TP: fix 0-sized tensor slices, AllReduce fallback (#21808)

* TP: fix 0-sized tensor slices, AllReduce fallback

* fix layer structure <-> GPU count aliasing

* add missing std::fill

* fix CUDA device set, max ggml ctx size

.devops/nix/package.nix

@@ -18,6 +18,7 @@
   vulkan-loader,
   openssl,
   shaderc,
+  spirv-headers,
   useBlas ?
     builtins.all (x: !x) [
       useCuda
@@ -145,6 +146,7 @@ effectiveStdenv.mkDerivation (finalAttrs: {
       ninja
       pkg-config
       git
+      spirv-headers
     ]
     ++ optionals useCuda [
       cudaPackages.cuda_nvcc

.github/workflows/build-android.yml

@@ -51,7 +51,7 @@ jobs:
           distribution: zulu
 
       - name: Setup Android SDK
-        uses: android-actions/setup-android@9fc6c4e9069bf8d3d10b2204b1fb8f6ef7065407 # v3
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
         with:
           log-accepted-android-sdk-licenses: false
 

.github/workflows/build-cross.yml

@@ -246,6 +246,7 @@ jobs:
           apt-get install -y --no-install-recommends \
                   build-essential \
                   glslc \
+                  spirv-headers \
                   gcc-14-loongarch64-linux-gnu \
                   g++-14-loongarch64-linux-gnu \
                   libvulkan-dev:loong64

.github/workflows/build-riscv.yml

@@ -47,22 +47,10 @@ jobs:
     steps:
       - name: Install dependencies
         run: |
-          sudo apt-get update
-
-          # Install necessary packages
-          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 cmake build-essential wget git-lfs
-
           # Set gcc-14 and g++-14 as the default compilers
           sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
           sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
 
-          if ! which rustc; then
-            # Install Rust stable version
-            sudo apt-get install -y rustup
-            rustup install stable
-            rustup default stable
-          fi
-
           git lfs install
 
       - name: GCC version check
@@ -74,12 +62,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # FIXME: Enable when ggml-org/ccache-action works on riscv64
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   with:
-      #     key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanytizer }}-${{ matrix.build_type }}
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+        with:
+          key: ubuntu-riscv64-native-sanitizer-${{ matrix.sanitizer }}-${{ matrix.build_type }}
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build

.github/workflows/build-self-hosted.yml

@@ -97,6 +97,36 @@ jobs:
           vulkaninfo --summary
           GG_BUILD_VULKAN=1 bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
 
+  # TODO: investigate slight precision issues in some operations for test-backend-ops on the WebGPU backend.
+  #ggml-ci-nvidia-webgpu:
+  #  runs-on: [self-hosted, Linux, NVIDIA]
+
+  #  steps:
+  #    - name: Clone
+  #      id: checkout
+  #      uses: actions/checkout@v6
+
+  #    - name: Dawn Dependency
+  #      id: dawn-depends
+  #      run: |
+  #        DAWN_VERSION="v20260317.182325"
+  #        DAWN_OWNER="google"
+  #        DAWN_REPO="dawn"
+  #        DAWN_ASSET_NAME="Dawn-18eb229ef5f707c1464cc581252e7603c73a3ef0-ubuntu-latest-Release"
+  #        echo "Fetching release asset from https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+  #        curl -L -o artifact.tar.gz \
+  #          "https://github.com/google/dawn/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.tar.gz"
+  #        mkdir dawn
+  #        tar -xvf artifact.tar.gz -C dawn --strip-components=1
+
+  #    - name: Test
+  #      id: ggml-ci
+  #      run: |
+  #        GG_BUILD_WEBGPU=1 \
+  #        GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+  #        GG_BUILD_WEBGPU_DAWN_DIR="$GITHUB_WORKSPACE/dawn/lib64/cmake/Dawn" \
+  #          bash ./ci/run.sh ~/results/llama.cpp /mnt/llama.cpp
+
   # TODO: provision AMX-compatible machine
   #ggml-ci-cpu-amx:
   #  runs-on: [self-hosted, Linux, CPU, AMX]

.github/workflows/build-vulkan.yml

@@ -93,4 +93,5 @@ jobs:
           export GGML_VK_DISABLE_F16=1
           export GGML_VK_DISABLE_COOPMAT=1
           # This is using llvmpipe and runs slower than other backends
-          ctest -L main --verbose --timeout 4800
+          # test-backend-ops is too slow on llvmpipe, skip it
+          ctest -L main -E test-backend-ops --verbose --timeout 900

.github/workflows/build.yml

@@ -267,6 +267,56 @@ jobs:
           wget https://huggingface.co/ggml-org/models/resolve/main/tinyllamas/stories260K-be.gguf
           ./bin/llama-completion -m stories260K-be.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
+  android-arm64:
+    runs-on: ubuntu-latest
+
+    env:
+      NDK_VERSION: "29.0.14206865"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: android-arm64
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: temurin
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Install NDK
+        run: |
+          sdkmanager "ndk;${{ env.NDK_VERSION }}"
+          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            -DGGML_RPC=ON
+          time cmake --build build --config Release -j $(nproc)
+
   ubuntu-latest-rpc:
     runs-on: ubuntu-latest
 
@@ -1001,22 +1051,14 @@ jobs:
     steps:
       - name: Install dependencies
         run: |
-          sudo apt-get update
-
           # Install necessary packages
-          sudo apt-get install -y libatomic1 libtsan2 gcc-14 g++-14 cmake build-essential libssl-dev wget git-lfs
+          sudo apt-get update
+          sudo apt-get install -y libssl-dev
 
           # Set gcc-14 and g++-14 as the default compilers
           sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-14 100
           sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-14 100
 
-          if ! which rustc; then
-            # Install Rust stable version
-            sudo apt-get install -y rustup
-            rustup install stable
-            rustup default stable
-          fi
-
           git lfs install
 
       - name: Check environment
@@ -1032,13 +1074,12 @@ jobs:
         id: checkout
         uses: actions/checkout@v6
 
-      # FIXME: Enable when ggml-org/ccache-action works on riscv64
-      # - name: ccache
-      #   uses: ggml-org/ccache-action@v1.2.21
-      #   with:
-      #     key: ubuntu-cpu-riscv64-native
-      #     evict-old-files: 1d
-      #     save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+      - name: ccache
+        uses: ggml-org/ccache-action@afde29e5b5422e5da23cb1f639e8baecadeadfc3 # https://github.com/ggml-org/ccache-action/pull/1
+        with:
+          key: ubuntu-cpu-riscv64-native
+          evict-old-files: 1d
+          save: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
 
       - name: Build
         id: cmake_build

.github/workflows/release.yml

@@ -236,6 +236,75 @@ jobs:
           path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
           name: llama-bin-ubuntu-vulkan-${{ matrix.build }}.tar.gz
 
+  android-arm64:
+    runs-on: ubuntu-latest
+
+    env:
+      NDK_VERSION: "29.0.14206865"
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+        with:
+          fetch-depth: 0
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.21
+        with:
+          key: android-arm64
+          evict-old-files: 1d
+
+      - name: Set up JDK
+        uses: actions/setup-java@v5
+        with:
+          java-version: 17
+          distribution: temurin
+
+      - name: Setup Android SDK
+        uses: android-actions/setup-android@40fd30fb8d7440372e1316f5d1809ec01dcd3699 # v4.0.1
+        with:
+          log-accepted-android-sdk-licenses: false
+
+      - name: Install NDK
+        run: |
+          sdkmanager "ndk;${{ env.NDK_VERSION }}"
+          echo "ANDROID_NDK=${ANDROID_SDK_ROOT}/ndk/${{ env.NDK_VERSION }}" >> $GITHUB_ENV
+
+      - name: Build
+        id: cmake_build
+        run: |
+          cmake -B build \
+            -DCMAKE_TOOLCHAIN_FILE=${ANDROID_NDK}/build/cmake/android.toolchain.cmake \
+            -DANDROID_ABI=arm64-v8a \
+            -DANDROID_PLATFORM=android-28 \
+            -DCMAKE_INSTALL_RPATH='$ORIGIN' \
+            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DGGML_OPENMP=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
+            ${{ env.CMAKE_ARGS }}
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Determine tag name
+        id: tag
+        uses: ./.github/actions/get-tag-name
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          cp LICENSE ./build/bin/
+          tar -czvf llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz --transform "s,./,llama-${{ steps.tag.outputs.name }}/," -C ./build/bin .
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v6
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz
+          name: llama-bin-android-arm64.tar.gz
+
   ubuntu-24-openvino:
     runs-on: ubuntu-24.04
 
@@ -618,6 +687,11 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Free up disk space
+        uses: ggml-org/free-disk-space@v1.3.1
+        with:
+          tool-cache: true
+
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.21
         with:
@@ -971,6 +1045,7 @@ jobs:
       - ubuntu-cpu
       - ubuntu-vulkan
       - ubuntu-24-openvino
+      - android-arm64
       - macOS-cpu
       - ios-xcode-build
       - openEuler-cann
@@ -1059,6 +1134,9 @@ jobs:
             - [Ubuntu x64 (ROCm 7.2)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.2-x64.tar.gz)
             - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
 
+            **Android:**
+            - [Android arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-android-arm64.tar.gz)
+
             **Windows:**
             - [Windows x64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-x64.zip)
             - [Windows arm64 (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-cpu-arm64.zip)

CMakeLists.txt

@@ -225,7 +225,7 @@ foreach(FILE_PATH ${EXTRA_LICENSES})
 endforeach()
 
 if (LLAMA_BUILD_COMMON)
-    license_generate(common)
+    license_generate(llama-common)
 endif()
 
