type));
-
-                [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-
+                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nrptg - 1)/nrptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, nrptg, 1)];
             } break;
         case GGML_OP_SET:
             {
@@ -5261,7 +5462,6 @@ static void ggml_backend_metal_buffer_free_buffer(ggml_backend_buffer_t buffer)
     }
 
     ggml_backend_metal_buffer_rset_free(ctx);
-    ggml_backend_metal_device_rel(buffer->buft->device->context);
 
     if (ctx->owned) {
 #if TARGET_OS_OSX
@@ -5370,7 +5570,10 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)buft->device->context;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     ctx->all_data = ggml_metal_host_malloc(size_aligned);
     ctx->all_size = size_aligned;
@@ -5393,14 +5596,12 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
     if (size_aligned > 0 && (ctx->all_data == NULL || ctx->buffers[0].metal == nil)) {
         GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f MiB\n", __func__, size_aligned / 1024.0 / 1024.0);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5411,17 +5612,14 @@ static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_ba
 
 static size_t ggml_backend_metal_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
     return 32;
+
     GGML_UNUSED(buft);
 }
 
 static size_t ggml_backend_metal_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
-    id<MTLDevice> device = ggml_backend_metal_device_acq(buft->device->context);
-    const size_t max_size = device.maxBufferLength;
-    ggml_backend_metal_device_rel(buft->device->context);
+    const size_t max_size = ((struct ggml_backend_metal_device_context *)buft->device->context)->max_size;
 
     return max_size;
-
-    GGML_UNUSED(buft);
 }
 
 static bool ggml_backend_metal_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
@@ -5494,7 +5692,10 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = &g_ggml_ctx_dev_main;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     // the buffer fits into the max buffer size allowed by the device
     if (size_aligned <= device.maxBufferLength) {
@@ -5550,7 +5751,6 @@ ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data, size_t siz
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5566,10 +5766,8 @@ static const char * ggml_backend_metal_name(ggml_backend_t backend) {
 }
 
 static void ggml_backend_metal_free(ggml_backend_t backend) {
-    struct ggml_backend_metal_context        * ctx     = backend->context;
-    struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
+    struct ggml_backend_metal_context * ctx = backend->context;
 
-    ggml_backend_metal_device_rel(ctx_dev);
     ggml_metal_free(ctx);
 
     free(backend);
@@ -5709,6 +5907,8 @@ bool ggml_backend_metal_supports_family(ggml_backend_t backend, int family) {
 
     struct ggml_backend_metal_device_context * ctx_dev = backend->device->context;
 
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
     return [ctx_dev->mtl_device supportsFamily:(MTLGPUFamilyApple1 + family - 1)];
 }
 
@@ -5728,10 +5928,7 @@ static const char * ggml_backend_metal_device_get_name(ggml_backend_dev_t dev) {
 }
 
 static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t dev) {
-    // acq/rel just to populate ctx->name in case it hasn't been done yet
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-    ggml_backend_metal_device_acq(ctx_dev);
-    ggml_backend_metal_device_rel(ctx_dev);
 
     return ctx_dev->name;
 }
@@ -5739,12 +5936,10 @@ static const char * ggml_backend_metal_device_get_description(ggml_backend_dev_t
 static void ggml_backend_metal_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
     if (@available(macOS 10.12, iOS 16.0, *)) {
         struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-        id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+        id<MTLDevice> device = ctx_dev->mtl_device;
 
         *total = device.recommendedMaxWorkingSetSize;
         *free  = *total - device.currentAllocatedSize;
-
-        ggml_backend_metal_device_rel(ctx_dev);
     } else {
         *free = 1;
         *total = 1;
@@ -5822,7 +6017,10 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back
     }
 
     struct ggml_backend_metal_device_context * ctx_dev = (struct ggml_backend_metal_device_context *)dev->context;
-    id<MTLDevice> device = ggml_backend_metal_device_acq(ctx_dev);
+
+    GGML_ASSERT(ctx_dev->mtl_device != nil);
+
+    id<MTLDevice> device = ctx_dev->mtl_device;
 
     // the buffer fits into the max buffer size allowed by the device
     if (size_aligned <= device.maxBufferLength) {
@@ -5878,7 +6076,6 @@ static ggml_backend_buffer_t ggml_backend_metal_device_buffer_from_ptr(ggml_back
     if (!ggml_backend_metal_buffer_rset_init(ctx, ctx_dev, device)) {
         GGML_LOG_ERROR("%s: error: failed to initialize residency set\n", __func__);
         free(ctx);
-        ggml_backend_metal_device_rel(ctx_dev);
         return NULL;
     }
 
@@ -5892,8 +6089,9 @@ static bool ggml_backend_metal_device_supports_op(ggml_backend_dev_t dev, const
 }
 
 static bool ggml_backend_metal_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return buft->iface.get_name == ggml_backend_metal_buffer_type_get_name ||
-            buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name;
+    return
+        buft->iface.get_name == ggml_backend_metal_buffer_type_get_name ||
+        buft->iface.get_name == ggml_backend_metal_buffer_from_ptr_type_get_name;
 
     GGML_UNUSED(dev);
 }
@@ -5978,8 +6176,19 @@ static struct ggml_backend_reg_i ggml_backend_metal_reg_i = {
     /* .get_proc_address = */ ggml_backend_metal_get_proc_address,
 };
 
+// called upon program exit
+static void ggml_metal_cleanup(void) {
+    ggml_backend_metal_device_rel(&g_ggml_ctx_dev_main);
+}
+
+// TODO: make thread-safe
 ggml_backend_reg_t ggml_backend_metal_reg(void) {
-    // TODO: make this thread-safe somehow?
+    ggml_backend_metal_device_acq(&g_ggml_ctx_dev_main);
+
+    // register cleanup callback
+    // TODO: not ideal, but not sure if there is a better way to do this in Objective-C
+    atexit(ggml_metal_cleanup);
+
     {
         g_ggml_backend_metal_reg = (struct ggml_backend_reg) {
             /* .api_version = */ GGML_BACKEND_API_VERSION,

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

@@ -35,6 +35,17 @@ constexpr constant static float kvalues_iq4nl_f[16] = {
     -127.f, -104.f, -83.f, -65.f, -49.f, -35.f, -22.f, -10.f, 1.f, 13.f, 25.f, 38.f, 53.f, 69.f, 89.f, 113.f
 };
 
+static inline int best_index_int8(int n, constant float * val, float x) {
+    if (x <= val[0]) return 0;
+    if (x >= val[n-1]) return n-1;
+    int ml = 0, mu = n-1;
+    while (mu-ml > 1) {
+        int mav = (ml+mu)/2;
+        if (x < val[mav]) mu = mav; else ml = mav;
+    }
+    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
 // NOTE: this is not dequantizing - we are simply fitting the template
 template <typename type4x4>
 void dequantize_f32(device const float4x4 * src, short il, thread type4x4 & reg) {
@@ -97,6 +108,173 @@ void dequantize_q4_0_t4(device const block_q4_0 * xb, short il, thread type4 & r
     }
 }
 
+void quantize_q4_0(device const float * src, device block_q4_0 & dst) {
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -8;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK4_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK4_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q4_1(device const float * src, device block_q4_1 & dst) {
+    float min = FLT_MAX;
+    float max = -FLT_MAX;
+
+    for (int j = 0; j < QK4_1; j++) {
+        const float v = src[j];
+        if (min > v) min = v;
+        if (max < v) max = v;
+    }
+
+    const float d = (max - min) / ((1 << 4) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    for (int j = 0; j < QK4_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK4_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+        const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+        dst.qs[j]  = xi0;
+        dst.qs[j] |= xi1 << 4;
+    }
+}
+
+void quantize_q5_0(device const float * src, device block_q5_0 & dst) {
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK5_0; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / -16;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_0/2; ++j) {
+        const float x0 = src[0       + j]*id;
+        const float x1 = src[QK5_0/2 + j]*id;
+
+        const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+        const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_q5_1(device const float * src, device block_q5_1 & dst) {
+    float max = src[0];
+    float min = src[0];
+
+    for (int j = 1; j < QK5_1; j++) {
+        const float v = src[j];
+        min = v < min ? v : min;
+        max = v > max ? v : max;
+    }
+
+    const float d = (max - min) / 31;
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+    dst.m = min;
+
+    uint32_t qh = 0;
+    for (int j = 0; j < QK5_1/2; ++j) {
+        const float x0 = (src[0       + j] - min)*id;
+        const float x1 = (src[QK5_1/2 + j] - min)*id;
+
+        const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+        const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+        dst.qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
+        qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+        qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
+    }
+
+    thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
+
+    for (int j = 0; j < 4; ++j) {
+        dst.qh[j] = qh8[j];
+    }
+}
+
+void quantize_iq4_nl(device const float * src, device block_iq4_nl & dst) {
+    float amax = 0.0f; // absolute max
+    float max  = 0.0f;
+
+    for (int j = 0; j < QK4_NL; j++) {
+        const float v = src[j];
+        if (amax < fabs(v)) {
+            amax = fabs(v);
+            max  = v;
+        }
+    }
+
+    const float d = max / kvalues_iq4nl_f[0];
+    const float id = d ? 1.0f/d : 0.0f;
+
+    float sumqx = 0, sumq2 = 0;
+    for (int j = 0; j < QK4_NL/2; ++j) {
+        const float x0 = src[0        + j]*id;
+        const float x1 = src[QK4_NL/2 + j]*id;
+
+        const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
+        const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
+
+        dst.qs[j] = xi0 | (xi1 << 4);
+
+        const float v0 = kvalues_iq4nl_f[xi0];
+        const float v1 = kvalues_iq4nl_f[xi1];
+        const float w0 = src[0        + j]*src[0        + j];
+        const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
+        sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
+        sumq2 += w0*v0*v0 + w1*v1*v1;
+
+    }
+
+    dst.d = sumq2 > 0 ? sumqx/sumq2 : d;
+}
+
 template <typename type4x4>
 void dequantize_q4_1(device const block_q4_1 * xb, short il, thread type4x4 & reg) {
     device const uint16_t * qs = ((device const uint16_t *)xb + 2);
@@ -279,6 +457,26 @@ void dequantize_q8_0_t4(device const block_q8_0 *xb, short il, thread type4 & re
     }
 }
 
+void quantize_q8_0(device const float * src, device block_q8_0 & dst) {
+    float amax = 0.0f; // absolute max
+
+    for (int j = 0; j < QK8_0; j++) {
+        const float v = src[j];
+        amax = MAX(amax, fabs(v));
+    }
+
+    const float d = amax / ((1 << 7) - 1);
+    const float id = d ? 1.0f/d : 0.0f;
+
+    dst.d = d;
+
+    for (int j = 0; j < QK8_0; ++j) {
+        const float x0 = src[j]*id;
+
+        dst.qs[j] = round(x0);
+    }
+}
+
 template <typename type4x4>
 void dequantize_q2_K(device const block_q2_K *xb, short il, thread type4x4 & reg) {
     const float d = xb->d;
@@ -993,31 +1191,61 @@ kernel void kernel_neg(
     dst[tpig] = -src0[tpig];
 }
 
+template <bool norm>
 kernel void kernel_sum_rows(
+        constant ggml_metal_kargs_sum_rows & args,
         device const float * src0,
         device       float * dst,
-        constant ggml_metal_kargs_sum_rows & args,
-        uint3 tpig[[thread_position_in_grid]]) {
-    int64_t i3 = tpig.z;
-    int64_t i2 = tpig.y;
-    int64_t i1 = tpig.x;
+        threadgroup  float * shmem_f32 [[threadgroup(0)]],
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort3 tpitg[[thread_position_in_threadgroup]],
+        ushort  sgitg[[simdgroup_index_in_threadgroup]],
+        ushort  tiisg[[thread_index_in_simdgroup]],
+        ushort3   ntg[[threads_per_threadgroup]]) {
+    int64_t i3 = tgpig.z;
+    int64_t i2 = tgpig.y;
+    int64_t i1 = tgpig.x;
 
     if (i3 >= args.ne03 || i2 >= args.ne02 || i1 >= args.ne01) {
         return;
     }
 
+    if (sgitg == 0) {
+        shmem_f32[tiisg] = 0.0f;
+    }
+
     device const float * src_row = (device const float *) ((device const char *) src0 + i1*args.nb01 + i2*args.nb02 + i3*args.nb03);
     device       float * dst_row = (device       float *) ((device       char *) dst  + i1*args.nb1  + i2*args.nb2  + i3*args.nb3);
 
-    float row_sum = 0;
+    float sumf = 0;
 
-    for (int64_t i0 = 0; i0 < args.ne00; i0++) {
-        row_sum += src_row[i0];
+    for (int64_t i0 = tpitg.x; i0 < args.ne00; i0 += ntg.x) {
+        sumf += src_row[i0];
     }
 
-    dst_row[0] = row_sum;
+    sumf = simd_sum(sumf);
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    if (tiisg == 0) {
+        shmem_f32[sgitg] = sumf;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    sumf = shmem_f32[tiisg];
+    sumf = simd_sum(sumf);
+
+    if (tpitg.x == 0) {
+        dst_row[0] = norm ? sumf / args.ne00 : sumf;
+    }
 }
 
+typedef decltype(kernel_sum_rows<false>) kernel_sum_rows_t;
+
+template [[host_name("kernel_sum_rows")]] kernel kernel_sum_rows_t kernel_sum_rows<false>;
+template [[host_name("kernel_mean")]]     kernel kernel_sum_rows_t kernel_sum_rows<true>;
+
 template<typename T>
 kernel void kernel_soft_max(
         device const  char * src0,
@@ -2502,6 +2730,70 @@ template [[host_name("kernel_mul_mv_bf16_f32")]]  kernel mul_mv_t kernel_mul_mv<
 template [[host_name("kernel_mul_mv_bf16_bf16")]] kernel mul_mv_t kernel_mul_mv<bfloat, bfloat4, bfloat, bfloat4>;
 #endif
 
+template<typename T04, typename T14, typename args_t>
+void kernel_mul_mv_c4_impl(
+        args_t args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig,
+        ushort tiisg) {
+    const int r0 = tgpig.x*32 + tiisg;
+    const int rb = tgpig.y*N_MV_T_T;
+    const int im = tgpig.z;
+
+    if (r0 >= args.ne01) {
+        return;
+    }
+
+    const uint i12 = im%args.ne12;
+    const uint i13 = im/args.ne12;
+
+    const uint64_t offset0 = r0*args.nb01 + (i12/args.r2)*args.nb02 + (i13/args.r3)*args.nb03;
+
+    device const T04 * x = (device const T04 *) (src0 + offset0);
+
+    device float * dst_f32 = (device float *) dst + (uint64_t)im*args.ne0*args.ne1;
+
+    for (int row = 0; row < N_MV_T_T; ++row) {
+        int r1 = rb + row;
+        if (r1 >= args.ne11) {
+            break;
+        }
+
+        const uint64_t offset1 = r1*args.nb11 + (i12   )*args.nb12 + (i13   )*args.nb13;
+
+        device const T14 * y = (device const T14 *) (src1 + offset1);
+
+        dst_f32[(uint64_t)r1*args.ne0 + r0] = dot((float4) x[0], (float4) y[0]);
+    }
+}
+
+template<typename T04, typename T14>
+kernel void kernel_mul_mv_c4(
+        constant ggml_metal_kargs_mul_mv & args,
+        device const char * src0,
+        device const char * src1,
+        device       char * dst,
+        uint3  tgpig[[threadgroup_position_in_grid]],
+        ushort tiisg[[thread_index_in_simdgroup]]) {
+    kernel_mul_mv_c4_impl<T04, T14, constant ggml_metal_kargs_mul_mv &>(
+        args,
+        src0,
+        src1,
+        dst,
+        tgpig,
+        tiisg);
+}
+
+typedef decltype(kernel_mul_mv_c4<half4, half4>) mul_mv_c4_t;
+
+template [[host_name("kernel_mul_mv_f32_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<float4,  float4>;
+template [[host_name("kernel_mul_mv_f16_f32_c4")]]  kernel mul_mv_c4_t kernel_mul_mv_c4<half4,   float4>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_mul_mv_bf16_f32_c4")]] kernel mul_mv_c4_t kernel_mul_mv_c4<bfloat4, float4>;
+#endif
+
 template<typename T, typename T4>
 kernel void kernel_mul_mv_1row(
         constant ggml_metal_kargs_mul_mv & args,
@@ -4276,11 +4568,16 @@ kernel void kernel_cpy(
         device  const char * src0,
         device        char * dst,
         uint3   tgpig[[threadgroup_position_in_grid]],
+        uint    tiitg[[thread_index_in_threadgroup]],
         ushort3 tpitg[[thread_position_in_threadgroup]],
-        ushort3   ntg[[threads_per_threadgroup]]) {
+        ushort3  tptg[[threads_per_threadgroup]]) {
     const int i03 = tgpig[2];
     const int i02 = tgpig[1];
-    const int i01 = tgpig[0];
+    const int i01 = tgpig[0]*tptg.y + tiitg/tptg.x;
+
+    if (i01 >= args.ne01) {
+        return;
+    }
 
     const int64_t n = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00;
 