 #
@@ -249,6 +249,10 @@ set_target_properties(llama
 
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
 
+if (LLAMA_BUILD_COMMON)
+    install(TARGETS llama-common LIBRARY)
+endif()
+
 configure_package_config_file(
         ${CMAKE_CURRENT_SOURCE_DIR}/cmake/llama-config.cmake.in
         ${CMAKE_CURRENT_BINARY_DIR}/llama-config.cmake

CODEOWNERS

@@ -1,5 +1,21 @@
 # collaborators can optionally add themselves here to indicate their availability for reviewing related PRs
-# multiplie collaborators per item can be specified
+# multiple collaborators per item can be specified
+#
+# ggml-org/ci               : CISC, danbev, ggerganov, netrunnereve, ngxson, taronaeo
+# ggml-org/ggml-cann        : hipudding
+# ggml-org/ggml-cuda        : JohannesGaessler, am17an, IMbackK, ORippler
+# ggml-org/ggml-hexagon     : lhez, max-krasnyansky
+# ggml-org/ggml-metal       : ggerganov
+# ggml-org/ggml-opencl      : lhez, max-krasnyansky
+# ggml-org/ggml-rpc         : rgerganov
+# ggml-org/ggml-sycl        : arthw
+# ggml-org/ggml-vulkan      : 0cc4m, jeffbolznv
+# ggml-org/ggml-webgpu      : reeselevine
+# ggml-org/ggml-zdnn        : taronaeo
+# ggml-org/llama-common     : ggerganov, aldehir, angt, danbev, ngxson, pwilkin
+# ggml-org/llama-mtmd       : ngxson
+# ggml-org/llama-server     : ggerganov, ngxson, allozaur, angt, ServeurpersoCom
+# ggml-org/llama-webui      : allozaur
 
 /.devops/*.Dockerfile                   @ngxson
 /.github/actions/                       @ggml-org/ci

common/CMakeLists.txt

@@ -1,9 +1,11 @@
-# common
-
 find_package(Threads REQUIRED)
 
 llama_add_compile_flags()
 
+#
+# llama-common-base
+#
+
 # Build info header
 
 if(EXISTS "${PROJECT_SOURCE_DIR}/.git")
@@ -33,17 +35,25 @@ endif()
 
 set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in")
 set(OUTPUT_FILE   "${CMAKE_CURRENT_BINARY_DIR}/build-info.cpp")
+
 configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
 
-set(TARGET build_info)
-add_library(${TARGET} OBJECT ${OUTPUT_FILE})
+set(TARGET llama-common-base)
+add_library(${TARGET} STATIC ${OUTPUT_FILE})
+
+target_include_directories(${TARGET} PUBLIC .)
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-set(TARGET common)
+#
+# llama-common
+#
 
-add_library(${TARGET} STATIC
+set(TARGET llama-common)
+
+add_library(${TARGET}
     arg.cpp
     arg.h
     base64.hpp
@@ -106,17 +116,24 @@ add_library(${TARGET} STATIC
     jinja/caps.h
     )
 
+set_target_properties(${TARGET} PROPERTIES
+    VERSION ${LLAMA_INSTALL_VERSION}
+    SOVERSION 0
+    MACHO_CURRENT_VERSION 0 # keep macOS linker from seeing oversized version number
+)
+
 target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+
+    # TODO: make fine-grained exports in the future
+    set_target_properties(${TARGET} PROPERTIES WINDOWS_EXPORT_ALL_SYMBOLS ON)
 endif()
 
-target_link_libraries(${TARGET} PRIVATE
-    build_info
-    cpp-httplib
-)
+target_link_libraries(${TARGET} PUBLIC  llama-common-base)
+target_link_libraries(${TARGET} PRIVATE cpp-httplib)
 
 if (LLAMA_LLGUIDANCE)
     include(ExternalProject)

common/arg.cpp

@@ -1,5 +1,6 @@
 #include "arg.h"
 
+#include "build-info.h"
 #include "chat.h"
 #include "common.h"
 #include "download.h"
@@ -291,7 +292,7 @@ static bool common_params_handle_remote_preset(common_params & params, llama_exa
         hf_tag = "default";
     }
 
-    std::string model_endpoint = get_model_endpoint();
+    std::string model_endpoint = common_get_model_endpoint();
     auto preset_url = model_endpoint + hf_repo + "/resolve/main/preset.ini";
 
     // prepare local path for caching
@@ -1044,8 +1045,8 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--version"},
         "show version and build info",
         [](common_params &) {
-            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            fprintf(stderr, "version: %d (%s)\n", llama_build_number(), llama_commit());
+            fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
             exit(0);
         }
     ));
@@ -1315,13 +1316,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED, LLAMA_EXAMPLE_BENCH, LLAMA_EXAMPLE_PARALLEL}));
     add_opt(common_arg(
-        {"--clear-idle"},
-        {"--no-clear-idle"},
+        {"--cache-idle-slots"},
+        {"--no-cache-idle-slots"},
         "save and clear idle slots on new task (default: enabled, requires unified KV and cache-ram)",
         [](common_params & params, bool value) {
-            params.clear_idle = value;
+            params.cache_idle_slots = value;
         }
-    ).set_env("LLAMA_ARG_CLEAR_IDLE").set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_env("LLAMA_ARG_CACHE_IDLE_SLOTS").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},

common/build-info.cpp.in

@@ -1,4 +1,35 @@
+#include "build-info.h"
+
+#include <cstdio>
+#include <string>
+
 int LLAMA_BUILD_NUMBER = @LLAMA_BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
-char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
-char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
+char const * LLAMA_COMMIT = "@LLAMA_BUILD_COMMIT@";
+char const * LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const * LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
+
+int llama_build_number(void) {
+    return LLAMA_BUILD_NUMBER;
+}
+
+const char * llama_commit(void) {
+    return LLAMA_COMMIT;
+}
+
+const char * llama_compiler(void) {
+    return LLAMA_COMPILER;
+}
+
+const char * llama_build_target(void) {
+    return LLAMA_BUILD_TARGET;
+}
+
+const char * llama_build_info(void) {
+    static std::string s = "b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT;
+    return s.c_str();
+}
+
+void llama_print_build_info(void) {
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, llama_build_number(), llama_commit());
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, llama_compiler(), llama_build_target());
+}

common/build-info.h

@@ -0,0 +1,11 @@
+#pragma once
+
+int llama_build_number(void);
+
+const char * llama_commit(void);
+const char * llama_compiler(void);
+
+const char * llama_build_target(void);
+const char * llama_build_info(void);
+
+void llama_print_build_info(void);

common/chat-auto-parser-generator.cpp

@@ -443,14 +443,14 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
     if (!format.per_call_start.empty()) {
         auto wrapped_call = format.per_call_start + p.space() + tool_choice + p.space() + format.per_call_end;
         if (inputs.parallel_tool_calls) {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call));
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.zero_or_more(p.space() + wrapped_call) + p.space());
         } else {
-            tool_calls = p.trigger_rule("tool-call", wrapped_call);
+            tool_calls = p.trigger_rule("tool-call", wrapped_call + p.space());
         }
         if (!format.section_start.empty()) {
             tool_calls = p.trigger_rule("tool-calls",
                                         p.literal(format.section_start) + p.space() + tool_calls + p.space() +
-                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end)));
+                                            (format.section_end.empty() ? p.end() : p.literal(format.section_end) + p.space()));
         }
     } else {
         std::string separator = ", ";  // Default

common/chat.cpp

@@ -2334,7 +2334,7 @@ common_chat_msg common_chat_peg_parse(const common_peg_arena &          src_pars
         ? input
         : params.generation_prompt + input;
 
-    LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
+    //LOG_DBG("Parsing PEG input with format %s: %s\n", common_chat_format_name(params.format), effective_input.c_str());
 
     common_peg_parse_flags flags = COMMON_PEG_PARSE_FLAG_LENIENT;
     if (params.debug) {

common/common.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "gguf.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "llama.h"
@@ -372,7 +373,7 @@ void common_init() {
     const char * build_type = " (debug)";
 #endif
 
-    LOG_DBG("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
+    LOG_DBG("build: %d (%s) with %s for %s%s\n", llama_build_number(), llama_commit(), llama_compiler(), llama_build_target(), build_type);
 }
 
 std::string common_params_get_system_info(const common_params & params) {
@@ -1381,7 +1382,7 @@ common_init_result_ptr common_init_from_params(common_params & params) {
 
 common_init_result::~common_init_result() = default;
 
-std::string get_model_endpoint() {
+std::string common_get_model_endpoint() {
     const char * model_endpoint_env = getenv("MODEL_ENDPOINT");
     // We still respect the use of environment-variable "HF_ENDPOINT" for backward-compatibility.
     const char * hf_endpoint_env = getenv("HF_ENDPOINT");
@@ -1396,6 +1397,42 @@ std::string get_model_endpoint() {
     return model_endpoint;
 }
 
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx) {
+    auto * mem = llama_get_memory(ctx);
+    if (mem == nullptr) {
+        return COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+    }
+
+    common_context_seq_rm_type res = COMMON_CONTEXT_SEQ_RM_TYPE_PART;
+
+    llama_memory_clear(mem, true);
+
+    // eval 2 tokens to check if the context is compatible
+    std::vector<llama_token> tmp;
+    tmp.push_back(0);
+    tmp.push_back(0);
+
+    int ret = llama_decode(ctx, llama_batch_get_one(tmp.data(), tmp.size()));
+    if (ret != 0) {
+        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_NO;
+        goto done;
+    }
+
+    // try to remove the last tokens
+    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
+        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
+        res = COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+        goto done;
+    }
+
+done:
+    llama_memory_clear(mem, true);
+    llama_synchronize(ctx);
+
+    return res;
+}
+
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora) {
     std::vector<llama_adapter_lora *> loras;
     std::vector<float> scales;

common/common.h

@@ -2,15 +2,15 @@
 
 #pragma once
 
+#include "llama-cpp.h"
+
 #include "ggml-opt.h"
 #include "ggml.h"
-#include "llama-cpp.h"
 
 #include <set>
 #include <sstream>
 #include <string>
 #include <string_view>
-#include <variant>
 #include <vector>
 #include <map>
 
@@ -27,11 +27,6 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
-} while(0)
-
 struct common_time_meas {
     common_time_meas(int64_t & t_acc, bool disable = false);
     ~common_time_meas();
@@ -53,14 +48,6 @@ struct common_adapter_lora_info {
 
 using llama_tokens = std::vector<llama_token>;
 
-// build info
-extern int LLAMA_BUILD_NUMBER;
-extern const char * LLAMA_COMMIT;
-extern const char * LLAMA_COMPILER;
-extern const char * LLAMA_BUILD_TARGET;
-
-const static std::string build_info("b" + std::to_string(LLAMA_BUILD_NUMBER) + "-" + LLAMA_COMMIT);
-
 struct common_control_vector_load_info;
 
 //
@@ -315,15 +302,15 @@ struct common_params_speculative {
     // general-purpose speculative decoding parameters
 
     int32_t n_max   = 16; // maximum number of tokens to draft during speculative decoding
-    int32_t n_min   = 0; // minimum number of draft tokens to use for speculative decoding
+    int32_t n_min   = 0;  // minimum number of draft tokens to use for speculative decoding
     float   p_split = 0.1f; // speculative decoding split probability
     float   p_min   = 0.75f; // minimum speculative decoding probability (greedy)
 
     // ngram-based speculative decoding
 
-    uint16_t ngram_size_n     = 12; // ngram size for lookup
-    uint16_t ngram_size_m     = 48; // mgram size for speculative tokens
-    uint16_t ngram_min_hits   =  1; // minimum hits at ngram/mgram lookup for mgram to be proposed
+    uint16_t ngram_size_n   = 12; // ngram size for lookup
+    uint16_t ngram_size_m   = 48; // mgram size for speculative tokens
+    uint16_t ngram_min_hits = 1; // minimum hits at ngram/mgram lookup for mgram to be proposed
 
     std::shared_ptr<common_ngram_mod> ngram_mod;
 
@@ -579,7 +566,7 @@ struct common_params {
     int32_t n_threads_http      = -1;    // number of threads to process HTTP requests (TODO: support threadpool)
     int32_t n_cache_reuse       = 0;     // min chunk size to reuse from the cache via KV shifting
     bool    cache_prompt        = true;  // whether to enable prompt caching
-    bool    clear_idle          = true;  // save and clear idle slots upon starting a new task
+    bool    cache_idle_slots    = true;  // save and clear idle slots upon starting a new task
     int32_t n_ctx_checkpoints   = 32;    // max number of context checkpoints per slot
     int32_t checkpoint_every_nt = 8192;  // make a checkpoint every n tokens during prefill
     int32_t cache_ram_mib       = 8192;  // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
@@ -859,7 +846,23 @@ struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_p
 // clear LoRA adapters from context, then apply new list of adapters
 void common_set_adapter_lora(struct llama_context * ctx, std::vector<common_adapter_lora_info> & lora);
 
-std::string                   get_model_endpoint();
+// model endpoint from env
+std::string common_get_model_endpoint();
+
+//
+// Context utils
+//
+
+enum common_context_seq_rm_type {
+    COMMON_CONTEXT_SEQ_RM_TYPE_NO   = 0, // seq_rm not supported (e.g. no memory module)
+    COMMON_CONTEXT_SEQ_RM_TYPE_PART = 1, // can seq_rm partial sequences
+    COMMON_CONTEXT_SEQ_RM_TYPE_FULL = 2, // can seq_rm full sequences only
+};
+
+// check if the llama_context can remove sequences
+// note: clears the memory of the context
+common_context_seq_rm_type common_context_can_seq_rm(llama_context * ctx);
+
 
 //
 // Batch utils

common/download.cpp

@@ -1,5 +1,6 @@
 #include "arg.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "download.h"
@@ -303,7 +304,7 @@ static int common_download_file_single_online(const std::string & url,
         headers.emplace(h.first, h.second);
     }
     if (headers.find("User-Agent") == headers.end()) {
-        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+        headers.emplace("User-Agent", "llama-cpp/" + std::string(llama_build_info()));
     }
     if (!opts.bearer_token.empty()) {
         headers.emplace("Authorization", "Bearer " + opts.bearer_token);
@@ -441,7 +442,7 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
         headers.emplace(h.first, h.second);
     }
     if (headers.find("User-Agent") == headers.end()) {
-        headers.emplace("User-Agent", "llama-cpp/" + build_info);
+        headers.emplace("User-Agent", "llama-cpp/" + std::string(llama_build_info()));
     }
 
     if (params.timeout > 0) {

common/hf-cache.cpp

@@ -1,5 +1,6 @@
 #include "hf-cache.h"
 
+#include "build-info.h"
 #include "common.h"
 #include "log.h"
 #include "http.h"
@@ -200,7 +201,7 @@ static nl::json api_get(const std::string & url,
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
-        {"User-Agent", "llama-cpp/" + build_info},
+        {"User-Agent", "llama-cpp/" + std::string(llama_build_info())},
         {"Accept", "application/json"}
     };
 
@@ -229,7 +230,7 @@ static nl::json api_get(const std::string & url,
 static std::string get_repo_commit(const std::string & repo_id,
                                    const std::string & token) {
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/refs", token);
 
         if (!json.is_object() ||
@@ -307,7 +308,7 @@ hf_files get_repo_files(const std::string & repo_id,
     hf_files files;
 
     try {
-        auto endpoint = get_model_endpoint();
+        auto endpoint = common_get_model_endpoint();
         auto json = api_get(endpoint + "api/models/" + repo_id + "/tree/" + commit + "?recursive=true", token);
 
         if (!json.is_array()) {

common/log.cpp

@@ -23,6 +23,10 @@
 
 int common_log_verbosity_thold = LOG_DEFAULT_LLAMA;
 
+int common_log_get_verbosity_thold(void) {
+    return common_log_verbosity_thold;
+}
+
 void common_log_set_verbosity_thold(int verbosity) {
     common_log_verbosity_thold = verbosity;
 }

common/log.h

@@ -38,7 +38,7 @@ enum log_colors {
 
 // needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
 // set via common_log_set_verbosity()
-extern int common_log_verbosity_thold;
+int  common_log_get_verbosity_thold(void);
 
 void common_log_set_verbosity_thold(int verbosity); // not thread-safe
 
@@ -98,7 +98,7 @@ void common_log_flush         (struct common_log * log);                    // f
 
 #define LOG_TMPL(level, verbosity, ...) \
     do { \
-        if ((verbosity) <= common_log_verbosity_thold) { \
+        if ((verbosity) <= common_log_get_verbosity_thold()) { \
             common_log_add(common_log_main(), (level), __VA_ARGS__); \
         } \
     } while (0)

common/ngram-map.cpp

@@ -208,7 +208,7 @@ void common_ngram_map_begin(
                 count_keys, count_keys_del, count_values_del, count_map_entries_upd);
     }
 
-    map.idx_last_check = (map.size_last_begin > 0) ? map.size_last_begin - 1 : 0;
+    map.idx_last_check = size_begin;
     map.size_last_begin = size_begin;
 }
 
@@ -231,7 +231,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         GGML_ABORT("%s: cur_len exceeds UINT32_MAX: %zu", __func__, cur_len);
     }
 
-    if (map.idx_last_check  > cur_len) {
+    if (map.idx_last_check > cur_len) {
         // Should not happen because of common_ngram_map_begin().
         GGML_ABORT("%s: map.idx_last_check > cur_len: %zu > %zu", __func__, map.idx_last_check, cur_len);
     }
@@ -386,7 +386,7 @@ void common_ngram_map_draft(common_ngram_map & map,
         LOG_DBG("%s: key_idx = %zu, key_offset = %zu, key_num = %d, draft.size = %zu\n", __func__,
                 curr_key.key_idx, key_offset, curr_key.key_num, draft.size());
 
-        map.last_draft_created   = false;
+        map.last_draft_created   = true;
         map.last_draft_key_idx   = key_offset;
         map.last_draft_value_idx = 0; // value 0 is used for simple mode
         return;
@@ -524,7 +524,7 @@ void common_ngram_map_accept(common_ngram_map & map, uint16_t n_accepted) {
     struct common_ngram_map_value & curr_value = curr_key.values[val_idx]; // value used for draft generation.
 