@@ -4291,7 +4588,7 @@ kernel void kernel_cpy(
 
     device T1 * dst_data = (device T1 *) (dst + i3*args.nb3 + i2*args.nb2 + i1*args.nb1 + i0*args.nb0);
 
-    for (int64_t i00 = tpitg.x; i00 < args.ne00; i00 += ntg.x) {
+    for (int64_t i00 = tiitg%tptg.x; i00 < args.ne00; i00 += tptg.x) {
         device const T0 * src = (device T0 *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
         dst_data[i00] = (T1) src[0];
     }
@@ -4311,6 +4608,7 @@ template [[host_name("kernel_cpy_bf16_f32")]]  kernel kernel_cpy_t kernel_cpy<bf
 template [[host_name("kernel_cpy_bf16_bf16")]] kernel kernel_cpy_t kernel_cpy<bfloat, bfloat>;
 #endif
 
+// TODO: templetify these kernels
 kernel void kernel_cpy_f32_q8_0(
         constant ggml_metal_kargs_cpy & args,
         device const char * src0,
@@ -4334,23 +4632,7 @@ kernel void kernel_cpy_f32_q8_0(
     for (int64_t i00 = tpitg.x*QK8_0; i00 < args.ne00; i00 += ntg.x*QK8_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-
-        for (int j = 0; j < QK8_0; j++) {
-            const float v = src[j];
-            amax = MAX(amax, fabs(v));
-        }
-
-        const float d = amax / ((1 << 7) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK8_0].d = d;
-
-        for (int j = 0; j < QK8_0; ++j) {
-            const float x0 = src[j]*id;
-
-            dst_data[i00/QK8_0].qs[j] = round(x0);
-        }
+        quantize_q8_0(src, dst_data[i00/QK8_0]);
     }
 }
 
@@ -4377,32 +4659,7 @@ kernel void kernel_cpy_f32_q4_0(
     for (int64_t i00 = tpitg.x*QK4_0; i00 < args.ne00; i00 += ntg.x*QK4_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -8;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_0].d = d;
-
-        for (int j = 0; j < QK4_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK4_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
-
-            dst_data[i00/QK4_0].qs[j]  = xi0;
-            dst_data[i00/QK4_0].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_0(src, dst_data[i00/QK4_0]);
     }
 }
 
@@ -4429,31 +4686,7 @@ kernel void kernel_cpy_f32_q4_1(
     for (int64_t i00 = tpitg.x*QK4_1; i00 < args.ne00; i00 += ntg.x*QK4_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float min = FLT_MAX;
-        float max = -FLT_MAX;
-
-        for (int j = 0; j < QK4_1; j++) {
-            const float v = src[j];
-            if (min > v) min = v;
-            if (max < v) max = v;
-        }
-
-        const float d = (max - min) / ((1 << 4) - 1);
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK4_1].d = d;
-        dst_data[i00/QK4_1].m = min;
-
-        for (int j = 0; j < QK4_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK4_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
-            const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
-
-            dst_data[i00/QK4_1].qs[j]  = xi0;
-            dst_data[i00/QK4_1].qs[j] |= xi1 << 4;
-        }
+        quantize_q4_1(src, dst_data[i00/QK4_1]);
     }
 }
 
@@ -4480,38 +4713,7 @@ kernel void kernel_cpy_f32_q5_0(
     for (int64_t i00 = tpitg.x*QK5_0; i00 < args.ne00; i00 += ntg.x*QK5_0) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK5_0; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / -16;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_0].d = d;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_0/2; ++j) {
-            const float x0 = src[0       + j]*id;
-            const float x1 = src[QK5_0/2 + j]*id;
-
-            const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
-            const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
-
-            dst_data[i00/QK5_0].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_0].qh[j] = qh8[j];
-        }
+        quantize_q5_0(src, dst_data[i00/QK5_0]);
     }
 }
 
@@ -4538,51 +4740,10 @@ kernel void kernel_cpy_f32_q5_1(
     for (int64_t i00 = tpitg.x*QK5_1; i00 < args.ne00; i00 += ntg.x*QK5_1) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float max = src[0];
-        float min = src[0];
-
-        for (int j = 1; j < QK5_1; j++) {
-            const float v = src[j];
-            min = v < min ? v : min;
-            max = v > max ? v : max;
-        }
-
-        const float d = (max - min) / 31;
-        const float id = d ? 1.0f/d : 0.0f;
-
-        dst_data[i00/QK5_1].d = d;
-        dst_data[i00/QK5_1].m = min;
-
-        uint32_t qh = 0;
-        for (int j = 0; j < QK5_1/2; ++j) {
-            const float x0 = (src[0       + j] - min)*id;
-            const float x1 = (src[QK5_1/2 + j] - min)*id;
-
-            const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
-            const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
-
-            dst_data[i00/QK5_1].qs[j] = (xi0 & 0xf) | ((xi1 & 0xf) << 4);
-            qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
-            qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_1/2);
-        }
-        thread const uint8_t * qh8 = (thread const uint8_t *)&qh;
-        for (int j = 0; j < 4; ++j) {
-            dst_data[i00/QK5_1].qh[j] = qh8[j];
-        }
+        quantize_q5_1(src, dst_data[i00/QK5_1]);
     }
 }
 
-static inline int best_index_int8(int n, constant float * val, float x) {
-    if (x <= val[0]) return 0;
-    if (x >= val[n-1]) return n-1;
-    int ml = 0, mu = n-1;
-    while (mu-ml > 1) {
-        int mav = (ml+mu)/2;
-        if (x < val[mav]) mu = mav; else ml = mav;
-    }
-    return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
-}
-
 kernel void kernel_cpy_f32_iq4_nl(
         constant ggml_metal_kargs_cpy & args,
         device const char * src0,
@@ -4606,40 +4767,7 @@ kernel void kernel_cpy_f32_iq4_nl(
     for (int64_t i00 = tpitg.x*QK4_NL; i00 < args.ne00; i00 += ntg.x*QK4_NL) {
         device const float * src = (device float *)(src0 + i03*args.nb03 + i02*args.nb02 + i01*args.nb01 + i00*args.nb00);
 
-        float amax = 0.0f; // absolute max
-        float max  = 0.0f;
-
-        for (int j = 0; j < QK4_NL; j++) {
-            const float v = src[j];
-            if (amax < fabs(v)) {
-                amax = fabs(v);
-                max  = v;
-            }
-        }
-
-        const float d = max / kvalues_iq4nl_f[0];
-        const float id = d ? 1.0f/d : 0.0f;
-
-        float sumqx = 0, sumq2 = 0;
-        for (int j = 0; j < QK4_NL/2; ++j) {
-            const float x0 = src[0        + j]*id;
-            const float x1 = src[QK4_NL/2 + j]*id;
-
-            const uint8_t xi0 = best_index_int8(16, kvalues_iq4nl_f, x0);
-            const uint8_t xi1 = best_index_int8(16, kvalues_iq4nl_f, x1);
-
-            dst_data[i00/QK4_NL].qs[j] = xi0 | (xi1 << 4);
-
-            const float v0 = kvalues_iq4nl_f[xi0];
-            const float v1 = kvalues_iq4nl_f[xi1];
-            const float w0 = src[0        + j]*src[0        + j];
-            const float w1 = src[QK4_NL/2 + j]*src[QK4_NL/2 + j];
-            sumqx += w0*v0*src[j] + w1*v1*src[QK4_NL/2 + j];
-            sumq2 += w0*v0*v0 + w1*v1*v1;
-
-        }
-
-        dst_data[i00/QK4_NL].d = sumq2 > 0 ? sumqx/sumq2 : d;
+        quantize_iq4_nl(src, dst_data[i00/QK4_NL]);
     }
 }
 
@@ -6320,10 +6448,10 @@ kernel void kernel_mul_mv_iq4_xs_f32(
 
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows_q(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6343,10 +6471,10 @@ kernel void kernel_get_rows_q(
 
 template<typename T>
 kernel void kernel_get_rows_f(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device       float * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6364,10 +6492,10 @@ kernel void kernel_get_rows_f(
 }
 
 kernel void kernel_get_rows_i32(
+        constant ggml_metal_kargs_get_rows & args,
         device const  void * src0,
         device const  void * src1,
         device     int32_t * dst,
-        constant ggml_metal_kargs_get_rows & args,
         uint3                tgpig[[threadgroup_position_in_grid]],
         uint                 tiitg[[thread_index_in_threadgroup]],
         uint3                tptg [[threads_per_threadgroup]]) {
@@ -6384,6 +6512,67 @@ kernel void kernel_get_rows_i32(
     }
 }
 
+template<typename block_q, void (*quantize_func)(device const float *, device block_q &)>
+kernel void kernel_set_rows_q32(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+          device block_q * dst_row = (      device block_q *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float   * src_row = (const device float   *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        quantize_func(src_row + 32*ind, dst_row[ind]);
+    }
+}
+
+template<typename T>
+kernel void kernel_set_rows_f(
+        constant ggml_metal_kargs_set_rows & args,
+        device const  void * src0,
+        device const  void * src1,
+        device       float * dst,
+        uint3                tgpig[[threadgroup_position_in_grid]],
+        uint                 tiitg[[thread_index_in_threadgroup]],
+        uint3                tptg [[threads_per_threadgroup]]) {
+    const int32_t i03 = tgpig.z;
+    const int32_t i02 = tgpig.y;
+
+    const int32_t i12 = i03%args.ne12;
+    const int32_t i11 = i02%args.ne11;
+
+    const int32_t i01 = tgpig.x*tptg.y + tiitg/tptg.x;
+    if (i01 >= args.ne01) {
+        return;
+    }
+
+    const int32_t i10 = i01;
+    const int64_t i1 = ((const device int64_t *) ((const device char *) src1 + i10*args.nb10 + i11*args.nb11 + i12*args.nb12))[0];
+
+    device T     * dst_row = (      device T     *) ((      device char *) dst  +  i1*args.nb1  + i02*args.nb2  + i03*args.nb3);
+    const device float * src_row = (const device float *) ((const device char *) src0 + i01*args.nb01 + i02*args.nb02 + i03*args.nb03);
+
+    for (int ind = tiitg%tptg.x; ind < args.nk0; ind += tptg.x) {
+        dst_row[ind] = (T) src_row[ind];
+    }
+}
 
 #define BLOCK_SIZE_M 64 // 8 simdgroup matrices from matrix A
 #define BLOCK_SIZE_N 32 // 4 simdgroup matrices from matrix B
@@ -6807,6 +6996,27 @@ template [[host_name("kernel_get_rows_iq1_m")]]   kernel get_rows_q_t kernel_get
 template [[host_name("kernel_get_rows_iq4_nl")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_nl,  2,     dequantize_iq4_nl>;
 template [[host_name("kernel_get_rows_iq4_xs")]]  kernel get_rows_q_t kernel_get_rows_q<block_iq4_xs,  QK_NL, dequantize_iq4_xs>;
 
+//
+// set rows
+//
+
+typedef decltype(kernel_set_rows_f<float>) set_rows_f_t;
+
+template [[host_name("kernel_set_rows_f32")]]  kernel set_rows_f_t kernel_set_rows_f<float>;
+template [[host_name("kernel_set_rows_f16")]]  kernel set_rows_f_t kernel_set_rows_f<half>;
+#if defined(GGML_METAL_USE_BF16)
+template [[host_name("kernel_set_rows_bf16")]] kernel set_rows_f_t kernel_set_rows_f<bfloat>;
+#endif
+
+typedef decltype(kernel_set_rows_q32<block_q8_0, quantize_q8_0>) set_rows_q32_t;
+
+template [[host_name("kernel_set_rows_q8_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q8_0,   quantize_q8_0>;
+template [[host_name("kernel_set_rows_q4_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_0,   quantize_q4_0>;
+template [[host_name("kernel_set_rows_q4_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q4_1,   quantize_q4_1>;
+template [[host_name("kernel_set_rows_q5_0")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_0,   quantize_q5_0>;
+template [[host_name("kernel_set_rows_q5_1")]]   kernel set_rows_q32_t kernel_set_rows_q32<block_q5_1,   quantize_q5_1>;
+template [[host_name("kernel_set_rows_iq4_nl")]] kernel set_rows_q32_t kernel_set_rows_q32<block_iq4_nl, quantize_iq4_nl>;
+
 //
 // matrix-matrix multiplication
 //

ggml/src/ggml-musa/mudnn.cuh

@@ -1,7 +1,7 @@
 #pragma once
 
-#include "../include/ggml.h"
-#include "../ggml-cuda/common.cuh"
+#include "ggml-cuda/common.cuh"
+#include "ggml.h"
 
 // Asynchronously copies data from src tensor to dst tensor using the provided context.
 // Returns a musaError_t indicating success or failure.

ggml/src/ggml-sycl/binbcast.cpp

@@ -225,9 +225,9 @@ struct bin_bcast_sycl {
                     dpct::has_capability_or_fail(stream->get_device(),
                                                  {sycl::aspect::fp16});
 
-                    stream->parallel_for(
-                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) *
-                                              sycl::range<3>(1, 1, block_size),
+                    sycl_parallel_for(
+                        stream,
+                        sycl::nd_range<3>(sycl::range<3>(1, 1, block_num) * sycl::range<3>(1, 1, block_size),
                                           sycl::range<3>(1, 1, block_size)),
                         [=](sycl::nd_item<3> item_ct1) {
                             k_bin_bcast_unravel<bin_op>(
@@ -246,9 +246,8 @@ struct bin_bcast_sycl {
                 dpct::has_capability_or_fail(stream->get_device(),
                                              {sycl::aspect::fp16});
 
-                stream->parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         k_bin_bcast<bin_op>(src0_dd, src1_dd, dst_dd, ne0, ne1,
                                             ne2, ne3, ne10, ne11, ne12, ne13,
                                             s1, s2, s3, s01, s02, s03, s11, s12, s13,

ggml/src/ggml-sycl/common.hpp

@@ -199,7 +199,7 @@ struct sycl_device_info {
     // size_t  smpb;               // max. shared memory per block
     bool    vmm;                // virtual memory support
     size_t  total_vram;
-    sycl_hw_info hw_info;
+    //sycl_hw_info hw_info;     \\ device id and aarch, currently not used
     optimize_feature opt_feature;
 };
 
@@ -286,29 +286,6 @@ struct ggml_tensor_extra_gpu {
 
 void release_extra_gpu(ggml_tensor_extra_gpu * extra, std::vector<queue_ptr> streams={});
 
-inline optimize_feature check_gpu_optimize_feature(syclex::architecture &arch) {
-    optimize_feature opt;
-
-    opt.reorder =
-        (arch == syclex::architecture::intel_gpu_dg1 ||
-         arch == syclex::architecture::intel_gpu_acm_g10 ||
-         arch == syclex::architecture::intel_gpu_acm_g11 ||
-         arch == syclex::architecture::intel_gpu_acm_g12 ||
-         arch == syclex::architecture::intel_gpu_pvc ||
-         arch == syclex::architecture::intel_gpu_pvc_vg ||
-         arch == syclex::architecture::intel_gpu_mtl_u ||
-         arch == syclex::architecture::intel_gpu_mtl_s ||
-         arch == syclex::architecture::intel_gpu_mtl_h ||
-         arch == syclex::architecture::intel_gpu_arl_u ||
-         arch == syclex::architecture::intel_gpu_arl_s ||
-         arch == syclex::architecture::intel_gpu_arl_h ||
-         arch == syclex::architecture::intel_gpu_bmg_g21 ||
-         arch == syclex::architecture::intel_gpu_lnl_m
-        );
-
-    return opt;
-}
-
 namespace sycl_ex = sycl::ext::oneapi::experimental;
 struct ggml_backend_sycl_context {
     int device;

ggml/src/ggml-sycl/concat.cpp

@@ -89,33 +89,24 @@ static void concat_f32_sycl(const float *x, const float *y, float *dst,
   sycl::range<3> gridDim(ne2, ne1, num_blocks);
   switch (dim) {
   case 0:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim0(x, y, dst, ne0, ne00, item_ct1); });
+      break;
   case 1:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim1(x, y, dst, ne0, ne01, item_ct1); });
+      break;
   // dim >=2 will be dispatched to the default path
   default:
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim *
-                              sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-          concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1);
-        });
-    break;
+      sycl_parallel_for(stream,
+                        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE),
+                                          sycl::range<3>(1, 1, SYCL_CONCAT_BLOCK_SIZE)),
+                        [=](sycl::nd_item<3> item_ct1) { concat_f32_dim2(x, y, dst, ne0, ne02, item_ct1); });
+      break;
   }
 }
 