     // update the value statistics
-    LOG_INF("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
+    LOG_DBG("common_ngram_map_send_accepted: n_accepted = %d, prev value_num = %d\n",
             n_accepted, curr_value.n_accepted);
     curr_value.n_accepted = n_accepted;
 }

common/speculative.cpp

@@ -13,6 +13,7 @@
 #include <cstring>
 #include <iomanip>
 #include <map>
+#include <cinttypes>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  128
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -144,10 +145,28 @@ struct common_speculative_state {
     virtual void accept(uint16_t n_accepted) = 0;
 };
 
+struct common_speculative_checkpoint {
+    llama_pos pos_min  = 0;
+    llama_pos pos_max  = 0;
+
+    int64_t   n_tokens = 0;
+
+    std::vector<uint8_t> data;
+
+    size_t size() const {
+        return data.size();
+    }
+
+    size_t ckpt_size   = 0;
+};
+
 struct common_speculative_state_draft : public common_speculative_state {
     llama_context * ctx_tgt; // only used for retokenizing from ctx_dft
     llama_context * ctx_dft;
 
+    bool use_ckpt = false;
+    struct common_speculative_checkpoint ckpt;
+
     common_sampler * smpl;
 
     llama_batch  batch;
@@ -160,10 +179,12 @@ struct common_speculative_state_draft : public common_speculative_state {
             enum common_speculative_type type,
             llama_context * ctx_tgt,
             llama_context * ctx_dft,
-            const std::vector<std::pair<std::string, std::string>> & replacements)
+            const std::vector<std::pair<std::string, std::string>> & replacements,
+            bool use_ckpt)
         : common_speculative_state(type)
         , ctx_tgt(ctx_tgt)
         , ctx_dft(ctx_dft)
+        , use_ckpt(use_ckpt)
     {
         batch = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
         smpl = nullptr;
@@ -218,7 +239,48 @@ struct common_speculative_state_draft : public common_speculative_state {
     }
 
     void begin(const llama_tokens & prompt) override {
-        GGML_UNUSED(prompt);
+        if (use_ckpt && ckpt.size() > 0) {
+            // delete checkpoint
+            LOG_DBG("%s: delete checkpoint, prompt.size=%zu, pos_min=%d, pos_max=%d, n_tokens=%" PRId64 ", size=%.3f MiB\n",
+                    __func__, prompt.size(), ckpt.pos_min, ckpt.pos_max, ckpt.n_tokens, (float) ckpt.data.size() / 1024 / 1024);
+            ckpt.pos_min   = 0;
+            ckpt.pos_max   = 0;
+            ckpt.n_tokens  = 0;
+            ckpt.ckpt_size = 0;
+            ckpt.data.clear();
+        }
+    }
+
+    size_t draft_create_checkpoint(int n_tokens_prompt, int n_tokens_batch) {
+        int slot_id = 0;
+        const size_t checkpoint_size = llama_state_seq_get_size_ext(ctx_dft, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+
+        ckpt.pos_min  = llama_memory_seq_pos_min(llama_get_memory(ctx_dft), slot_id);
+        ckpt.pos_max  = llama_memory_seq_pos_max(llama_get_memory(ctx_dft), slot_id);
+        ckpt.n_tokens = n_tokens_prompt - n_tokens_batch;
+        ckpt.data.resize(checkpoint_size);
+
+        const size_t n = llama_state_seq_get_data_ext(ctx_dft, ckpt.data.data(), checkpoint_size, slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != checkpoint_size) {
+            GGML_ABORT("checkpoint size mismatch: expected %zu, got %zu\n", checkpoint_size, n);
+        }
+
+        LOG_DBG("%s: pos_min = %d, pos_max = %d, size = %.3f MiB\n", __func__,
+                ckpt.pos_min, ckpt.pos_max, (float) ckpt.data.size() / 1024 / 1024);
+        return n;
+    }
+
+    size_t draft_restore_checkpoint(size_t ckpt_size_part_expected) {
+        int slot_id = 0;
+        LOG_DBG("%s: pos_min = %d, pos_max = %d\n", __func__, ckpt.pos_min, ckpt.pos_max);
+        const size_t n = llama_state_seq_set_data_ext(ctx_dft, ckpt.data.data(), ckpt.size(), slot_id, LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY);
+        if (n != ckpt_size_part_expected) {
+            GGML_ABORT("%s: failed to restore context checkpoint (pos_min=%d, pos_max=%d, size=%zu, get_data_ext->%zu, set_data_ext->%zu",
+                        __func__, ckpt.pos_min, ckpt.pos_max, ckpt.size(), ckpt_size_part_expected, n);
+        }
+        llama_memory_seq_rm(llama_get_memory(ctx_dft), slot_id, ckpt.pos_max + 1, -1);
+
+        return n;
     }
 
     void draft(
@@ -236,8 +298,8 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         auto * mem_dft = llama_get_memory(ctx_dft);
 
-        int reuse_i = 0;
-        int reuse_n = 0;
+        int reuse_i = 0; // index of part to be reused in prompt_dft
+        int reuse_n = 0; // length of part to be reused in prompt_dft
 
         const int n_ctx = llama_n_ctx(ctx_dft) - params.n_max;
 
@@ -287,18 +349,26 @@ struct common_speculative_state_draft : public common_speculative_state {
             }
         }
 
-        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, prompt = %d\n", __func__, reuse_i, reuse_n, (int) prompt_dft.size());
+        LOG_DBG("%s: reuse_i = %d, reuse_n = %d, #prompt_dft = %zu, #prompt_cur = %zu\n",
+                __func__, reuse_i, reuse_n, prompt_dft.size(), prompt_cur.size());
+        if (use_ckpt && ckpt.ckpt_size == 0 && reuse_n > 0) {
+            LOG_DBG("%s: no checkpoint available, no reuse, (reuse_i=%d, reuse_n=%d) -> (0, 0)\n",
+                    __func__, reuse_i, reuse_n);
+            reuse_i = 0;
+            reuse_n = 0;
+        }
 
         result.clear();
         result.reserve(params.n_max);
 
-        if (reuse_n == 0) {
+        bool needs_ckpt = use_ckpt && prompt_dft.size() > 0;
+        if (reuse_n == 0 || (use_ckpt && reuse_i > 0)) {
             llama_memory_clear(mem_dft, false);
             prompt_dft.clear();
         } else {
             // this happens when a previous draft has been discarded (for example, due to being too small), but the
             // target model agreed with it. in this case, we simply pass back the previous results to save compute
-            if (reuse_i + reuse_n < (int) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
+            if (reuse_i + reuse_n < (int64_t) prompt_dft.size() && prompt_dft[reuse_i + reuse_n] == id_last) {
                 for (int i = reuse_i + reuse_n + 1; i < (int) prompt_dft.size(); ++i) {
                     result.push_back(prompt_dft[i]);
 
@@ -310,19 +380,50 @@ struct common_speculative_state_draft : public common_speculative_state {
                 return;
             }
 
+            bool do_restore = false;
+            if (prompt_dft.size() > prompt_cur.size() && reuse_i + reuse_n < (int64_t) prompt_dft.size()) {
+                // This can happen after a partial acceptance (speculative decoding with checkpoints)
+                LOG_DBG("%s: #prompt_dft=%zu, #prompt_cur=%zu, shorten draft\n",
+                        __func__, prompt_dft.size(), prompt_cur.size());
+                prompt_dft.resize(prompt_cur.size());
+                do_restore = true;
+            }
+
             if (reuse_i > 0) {
-                llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                bool is_removed = llama_memory_seq_rm (mem_dft, 0, 0, reuse_i);
+                if (!is_removed) {
+                    LOG_ERR("%s: llama_memory_seq_rm failed, reuse_i=%d\n", __func__, reuse_i);
+                }
                 llama_memory_seq_add(mem_dft, 0, reuse_i, -1, -reuse_i);
 
                 prompt_dft.erase(prompt_dft.begin(), prompt_dft.begin() + reuse_i);
             }
 
-            if (reuse_n < (int) prompt_dft.size()) {
-                llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
-                prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+            if (reuse_n < (int) prompt_dft.size() || do_restore) {
+                if (use_ckpt) {
+                    if (ckpt.n_tokens > (int64_t) prompt_dft.size()) {
+                        LOG_INF("%s: checkpoint is too large, prompt_tgt.size=%zu, ckpt.n_tokens=%" PRId64 ", reuse_n=%d, prompt_dft.size=%zu\n",
+                                __func__, prompt_tgt.size(), ckpt.n_tokens, reuse_n, prompt_dft.size());
+                    }
+                    draft_restore_checkpoint(ckpt.ckpt_size);
+                    reuse_n = ckpt.n_tokens;
+                    prompt_dft.resize(reuse_n);
+                    needs_ckpt = false;
+                } else {
+                    bool is_removed = llama_memory_seq_rm (mem_dft, 0, reuse_n, -1);
+                    if (!is_removed) {
+                        LOG_ERR("%s: llama_memory_seq_rm failed, reuse_n=%d, prompt_dft.size=%zu\n",
+                                __func__, reuse_n, prompt_dft.size());
+                    }
+                    prompt_dft.erase(prompt_dft.begin() + reuse_n, prompt_dft.end());
+                }
             }
         }
 
+        if (needs_ckpt) {
+            ckpt.ckpt_size = draft_create_checkpoint(prompt_dft.size(), batch.n_tokens);
+        }
+
         // prepare a batch to evaluate any new tokens in the prompt
         common_batch_clear(batch);
 
@@ -337,7 +438,11 @@ struct common_speculative_state_draft : public common_speculative_state {
         if (batch.n_tokens > 0) {
             //LOG_DBG("%s: draft prompt batch: %s\n", __func__, string_from(ctx, batch).c_str());
 
-            llama_decode(ctx_dft, batch);
+            int ret = llama_decode(ctx_dft, batch);
+            if (ret != 0 && ret != 1) {
+                LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu\n",
+                        __func__, ret, prompt_cur.size());
+            }
         }
 
         const llama_pos n_past = prompt_dft.size();
@@ -351,7 +456,11 @@ struct common_speculative_state_draft : public common_speculative_state {
 
         LOG_DBG("%s: draft prompt: %s\n", __func__, string_from(ctx_dft, prompt_dft).c_str());
 
-        llama_decode(ctx_dft, batch);
+        int ret = llama_decode(ctx_dft, batch);
+        if (ret != 0 && ret != 1) {
+            LOG_WRN("%s: llama_decode returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                    __func__, ret, prompt_cur.size(), prompt_dft.size());
+        }
 
         common_sampler_reset(smpl);
 
@@ -387,7 +496,11 @@ struct common_speculative_state_draft : public common_speculative_state {
             common_batch_add(batch, id, n_past + i + 1, { 0 }, true);
 
             // evaluate the drafted tokens on the draft model
-            llama_decode(ctx_dft, batch);
+            ret = llama_decode(ctx_dft, batch);
+            if (ret != 0) {
+                LOG_WRN("%s: llama_decode[%d] returned %d, prompt_cur.size=%zu, prompt_dft.size=%zu\n",
+                        __func__, i, ret, prompt_cur.size(), prompt_dft.size());
+            }
 
             prompt_dft.push_back(id);
         }
@@ -739,6 +852,7 @@ struct common_speculative_state_ngram_cache : public common_speculative_state {
 
 struct common_speculative {
     std::vector<std::unique_ptr<common_speculative_state>> impls; // list of implementations to use and their states
+
     common_speculative_state * curr_impl = nullptr; // current implementation in use (for stats)
 };
 
@@ -798,42 +912,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
     return it->second;
 }
 
-bool common_speculative_is_compat(llama_context * ctx_tgt) {
-    auto * mem = llama_get_memory(ctx_tgt);
-    if (mem == nullptr) {
-        return false;
-    }
-
-    bool res = true;
-
-    llama_memory_clear(mem, true);
-
-    // eval 2 tokens to check if the context is compatible
-    std::vector<llama_token> tmp;
-    tmp.push_back(0);
-    tmp.push_back(0);
-
-    int ret = llama_decode(ctx_tgt, llama_batch_get_one(tmp.data(), tmp.size()));
-    if (ret != 0) {
-        LOG_ERR("%s: llama_decode() failed: %d\n", __func__, ret);
-        res = false;
-        goto done;
-    }
-
-    // try to remove the last tokens
-    if (!llama_memory_seq_rm(mem, 0, 1, -1)) {
-        LOG_WRN("%s: the target context does not support partial sequence removal\n", __func__);
-        res = false;
-        goto done;
-    }
-
-done:
-    llama_memory_clear(mem, true);
-    llama_synchronize(ctx_tgt);
-
-    return res;
-}
-
 // initialization of the speculative decoding system
 //
 common_speculative * common_speculative_init(
@@ -908,10 +986,13 @@ common_speculative * common_speculative_init(
             case COMMON_SPECULATIVE_TYPE_NONE:
                 break;
             case COMMON_SPECULATIVE_TYPE_DRAFT: {
+                const bool use_ckpt = common_context_can_seq_rm(ctx_dft) == COMMON_CONTEXT_SEQ_RM_TYPE_FULL;
+
                 impls.push_back(std::make_unique<common_speculative_state_draft>(config.type,
                     /* .ctx_tgt      = */ ctx_tgt,
                     /* .ctx_dft      = */ ctx_dft,
-                    /* .replacements = */ params.replacements
+                    /* .replacements = */ params.replacements,
+                    /* .use_ckpt     = */ use_ckpt
                 ));
                 break;
             }
@@ -966,7 +1047,8 @@ common_speculative * common_speculative_init(
     }
 
     auto * result = new common_speculative {
-        /* .impls = */ std::move(impls)
+        /* .impls     = */ std::move(impls),
+        /* .curr_impl = */ nullptr,
     };
 
     return result;

common/speculative.h

@@ -14,10 +14,6 @@ enum common_speculative_type common_speculative_type_from_name(const std::string
 // convert type to string
 std::string common_speculative_type_to_str(enum common_speculative_type type);
 
-// check if the llama_context is compatible for speculative decoding
-// note: clears the memory of the context
-bool common_speculative_is_compat(llama_context * ctx_tgt);
-
 common_speculative * common_speculative_init(
         common_params_speculative & params,
         llama_context             * ctx_tgt);
@@ -39,3 +35,9 @@ void common_speculative_accept(common_speculative * spec, uint16_t n_accepted);
 
 // print statistics about the speculative decoding
 void common_speculative_print_stats(const common_speculative * spec);
+
+struct common_speculative_deleter {
+    void operator()(common_speculative * s) { common_speculative_free(s); }
+};
+
+typedef std::unique_ptr<common_speculative, common_speculative_deleter> common_speculative_ptr;

convert_hf_to_gguf.py

@@ -1850,20 +1850,28 @@ class TextModel(ModelBase):
             with open(module_path, encoding="utf-8") as f:
                 modules = json.load(f)
             for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
+                if mod["type"].endswith("Pooling"):
                     pooling_path = mod["path"]
                     break
 
+        mode_mapping = {
+            "mean": gguf.PoolingType.MEAN,
+            "cls": gguf.PoolingType.CLS,
+            "lasttoken": gguf.PoolingType.LAST,
+        }
+
         # get pooling type
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
+            if pooling.get("pooling_mode_mean_tokens"):
                 pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
+            elif pooling.get("pooling_mode_cls_token"):
                 pooling_type = gguf.PoolingType.CLS
-            elif pooling["pooling_mode_lasttoken"]:
+            elif pooling.get("pooling_mode_lasttoken"):
                 pooling_type = gguf.PoolingType.LAST
+            elif (pooling_mode := pooling.get("pooling_mode")) in mode_mapping:
+                pooling_type = mode_mapping[pooling_mode]
             else:
                 raise NotImplementedError("Only MEAN, CLS, and LAST pooling types supported")
             self.gguf_writer.add_pooling_type(pooling_type)
@@ -7180,7 +7188,7 @@ class EmbeddingGemma(Gemma3Model):
                 with open(modules_file, encoding="utf-8") as modules_json_file:
                     mods = json.load(modules_json_file)
                 for mod in mods:
-                    if mod["type"] == "sentence_transformers.models.Dense":
+                    if mod["type"].endswith("Dense"):
                         mod_path = mod["path"]
                         # check if model.safetensors file for Dense layer exists
                         model_tensors_file = self.dir_model / mod_path / "model.safetensors"
@@ -10893,7 +10901,64 @@ class NemotronHModel(GraniteHybridModel):
                 self.gguf_writer.add_moe_latent_size(latent_size)
 
     def set_vocab(self):
-        super().set_vocab()
+        # The NemotronH config uses pattern characters (e.g. '-') that may not
+        # be supported by the installed transformers version. AutoTokenizer
+        # internally calls AutoConfig which triggers this parsing failure.
+        # Using trust_remote_code=True to load the model's own config class.
+        tokens: list[str] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)
+
+        # Pad vocab size (from Mamba2Model/GraniteHybridModel)
+        self.hparams["pad_vocab_size_multiple"] = 8 # Setting this here since GraniteHybridModel.set_vocab() isn't being invoked now.
+        # From Mamba2Model.set_vocab():
+        vocab_size = self.hparams["vocab_size"]
+        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 16)
+        # ref: https://stackoverflow.com/a/17511341/22827863
+        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
+        self.hparams["vocab_size"] = vocab_size
+
+        assert max(tokenizer.vocab.values()) < vocab_size  # ty: ignore[unresolved-attribute]
+
+        tokpre = self.get_vocab_base_pre(tokenizer)
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]
+        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]
+
+        added_tokens_decoder = tokenizer.added_tokens_decoder  # ty: ignore[unresolved-attribute]
+
+        for i in range(vocab_size):
+            if i not in reverse_vocab:
+                tokens.append(f"[PAD{i}]")
+                toktypes.append(gguf.TokenType.UNUSED)
+            else:
+                token: str = reverse_vocab[i]
+                if token in added_vocab:
+                    if not added_tokens_decoder[i].normalized:
+                        previous_token = token
+                        token = tokenizer.decode(tokenizer.encode(token, add_special_tokens=False))  # ty: ignore[unresolved-attribute, invalid-assignment]
+                        if previous_token != token:
+                            logger.info(f"{repr(previous_token)} is encoded and decoded back to {repr(token)} using AutoTokenizer")
+
+                    if added_tokens_decoder[i].special or self.does_token_look_special(token):
+                        toktypes.append(gguf.TokenType.CONTROL)
+                    else:
+                        token = token.replace(b"\xe2\x96\x81".decode("utf-8"), " ")  # pre-normalize user-defined spaces
+                        toktypes.append(gguf.TokenType.USER_DEFINED)
+                else:
+                    toktypes.append(gguf.TokenType.NORMAL)
+                tokens.append(token)
+
+        # From TextModel.set_vocab_gpt2():
+        self.gguf_writer.add_tokenizer_model("gpt2")
+        self.gguf_writer.add_tokenizer_pre(tokpre)
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
+        special_vocab.add_to_gguf(self.gguf_writer)
 
         # The tokenizer _does_ add a BOS token (via post_processor type
         # TemplateProcessing) but does not set add_bos_token to true in the

docs/backend/SYCL.md

@@ -689,6 +689,7 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path. (1.) |
 | GGML_SYCL_GRAPH    | OFF *(default)* \|ON *(Optional)*     | Enable build with [SYCL Graph extension](https://github.com/intel/llvm/blob/sycl/sycl/doc/extensions/experimental/sycl_ext_oneapi_graph.asciidoc). |
 | GGML_SYCL_DNN      | ON *(default)* \|OFF *(Optional)*     | Enable build with oneDNN.                   |
+| GGML_SYCL_HOST_MEM_FALLBACK | ON *(default)* \|OFF *(Optional)* | Allow host memory fallback when device memory is full during quantized weight reorder. Enables inference to continue at reduced speed (reading over PCIe) instead of failing. Requires Linux kernel 6.8+. |
 | CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
 | CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 

docs/build.md

@@ -281,6 +281,12 @@ Use `GGML_CUDA_FORCE_CUBLAS_COMPUTE_16F` environment variable to force use FP16
 
 The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
 
+### Peer Access
+
+The environment variable `GGML_CUDA_P2P` can be set to enable peer-to-peer access between multiple GPUs, allowing them to transfer data directly rather than to go through system memory.
+Requires driver support (usually restricted to workstation/datacenter GPUs).
+May cause crashes or corrupted outputs for some motherboards and BIOS settings (e.g. IOMMU).
+
 ### Performance Tuning
 
 The following compilation options are also available to tweak performance:

docs/development/HOWTO-add-model.md

@@ -130,6 +130,23 @@ Note:
 - Adding a model-specific API or CLI is an anti-pattern in `libmtmd`. The goal of `libmtmd` is to provide an easy-to-use, model-agnostic library for multimodal pipeline.
 - In most cases, `llama-mtmd-cli` should not be modified. If a model requires a specific prompt, either let the user provide it or bake it into the Jinja chat template.
 