@@ -129,33 +120,29 @@ static void concat_f32_sycl_non_cont(
     int64_t ne2, int64_t ne3, uint64_t nb0, uint64_t nb1, uint64_t nb2,
     uint64_t nb3, int32_t dim) {
   sycl::range<3> gridDim(ne3, ne2, ne1);
-  stream->parallel_for(
-      sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)),
-      [=](sycl::nd_item<3> item_ct1) {
-        int64_t i3 = item_ct1.get_group(0);
-        int64_t i2 = item_ct1.get_group(1);
-        int64_t i1 = item_ct1.get_group(2);
+  sycl_parallel_for(stream, sycl::nd_range<3>(gridDim, sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+      int64_t i3 = item_ct1.get_group(0);
+      int64_t i2 = item_ct1.get_group(1);
+      int64_t i1 = item_ct1.get_group(2);
 
-        int64_t o[4] = {0, 0, 0, 0};
-        o[dim] = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
+      int64_t o[4] = { 0, 0, 0, 0 };
+      o[dim]       = dim == 0 ? ne00 : (dim == 1 ? ne01 : (dim == 2 ? ne02 : ne03));
 
-        const float *x;
+      const float * x;
 
-        for (int i0 = item_ct1.get_local_id(2); i0 < ne0;
-             i0 += item_ct1.get_local_range(2)) {
+      for (int i0 = item_ct1.get_local_id(2); i0 < ne0; i0 += item_ct1.get_local_range(2)) {
           if (i0 < ne00 && i1 < ne01 && i2 < ne02 && i3 < ne03) {
-            x = (const float *)(src0 + (i3)*nb03 + (i2)*nb02 + (i1)*nb01 +
-                                (i0)*nb00);
+              x = (const float *) (src0 + (i3) *nb03 + (i2) *nb02 + (i1) *nb01 + (i0) *nb00);
           } else {
-            x = (const float *)(src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 +
-                                (i1 - o[1]) * nb11 + (i0 - o[0]) * nb10);
+              x = (const float *) (src1 + (i3 - o[3]) * nb13 + (i2 - o[2]) * nb12 + (i1 - o[1]) * nb11 +
+                                   (i0 - o[0]) * nb10);
           }
 
           float *y = (float *)(dst + i3 * nb3 + i2 * nb2 + i1 * nb1 + i0 * nb0);
 
           *y = *x;
-        }
-      });
+      }
+  });
 }
 
 void ggml_sycl_op_concat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {

ggml/src/ggml-sycl/conv.cpp

@@ -59,16 +59,10 @@ static void conv_transpose_1d_f32_f32_sycl(
     const int num_blocks = (output_size + SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE - 1) / SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE;
     const sycl::range<3> block_dims(1, 1, SYCL_CONV_TRANPOSE_1D_BLOCK_SIZE);
     const sycl::range<3> block_nums(1, 1, num_blocks);
-    stream->parallel_for(
-        sycl::nd_range<3>(
-            block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) {
-            conv_transpose_1d_kernel(
-                s0, output_size,
-                src0_ne0, src0_ne1, src0_ne2,
-                src1_ne0, dst_ne0,
-                src0, src1, dst, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        conv_transpose_1d_kernel(s0, output_size, src0_ne0, src0_ne1, src0_ne2, src1_ne0, dst_ne0, src0, src1, dst,
+                                 item_ct1);
+    });
 }
 
 void ggml_sycl_op_conv_transpose_1d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {

ggml/src/ggml-sycl/convert.cpp

@@ -33,14 +33,11 @@ static void dequantize_block_sycl(const void *__restrict__ vx,
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
-        stream->parallel_for(
-            sycl::nd_range<3>(
-                sycl::range<3>(1, 1, num_blocks) *
-                    sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
-                sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
-            [=](sycl::nd_item<3> item_ct1) {
-                dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1);
-            });
+        sycl_parallel_for(
+            stream,
+            sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_DEQUANTIZE_BLOCK_SIZE)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block<qk, qr, dequantize_kernel>(vx, y, k, item_ct1); });
     }
 }
 
@@ -53,24 +50,18 @@ static void dequantize_row_q2_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q2_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q2_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q2_K(vx, y, item_ct1); });
     }
 
 #endif
@@ -85,24 +76,18 @@ static void dequantize_row_q3_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q3_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q3_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q3_K(vx, y, item_ct1); });
     }
 #endif
 }
@@ -116,12 +101,9 @@ static void dequantize_row_q4_0_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_0(vx, y, nb32, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_0(vx, y, nb32, item_ct1); });
     }
 }
 
@@ -135,13 +117,12 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int
     int constexpr WARP_K = WARP_SIZE * QK4_0;
     const int n_warp = (k + WARP_K - 1) / WARP_K;
     GGML_ASSERT(k % 2 == 0);
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) *
-        sycl::range<3>(1, 1, WARP_SIZE),
-        sycl::range<3>(1, 1, WARP_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
-            dequantize_block_q4_0_reorder(vx, y, k, item_ct1);
-        });
-
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, n_warp) * sycl::range<3>(1, 1, WARP_SIZE),
+                                        sycl::range<3>(1, 1, WARP_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                          dequantize_block_q4_0_reorder(vx, y, k, item_ct1);
+                      });
 }
 
 template <typename dst_t>
@@ -153,12 +134,9 @@ static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_1(vx, y, nb32, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q4_1(vx, y, nb32, item_ct1); });
     }
 }
 
@@ -171,14 +149,13 @@ static void dequantize_row_q4_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
-                             });
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_q4_K(vx, y, get_pointer(scale_local_acc), item_ct1);
+                });
         });
     }
 }
@@ -191,13 +168,13 @@ static void dequantize_row_q4_K_sycl_reorder(const void * vx, dst_t * y, const i
 
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-    stream->submit([&](sycl::handler & cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<uint8_t, 1> scale_local_acc(sycl::range<1>(12), cgh);
 
-        cgh.parallel_for(sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
-                         [=](sycl::nd_item<1> item_ct1) {
-                             dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
-                         });
+        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(sycl::range<1>(global_size), sycl::range<1>(local_size)),
+                             [=](sycl::nd_item<1> item_ct1) {
+                                 dequantize_block_q4_K_reorder(vx, y, get_pointer(scale_local_acc), item_ct1, nb);
+                             });
     });
 }
 
@@ -210,24 +187,18 @@ static void dequantize_row_q5_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q5_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q5_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q5_K(vx, y, item_ct1); });
     }
 
 #endif
@@ -242,24 +213,18 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 64),
-                                               sycl::range<3>(1, 1, 64)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q6_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
     }
 #else
     {
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_q6_K(vx, y, item_ct1);
-                             });
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+            [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K(vx, y, item_ct1); });
     }
 
 #endif
@@ -271,9 +236,9 @@ static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const i
 
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
-        [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+                      [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
 }
 
 template <typename dst_t>
@@ -284,15 +249,10 @@ static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq1_s(
-                                     vx, y, item_ct1, iq1s_grid_gpu
-                                     );
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_s(vx, y, item_ct1, iq1s_grid_gpu); });
         });
     }
 }
@@ -305,15 +265,10 @@ static void dequantize_row_iq1_m_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq1_m(
-                                     vx, y, item_ct1, iq1s_grid_gpu
-                                     );
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq1_m(vx, y, item_ct1, iq1s_grid_gpu); });
         });
     }
 }
@@ -326,15 +281,12 @@ static void dequantize_row_iq2_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_xxs(
-                                     vx, y, item_ct1, iq2xxs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq2_xxs(vx, y, item_ct1, iq2xxs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -347,15 +299,12 @@ static void dequantize_row_iq2_xs_sycl(const void *vx, dst_t *y, const int64_t k
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_xs(
-                                     vx, y, item_ct1, iq2xs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq2_xs(vx, y, item_ct1, iq2xs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -368,13 +317,10 @@ static void dequantize_row_iq2_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq2_s(vx, y, item_ct1);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq2_s(vx, y, item_ct1); });
         });
     }
 }
@@ -388,15 +334,12 @@ static void dequantize_row_iq3_xxs_sycl(const void *vx, dst_t *y, const int64_t
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq3_xxs(
-                                     vx, y, item_ct1, iq3xxs_grid,
-                                     ksigns_iq2xs, kmask_iq2xs);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) {
+                    dequantize_block_iq3_xxs(vx, y, item_ct1, iq3xxs_grid, ksigns_iq2xs, kmask_iq2xs);
+                });
         });
     }
 }
@@ -409,14 +352,10 @@ static void dequantize_row_iq3_s_sycl(const void *vx, dst_t *y, const int64_t k,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                                   sycl::range<3>(1, 1, 32),
-                                               sycl::range<3>(1, 1, 32)),
-                             [=](sycl::nd_item<3> item_ct1) {
-                                 dequantize_block_iq3_s(
-                                     vx, y, item_ct1, kmask_iq2xs, iq3s_grid);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq3_s(vx, y, item_ct1, kmask_iq2xs, iq3s_grid); });
         });
     }
 }
@@ -432,14 +371,11 @@ static void dequantize_row_iq4_xs_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
-                  cgh.parallel_for(
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                            sycl::range<3>(1, 1, 32),
-                                        sycl::range<3>(1, 1, 32)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                            dequantize_block_iq4_xs(vx, y, item_ct1);
-                      });
+            sycl_launch(stream, [&](sycl::handler & cgh) {
+                sycl_parallel_for(
+                    cgh,
+                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_xs(vx, y, item_ct1); });
             });
       }
 #endif
@@ -453,14 +389,11 @@ static void dequantize_row_iq4_nl_sycl(const void *vx, dst_t *y, const int64_t k
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
-                  cgh.parallel_for(
-                      sycl::nd_range<3>(sycl::range<3>(1, 1, nb) *
-                                            sycl::range<3>(1, 1, 32),
-                                        sycl::range<3>(1, 1, 32)),
-                      [=](sycl::nd_item<3> item_ct1) {
-                            dequantize_block_iq4_nl(vx, y, item_ct1);
-                      });
+            sycl_launch(stream, [&](sycl::handler & cgh) {
+                sycl_parallel_for(
+                    cgh,
+                    sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 32), sycl::range<3>(1, 1, 32)),
+                    [=](sycl::nd_item<3> item_ct1) { dequantize_block_iq4_nl(vx, y, item_ct1); });
             });
       }
 }

ggml/src/ggml-sycl/cpy.cpp

@@ -413,7 +413,8 @@ static void ggml_cpy_f16_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -431,7 +432,8 @@ static void ggml_cpy_f32_f32_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -449,7 +451,8 @@ static void ggml_cpy_f32_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -465,11 +468,11 @@ static void ggml_cpy_f32_q8_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK8_0 == 0);
     const int num_blocks = ne / QK8_0;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -477,11 +480,11 @@ static void ggml_cpy_q8_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_q_f32<cpy_blck_q8_0_f32, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -490,11 +493,11 @@ static void ggml_cpy_f32_q4_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_0 == 0);
     const int num_blocks = ne / QK4_0;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -502,8 +505,9 @@ static void ggml_cpy_q4_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(
+        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_0, QK4_0>, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -516,11 +520,11 @@ static void ggml_cpy_f32_q4_1_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_1 == 0);
     const int num_blocks = ne / QK4_1;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -528,8 +532,9 @@ static void ggml_cpy_q4_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(
+        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q4_1, QK4_1>, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -542,11 +547,11 @@ static void ggml_cpy_f32_q5_0_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK5_0 == 0);
     const int num_blocks = ne / QK5_0;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -554,8 +559,9 @@ static void ggml_cpy_q5_0_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(
+        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_0, QK5_0>, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -568,11 +574,11 @@ static void ggml_cpy_f32_q5_1_sycl(const char * cx, char * cdst, const int ne, c
                                    const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK5_1 == 0);
     const int num_blocks = ne / QK5_1;
-    stream->parallel_for(sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
-                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                              ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -580,8 +586,9 @@ static void ggml_cpy_q5_1_f32_sycl(const char * cx, char * cdst, const int ne, c
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ne;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(
+        stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+        [=](sycl::nd_item<3> item_ct1) {
             cpy_q_f32<cpy_blck_q_f32<dequantize_q5_1, QK5_1>, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02,
                                                                      nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13,
                                                                      item_ct1);
@@ -594,11 +601,11 @@ static void ggml_cpy_f32_iq4_nl_sycl(const char * cx, char * cdst, const int ne,
                                      const int nb12, const int nb13, queue_ptr stream) {
     GGML_ASSERT(ne % QK4_NL == 0);
     const int num_blocks = ne / QK4_NL;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
-                                                   ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks), sycl::range<3>(1, 1, 1)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_f32_q<cpy_blck_f32_iq4_nl, QK4_NL>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03,
+                                                                 ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
@@ -609,7 +616,8 @@ static void ggml_cpy_f16_f16_sycl(const char * cx, char * cdst, const int ne, co
     {
         dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -628,7 +636,8 @@ static void ggml_cpy_i16_i16_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -647,7 +656,8 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
         // dpct::has_capability_or_fail(stream->get_device(),
         //                              {sycl::aspect::fp16});
 
-        stream->parallel_for(
+        sycl_parallel_for(
+            stream,
             sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
                               sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
             [=](sycl::nd_item<3> item_ct1) {
@@ -662,11 +672,13 @@ static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 
@@ -675,11 +687,13 @@ static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 
@@ -689,11 +703,13 @@ static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
 
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
-                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 
@@ -702,10 +718,13 @@ static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const
                                    const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
                                    const int nb12, const int nb13, queue_ptr stream) {
     const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                     ne12, nb10, nb11, nb12, nb13, item_ct1);
+                      });
 }
 
 
@@ -715,10 +734,13 @@ static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const
                                    const int nb12, const int nb13, queue_ptr stream) {
 
    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
-   stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
-            cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
-        });
+   sycl_parallel_for(stream,
+                     sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                                       sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)),
+                     [=](sycl::nd_item<3> item_ct1) {
+                         cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11,
+                                                    ne12, nb10, nb11, nb12, nb13, item_ct1);
+                     });
 }
 
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {

ggml/src/ggml-sycl/dmmv.cpp

@@ -208,12 +208,10 @@ static void convert_mul_mat_vec_f16_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols,
-                                                          nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<1, 1, convert_f16>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -877,12 +875,11 @@ static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloa
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(vx, y, dst, ncols,
+                                                                                                    nrows, item_ct1);
+                          });
     }
 }
 
@@ -900,12 +897,10 @@ static void dequantize_mul_mat_vec_q4_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<QK4_0, QR4_0, dequantize_q4_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -921,12 +916,10 @@ static void dequantize_mul_mat_vec_q4_1_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<QK4_1, QR4_1, dequantize_q4_1>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -942,12 +935,10 @@ static void dequantize_mul_mat_vec_q5_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<QK5_0, QR5_0, dequantize_q5_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -963,12 +954,10 @@ static void dequantize_mul_mat_vec_q5_1_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<QK5_1, QR5_1, dequantize_q5_1>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -984,12 +973,10 @@ static void dequantize_mul_mat_vec_q8_0_sycl(const void *vx, const dfloat *y,
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>(
-                    vx, y, dst, ncols, nrows, item_ct1);
-            });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              dequantize_mul_mat_vec<QK8_0, QR8_0, dequantize_q8_0>(vx, y, dst, ncols, nrows, item_ct1);
+                          });
     }
 }
 