+## Tips and tricks
+
+### Working with ggml_rope_ext
+
+PyTorch implementations usually prefer explicitly calculating `freq_cis`/`sin`/`cos` components. However, in llama.cpp, most RoPE operations can be handled via `ggml_rope_ext`, which does not require a sin/cos matrix. This saves memory while allowing the GGML RoPE kernel to be fused with other ops.
+
+However, since `ggml_rope_ext` only provides a subset of the RoPE implementations that models use, converting models from PyTorch to llama.cpp may require some creative adaptations.
+
+For more information about `ggml_rope_ext`, please refer to the in-code documentation in `ggml.h`.
+
+Examples:
+- `libmtmd` implements 2D RoPE with `GGML_ROPE_TYPE_NORMAL` ordering by splitting the input tensor in half, applying `ggml_rope_ext` separately to each half, then joining them back together using `ggml_concat`.
+- The [Kimi-K2.5](https://github.com/ggml-org/llama.cpp/pull/19170) vision encoder uses vision RoPE with interleaved frequencies. The weights must be permuted during conversion in order to reuse the `build_rope_2d()` function.
+- [Gemma 4](https://github.com/ggml-org/llama.cpp/pull/21309) uses "proportional" RoPE. We employ a trick where `rope_freqs` is set to a very large value in the last dimensions to prevent those dimensions from being rotated. See the `Gemma4Model` class in `convert_hf_to_gguf.py`.
+- Some models require scaling the input position. For example, `[0, 1, 2, ...]` becomes `[0, 0.5, 1, ...]`. In this case, you can provide the scaling via `freq_scale = 0.5f`.
+- Some models use learned RoPE frequencies instead of relying on `powf(freq_base, -2.0 * i / n_dims)`. In this case, you can provide the learned frequencies via the `rope_freqs` tensor (corresponding to the `c` argument in `ggml_rope_ext`), then set `freq_base = 1.0f`. An important note is that `rope_freqs` in GGML is the **inverse** (`theta = pos[i] / rope_freqs`), so you may need to invert `rope_freqs` during conversion.
+
 ## GGUF specification
 
 https://github.com/ggml-org/ggml/blob/master/docs/gguf.md

docs/ops.md

@@ -22,13 +22,13 @@ Legend:
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              COS | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
@@ -46,7 +46,7 @@ Legend:
 |                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ |
 |                             FILL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | 🟡 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
@@ -84,10 +84,10 @@ Legend:
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -116,6 +116,6 @@ Legend:
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                            TOP_K | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |

examples/batched/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-batched)
 add_executable(${TARGET} batched.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/convert-llama2c-to-ggml/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-convert-llama2c-to-ggml)
 add_executable(${TARGET} convert-llama2c-to-ggml.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/debug/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-debug)
 add_executable(${TARGET} debug.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/diffusion/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-diffusion-cli)
 add_executable(${TARGET} diffusion-cli.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama llama-common ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/embedding/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-embedding)
 add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/eval-callback/CMakeLists.txt

@@ -1,7 +1,7 @@
 set(TARGET llama-eval-callback)
 add_executable(${TARGET} eval-callback.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 if(LLAMA_BUILD_TESTS)

examples/gen-docs/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-gen-docs)
 add_executable(${TARGET} gen-docs.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/idle/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-idle)
 add_executable(${TARGET} idle.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama llama-common ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/llama.android/lib/src/main/cpp/CMakeLists.txt

@@ -51,6 +51,6 @@ target_include_directories(${CMAKE_PROJECT_NAME} PRIVATE
 
 target_link_libraries(${CMAKE_PROJECT_NAME}
         llama
-        common
+        llama-common
         android
         log)

examples/lookahead/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-lookahead)
 add_executable(${TARGET} lookahead.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/lookup/CMakeLists.txt

@@ -1,23 +1,23 @@
 set(TARGET llama-lookup)
 add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-create)
 add_executable(${TARGET} lookup-create.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-merge)
 add_executable(${TARGET} lookup-merge.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TARGET llama-lookup-stats)
 add_executable(${TARGET} lookup-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/parallel/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-parallel)
 add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/passkey/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-passkey)
 add_executable(${TARGET} passkey.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/retrieval/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-retrieval)
 add_executable(${TARGET} retrieval.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/save-load-state/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-save-load-state)
 add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/speculative-simple/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-speculative-simple)
 add_executable(${TARGET} speculative-simple.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/speculative/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-speculative)
 add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/sycl/CMakeLists.txt

@@ -5,5 +5,5 @@
 set(TARGET llama-ls-sycl-device)
 add_executable(${TARGET} ls-sycl-device.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_17)

examples/training/CMakeLists.txt

@@ -1,5 +1,5 @@
 set(TARGET llama-finetune)
 add_executable(${TARGET} finetune.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama-common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

ggml/CMakeLists.txt

@@ -1,11 +1,5 @@
 cmake_minimum_required(VERSION 3.14...3.28) # for add_link_options and implicit target directories.
 
-# ref: https://cmake.org/cmake/help/latest/policy/CMP0194.html
-# MSVC is not a valid assembler for the ASM language.
-# Set to NEW to avoid a warning on CMake 4.1+ with MSVC.
-if (POLICY CMP0194)
-    cmake_policy(SET CMP0194 NEW)
-endif()
 project("ggml" C CXX ASM)
 
 ### GGML Version
@@ -254,6 +248,7 @@ option(GGML_RPC                             "ggml: use RPC"
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 option(GGML_SYCL_GRAPH                      "ggml: enable graphs in the SYCL backend"         ON)
+option(GGML_SYCL_HOST_MEM_FALLBACK          "ggml: allow host memory fallback in SYCL reorder (requires kernel 6.8+)" ON)
 option(GGML_SYCL_DNN                        "ggml: enable oneDNN in the SYCL backend"         ON)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
                                             "ggml: sycl target device")

ggml/include/ggml-backend.h

@@ -202,8 +202,11 @@ extern "C" {
 
     // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // AllReduce operation for tensor parallelism (meta backend)
-    typedef bool                         (*ggml_backend_allreduce_tensor_t)(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends);
+    // Context management and operations for faster communication between backends, used for tensor parallelism (meta backend)
+    typedef void * (*ggml_backend_comm_init_t)(ggml_backend_t * backends, size_t n_backends);
+    typedef void   (*ggml_backend_comm_free_t)(void * comm_ctx);
+    typedef bool   (*ggml_backend_comm_allreduce_tensor_t)(void * comm_ctx, struct ggml_tensor ** tensors);
+
     // Split buffer type for tensor parallelism (old)
     typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
     // Set the number of threads for the backend

ggml/include/ggml-rpc.h

@@ -6,9 +6,9 @@
 extern "C" {
 #endif
 
-#define RPC_PROTO_MAJOR_VERSION    3
-#define RPC_PROTO_MINOR_VERSION    6
-#define RPC_PROTO_PATCH_VERSION    1
+#define RPC_PROTO_MAJOR_VERSION    4
+#define RPC_PROTO_MINOR_VERSION    0
+#define RPC_PROTO_PATCH_VERSION    0
 
 #ifdef  __cplusplus
 static_assert(GGML_OP_COUNT == 96, "GGML_OP_COUNT has changed - update RPC_PROTO_PATCH_VERSION");

ggml/include/ggml.h

@@ -1773,8 +1773,32 @@ extern "C" {
             int                   n_dims,
             int                   mode);
 
-    // custom RoPE
+    // RoPE operations with extended options
+    // a is the input tensor to apply RoPE to, shape [n_embd, n_head, n_token]
+    // b is an int32 vector with size n_token
     // c is freq factors (e.g. phi3-128k), (optional)
+    // mode can be GGML_ROPE_TYPE_NORMAL or NEOX; for MROPE and VISION mode, use ggml_rope_multi
+    //
+    // pseudo-code for computing theta:
+    //   for i in [0, n_dims/2):
+    //     theta[i] = b[i] * powf(freq_base, -2.0 * i / n_dims);
+    //     theta[i] = theta[i] / c[i];  # if c is provided, divide theta by c
+    //     theta[i] = rope_yarn(theta[i], ...);  # note: theta = theta * freq_scale is applied here
+    //
+    // other params are used by YaRN RoPE scaling, these default values will disable YaRN:
+    //   freq_scale  = 1.0f
+    //   ext_factor  = 0.0f
+    //   attn_factor = 1.0f
+    //   beta_fast   = 0.0f
+    //   beta_slow   = 0.0f
+    //
+    // example:
+    //   (marking: c = cos, s = sin, 0 = unrotated)
+    //   given a single head with size = 8 --> [00000000]
+    //   GGML_ROPE_TYPE_NORMAL  n_dims = 4 --> [cscs0000]
+    //   GGML_ROPE_TYPE_NORMAL  n_dims = 8 --> [cscscscs]
+    //   GGML_ROPE_TYPE_NEOX    n_dims = 4 --> [ccss0000]
+    //   GGML_ROPE_TYPE_NEOX    n_dims = 8 --> [ccccssss]
     GGML_API struct ggml_tensor * ggml_rope_ext(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -1790,6 +1814,36 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    // multi-dimensional RoPE, for Qwen-VL and similar vision models
+    // mode can be either VISION, MROPE, IMROPE, cannot be combined with NORMAL or NEOX
+    // sections specify how many dimensions to rotate in each section:
+    //   section length is equivalent to number of cos/sin pairs, NOT the number of dims
+    //   (i.e. sum of 4 sections are expected to be n_dims/2)
+    //   last sections can be 0, means ignored
+    // all other options are identical to ggml_rope_ext
+    //
+    // important note:
+    //   - NEOX ordering is automatically applied and cannot be disabled for MROPE and VISION
+    //     if you need normal ordering, there are 2 methods:
+    //     (1) split the tensor manually using ggml_view
+    //     (2) permute the weight upon conversion
+    //   - for VISION, n_dims must be head_size/2
+    //
+    // example M-RoPE:
+    //  given sections = [t=4, y=2, x=2, 0]
+    //  given a single head with size = 18 --> [000000000000000000]
+    //  GGML_ROPE_TYPE_MROPE   n_dims = 16 --> [ttttyyxxttttyyxx00] (cos/sin are applied in NEOX ordering)
+    //  GGML_ROPE_TYPE_IMROPE  n_dims = 16 --> [ttyxttyxttyxttyx00] (interleaved M-RoPE, still NEOX ordering)
+    //  note: the theta for each dim is computed the same way as ggml_rope_ext, no matter the section
+    //        in other words, idx used for theta: [0123456789... until n_dims/2], not reset for each section
+    //
+    // example vision RoPE:
+    //  given sections = [y=4, x=4, 0, 0] (last 2 sections are ignored)
+    //  given a single head with size = 8 --> [00000000]
+    //  GGML_ROPE_TYPE_VISION  n_dims = 4 --> [yyyyxxxx]
+    //  other values of n_dims are untested and is undefined behavior
+    //  note: unlike MROPE, the theta for each dim is computed differently for each section
+    //        in other words, idx used for theta: [0123] for y section, then [0123] for x section
     GGML_API struct ggml_tensor * ggml_rope_multi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,

ggml/src/ggml-backend-meta.cpp

@@ -1133,7 +1133,7 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
         if (t_ij->view_src != nullptr && ggml_backend_buffer_is_meta(t_ij->view_src->buffer)) {
             t_ij->view_src = ggml_backend_meta_buffer_simple_tensor(tensor->view_src, j);
             if (t_ij->view_offs > 0 && split_dim >= 0 && split_dim < GGML_MAX_DIMS) {
-                GGML_ASSERT(ne[split_dim] != 0 && tensor->ne[split_dim] != 0);
+                GGML_ASSERT(tensor->ne[split_dim] != 0);
                 const int split_dim_view_src = ggml_backend_meta_get_split_state(tensor->view_src, /*assume_sync =*/ true).axis;
                 GGML_ASSERT(split_dim_view_src >= 0 && split_dim_view_src < GGML_MAX_DIMS);
 
@@ -1170,6 +1170,28 @@ static enum ggml_status ggml_backend_meta_buffer_init_tensor(ggml_backend_buffer
 
         simple_tensors.push_back(t_ij);
     }
+
+    // If one of the sources has a zero-sized slice, disable the computation:
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (tensor->src[i] == nullptr || !ggml_backend_buffer_is_meta(tensor->src[i]->buffer)) {
+            continue;
+        }
+
+        const ggml_backend_meta_split_state split_state_src = ggml_backend_meta_get_split_state(tensor->src[i], /*assume_sync =*/ true);
+        if (split_state_src.axis < 0 || split_state_src.axis >= GGML_MAX_DIMS) {
+            continue;
+        }
+        for (size_t j = 0; j < n_simple_bufs; j++) {
+            int64_t ne_sum = 0;
+            for (size_t s = 0; s < split_state_src.n_segments; s++) {
+                ne_sum += split_state_src.ne[s*n_simple_bufs + j];
+            }
+            if (ne_sum == 0) {
+                simple_tensors[j]->flags &= ~GGML_TENSOR_FLAG_COMPUTE;
+            }
+        }
+    }
+
     buf_ctx->simple_tensors[tensor] = simple_tensors;
 
     return GGML_STATUS_SUCCESS;
@@ -1270,7 +1292,45 @@ static void ggml_backend_meta_buffer_get_tensor(ggml_backend_buffer_t buffer, co
     GGML_ASSERT(ggml_is_contiguous(tensor));
 
     const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(tensor, /*assume_sync =*/ false);
-    GGML_ASSERT(split_state.n_segments == 1);
+
+    if (split_state.n_segments != 1) {
+        GGML_ASSERT(split_state.axis >= 0 && split_state.axis < GGML_MAX_DIMS);
+        GGML_ASSERT(offset == 0);
+        GGML_ASSERT(size == ggml_nbytes(tensor));
+        GGML_ASSERT(tensor->ne[3] == 1);
+        size_t offset_data = 0;
+        std::vector<size_t> simple_offsets(n_bufs, 0);
+        if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_0) {
+            GGML_ASSERT(tensor->ne[2] == 1);
+            const int64_t blck_size = ggml_blck_size(tensor->type);
+            for (size_t s = 0; s < split_state.n_segments; s++) {
+                for (size_t j = 0; j < n_bufs; j++) {
+                    const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                    GGML_ASSERT(split_state.ne[s*n_bufs + j] % blck_size == 0);
+                    const size_t nbytes = split_state.ne[s*n_bufs + j]/blck_size * tensor->nb[0];
+                    ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
+                        tensor->ne[1], simple_tensor->nb[1], tensor->nb[1]);
+                    offset_data       += nbytes;
+                    simple_offsets[j] += nbytes;
+                }
+            }
+            GGML_ASSERT(offset_data*tensor->ne[1] == size);
+            return;
+        }
+        GGML_ASSERT(split_state.axis == GGML_BACKEND_SPLIT_AXIS_1);
+        for (size_t s = 0; s < split_state.n_segments; s++) {
+            for (size_t j = 0; j < n_bufs; j++) {
+                const ggml_tensor * simple_tensor = ggml_backend_meta_buffer_simple_tensor(tensor, j);
+                const size_t nbytes = split_state.ne[s*n_bufs + j] * tensor->nb[1];
+                ggml_backend_tensor_get_2d(simple_tensor, (char *) data + offset_data, simple_offsets[j], nbytes,
+                    tensor->ne[2], simple_tensor->nb[2], tensor->nb[2]);
+                offset_data       += nbytes;
+                simple_offsets[j] += nbytes;
+            }
+        }
+        GGML_ASSERT(offset_data*tensor->ne[2] == size);
+        return;
+    }
 
     switch (split_state.axis) {
         case GGML_BACKEND_SPLIT_AXIS_0:
@@ -1404,45 +1464,73 @@ struct ggml_backend_meta_context {
     struct backend_config {
         ggml_backend_t backend;
 
-        std::vector<cgraph_config> cgraphs;
-        std::vector<ggml_tensor *> nodes;
-        ggml_backend_buffer_ptr    buf;
+        std::vector<cgraph_config>           cgraphs;
+        std::vector<ggml_tensor *>           nodes;
+        std::vector<ggml_backend_buffer_ptr> bufs;
 
-        backend_config(ggml_backend_t backend) : backend(backend) {}
+        backend_config(ggml_backend_t backend, const size_t n_reduce_steps) : backend(backend) {
+            bufs.resize(n_reduce_steps);
+        }
     };
     std::string                 name;
     std::vector<backend_config> backend_configs;
     ggml_context_ptr            ctx;
     std::vector<ggml_cgraph *>  cgraphs_aux;
     std::vector<ggml_tensor *>  nodes_aux;
+    size_t                      n_reduce_steps;
     int                         max_nnodes    = 0;
     size_t                      max_tmp_size  = 0;
     size_t                      max_subgraphs = 0;
+    size_t                      n_subgraphs   = 0;
+    uint64_t                    uid           = 0;
+
+    void *                               comm_ctx       = nullptr;
+    ggml_backend_comm_allreduce_tensor_t comm_allreduce = nullptr;
 
     ggml_backend_meta_context(ggml_backend_dev_t meta_dev, const char * params) {
         const size_t n_devs = ggml_backend_meta_dev_n_devs(meta_dev);
+        n_reduce_steps = std::ceil(std::log2(n_devs));
         name = "Meta(";
+        std::vector<ggml_backend_t> simple_backends;
         backend_configs.reserve(n_devs);
+        simple_backends.reserve(n_devs);
         for (size_t i = 0; i < n_devs; i++) {
             ggml_backend_dev_t simple_dev = ggml_backend_meta_dev_simple_dev(meta_dev, i);
             if (i > 0) {
                 name += ",";
             }
             name += ggml_backend_dev_name(simple_dev);
-            backend_configs.emplace_back(ggml_backend_dev_init(simple_dev, params));
+            simple_backends.push_back(ggml_backend_dev_init(simple_dev, params));
+            backend_configs.emplace_back(simple_backends.back(), n_reduce_steps);
         }
         name += ")";
+
+        if (n_devs > 1) {
+            ggml_backend_comm_init_t comm_init = (ggml_backend_comm_init_t) ggml_backend_reg_get_proc_address(
+                ggml_backend_dev_backend_reg(ggml_backend_get_device(simple_backends[0])), "ggml_backend_comm_init");
+            if (comm_init != nullptr) {
+                comm_ctx = comm_init(simple_backends.data(), simple_backends.size());
+            }
+        }
+        if (comm_ctx != nullptr) {
+            comm_allreduce = (ggml_backend_comm_allreduce_tensor_t)
+                ggml_backend_reg_get_proc_address(ggml_backend_dev_backend_reg(
+                    ggml_backend_get_device(simple_backends[0])), "ggml_backend_comm_allreduce_tensor");
+            GGML_ASSERT(comm_allreduce != nullptr);
+        }
     }
 