@@ -1002,11 +989,10 @@ static void dequantize_mul_mat_vec_q2_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q2_k(vx, y, dst, ncols, nrows, item_ct1);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
@@ -1018,11 +1004,10 @@ static void dequantize_mul_mat_vec_q3_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q3_k(vx, y, dst, ncols, nrows, item_ct1);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
@@ -1034,11 +1019,10 @@ static void dequantize_mul_mat_vec_q4_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q4_k(vx, y, dst, ncols, nrows, item_ct1);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
@@ -1047,11 +1031,10 @@ static void dequantize_mul_mat_vec_q5_K_sycl(const void *vx, const float *y,
                                              dpct::queue_ptr stream) {
     GGML_ASSERT(ncols % QK_K == 0);
     const sycl::range<3> block_dims(1, 1, QK_WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q5_k(vx, y, dst, ncols, item_ct1);
+                      });
 }
 
 static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
@@ -1063,11 +1046,10 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
     const int block_num_y = (nrows + ny - 1) / ny;
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, ny, QK_WARP_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(block_nums * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
-            dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                      [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(QK_WARP_SIZE)]] {
+                          dequantize_mul_mat_vec_q6_k(vx, y, dst, ncols, nrows, item_ct1);
+                      });
 }
 
 void ggml_sycl_op_dequantize_mul_mat_vec(

ggml/src/ggml-sycl/dpct/helper.hpp

@@ -13,10 +13,10 @@
 #ifndef GGML_SYCL_DPCT_HELPER_HPP
 #define GGML_SYCL_DPCT_HELPER_HPP
 
+#include <map>
 #include <sycl/sycl.hpp>
 #include <sycl/half_type.hpp>
 #include <syclcompat/math.hpp>
-#include <map>
 
 #ifdef GGML_SYCL_USE_INTEL_ONEMKL
 #include <oneapi/mkl.hpp>
@@ -118,6 +118,36 @@ inline auto get_onemath_backend(sycl::queue& queue)
 #endif
 }
 
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    namespace syclex = sycl::ext::oneapi::experimental;
+#endif
+
+template <int NR, typename Func>
+__dpct_inline__ void sycl_parallel_for(sycl::handler & cgh, sycl::nd_range<NR> nd_range, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::nd_launch(cgh, nd_range, func);
+#else
+    cgh.parallel_for(nd_range, func);
+#endif
+}
+
+template <int NR, typename Func>
+__dpct_inline__ void sycl_parallel_for(sycl::queue * q, sycl::nd_range<NR> nd_range, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::nd_launch(*q, nd_range, func);
+#else
+    q->parallel_for(nd_range, func);
+#endif
+}
+
+template <typename Func> __dpct_inline__ void sycl_launch(sycl::queue * stream, Func && func) {
+#ifdef SYCL_EXT_ONEAPI_ENQUEUE_FUNCTIONS
+    syclex::submit(*stream, func);
+#else
+    stream->submit(func);
+#endif
+}
+
 namespace dpct
 {
     typedef sycl::queue *queue_ptr;

ggml/src/ggml-sycl/element_wise.cpp

@@ -329,60 +329,51 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst,
                          const int ne12, const int nb1, const int nb2,
                          const int offset, queue_ptr stream) {
     int num_blocks = (n_elements + SYCL_ACC_BLOCK_SIZE - 1) / SYCL_ACC_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset,
-                    item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_ACC_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) {
+                          acc_f32(x, y, dst, n_elements, ne10, ne11, ne12, nb1, nb2, offset, item_ct1);
+                      });
 }
 
 template<typename T>
 static void gelu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            gelu(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { gelu(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void silu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_SILU_BLOCK_SIZE - 1) / SYCL_SILU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            silu(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_SILU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { silu(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void sgn_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            sgn(x, dst, k, item_ct1);
-            });
+    sycl_parallel_for(
+        stream, sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range(1, 1, 256)), sycl::range(1, 1, 256)),
+        [=](sycl::nd_item<3> item_ct1) { sgn(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void abs_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            abs_op(x, dst, k, item_ct1);
-            });
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)),
+        [=](sycl::nd_item<3> item_ct1) { abs_op(x, dst, k, item_ct1); });
 }
 
 
@@ -390,23 +381,20 @@ template<typename T>
 static void elu_sycl(const T * x, T * dst, const int k, queue_ptr stream) {
     // hard code for now
     const int num_blocks = ceil_div(k, 256);
-    stream->parallel_for(
-            sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)), [=](sycl::nd_item<3> item_ct1) {
-            elu_op(x, dst, k, item_ct1);
-            });
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<3>((sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, 256)), sycl::range<3>(1, 1, 256)),
+        [=](sycl::nd_item<3> item_ct1) { elu_op(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void gelu_quick_sycl(const T *x, T *dst, const int k,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_GELU_BLOCK_SIZE - 1) / SYCL_GELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            gelu_quick(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { gelu_quick(x, dst, k, item_ct1); });
 }
 
 
@@ -414,169 +402,133 @@ template<typename T>
 static void gelu_erf_sycl(const T *x, T *dst, const int k,
                                 queue_ptr stream) {
     const int num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            gelu_erf(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_GELU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { gelu_erf(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void tanh_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_TANH_BLOCK_SIZE - 1) / SYCL_TANH_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            tanh(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_TANH_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { tanh(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void relu_sycl(const T *x, T *dst, const int k,
                           queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            relu(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { relu(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void hardsigmoid_sycl(const T *x, T *dst, const int k,
                                  queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSIGMOID_BLOCK_SIZE - 1) / SYCL_HARDSIGMOID_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_HARDSIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            hardsigmoid(x, dst, k, item_ct1);
-        });
+        [=](sycl::nd_item<3> item_ct1) { hardsigmoid(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void hardswish_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_HARDSWISH_BLOCK_SIZE - 1) / SYCL_HARDSWISH_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_HARDSWISH_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            hardswish(x, dst, k, item_ct1);
-        });
+        [=](sycl::nd_item<3> item_ct1) { hardswish(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void exp_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            exp(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { exp(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void log_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_EXP_BLOCK_SIZE - 1) / SYCL_EXP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            log(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_EXP_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { log(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void neg_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            neg(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { neg(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void step_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_NEG_BLOCK_SIZE - 1) / SYCL_NEG_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            step(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_NEG_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { step(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void sigmoid_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIGMOID_BLOCK_SIZE - 1) / SYCL_SIGMOID_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE),
+    sycl_parallel_for(
+        stream,
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE),
                           sycl::range<3>(1, 1, SYCL_SIGMOID_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sigmoid(x, dst, k, item_ct1);
-        });
+        [=](sycl::nd_item<3> item_ct1) { sigmoid(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void sqrt_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQRT_BLOCK_SIZE - 1) / SYCL_SQRT_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sqrt(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_SQRT_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { sqrt(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void sin_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sin(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { sin(x, dst, k, item_ct1); });
 }
 
 template<typename T>
 static void cos_sycl(const T *x, T *dst, const int k,
                                queue_ptr stream) {
     const int num_blocks = (k + SYCL_SIN_BLOCK_SIZE - 1) / SYCL_SIN_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            cos(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_SIN_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { cos(x, dst, k, item_ct1); });
 }
 
 template<typename T>
@@ -584,26 +536,20 @@ static void leaky_relu_sycl(const T *x, T *dst, const int k,
                                 const float negative_slope,
                                 queue_ptr stream) {
     const int num_blocks = (k + SYCL_RELU_BLOCK_SIZE - 1) / SYCL_RELU_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            leaky_relu(x, dst, k, negative_slope, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_RELU_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { leaky_relu(x, dst, k, negative_slope, item_ct1); });
 }
 
 template<typename T>
 static void sqr_sycl(const T *x, T *dst, const int k,
                          queue_ptr stream) {
     const int num_blocks = (k + SYCL_SQR_BLOCK_SIZE - 1) / SYCL_SQR_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            sqr(x, dst, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_SQR_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { sqr(x, dst, k, item_ct1); });
 }
 
 template<typename T>
@@ -614,9 +560,8 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
     int dst_size = ne10 * ne11 * ne12 * ne13;
     int num_blocks = (dst_size + SYCL_UPSCALE_BLOCK_SIZE - 1) / SYCL_UPSCALE_BLOCK_SIZE;
     sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
-    stream->parallel_for(
-        sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)),
-        [=](sycl::nd_item<1> item_ct1) {
+    sycl_parallel_for<1>(
+        stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
             upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
         });
 }
@@ -627,12 +572,10 @@ static void pad_sycl(const T *x, T *dst, const int ne00,
                          const int ne1, const int ne2, queue_ptr stream) {
     int num_blocks = (ne0 + SYCL_PAD_BLOCK_SIZE - 1) / SYCL_PAD_BLOCK_SIZE;
     sycl::range<3> gridDim(ne2, ne1, num_blocks);
-    stream->parallel_for(
-        sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            pad(x, dst, ne0, ne00, ne01, ne02, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
 }
 
 template<typename T>
@@ -640,13 +583,10 @@ static void clamp_sycl(const T *x, T *dst, const float min,
                            const float max, const int k,
                            queue_ptr stream) {
     const int num_blocks = (k + SYCL_CLAMP_BLOCK_SIZE - 1) / SYCL_CLAMP_BLOCK_SIZE;
-    stream->parallel_for(
-        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) *
-                              sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
-                          sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
-        [=](sycl::nd_item<3> item_ct1) {
-            clamp(x, dst, min, max, k, item_ct1);
-        });
+    sycl_parallel_for(stream,
+                      sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE),
+                                        sycl::range<3>(1, 1, SYCL_CLAMP_BLOCK_SIZE)),
+                      [=](sycl::nd_item<3> item_ct1) { clamp(x, dst, min, max, k, item_ct1); });
 }
 
 inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-sycl/getrows.cpp

@@ -60,54 +60,6 @@ static void k_get_rows(
     dst_row[iybs + iqs + y_offset] = v.y();
 }
 
-template<int qk, int qr, dequantize_kernel_t_reorder dequantize_kernel_recorder, typename dst_t>
-static void k_get_rows_reorder(
-            const void * src0, const void *src0_dq, const int32_t * src1, dst_t * dst,
-            int64_t ne00, /*int64_t ne01, int64_t ne02, int64_t ne03,*/
-            /*int64_t ne10, int64_t ne11,*/ int64_t ne12, /*int64_t ne13,*/
-            /*size_t s0,*/ size_t s1, size_t s2, size_t s3,
-            /*size_t nb00,*/ size_t nb01, size_t nb02, size_t nb03,
-            size_t s10, size_t s11, size_t s12,
-            const sycl::nd_item<3> &item_ct1/*, size_t s13*/) {
-
-    const int i00 = (item_ct1.get_group(2) * item_ct1.get_local_range(2) +
-                     item_ct1.get_local_id(2)) *
-                    2;
-    const int i10 = item_ct1.get_local_range(1) * item_ct1.get_group(1) +
-                    item_ct1.get_local_id(1);
-    const int i11 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) /
-                    ne12;
-    const int i12 = (item_ct1.get_group(0) * item_ct1.get_local_range(0) +
-                     item_ct1.get_local_id(0)) %
-                    ne12;
-
-    if (i00 >= ne00) {
-        return;
-    }
-    auto ncols = ne00;
-    const int i01 = src1[i10*s10 + i11*s11 + i12*s12];
-
-    dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
-
-    const int src0_off = i01 * ncols + i00;
-    const int ib = src0_off / QK4_0; // block index
-    const int iqs = (i00%qk)/qr; // x quant index
-    const int iybs = i00 - i00%qk; // dst block start index
-    const int y_offset = qr == 1 ? 1 : qk/2;
-
-    // dequantize
-    dfloat2 v;
-    dequantize_kernel_recorder((const void *)src0_dq, ib, (const void *)src0, src0_off/2, v);
-
-    dst_row[iybs + iqs + 0] = v.x();
-    dst_row[iybs + iqs + y_offset] = v.y();
-
-    GGML_UNUSED(nb01);
-    GGML_UNUSED(nb02);
-    GGML_UNUSED(nb03);
-}
-
 template<typename src0_t, typename dst_t>
 static void k_get_rows_float(
             const src0_t * src0, const int32_t * src1, dst_t * dst,
@@ -166,58 +118,15 @@ static void get_rows_sycl(ggml_backend_sycl_context & ctx, const ggml_tensor *sr
 
     GGML_ASSERT(ne00 % 2 == 0);
 
-    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                         [=](sycl::nd_item<3> item_ct1) {
-                             k_get_rows<qk, qr, dq>(
-                                 src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
-                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
-                         });
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        k_get_rows<qk, qr, dq>(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2, s3, nb01, nb02, nb03, s10, s11, s12,
+                               item_ct1);
+    });
 
     GGML_UNUSED(dst);
     GGML_UNUSED(ctx);
 }
 
-template <int qk, int qr, dequantize_kernel_t_reorder dq_reorder>
-static void get_rows_sycl_reorder(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
-                          ggml_tensor *dst, const void *src0_dd,
-                          const int32_t *src1_dd, float *dst_dd,
-                          queue_ptr stream) {
-
-    GGML_TENSOR_BINARY_OP_LOCALS
-
-    const sycl::range<3> block_dims(1, 1, SYCL_GET_ROWS_BLOCK_SIZE);
-    const int block_num_x = (ne00 + 2*SYCL_GET_ROWS_BLOCK_SIZE - 1) / (2*SYCL_GET_ROWS_BLOCK_SIZE);
-    const sycl::range<3> block_nums(ne11 * ne12, ne10, block_num_x);
-
-    // strides in elements
-    //const size_t s0 = nb0 / ggml_element_size(dst);
-    const size_t s1 = nb1 / ggml_element_size(dst);
-    const size_t s2 = nb2 / ggml_element_size(dst);
-    const size_t s3 = nb3 / ggml_element_size(dst);
-
-    const size_t s10 = nb10 / ggml_element_size(src1);
-    const size_t s11 = nb11 / ggml_element_size(src1);
-    const size_t s12 = nb12 / ggml_element_size(src1);
-    //const size_t s13 = nb13 / ggml_element_size(src1);
-
-    GGML_ASSERT(ne00 % 2 == 0);
-
-    const uint8_t* src0_q = (const uint8_t*)src0_dd;
-    const size_t ncols = ne00;
-    const size_t nrows = ne01;
-    const sycl::half* src0_dq = (const sycl::half*)(src0_q + nrows * ncols / 2);
-    stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                         [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]]{
-                             k_get_rows_reorder<qk, qr, dq_reorder>(
-                                 src0_dd, src0_dq, src1_dd, dst_dd, ne00, ne12, s1, s2,
-                                 s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
-                         });
-
-    GGML_UNUSED(dst);
-    GGML_UNUSED(ctx);
-}
-
-
 template <typename src0_t>
 static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
                                 const ggml_tensor *src1, ggml_tensor *dst,
@@ -245,9 +154,8 @@ static void get_rows_sycl_float(ggml_backend_sycl_context & ctx, const ggml_tens
         dpct::has_capability_or_fail(stream->get_device(),
                                      {sycl::aspect::fp16});
 
-        stream->parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(
+            stream, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                 k_get_rows_float(src0_dd, src1_dd, dst_dd, ne00, ne12, s1, s2,
                                  s3, nb01, nb02, nb03, s10, s11, s12, item_ct1);
             });
@@ -277,13 +185,8 @@ void ggml_sycl_op_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
             src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q4_0:
-            if (ctx.opt_feature.reorder && dst->op == GGML_OP_MUL_MAT) {
-                get_rows_sycl_reorder<QK4_0, QR4_0, dequantize_q4_0_reorder>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
-                src1_i32, (float *)dst->data, ctx.stream());
-            } else {
-                get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
-                src1_i32, (float *)dst->data, ctx.stream());
-            }
+            get_rows_sycl<QK4_0, QR4_0, dequantize_q4_0>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,
+            src1_i32, (float *)dst->data, ctx.stream());
             break;
         case GGML_TYPE_Q4_1:
             get_rows_sycl<QK4_1, QR4_1, dequantize_q4_1>(ctx, dst->src[0], dst->src[1], dst, (const float *)dst->src[0]->data,

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

@@ -83,9 +83,7 @@ static ggml_sycl_device_info ggml_sycl_init() {
 
         info.devices[i].cc =
             100 * prop.get_major_version() + 10 * prop.get_minor_version();
-        info.devices[i].hw_info = get_device_hw_info(&device);
-        info.devices[i].opt_feature = check_gpu_optimize_feature(info.devices[i].hw_info.arch);
-
+        info.devices[i].opt_feature.reorder = !device.ext_oneapi_architecture_is(syclex::arch_category::intel_gpu);
         info.max_work_group_sizes[i] = prop.get_max_work_group_size();
     }
 
@@ -195,7 +193,7 @@ static void ggml_check_sycl() try {
 
     if (!initialized) {
         g_ggml_sycl_debug = get_sycl_env("GGML_SYCL_DEBUG", 0);
-        g_ggml_sycl_disable_optimize= get_sycl_env("GGML_SYCL_DISABLE_OPT", 1);
+        g_ggml_sycl_disable_optimize = get_sycl_env("GGML_SYCL_DISABLE_OPT", 0);
         g_ggml_sycl_disable_graph = get_sycl_env("GGML_SYCL_DISABLE_GRAPH", 1);
         g_ggml_sycl_disable_dnn = get_sycl_env("GGML_SYCL_DISABLE_DNN", 0);
         g_ggml_sycl_prioritize_dmmv = get_sycl_env("GGML_SYCL_PRIORITIZE_DMMV", 0);
@@ -1887,13 +1885,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const size_t shared_mem = ncols_pad * sizeof(int);
 
     if (order == GGML_SORT_ORDER_ASC) {
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_ASC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1901,13 +1898,12 @@ static void argsort_f32_i32_sycl(const float *x, int *dst, const int ncols,
                 });
         });
     } else if (order == GGML_SORT_ORDER_DESC) {
-        stream->submit([&](sycl::handler &cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<uint8_t, 1> dpct_local_acc_ct1(
                 sycl::range<1>(shared_mem), cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_argsort_f32_i32<GGML_SORT_ORDER_DESC>(
                         x, dst, ncols, ncols_pad, item_ct1,
                         dpct_local_acc_ct1.get_multi_ptr<sycl::access::decorated::no>()
@@ -1925,50 +1921,47 @@ static void argmax_f32_i32_sycl(const float *x, int *dst, const int ncols,
     const sycl::range<3> block_nums(1, nrows, 1);
     const size_t shared_mem = 256 * sizeof(float);
 
-    stream->submit([&](sycl::handler &cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<float, 1> shared_data(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
         sycl::local_accessor<int, 1> shared_indices(
             sycl::range<1>(shared_mem/sizeof(float)), cgh);
 
-        cgh.parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                const int tid = item_ct1.get_local_id(2);
-                const int row = item_ct1.get_global_id(1);
+        sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+            const int tid = item_ct1.get_local_id(2);
+            const int row = item_ct1.get_global_id(1);
 
-                float max_val = -INFINITY;
-                int max_idx = -1;
+            float max_val = -INFINITY;
+            int   max_idx = -1;
 
-                for (int col = tid; col < ncols; col += 256) {
-                    float val = x[row * ncols + col];
-                    if (val > max_val) {
-                        max_val = val;
-                        max_idx = col;
+            for (int col = tid; col < ncols; col += 256) {
+                float val = x[row * ncols + col];
+                if (val > max_val) {
+                    max_val = val;
+                    max_idx = col;
+                }
+            }
+
+            shared_data[tid]    = max_val;
+            shared_indices[tid] = max_idx;
+            item_ct1.barrier(sycl::access::fence_space::local_space);
+
+            for (int stride = 256 / 2; stride > 0; stride >>= 1) {
+                if (tid < stride) {
+                    float val1 = shared_data[tid];
+                    float val2 = shared_data[tid + stride];
+                    if (val2 > val1) {
+                        shared_data[tid]    = val2;
+                        shared_indices[tid] = shared_indices[tid + stride];
                     }
                 }
-
-                shared_data[tid] = max_val;
-                shared_indices[tid] = max_idx;
                 item_ct1.barrier(sycl::access::fence_space::local_space);
+            }
 
-                for (int stride = 256/2; stride > 0; stride >>= 1) {
-                    if (tid < stride) {
-                        float val1 = shared_data[tid];
-                        float val2 = shared_data[tid + stride];
-                        if (val2 > val1) {
-                            shared_data[tid] = val2;
-                            shared_indices[tid] = shared_indices[tid + stride];
-                        }
-                    }
-                    item_ct1.barrier(sycl::access::fence_space::local_space);
-                }
-
-
-                if (tid == 0) {
-                    dst[row] = shared_indices[0];
-                }
-            });
+            if (tid == 0) {
+                dst[row] = shared_indices[0];
+            }
+        });
     });
 }
 static void diag_mask_inf_f32_sycl(const float *x, float *dst,
@@ -2952,7 +2945,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 void **       ptrs_dst_get = ptrs_dst.get();
                 size_t        nb12_scaled  = src1->type == GGML_TYPE_F16 ? nb12 : s12 * sizeof(sycl::half);
                 size_t        nb13_scaled  = src1->type == GGML_TYPE_F16 ? nb13 : s13 * sizeof(sycl::half);
-                cgh.parallel_for(sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(cgh, sycl::nd_range<3>(block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     k_compute_batched_ptrs(src0_f16, src1_f16, dst_ddf, ptrs_src_get, ptrs_dst_get, ne12, ne13, ne23, nb02,
                                            nb03, nb12_scaled, nb13_scaled, nbd2, nbd3, r2, r3, item_ct1);
                 });
@@ -3456,7 +3449,7 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne10, 768u));
                 sycl::range<3> grid_dims(1, n_ids, ids->ne[1]);
-                stream->submit([&](sycl::handler &cgh) {
+                sycl_launch(stream, [&](sycl::handler & cgh) {
                     sycl::local_accessor<int, 0> src1_row_acc(cgh);
 
                     char *__restrict src1_contiguous_get =
@@ -3468,9 +3461,8 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
                     size_t ids_nb_ct6 = ids->nb[1];
                     size_t ids_nb_ct7 = ids->nb[0];
 
-                    cgh.parallel_for(
-                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                        [=](sycl::nd_item<3> item_ct1) {
+                    sycl_parallel_for(
+                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                             k_copy_src1_to_contiguous(
                                 src1_original, src1_contiguous_get,
                                 dev_cur_src1_row_get,
@@ -3501,15 +3493,14 @@ static void ggml_sycl_mul_mat_id(ggml_backend_sycl_context & ctx,
             {
                 sycl::range<3> block_dims(1, 1, std::min((unsigned int)ne0, 768u));
                 sycl::range<3> grid_dims(1, 1, num_src1_rows);
-                stream->submit([&](sycl::handler &cgh) {
+                sycl_launch(stream, [&](sycl::handler & cgh) {
                     const char *__restrict dst_contiguous_get =
                         dst_contiguous.get();
                     const mmid_row_mapping *__restrict dev_row_mapping_get =
                         dev_row_mapping.get();
 
-                    cgh.parallel_for(
-                        sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                        [=](sycl::nd_item<3> item_ct1) {
+                    sycl_parallel_for(
+                        cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                             k_copy_dst_from_contiguous(dst_original,
                                                        dst_contiguous_get,
                                                        dev_row_mapping_get,

ggml/src/ggml-sycl/gla.cpp

@@ -11,13 +11,13 @@ static void gated_linear_attn_f32_kernel(const dpct::queue_ptr stream, u_int B,
     const u_int n_seq_tokens = T / B;
     sycl::range<1> block_dims((C / H));
     sycl::range<1> grid_dims((B * H));
-    stream->submit([&](sycl::handler & cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         /* local memory accessors*/
         auto _k  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _r  = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
         auto _td = sycl::local_accessor<float, 1>(sycl::range<1>(head_size), cgh);
 
-        cgh.parallel_for(sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
+        sycl_parallel_for<1>(cgh, sycl::nd_range<1>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<1> item) {
             u_int tid = item.get_local_id(0);
             u_int bid = item.get_group(0);
 

ggml/src/ggml-sycl/im2col.cpp

@@ -70,7 +70,7 @@ static void im2col_sycl_internal(const float * x, T * dst, int64_t IW, int64_t I
 
     const int64_t CHW = IC * KH * KW;
 
-    stream->parallel_for(sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
+    sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * local_range, local_range), [=](sycl::nd_item<3> item_ct1) {
         im2col_kernel<T>(x, dst, batch_offset, offset_delta, IC, IW, IH, OH, OW, KW, KH, parallel_elements, CHW, s0, s1,
                          p0, p1, d0, d1, item_ct1);
     });

ggml/src/ggml-sycl/mmq.cpp

@@ -1818,7 +1818,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1829,9 +1829,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1853,7 +1852,7 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q4_0_acc_ct1(
@@ -1864,9 +1863,8 @@ static void ggml_mul_mat_q4_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1933,7 +1931,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1944,9 +1942,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -1968,7 +1965,7 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q4_1_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + +mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_1_acc_ct1(
@@ -1979,9 +1976,8 @@ static void ggml_mul_mat_q4_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2048,7 +2044,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2059,9 +2055,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2083,7 +2078,7 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_0_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q5_0_acc_ct1(
@@ -2094,9 +2089,8 @@ static void ggml_mul_mat_q5_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2163,7 +2157,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2174,9 +2168,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2198,7 +2191,7 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_1_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_1_acc_ct1(
@@ -2209,9 +2202,8 @@ static void ggml_mul_mat_q5_1_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_1<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2278,7 +2270,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2289,9 +2281,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2313,7 +2304,7 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_qs_q8_0_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<float, 1> tile_x_d_q8_0_acc_ct1(
@@ -2324,9 +2315,8 @@ static void ggml_mul_mat_q8_0_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q8_0<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2393,7 +2383,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2406,9 +2396,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2431,7 +2420,7 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q2_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q2_K_acc_ct1(
@@ -2444,9 +2433,8 @@ static void ggml_mul_mat_q2_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q2_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2516,7 +2504,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2531,9 +2519,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2557,7 +2544,7 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q3_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q3_K_acc_ct1(
@@ -2572,9 +2559,8 @@ static void ggml_mul_mat_q3_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q3_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2644,7 +2630,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2657,9 +2643,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2682,7 +2667,7 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q4_K_acc_ct1(
                     sycl::range<1>(mmq_y * (WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q4_K_acc_ct1(
@@ -2695,9 +2680,8 @@ static void ggml_mul_mat_q4_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q4_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2765,7 +2749,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2778,9 +2762,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2803,7 +2786,7 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_q5_K_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_q5_K_acc_ct1(
@@ -2816,9 +2799,8 @@ static void ggml_mul_mat_q5_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q5_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2886,7 +2868,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2899,9 +2881,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,
@@ -2924,7 +2905,7 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
             dpct::has_capability_or_fail(stream->get_device(),
                                          {sycl::aspect::fp16});
 
-            stream->submit([&](sycl::handler &cgh) {
+            sycl_launch(stream, [&](sycl::handler & cgh) {
                 sycl::local_accessor<int, 1> tile_x_ql_acc_ct1(
                     sycl::range<1>(mmq_y * (2 * WARP_SIZE) + mmq_y), cgh);
                 sycl::local_accessor<sycl::half2, 1> tile_x_dm_acc_ct1(
@@ -2937,9 +2918,8 @@ static void ggml_mul_mat_q6_K_q8_1_sycl(const void *vx, const void *vy,
                 sycl::local_accessor<sycl::half2, 1> tile_y_ds_acc_ct1(
                     sycl::range<1>(mmq_x * WARP_SIZE / QI8_1), cgh);
 
-                cgh.parallel_for(
-                    sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                    [=](sycl::nd_item<3> item_ct1) {
+                sycl_parallel_for(
+                    cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                         mul_mat_q6_K<need_check>(
                             vx, vy, dst, ncols_x, nrows_x, ncols_y, nrows_y,
                             nrows_dst, item_ct1,

ggml/src/ggml-sycl/mmvq.cpp

@@ -544,12 +544,12 @@ static void reorder_mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, (block_num_y * WARP_SIZE));
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    stream->submit([&](sycl::handler & cgh) {
-        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
-                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
-                                                                                           nd_item);
-                         });
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
     });
 }
 
@@ -561,12 +561,12 @@ static void mul_mat_vec_q4_0_q8_1_sycl(const void * vx, const void * vy, float *
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
 
     {
-        stream->submit([&](sycl::handler & cgh) {
-            cgh.parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                 mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
-                                     vx, vy, dst, ncols, nrows, item_ct1);
-                             });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK4_0, QI4_0, block_q4_0, VDR_Q4_0_Q8_1_MMVQ, vec_dot_q4_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -580,17 +580,12 @@ static void mul_mat_vec_q4_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK4_0, QI4_1, block_q4_1,
-                                      VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK4_0, QI4_1, block_q4_1, VDR_Q4_1_Q8_1_MMVQ, vec_dot_q4_1_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -604,17 +599,12 @@ static void mul_mat_vec_q5_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK5_0, QI5_0, block_q5_0,
-                                      VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK5_0, QI5_0, block_q5_0, VDR_Q5_0_Q8_1_MMVQ, vec_dot_q5_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -628,17 +618,12 @@ static void mul_mat_vec_q5_1_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK5_1, QI5_1, block_q5_1,
-                                      VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK5_1, QI5_1, block_q5_1, VDR_Q5_1_Q8_1_MMVQ, vec_dot_q5_1_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -652,17 +637,12 @@ static void mul_mat_vec_q8_0_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK8_0, QI8_0, block_q8_0,
-                                      VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK8_0, QI8_0, block_q8_0, VDR_Q8_0_Q8_1_MMVQ, vec_dot_q8_0_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -676,17 +656,12 @@ static void mul_mat_vec_q2_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI2_K, block_q2_K,
-                                      VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI2_K, block_q2_K, VDR_Q2_K_Q8_1_MMVQ, vec_dot_q2_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -700,17 +675,12 @@ static void mul_mat_vec_q3_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI3_K, block_q3_K,
-                                      VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI3_K, block_q3_K, VDR_Q3_K_Q8_1_MMVQ, vec_dot_q3_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -724,17 +694,12 @@ static void mul_mat_vec_q4_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI4_K, block_q4_K,
-                                      VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI4_K, block_q4_K, VDR_Q4_K_Q8_1_MMVQ, vec_dot_q4_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -750,12 +715,12 @@ static void reorder_mul_mat_vec_q4_k_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    stream->submit([&](sycl::handler & cgh) {
-        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
-                            [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                                mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols,
-                                                                                            nrows, nd_item);
-                            });
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
     });
 }
 
@@ -769,17 +734,12 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI5_K, block_q5_K,
-                                      VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI5_K, block_q5_K, VDR_Q5_K_Q8_1_MMVQ, vec_dot_q5_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -794,12 +754,12 @@ static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy,
     const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
     const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
 
-    stream->submit([&](sycl::handler & cgh) {
-        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
-                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
-                                                                                           nd_item);
-                         });
+    sycl_launch(stream, [&](sycl::handler & cgh) {
+        sycl_parallel_for(cgh, sycl::nd_range<3>(global_size, workgroup_size),
+                          [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                              mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
+                                                                                            nd_item);
+                          });
     });
 }
 static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
@@ -811,17 +771,12 @@ static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q<QK_K, QI6_K, block_q6_K,
-                                      VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q<QK_K, QI6_K, block_q6_K, VDR_Q6_K_Q8_1_MMVQ, vec_dot_q6_K_q8_1>(
+                                      vx, vy, dst, ncols, nrows, item_ct1);
+                              });
         });
     }
 }
@@ -836,14 +791,12 @@ static void mul_mat_vec_iq2_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS/2, block_iq2_xxs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_xxs_q8_1<QK_K, QI2_XXS / 2, block_iq2_xxs, 1>(vx, vy, dst, ncols,
+                                                                                                  nrows, item_ct1);
+                              });
         });
     }
 }
@@ -857,14 +810,12 @@ static void mul_mat_vec_iq2_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        stream->submit([&](sycl::handler & cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS/2, block_iq2_xs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_xs_q8_1<QK_K, QI2_XS / 2, block_iq2_xs, 1>(vx, vy, dst, ncols,
+                                                                                               nrows, item_ct1);
+                              });
         });
     }
 }
@@ -878,15 +829,12 @@ static void mul_mat_vec_iq2_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S/2, block_iq2_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq2_s_q8_1<QK_K, QI2_S / 2, block_iq2_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                            item_ct1);
+                              });
         });
     }
 }
@@ -900,15 +848,12 @@ static void mul_mat_vec_iq3_xxs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS/2, block_iq3_xxs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq3_xxs_q8_1<QK_K, QI3_XXS / 2, block_iq3_xxs, 1>(vx, vy, dst, ncols,
+                                                                                                  nrows, item_ct1);
+                              });
         });
     }
 }
@@ -922,15 +867,12 @@ static void mul_mat_vec_iq3_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S/2, block_iq3_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq3_s_q8_1<QK_K, QI3_S / 2, block_iq3_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                            item_ct1);
+                              });
         });
     }
 }
@@ -944,15 +886,12 @@ static void mul_mat_vec_iq1_s_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq1_s_q8_1<QK_K, QI1_S, block_iq1_s, 1>(vx, vy, dst, ncols, nrows,
+                                                                                        item_ct1);
+                              });
         });
     }
 }
@@ -966,14 +905,12 @@ static void mul_mat_vec_iq1_m_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq1_m_q8_1<QK_K, QI1_S, block_iq1_m, 1>(vx, vy, dst, ncols, nrows,
+                                                                                        item_ct1);
+                              });
         });
     }
 }
@@ -987,15 +924,12 @@ static void mul_mat_vec_iq4_nl_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq4_nl_q8_1<QK4_NL, QI4_NL, block_iq4_nl, 2>(vx, vy, dst, ncols, nrows,
+                                                                                             item_ct1);
+                              });
         });
     }
 }
@@ -1009,15 +943,12 @@ static void mul_mat_vec_iq4_xs_q8_1_sycl(const void *vx, const void *vy,
     const sycl::range<3> block_nums(1, 1, block_num_y);
     const sycl::range<3> block_dims(1, GGML_SYCL_MMV_Y, WARP_SIZE);
     {
-
-        stream->submit([&](sycl::handler &cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(block_nums * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                    [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                        mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS/4, block_iq4_xs, 1>(
-                            vx, vy, dst, ncols, nrows, item_ct1);
-                    });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  mul_mat_vec_q_iq4_xs_q8_1<QK_K, QI4_XS / 4, block_iq4_xs, 1>(vx, vy, dst, ncols,
+                                                                                               nrows, item_ct1);
+                              });
         });
     }
 }

ggml/src/ggml-sycl/norm.cpp

@@ -254,14 +254,13 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
     GGML_ASSERT(ncols % WARP_SIZE == 0);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        stream->submit([&](sycl::handler& cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                           nullptr, WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -272,16 +271,15 @@ static void norm_f32_sycl(const float * x, float * dst, const int ncols, const i
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<sycl::float2, 1> s_sum_acc_ct1(
                             sycl::range<1>(work_group_size / WARP_SIZE), cgh);
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                           get_pointer(s_sum_acc_ct1), work_group_size);
+                              });
+        });
     }
 }
 
@@ -290,18 +288,14 @@ static void group_norm_f32_sycl(const float* x, float* dst,
     const int ne_elements, queue_ptr stream, int device) {
     if (group_size < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             const float eps_ct4 = eps;
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    group_norm_f32(
-                        x, dst, group_size, ne_elements, eps_ct4, item_ct1,
-                        nullptr, WARP_SIZE);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1, nullptr,
+                                                 WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -313,22 +307,18 @@ static void group_norm_f32_sycl(const float* x, float* dst,
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
 
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
 
             const float eps_ct4 = eps;
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    group_norm_f32(x, dst, group_size, ne_elements,
-                        eps_ct4, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, num_groups) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  group_norm_f32(x, dst, group_size, ne_elements, eps_ct4, item_ct1,
+                                                 get_pointer(s_sum_acc_ct1), work_group_size);
+                              });
+        });
     }
 }
 
@@ -340,14 +330,13 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
     const sycl::range<3> global_dims(nsamples, nchannels, nrows);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        stream->submit([&](sycl::handler& cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, nullptr, WARP_SIZE);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                               nullptr, WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -358,16 +347,15 @@ static void rms_norm_f32_sycl(const float* x, float* dst, const int ncols, const
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
-            cgh.parallel_for(
-                sycl::nd_range<3>(global_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1, get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(global_dims * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  rms_norm_f32(x, dst, ncols, stride_row, stride_channel, stride_sample, eps, item_ct1,
+                                               get_pointer(s_sum_acc_ct1), work_group_size);
+                              });
+        });
     }
 }
 
@@ -378,16 +366,12 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
     // printf("%s ncols=%d, nrows=%d, WARP_SIZE=%d\n", __func__, ncols, nrows, WARP_SIZE);
     if (ncols < 1024) {
         const sycl::range<3> block_dims(1, 1, WARP_SIZE);
-        stream->submit([&](sycl::handler& cgh) {
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
-                        nullptr, WARP_SIZE);
-                });
-            });
+        sycl_launch(stream, [&](sycl::handler & cgh) {
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  l2_norm_f32(x, dst, ncols, eps, item_ct1, nullptr, WARP_SIZE);
+                              });
+        });
     }
     else {
         const int work_group_size = ggml_sycl_info().max_work_group_sizes[device];
@@ -398,18 +382,15 @@ static void l2_norm_f32_sycl(const float* x, float* dst, const int ncols,
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> s_sum_acc_ct1(sycl::range<1>(work_group_size / WARP_SIZE),
                 cgh);
-            cgh.parallel_for(
-                sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims,
-                    block_dims),
-                [=](sycl::nd_item<3> item_ct1)
-                [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                    l2_norm_f32(x, dst, ncols, eps, item_ct1,
-                        get_pointer(s_sum_acc_ct1), work_group_size);
-                });
-            });
+            sycl_parallel_for(cgh, sycl::nd_range<3>(sycl::range<3>(1, 1, nrows) * block_dims, block_dims),
+                              [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                  l2_norm_f32(x, dst, ncols, eps, item_ct1, get_pointer(s_sum_acc_ct1),
+                                              work_group_size);
+                              });
+        });
     }
 }
 

ggml/src/ggml-sycl/rope.cpp

@@ -235,20 +235,22 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                                theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_norm<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     } else {
         /*
         DPCT1049:41: The work-group size passed to the SYCL kernel may exceed
         the limit. To get the device limit, query
         info::device::max_work_group_size. Adjust the work-group size if needed.
         */
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                               theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_norm<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     }
 }
 
@@ -267,15 +269,17 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
     dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
 
     if (freq_factors == nullptr) {
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                                theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_neox<T, false>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                  attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     } else {
-        stream->parallel_for(sycl::nd_range<3>(block_nums * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
-            rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor, corr_dims,
-                               theta_scale, freq_factors, item_ct1);
-        });
+        sycl_parallel_for(stream, sycl::nd_range<3>(block_nums * block_dims, block_dims),
+                          [=](sycl::nd_item<3> item_ct1) {
+                              rope_neox<T, true>(x, dst, ne0, ne1, s1, s2, n_dims, pos, freq_scale, ext_factor,
+                                                 attn_factor, corr_dims, theta_scale, freq_factors, item_ct1);
+                          });
     }
 }
 
@@ -298,12 +302,12 @@ static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
@@ -333,12 +337,12 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1,
     }
     // launch kernel
     if (freq_factors == nullptr) {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                   corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });
     } else {
-        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+        sycl_parallel_for(stream, nd_range, [=](sycl::nd_item<3> item_ct1) {
             rope_vision<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
         });

ggml/src/ggml-sycl/softmax.cpp

@@ -127,11 +127,11 @@ static void soft_max_f32_submitter(const float * x, const T * mask, float * dst,
                                    const int nrows_y, const float scale, const float max_bias, const float m0,
                                    const float m1, uint32_t n_head_log2, sycl::range<3> block_nums, sycl::range<3> block_dims,
                                    const size_t n_local_scratch, queue_ptr stream) {
-    stream->submit([&](sycl::handler &cgh) {
+    sycl_launch(stream, [&](sycl::handler & cgh) {
         sycl::local_accessor<float, 1> local_buf_acc(n_local_scratch, cgh);
 
-        cgh.parallel_for(
-            sycl::nd_range<3>(block_nums * block_dims, block_dims),
+        sycl_parallel_for(
+            cgh, sycl::nd_range<3>(block_nums * block_dims, block_dims),
             [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
                 soft_max_f32<vals_smem, ncols_template, block_size_template>(x, mask, dst, ncols_par,
                                                                              nrows_y, scale, max_bias, m0,

ggml/src/ggml-sycl/sycl_hw.cpp

@@ -1,6 +1,7 @@
 #include "sycl_hw.hpp"
 
-
+// TODO: currently not used
+/*
 sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
   sycl_hw_info res;
   int32_t id = device_ptr->get_info<sycl::ext::intel::info::device::device_id>();
@@ -11,3 +12,4 @@ sycl_hw_info get_device_hw_info(sycl::device *device_ptr) {
 
   return res;
 }
+*/

ggml/src/ggml-sycl/sycl_hw.hpp

@@ -10,6 +10,8 @@
 
 namespace syclex = sycl::ext::oneapi::experimental;
 
+// TODO: currently not used
+/*
 struct sycl_hw_info {
   syclex::architecture arch;
   int32_t device_id;
@@ -18,6 +20,7 @@ struct sycl_hw_info {
 bool is_in_vector(std::vector<int> &vec, int item);
 
 sycl_hw_info get_device_hw_info(sycl::device *device_ptr);
+*/
 
 
 #endif // SYCL_HW_HPP

ggml/src/ggml-sycl/tsembd.cpp

@@ -45,14 +45,9 @@ static void timestep_embedding_f32_sycl(
     int num_blocks = (half_ceil + SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE - 1) / SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE;
     sycl::range<3> block_dims(1, 1, SYCL_TIMESTEP_EMBEDDING_BLOCK_SIZE);
     sycl::range<3> gridDim(1, ne00, num_blocks);
-    stream->parallel_for(
-        sycl::nd_range<3>(
-            gridDim * block_dims, block_dims),
-        [=](sycl::nd_item<3> item_ct1) {
-            timestep_embedding_f32(
-                x, dst, nb1, dim, max_period, item_ct1
-            );
-        });
+    sycl_parallel_for(stream, sycl::nd_range<3>(gridDim * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
+        timestep_embedding_f32(x, dst, nb1, dim, max_period, item_ct1);
+    });
 }
 
 void ggml_sycl_op_timestep_embedding(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-sycl/wkv.cpp

@@ -207,12 +207,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE>(
                         B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -220,12 +219,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -264,12 +262,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
 
     // Submit kernel
     if (C / H == WKV_BLOCK_SIZE) {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE>(
                         B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -277,12 +274,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
                 });
         });
     } else {
-        stream->submit([&](sycl::handler& cgh) {
+        sycl_launch(stream, [&](sycl::handler & cgh) {
             sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
 
-            cgh.parallel_for(
-                sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-                [=](sycl::nd_item<3> item_ct1) {
+            sycl_parallel_for(
+                cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
                     rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE * 2>(
                         B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
                         item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()

ggml/src/ggml-vulkan/CMakeLists.txt

@@ -99,6 +99,7 @@ if (Vulkan_FOUND)
 
     if (GGML_VULKAN_SHADER_DEBUG_INFO)
         add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+        list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -DGGML_VULKAN_SHADER_DEBUG_INFO=ON)
     endif()
 
     if (GGML_VULKAN_VALIDATE)
@@ -143,7 +144,8 @@ if (Vulkan_FOUND)
                    -DCMAKE_BUILD_TYPE=$<CONFIG>
                    ${VULKAN_SHADER_GEN_CMAKE_ARGS}
 
-        BUILD_COMMAND   ${CMAKE_COMMAND} --build   . --config $<CONFIG>
+        BUILD_COMMAND ${CMAKE_COMMAND} --build . --config $<CONFIG>
+        BUILD_ALWAYS  TRUE
 
         # NOTE: When DESTDIR is set using Makefile generators and
         # "make install" triggers the build step, vulkan-shaders-gen
@@ -164,6 +166,14 @@ if (Vulkan_FOUND)
 
     file(GLOB _ggml_vk_shader_files CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.comp")
 
+    # Because external projects do not provide source-level tracking,
+    # the vulkan-shaders-gen sources need to be explicitly added to
+    # ensure that changes will cascade into shader re-generation.
+
+    file(GLOB _ggml_vk_shaders_gen_sources
+              CONFIGURE_DEPENDS "${_ggml_vk_input_dir}/*.cpp"
+                                "${_ggml_vk_input_dir}/*.h")
+
     add_custom_command(
         OUTPUT ${_ggml_vk_header}
                ${_ggml_vk_source}
@@ -177,6 +187,7 @@ if (Vulkan_FOUND)
             --no-clean
 
         DEPENDS ${_ggml_vk_shader_files}
+                ${_ggml_vk_shaders_gen_sources}
                 vulkan-shaders-gen
 
         COMMENT "Generate vulkan shaders"

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

@@ -1041,6 +1041,14 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
 struct vk_instance_t {
     vk::Instance instance;
 
+    bool debug_utils_support = false;  // VK_EXT_debug_utils enabled
+    PFN_vkSetDebugUtilsObjectNameEXT pfn_vkSetDebugUtilsObjectNameEXT = {};
+    PFN_vkQueueBeginDebugUtilsLabelEXT pfn_vkQueueBeginDebugUtilsLabelEXT = {};
+    PFN_vkQueueEndDebugUtilsLabelEXT   pfn_vkQueueEndDebugUtilsLabelEXT   = {};
+    PFN_vkCmdBeginDebugUtilsLabelEXT   pfn_vkCmdBeginDebugUtilsLabelEXT   = {};
+    PFN_vkCmdEndDebugUtilsLabelEXT pfn_vkCmdEndDebugUtilsLabelEXT = {};
+    PFN_vkCmdInsertDebugUtilsLabelEXT  pfn_vkCmdInsertDebugUtilsLabelEXT  = {};
+
     std::vector<size_t> device_indices;
     vk_device devices[GGML_VK_MAX_DEVICES];
 };
@@ -1180,6 +1188,14 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
     }
     pipeline->compiled = true;
 
+    if (vk_instance.debug_utils_support) {
+        vk::DebugUtilsObjectNameInfoEXT duoni;
+        duoni.objectType = vk::ObjectType::ePipeline;
+        duoni.pObjectName = pipeline->name.c_str();
+        duoni.objectHandle = reinterpret_cast<uint64_t>(static_cast<VkPipeline_T*>(pipeline->pipeline));
+        vk_instance.pfn_vkSetDebugUtilsObjectNameEXT(device->device, &static_cast<VkDebugUtilsObjectNameInfoEXT &>(duoni));
+    }
+
     {
         std::lock_guard<std::mutex> guard(device->mutex);
         device->pipelines.insert({ pipeline->name, pipeline });
@@ -3561,6 +3577,8 @@ static void ggml_vk_print_gpu_info(size_t idx) {
 static bool ggml_vk_instance_validation_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
 static bool ggml_vk_instance_portability_enumeration_ext_available(const std::vector<vk::ExtensionProperties>& instance_extensions);
 
+static bool ggml_vk_instance_debug_utils_ext_available(const std::vector<vk::ExtensionProperties> & instance_extensions);
+
 static void ggml_vk_instance_init() {
     if (vk_instance_initialized) {
         return;
@@ -3581,7 +3599,7 @@ static void ggml_vk_instance_init() {
 #ifdef __APPLE__
     const bool portability_enumeration_ext = ggml_vk_instance_portability_enumeration_ext_available(instance_extensions);
 #endif
-
+    const bool debug_utils_ext = ggml_vk_instance_debug_utils_ext_available(instance_extensions) && getenv("GGML_VK_DEBUG_MARKERS") != nullptr;
     std::vector<const char*> layers;
 
     if (validation_ext) {
@@ -3596,6 +3614,9 @@ static void ggml_vk_instance_init() {
         extensions.push_back("VK_KHR_portability_enumeration");
     }
 #endif
+    if (debug_utils_ext) {
+        extensions.push_back("VK_EXT_debug_utils");
+    }
     vk::InstanceCreateInfo instance_create_info(vk::InstanceCreateFlags{}, &app_info, layers, extensions);
 #ifdef __APPLE__
     if (portability_enumeration_ext) {
@@ -3619,6 +3640,18 @@ static void ggml_vk_instance_init() {
     vk_instance.instance = vk::createInstance(instance_create_info);
     vk_instance_initialized = true;
 
+    if (debug_utils_ext) {
+        vk_instance.debug_utils_support              = true;
+        vk_instance.pfn_vkSetDebugUtilsObjectNameEXT = (PFN_vkSetDebugUtilsObjectNameEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkSetDebugUtilsObjectNameEXT");
+        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT = (PFN_vkQueueBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueBeginDebugUtilsLabelEXT");
+        vk_instance.pfn_vkQueueEndDebugUtilsLabelEXT = (PFN_vkQueueEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkQueueEndDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdBeginDebugUtilsLabelEXT = (PFN_vkCmdBeginDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdBeginDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdEndDebugUtilsLabelEXT =   (PFN_vkCmdEndDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdEndDebugUtilsLabelEXT");
+        vk_instance.pfn_vkCmdInsertDebugUtilsLabelEXT = (PFN_vkCmdInsertDebugUtilsLabelEXT) vkGetInstanceProcAddr(vk_instance.instance, "vkCmdInsertDebugUtilsLabelEXT");
+
+    }
+
+    size_t num_available_devices = vk_instance.instance.enumeratePhysicalDevices().size();
     vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
 
     // Emulate behavior of CUDA_VISIBLE_DEVICES for Vulkan
@@ -9495,6 +9528,12 @@ static size_t ggml_backend_vk_host_buffer_type_get_alignment(ggml_backend_buffer
     UNUSED(buft);
 }
 
+static size_t ggml_backend_vk_host_buffer_type_get_max_size(ggml_backend_buffer_type_t buft) {
+    return vk_instance.devices[0]->suballocation_block_size;
+
+    UNUSED(buft);
+}
+
 // Should be changed to return device-specific host buffer type
 // but that probably requires changes in llama.cpp
 ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
@@ -9503,7 +9542,7 @@ ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type() {
             /* .get_name         = */ ggml_backend_vk_host_buffer_type_name,
             /* .alloc_buffer     = */ ggml_backend_vk_host_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_vk_host_buffer_type_get_alignment,
-            /* .get_max_size     = */ NULL, // defaults to SIZE_MAX
+            /* .get_max_size     = */ ggml_backend_vk_host_buffer_type_get_max_size,
             /* .get_alloc_size   = */ ggml_backend_cpu_buffer_type()->iface.get_alloc_size,
             /* .is_host          = */ ggml_backend_cpu_buffer_type()->iface.is_host,
         },
@@ -9650,6 +9689,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
 
+    if (vk_instance.debug_utils_support) {
+        vk::DebugUtilsLabelEXT dul = {};
+        dul.pLabelName = "ggml_backend_vk_graph_compute";
+        dul.color = std::array<float,4>{1.0f, 1.0f, 1.0f, 1.0f};
+        vk_instance.pfn_vkQueueBeginDebugUtilsLabelEXT(ctx->device->compute_queue.queue, reinterpret_cast<VkDebugUtilsLabelEXT*>(&dul));
+    }
+
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
         ggml_vk_build_graph(ctx, cgraph->nodes[i], i, nullptr, 0, true, false, false, false);
@@ -10339,6 +10385,22 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
     UNUSED(instance_extensions);
 }
 
+// Extension availability
+static bool ggml_vk_instance_debug_utils_ext_available(
+    const std::vector<vk::ExtensionProperties> & instance_extensions) {
+    // Check for portability enumeration extension for MoltenVK support
+    for (const auto & properties : instance_extensions) {
+        if (strcmp("VK_EXT_debug_utils", properties.extensionName) == 0) {
+            return true;
+        }
+    }
+
+    std::cerr << "ggml_vulkan: WARNING: Instance extension VK_EXT_debug_utils not found." << std::endl;
+    return false;
+
+    UNUSED(instance_extensions);
+}
+
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
     switch (props.vendorID) {
     case VK_VENDOR_ID_INTEL:

ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt

@@ -19,6 +19,10 @@ if (GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     add_compile_definitions(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
     message(STATUS "Enabling bfloat16 glslc support")
 endif()
+if (GGML_VULKAN_SHADER_DEBUG_INFO)
+    add_compile_definitions(GGML_VULKAN_SHADER_DEBUG_INFO)
+    message(STATUS "Enabling shader debug info")
+endif()
 
 set(TARGET vulkan-shaders-gen)
 add_executable(${TARGET} vulkan-shaders-gen.cpp)

ggml/src/ggml.c

@@ -61,9 +61,6 @@
 #define m512i(p) (__m512i)(p)
 #endif
 
-// precomputed f32 table for f16 (256 KB) (ggml-impl.h)
-float ggml_table_f32_f16[1 << 16];
-
 #if defined(__linux__) || \
     defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
     (defined(__APPLE__) && !TARGET_OS_TV && !TARGET_OS_WATCH)
@@ -936,6 +933,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "TRANSPOSE",
     "GET_ROWS",
     "GET_ROWS_BACK",
+    "SET_ROWS",
     "DIAG",
     "DIAG_MASK_INF",
     "DIAG_MASK_ZERO",
@@ -955,6 +953,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "UPSCALE",
     "PAD",
     "PAD_REFLECT_1D",
+    "ROLL",
     "ARANGE",
     "TIMESTEP_EMBEDDING",
     "ARGSORT",
@@ -985,7 +984,7 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "OPT_STEP_ADAMW",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 84, "GGML_OP_COUNT != 84");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1031,6 +1030,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "transpose(x)",
     "get_rows(x)",
     "get_rows_back(x)",
+    "set_rows(x)",
     "diag(x)",
     "diag_mask_inf(x)",
     "diag_mask_zero(x)",
@@ -1050,6 +1050,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "upscale(x)",
     "pad(x)",
     "pad_reflect_1d(x)",
+    "roll(x)",
     "arange(start, stop, step)",
     "timestep_embedding(timesteps, dim, max_period)",
     "argsort(x)",
@@ -1080,7 +1081,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "adamw(x)",
 };
 
-static_assert(GGML_OP_COUNT == 82, "GGML_OP_COUNT != 82");
+static_assert(GGML_OP_COUNT == 84, "GGML_OP_COUNT != 84");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
@@ -1349,6 +1350,12 @@ bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor) {
         tensor->nb[2] == ggml_type_size(tensor->type);
 }
 
+bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor) {
+    return
+        tensor->ne[0] == ggml_blck_size(tensor->type) ||
+        tensor->nb[0] == ggml_type_size(tensor->type);
+}
+
 static inline bool ggml_is_padded_1d(const struct ggml_tensor * tensor) {
     static_assert(GGML_MAX_DIMS == 4, "GGML_MAX_DIMS is not 4 - update this function");
 
@@ -1420,14 +1427,6 @@ struct ggml_context * ggml_init(struct ggml_init_params params) {
         // initialize time system (required on Windows)
         ggml_time_init();
 
-        for (int i = 0; i < (1 << 16); ++i) {
-            union {
-                uint16_t u16;
-                ggml_fp16_t fp16;
-            } u = {i};
-            ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32(u.fp16);
-        }
-
         is_first_call = false;
     }
 
@@ -3393,6 +3392,35 @@ struct ggml_tensor * ggml_get_rows_back(
     return result;
 }
 
+// ggml_set_rows
+
+struct ggml_tensor * ggml_set_rows(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c) {
+    GGML_ASSERT(a->ne[0] == b->ne[0]);
+    GGML_ASSERT(a->ne[2] == b->ne[2]);
+    GGML_ASSERT(a->ne[3] == b->ne[3]);
+    GGML_ASSERT(b->ne[1] == c->ne[0]);
+    GGML_ASSERT(b->ne[2] % c->ne[1] == 0);
+    GGML_ASSERT(b->ne[3] % c->ne[2] == 0);
+    GGML_ASSERT(c->ne[3] == 1);
+    GGML_ASSERT(b->type == GGML_TYPE_F32);
+    GGML_ASSERT(c->type == GGML_TYPE_I64);
+
+    GGML_ASSERT(ggml_is_contiguous_rows(a));
+    GGML_ASSERT(ggml_is_contiguous_rows(b));
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_SET_ROWS;
+    result->src[0] = b;
+    result->src[1] = c;
+
+    return result;
+}
+
 // ggml_diag
 
 struct ggml_tensor * ggml_diag(
@@ -4341,6 +4369,34 @@ struct ggml_tensor * ggml_pad_reflect_1d(
     return result;
 }
 
+// ggml_roll
+
+struct ggml_tensor * ggml_roll(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   shift0,
+        int                   shift1,
+        int                   shift2,
+        int                   shift3) {
+    GGML_ASSERT(a->nb[0] == ggml_type_size(a->type));
+    GGML_ASSERT(abs(shift0) < a->ne[0]);
+    GGML_ASSERT(abs(shift1) < a->ne[1]);
+    GGML_ASSERT(abs(shift2) < a->ne[2]);
+    GGML_ASSERT(abs(shift3) < a->ne[3]);
+
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
+
+    ggml_set_op_params_i32(result, 0, shift0);
+    ggml_set_op_params_i32(result, 1, shift1);
+    ggml_set_op_params_i32(result, 2, shift2);
+    ggml_set_op_params_i32(result, 3, shift3);
+
+    result->op     = GGML_OP_ROLL;
+    result->src[0] = a;
+
+    return result;
+}
+
 // ggml_arange
 
 struct ggml_tensor * ggml_arange(

ggml/src/gguf.cpp

@@ -335,7 +335,11 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
 
         for (uint32_t i = 0; i < magic.size(); i++) {
             if (magic[i] != GGUF_MAGIC[i]) {
-                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, magic[0], magic[1], magic[2], magic[3]);
+                char c0 = isprint(magic[0]) ? magic[0] : '?';
+                char c1 = isprint(magic[1]) ? magic[1] : '?';
+                char c2 = isprint(magic[2]) ? magic[2] : '?';
+                char c3 = isprint(magic[3]) ? magic[3] : '?';
+                GGML_LOG_ERROR("%s: invalid magic characters: '%c%c%c%c', expected 'GGUF'\n", __func__, c0, c1, c2, c3);
                 gguf_free(ctx);
                 return nullptr;
             }

gguf-py/gguf/constants.py

@@ -118,6 +118,10 @@ class Keys:
         EMBEDDING_SCALE                   = "{arch}.embedding_scale"
         TOKEN_SHIFT_COUNT                 = "{arch}.token_shift_count"
         INTERLEAVE_MOE_LAYER_STEP         = "{arch}.interleave_moe_layer_step"
+        ACTIVATION_SPARSITY_SCALE         = "{arch}.activation_sparsity_scale"
+        ALTUP_ACTIVE_IDX                  = "{arch}.altup.active_idx"
+        ALTUP_NUM_INPUTS                  = "{arch}.altup.num_inputs"
+        EMBD_LENGTH_PER_LAYER_INP         = "{arch}.embedding_length_per_layer_input"
 
     class Attention:
         HEAD_COUNT                   = "{arch}.attention.head_count"
@@ -142,6 +146,8 @@ class Keys:
         SCALE                        = "{arch}.attention.scale"
         KEY_LENGTH_MLA               = "{arch}.attention.key_length_mla"
         VALUE_LENGTH_MLA             = "{arch}.attention.value_length_mla"
+        SHARED_KV_LAYERS             = "{arch}.attention.shared_kv_layers"
+        SLIDING_WINDOW_PATTERN       = "{arch}.attention.sliding_window_pattern"
 
     class Rope:
         DIMENSION_COUNT         = "{arch}.rope.dimension_count"
@@ -198,6 +204,7 @@ class Keys:
         MASK_ID              = "tokenizer.ggml.mask_token_id"
         ADD_BOS              = "tokenizer.ggml.add_bos_token"
         ADD_EOS              = "tokenizer.ggml.add_eos_token"
+        ADD_SEP              = "tokenizer.ggml.add_sep_token"
         ADD_PREFIX           = "tokenizer.ggml.add_space_prefix"
         REMOVE_EXTRA_WS      = "tokenizer.ggml.remove_extra_whitespaces"
         PRECOMPILED_CHARSMAP = "tokenizer.ggml.precompiled_charsmap"
@@ -313,6 +320,7 @@ class MODEL_ARCH(IntEnum):
     GEMMA            = auto()
     GEMMA2           = auto()
     GEMMA3           = auto()
+    GEMMA3N          = auto()
     STARCODER2       = auto()
     RWKV6            = auto()
     RWKV6QWEN2       = auto()
@@ -346,6 +354,7 @@ class MODEL_ARCH(IntEnum):
     BAILINGMOE       = auto()
     DOTS1            = auto()
     ARCEE            = auto()
+    ERNIE4_5         = auto()
 
 
 class VISION_PROJECTOR_TYPE(IntEnum):
@@ -398,6 +407,22 @@ class MODEL_TENSOR(IntEnum):
     ATTN_Q_NORM          = auto()
     ATTN_K_NORM          = auto()
     LAYER_OUT_NORM       = auto()
+    PER_LAYER_TOKEN_EMBD = auto() # gemma3n
+    PER_LAYER_MODEL_PROJ = auto() # gemma3n
+    PER_LAYER_INP_GATE   = auto() # gemma3n
+    PER_LAYER_PROJ       = auto() # gemma3n
+    PER_LAYER_PROJ_NORM  = auto() # gemma3n
+    PER_LAYER_POST_NORM  = auto() # gemma3n
+    ALTUP_PROJ           = auto() # gemma3n
+    ALTUP_UNEMBD_PROJ    = auto() # gemma3n
+    ALTUP_CORRECT_COEF   = auto() # gemma3n
+    ALTUP_CORRECT_SCALE  = auto() # gemma3n
+    ALTUP_PREDICT_COEF   = auto() # gemma3n
+    ALTUP_ROUTER         = auto() # gemma3n
+    ALTUP_ROUTER_NORM    = auto() # gemma3n
+    LAUREL_L             = auto() # gemma3n
+    LAUREL_R             = auto() # gemma3n
+    LAUREL_POST_NORM     = auto() # gemma3n
     SSM_IN               = auto()
     SSM_CONV1D           = auto()
     SSM_X                = auto()
@@ -596,6 +621,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.GEMMA:            "gemma",
     MODEL_ARCH.GEMMA2:           "gemma2",
     MODEL_ARCH.GEMMA3:           "gemma3",
+    MODEL_ARCH.GEMMA3N:          "gemma3n",
     MODEL_ARCH.STARCODER2:       "starcoder2",
     MODEL_ARCH.RWKV6:            "rwkv6",
     MODEL_ARCH.RWKV6QWEN2:       "rwkv6qwen2",
@@ -629,6 +655,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = {
     MODEL_ARCH.BAILINGMOE:       "bailingmoe",
     MODEL_ARCH.DOTS1:            "dots1",
     MODEL_ARCH.ARCEE:            "arcee",
+    MODEL_ARCH.ERNIE4_5:         "ernie4_5",
 }
 
 VISION_PROJECTOR_TYPE_NAMES: dict[VISION_PROJECTOR_TYPE, str] = {
@@ -681,6 +708,22 @@ TENSOR_NAMES: dict[MODEL_TENSOR, str] = {
     MODEL_TENSOR.FFN_UP_EXP:                "blk.{bid}.ffn_up_exps",
     MODEL_TENSOR.FFN_EXP_PROBS_B:           "blk.{bid}.exp_probs_b",
     MODEL_TENSOR.LAYER_OUT_NORM:            "blk.{bid}.layer_output_norm",
+    MODEL_TENSOR.PER_LAYER_TOKEN_EMBD:      "per_layer_token_embd",           # gemma3n
+    MODEL_TENSOR.PER_LAYER_MODEL_PROJ:      "per_layer_model_proj",           # gemma3n
+    MODEL_TENSOR.PER_LAYER_PROJ_NORM:       "per_layer_proj_norm",            # gemma3n
+    MODEL_TENSOR.ALTUP_UNEMBD_PROJ:         "altup_unembd_proj",              # gemma3n
+    MODEL_TENSOR.ALTUP_PROJ:                "altup_proj",                     # gemma3n
+    MODEL_TENSOR.PER_LAYER_INP_GATE:        "blk.{bid}.inp_gate",             # gemma3n
+    MODEL_TENSOR.PER_LAYER_PROJ:            "blk.{bid}.proj",                 # gemma3n
+    MODEL_TENSOR.PER_LAYER_POST_NORM:       "blk.{bid}.post_norm",            # gemma3n
+    MODEL_TENSOR.ALTUP_CORRECT_COEF:        "blk.{bid}.altup_correct_coef",   # gemma3n
+    MODEL_TENSOR.ALTUP_CORRECT_SCALE:       "blk.{bid}.altup_correct_scale",  # gemma3n
+    MODEL_TENSOR.ALTUP_PREDICT_COEF:        "blk.{bid}.altup_predict_coef",   # gemma3n
+    MODEL_TENSOR.ALTUP_ROUTER:              "blk.{bid}.altup_router",         # gemma3n
+    MODEL_TENSOR.ALTUP_ROUTER_NORM:         "blk.{bid}.altup_router_norm",    # gemma3n
+    MODEL_TENSOR.LAUREL_L:                  "blk.{bid}.laurel_l",             # gemma3n
+    MODEL_TENSOR.LAUREL_R:                  "blk.{bid}.laurel_r",             # gemma3n
+    MODEL_TENSOR.LAUREL_POST_NORM:          "blk.{bid}.laurel_post_norm",     # gemma3n
     MODEL_TENSOR.SSM_IN:                    "blk.{bid}.ssm_in",
     MODEL_TENSOR.SSM_CONV1D:                "blk.{bid}.ssm_conv1d",
     MODEL_TENSOR.SSM_X:                     "blk.{bid}.ssm_x",
@@ -1485,6 +1528,41 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_PRE_NORM,
         MODEL_TENSOR.FFN_POST_NORM,
     ],
+    MODEL_ARCH.GEMMA3N: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_Q_NORM,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_K_NORM,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_POST_NORM,
+        MODEL_TENSOR.FFN_PRE_NORM,
+        MODEL_TENSOR.FFN_POST_NORM,
+        # altup / laurel
+        MODEL_TENSOR.PER_LAYER_TOKEN_EMBD,
+        MODEL_TENSOR.PER_LAYER_MODEL_PROJ,
+        MODEL_TENSOR.PER_LAYER_INP_GATE,
+        MODEL_TENSOR.PER_LAYER_PROJ,
+        MODEL_TENSOR.PER_LAYER_PROJ_NORM,
+        MODEL_TENSOR.PER_LAYER_POST_NORM,
+        MODEL_TENSOR.ALTUP_PROJ,
+        MODEL_TENSOR.ALTUP_UNEMBD_PROJ,
+        MODEL_TENSOR.ALTUP_CORRECT_COEF,
+        MODEL_TENSOR.ALTUP_CORRECT_SCALE,
+        MODEL_TENSOR.ALTUP_PREDICT_COEF,
+        MODEL_TENSOR.ALTUP_ROUTER,
+        MODEL_TENSOR.ALTUP_ROUTER_NORM,
+        MODEL_TENSOR.LAUREL_L,
+        MODEL_TENSOR.LAUREL_R,
+        MODEL_TENSOR.LAUREL_POST_NORM,
+    ],
     MODEL_ARCH.STARCODER2: [
         MODEL_TENSOR.TOKEN_EMBD,
         MODEL_TENSOR.OUTPUT_NORM,
@@ -2101,6 +2179,20 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
         MODEL_TENSOR.FFN_DOWN,
         MODEL_TENSOR.FFN_UP,
     ],
+    MODEL_ARCH.ERNIE4_5: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_Q,
+        MODEL_TENSOR.ATTN_K,
+        MODEL_TENSOR.ATTN_V,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_NORM,
+        MODEL_TENSOR.FFN_GATE,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ],
     # TODO
 }
 

gguf-py/gguf/gguf_writer.py

@@ -672,6 +672,18 @@ class GGUFWriter:
     def add_decoder_start_token_id(self, id: int) -> None:
         self.add_uint32(Keys.LLM.DECODER_START_TOKEN_ID.format(arch=self.arch), id)
 
+    def add_embedding_length_per_layer_input(self, value: int) -> None:
+        self.add_uint32(Keys.LLM.EMBD_LENGTH_PER_LAYER_INP.format(arch=self.arch), value)
+
+    def add_altup_active_idx(self, val: int) -> None:
+        self.add_uint32(Keys.LLM.ALTUP_ACTIVE_IDX.format(arch=self.arch), val)
+
+    def add_altup_num_inputs(self, val: int) -> None:
+        self.add_uint32(Keys.LLM.ALTUP_NUM_INPUTS.format(arch=self.arch), val)
+
+    def add_activation_sparsity_scale(self, values: Sequence[float]) -> None:
+        self.add_array(Keys.LLM.ACTIVATION_SPARSITY_SCALE.format(arch=self.arch), values)
+
     def add_head_count(self, count: int | Sequence[int]) -> None:
         if isinstance(count, int):
             self.add_uint32(Keys.Attention.HEAD_COUNT.format(arch=self.arch), count)
@@ -702,6 +714,12 @@ class GGUFWriter:
     def add_clamp_kqv(self, value: float) -> None:
         self.add_float32(Keys.Attention.CLAMP_KQV.format(arch=self.arch), value)
 
+    def add_shared_kv_layers(self, value: float) -> None:
+        self.add_float32(Keys.Attention.SHARED_KV_LAYERS.format(arch=self.arch), value)
+
+    def add_sliding_window_pattern(self, value: Sequence[bool]) -> None:
+        self.add_array(Keys.Attention.SLIDING_WINDOW_PATTERN.format(arch=self.arch), value)
+
     def add_logit_scale(self, value: float) -> None:
         self.add_float32(Keys.LLM.LOGIT_SCALE.format(arch=self.arch), value)
 
@@ -891,6 +909,9 @@ class GGUFWriter:
     def add_add_eos_token(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_EOS, value)
 
+    def add_add_sep_token(self, value: bool) -> None:
+        self.add_bool(Keys.Tokenizer.ADD_SEP, value)
+
     def add_add_space_prefix(self, value: bool) -> None:
         self.add_bool(Keys.Tokenizer.ADD_PREFIX, value)
 

gguf-py/gguf/tensor_mapping.py

@@ -480,6 +480,70 @@ class TensorNameMap:
             "encoder.layer.{bid}.layer_norm_2"              # jina-v2-code
         ),
 
+        MODEL_TENSOR.PER_LAYER_TOKEN_EMBD: (
+            "model.embed_tokens_per_layer",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_MODEL_PROJ: (
+            "model.per_layer_model_projection",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_PROJ_NORM: (
+            "model.per_layer_projection_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_PROJ: (
+            "model.altup_projections",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_UNEMBD_PROJ: (
+            "model.altup_unembed_projections",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_INP_GATE: (
+            "model.layers.{bid}.per_layer_input_gate",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_PROJ: (
+            "model.layers.{bid}.per_layer_projection",  # gemma3n
+        ),
+
+        MODEL_TENSOR.PER_LAYER_POST_NORM: (
+            "model.layers.{bid}.post_per_layer_input_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_CORRECT_COEF: (
+            "model.layers.{bid}.altup.correction_coefs",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_CORRECT_SCALE: (
+            "model.layers.{bid}.altup.correct_output_scale",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_PREDICT_COEF: (
+            "model.layers.{bid}.altup.prediction_coefs",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_ROUTER: (
+            "model.layers.{bid}.altup.modality_router",  # gemma3n
+        ),
+
+        MODEL_TENSOR.ALTUP_ROUTER_NORM: (
+            "model.layers.{bid}.altup.router_norm",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_L: (
+            "model.layers.{bid}.laurel.linear_left",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_R: (
+            "model.layers.{bid}.laurel.linear_right",  # gemma3n
+        ),
+
+        MODEL_TENSOR.LAUREL_POST_NORM: (
+            "model.layers.{bid}.laurel.post_laurel_norm",  # gemma3n
+        ),
+
         MODEL_TENSOR.SSM_IN: (
             "model.layers.{bid}.in_proj",
             "backbone.layers.{bid}.mixer.in_proj",

gguf-py/gguf/vocab.py

@@ -7,7 +7,10 @@ import os
 from pathlib import Path
 from typing import Any, Callable, Sequence, Mapping, Iterable, Protocol, ClassVar, runtime_checkable
 
-from sentencepiece import SentencePieceProcessor
+try:
+    from sentencepiece import SentencePieceProcessor
+except ImportError:
+    SentencePieceProcessor = None
 
 import gguf
 
@@ -116,6 +119,7 @@ class SpecialVocab:
         logger.warning(f'Special token type {typ}, id {tid} out of range, must be under {self.n_vocab} - skipping')
 
     def _try_load_from_tokenizer_json(self, path: Path) -> bool:
+        tokenizer = None
         tokenizer_file = path / 'tokenizer.json'
         if tokenizer_file.is_file():
             with open(tokenizer_file, encoding = 'utf-8') as f:
@@ -149,11 +153,97 @@ class SpecialVocab:
             added_tokens = tokenizer.get('added_tokens', {})
         else:
             added_tokens = {}
+        tokenizer_config = None
         tokenizer_config_file = path / 'tokenizer_config.json'
-        if not tokenizer_config_file.is_file():
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, encoding = 'utf-8') as f:
+                tokenizer_config = json.load(f)
+        if tokenizer:
+            special_bos = (tokenizer_config or {}).get('bos_token')
+            special_cls = (tokenizer_config or {}).get('cls_token')
+            special_eos = (tokenizer_config or {}).get('eos_token')
+            special_sep = (tokenizer_config or {}).get('sep_token')
+            if not special_bos and special_cls and tokenizer_config:
+                tokenizer_config['bos_token'] = special_bos = special_cls
+            if not special_eos and special_sep and tokenizer_config:
+                tokenizer_config['eos_token'] = special_eos = special_sep
+            if post_processor := tokenizer.get('post_processor'):
+                for processor in post_processor.get('processors', [post_processor]):
+                    if processor.get('type') == 'RobertaProcessing':
+                        self.add_special_token['bos'] = True
+                        self.add_special_token['eos'] = True
+                        self.add_special_token['sep'] = True
+                        if not special_cls and tokenizer_config:
+                            special_cls = processor.get('cls', [special_bos])[0]
+                            tokenizer_config['cls_token'] = special_cls
+                        if not special_sep and tokenizer_config:
+                            special_sep = processor.get('sep', [special_eos])[0]
+                            tokenizer_config['sep_token'] = special_sep
+                        continue
+                    # Crude parsing of TemplateProcessing to determine if BOS/SEP/EOS should be added
+                    # Only works with simple templates, **will** get it wrong on unusual sequences
+                    if processor.get('type') == 'TemplateProcessing':
+                        tmpl_single = processor.get('single', [])
+                        tmpl_pair = processor.get('pair', [])
+                        special_first = None
+                        special_last = None
+                        if len(tmpl_single) > 1:
+                            if special_first := tmpl_single[0].get('SpecialToken', {}).get('id'):
+                                if not tokenizer_config:
+                                    special_bos = special_first
+                                self.add_special_token['bos'] = True if special_first in (special_bos, special_cls) else False
+                                if special_first not in (special_bos, special_cls):
+                                    logger.warning(f'Unknown leading special token {special_first!r} in TemplateProcessing<single>')
+                            if special_last := tmpl_single[-1].get('SpecialToken', {}).get('id'):
+                                if not tokenizer_config:
+                                    special_eos = special_last
+                                elif special_last != special_eos:
+                                    if 'eot' not in self.special_token_types:
+                                        self.special_token_types = tuple(self.special_token_types) + ('eot', )
+                                        tokenizer_config['eot_token'] = special_eos
+                                    elif 'eom' not in self.special_token_types:
+                                        self.special_token_types = tuple(self.special_token_types) + ('eom', )
+                                        tokenizer_config['eom_token'] = special_eos
+                                    else:
+                                        logger.warning(f'Overriding EOS token {special_eos!r} with {special_last!r} without EOT/EOM fallback!')
+                                    tokenizer_config['eos_token'] = special_eos = special_last
+                                self.add_special_token['eos'] = True if special_last == special_eos else False
+                                if special_last != special_eos:
+                                    logger.warning(f'Unknown trailing special token {special_last!r} in TemplateProcessing<single>')
+                        if tmpl_pair:
+                            seq_start = 1 if special_first and tmpl_pair[0].get('SpecialToken', {}).get('id') == special_first else 0
+                            seq_stop = -1 if special_last and tmpl_pair[-1].get('SpecialToken', {}).get('id') == special_last else None
+                            if (special_first and seq_start == 0) or (special_last and seq_stop is None):
+                                logger.warning('TemplateProcessing<single> leading/trailing special tokens do not match TemplateProcessing<pair>')
+                            if tmpl_pair := tmpl_pair[slice(seq_start, seq_stop)]:
+                                tmpl_a = tmpl_pair[0].get('Sequence', {}).get('id')
+                                tmpl_b = tmpl_pair[-1].get('Sequence', {}).get('id')
+                                if tmpl_a != 'A' or tmpl_b != 'B':
+                                    logger.warning(f'Unknown sequence {tmpl_a}...{tmpl_b} in TemplateProcessing<pair>')
+                                # A [sep] [eos] B
+                                if tmpl_a == 'A' and tmpl_b == 'B' and (tmpl_pair := tmpl_pair[1:-1]):
+                                    add_sep = False
+                                    if special_entry := tmpl_pair[0].get('SpecialToken', {}).get('id'):
+                                        if special_entry in (special_sep, special_eos) and not special_last:
+                                            add_sep = True
+                                        if special_entry not in (special_sep, special_eos):
+                                            logger.warning(f'Unknown separator token {special_entry!r} in TemplateProcessing<pair>')
+                                    else:
+                                        logger.warning(f'Unknown middle sequence {tmpl_pair[0]!r} in TemplateProcessing<pair>')
+                                    if len(tmpl_pair) == 2:
+                                        if special_entry := tmpl_pair[1].get('SpecialToken', {}).get('id'):
+                                            if special_entry in (special_sep, special_eos):
+                                                add_sep = True
+                                            if special_entry not in (special_sep, special_eos):
+                                                logger.warning(f'Unknown second separator token {special_entry!r} in TemplateProcessing<pair>')
+                                        else:
+                                            logger.warning(f'Unknown second middle sequence {tmpl_pair[1]!r} in TemplateProcessing<pair>')
+                                    self.add_special_token['sep'] = add_sep
+                                    if add_sep and not special_sep and tokenizer_config:
+                                        tokenizer_config['sep_token'] = special_eos
+                        continue
+        if not tokenizer_config:
             return True
-        with open(tokenizer_config_file, encoding = 'utf-8') as f:
-            tokenizer_config = json.load(f)
         chat_template_alt = None
         chat_template_file = path / 'chat_template.json'
         if chat_template_file.is_file():
@@ -302,6 +392,9 @@ class SentencePieceVocab(Vocab):
     name = "spm"
 
     def __init__(self, base_path: Path):
+        if SentencePieceProcessor is None:
+            raise RuntimeError("sentencepiece is not installed")
+
         added_tokens: dict[str, int] = {}
         if (fname_tokenizer := base_path / 'tokenizer.model').exists():
             # normal location

gguf-py/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.17.0"
+version = "0.17.1"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
@@ -22,7 +22,7 @@ python = ">=3.8"
 numpy = ">=1.17"
 tqdm = ">=4.27"
 pyyaml = ">=5.1"
-sentencepiece = ">=0.1.98,<=0.2.0"
+sentencepiece = { version = ">=0.1.98,<=0.2.0", optional = true }
 PySide6 = { version = "^6.9", python = ">=3.9,<3.14", optional = true }
 
 [tool.poetry.dev-dependencies]

include/llama.h

@@ -390,6 +390,7 @@ extern "C" {
         void * imatrix;                       // pointer to importance matrix data
         void * kv_overrides;                  // pointer to vector containing overrides
         void * tensor_types;                  // pointer to vector containing tensor types
+        void * prune_layers;                  // pointer to vector containing layer indices to prune
     } llama_model_quantize_params;
 
     typedef struct llama_logit_bias {
@@ -943,12 +944,14 @@ extern "C" {
     // Requires the context to have a memory.
     // For encode-decoder contexts, processes the batch using the decoder.
     // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the memory state is restored to the state before this call
+    // Upon fatal-error or abort, the ubatches that managed to be been processed will remain in the memory state of the context
+    //   To handle this correctly, query the memory state using llama_memory_seq_pos_min() and llama_memory_seq_pos_max()
+    // Upon other return values, the memory state is restored to the state before this call
     //    0 - success
     //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
-    //    2 - aborted
+    //    2 - aborted     (processed ubatches will remain in the context's memory)
     //   -1 - invalid input batch
-    // < -1 - error
+    // < -1 - fatal error (processed ubatches will remain in the context's memory)
     LLAMA_API int32_t llama_decode(
             struct llama_context * ctx,
               struct llama_batch   batch);
@@ -1044,6 +1047,7 @@ extern "C" {
 
     LLAMA_API bool llama_vocab_get_add_bos(const struct llama_vocab * vocab);
     LLAMA_API bool llama_vocab_get_add_eos(const struct llama_vocab * vocab);
+    LLAMA_API bool llama_vocab_get_add_sep(const struct llama_vocab * vocab);
 
     LLAMA_API llama_token llama_vocab_fim_pre(const struct llama_vocab * vocab);
     LLAMA_API llama_token llama_vocab_fim_suf(const struct llama_vocab * vocab);
@@ -1087,6 +1091,7 @@ extern "C" {
     /// @param tokens The tokens pointer must be large enough to hold the resulting tokens.
     /// @return Returns the number of tokens on success, no more than n_tokens_max
     /// @return Returns a negative number on failure - the number of tokens that would have been returned
+    /// @return Returns INT32_MIN on overflow (e.g., tokenization result size exceeds int32_t limit)
     /// @param add_special Allow to add BOS and EOS tokens if model is configured to do so.
     /// @param parse_special Allow tokenizing special and/or control tokens which otherwise are not exposed and treated
     ///                      as plaintext. Does not insert a leading space.