     ~ggml_backend_meta_context() {
+        if (comm_ctx != nullptr) {
+            ggml_backend_comm_free_t comm_free = (ggml_backend_comm_free_t) ggml_backend_reg_get_proc_address(
+                ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_configs[0].backend)), "ggml_backend_comm_free");
+            GGML_ASSERT(comm_free != nullptr);
+            comm_free(comm_ctx);
+        }
         for (auto & bc : backend_configs) {
             ggml_backend_free(bc.backend);
         }
     }
-
-    size_t n_reduce_steps() const {
-        return std::ceil(std::log2(backend_configs.size()));
-    }
 };
 
 static const char * ggml_backend_meta_get_name(ggml_backend_t backend) {
@@ -1552,6 +1640,9 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     const size_t n_backends = ggml_backend_meta_n_backends(backend);
     ggml_backend_meta_context * backend_ctx = (ggml_backend_meta_context *) backend->context;
 
+    // If the previous cgraph had a defined UID it can be used to skip rebuilding the subgraphs per simple backend.
+    const bool needs_rebuild = (cgraph->uid == 0) || (cgraph->uid != backend_ctx->uid);
+
     bool max_nnodes_raised = false;
     if (cgraph->n_nodes > backend_ctx->max_nnodes) {
         for (size_t j = 0; j < n_backends; j++) {
@@ -1561,173 +1652,182 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         backend_ctx->max_nnodes = cgraph->n_nodes;
         max_nnodes_raised = true;
+        assert(needs_rebuild);
     }
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
 
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
-                // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
-                bcj.nodes[i] = node;
-                continue;
+    if (needs_rebuild) {
+        size_t n_subgraphs  = 0;
+        size_t max_tmp_size = 0;
+
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    // FIXME s_copy_main is on the CPU and its view seems to be incorrectly added to the graph nodes.
+                    // For regular usage this doesn't matter since it's a noop but trying to call ggml_backend_meta_buffer_simple_tensor results in a crash.
+                    bcj.nodes[i] = node;
+                    continue;
+                }
+                bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
+                GGML_ASSERT(bcj.nodes[i]);
             }
-            bcj.nodes[i] = ggml_backend_meta_buffer_simple_tensor(node, j);
-            GGML_ASSERT(bcj.nodes[i]);
         }
-    }
 
-    size_t n_subgraphs  = 0;
-    size_t max_tmp_size = 0;
-    {
-        // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
-        auto get_i_delayed = [&](const int i) -> int {
-            int id = i; // i_delayed
-            int idr = i; // i_delayed return, last safe return value
+        {
+            // For MoE models it may make sense to delay the AllReduce in order to reduce I/O:
+            auto get_i_delayed = [&](const int i) -> int {
+                int id = i; // i_delayed
+                int idr = i; // i_delayed return, last safe return value
 
-            ggml_tensor * node = cgraph->nodes[id];
-            int32_t n_used = ggml_node_get_use_count(cgraph, id);
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
-                        ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                ggml_tensor * node = cgraph->nodes[id];
+                int32_t n_used = ggml_node_get_use_count(cgraph, id);
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_ADD_ID && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL &&
+                            ggml_backend_meta_get_split_state(next->src[2], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+                if (id + 1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op == GGML_OP_MUL && next->src[0] == node &&
+                            ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
+                        node = next;
+                        id++;
+                        idr = id;
+                        n_used = ggml_node_get_use_count(cgraph, id);
+                    }
+                }
+
+                if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                    return idr;
+                }
+                for (int32_t k = 0; k < n_used; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
+                            next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
+                            ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-            if (id + 1 >= cgraph->n_nodes) {
-                return idr;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op == GGML_OP_MUL && next->src[0] == node &&
-                        ggml_backend_meta_get_split_state(next->src[1], false).axis == GGML_BACKEND_SPLIT_AXIS_MIRRORED) {
-                    node = next;
+                {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
                     id++;
-                    idr = id;
-                    n_used = ggml_node_get_use_count(cgraph, id);
                 }
-            }
-
-            if (n_used != node->ne[1] || id + 2*n_used-1 >= cgraph->n_nodes) {
+                for (int32_t k = 0; k < n_used - 2; k++) {
+                    ggml_tensor * next = cgraph->nodes[id+1];
+                    if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
+                            next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
+                        return idr;
+                    }
+                    id++;
+                }
+                idr = id;
                 return idr;
-            }
-            for (int32_t k = 0; k < n_used; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_VIEW || next->view_src != node || next->view_offs != k*node->nb[1] ||
-                        next->ne[0] != node->ne[0] || next->ne[1] != node->ne[2] || next->nb[1] != node->nb[2] ||
-                        ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id - (n_used-1)] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            for (int32_t k = 0; k < n_used - 2; k++) {
-                ggml_tensor * next = cgraph->nodes[id+1];
-                if (next->op != GGML_OP_ADD || next->src[0] != cgraph->nodes[id] ||
-                        next->src[1] != cgraph->nodes[id - (n_used-2)] || ggml_node_get_use_count(cgraph, id+1) != 1) {
-                    return idr;
-                }
-                id++;
-            }
-            idr = id;
-            return idr;
-        };
+            };
 
-        int i_start = 0;
-        for (int i = 0; i < cgraph->n_nodes; i++) {
-            ggml_tensor * node = cgraph->nodes[i];
-            if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
-                continue;
-            }
-            const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
-            if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
-                max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
-            }
-            const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
-            if (!new_subgraph) {
-                continue;
-            }
+            int i_start = 0;
+            for (int i = 0; i < cgraph->n_nodes; i++) {
+                ggml_tensor * node = cgraph->nodes[i];
+                if (node->view_src != nullptr && node->view_src->op == GGML_OP_NONE && ggml_backend_buffer_is_host(node->view_src->buffer)) {
+                    continue;
+                }
+                const ggml_backend_meta_split_state split_state = ggml_backend_meta_get_split_state(node, /*assume_sync =*/ false);
+                if (split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL) {
+                    max_tmp_size = std::max(max_tmp_size, ggml_nbytes(node));
+                }
+                const bool new_subgraph = i + 1 == cgraph->n_nodes || split_state.axis == GGML_BACKEND_SPLIT_AXIS_PARTIAL;
+                if (!new_subgraph) {
+                    continue;
+                }
 
-            i = get_i_delayed(i);
+                i = get_i_delayed(i);
 
+                for (size_t j = 0; j < n_backends; j++) {
+                    auto & bcj = backend_ctx->backend_configs[j];
+                    bcj.cgraphs[n_subgraphs].offset = i_start;
+                }
+                n_subgraphs++;
+                i_start = i + 1;
+            }
+            GGML_ASSERT(i_start == cgraph->n_nodes);
+        }
+
+        backend_ctx->uid         = cgraph->uid;
+        backend_ctx->n_subgraphs = n_subgraphs;
+
+        if (max_tmp_size > backend_ctx->max_tmp_size) {
             for (size_t j = 0; j < n_backends; j++) {
                 auto & bcj = backend_ctx->backend_configs[j];
-                bcj.cgraphs[n_subgraphs].offset = i_start;
+                for (size_t i = 0; i < backend_ctx->n_reduce_steps; i++) {
+                    bcj.bufs[i].reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
+                }
             }
-            n_subgraphs++;
-            i_start = i + 1;
+            backend_ctx->max_tmp_size = max_tmp_size;
         }
-        GGML_ASSERT(i_start == cgraph->n_nodes);
-    }
 
-    if (max_tmp_size > backend_ctx->max_tmp_size) {
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            bcj.buf.reset(ggml_backend_alloc_buffer(bcj.backend, max_tmp_size));
-        }
-        backend_ctx->max_tmp_size = max_tmp_size;
-    }
-
-
-    if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
-        backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
-        const size_t n_reduce_steps = backend_ctx->n_reduce_steps();
-        const size_t n_nodes_per_device = 2 * n_reduce_steps; // tmp + ADD per step
-        const size_t n_cgraphs_per_device = n_reduce_steps;    // 1 ADD graph per step
-        const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
-        const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
-        const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
-        ggml_init_params params = {
-            /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
-            /*.mem_buffer =*/ nullptr,
-            /*.no_alloc   =*/ true,
-        };
-        backend_ctx->ctx.reset(ggml_init(params));
-        for (size_t j = 0; j < n_backends; j++) {
-            auto & bcj = backend_ctx->backend_configs[j];
-            for (size_t i = 0; i < n_subgraphs; i++) {
-                bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+        if (max_nnodes_raised || n_subgraphs > backend_ctx->max_subgraphs) {
+            backend_ctx->max_subgraphs = std::max(backend_ctx->max_subgraphs, n_subgraphs);
+            const size_t n_nodes_per_device = 3 * backend_ctx->n_reduce_steps; // tmp + ADD (+zeroing) graph per step and device
+            const size_t n_cgraphs_per_device = 2 * backend_ctx->n_reduce_steps; // ADD ( + zeroing) graph per step and device
+            const size_t mem_per_device_graphs_main = backend_ctx->max_subgraphs*ggml_graph_overhead_custom(backend_ctx->max_nnodes, cgraph->grads);
+            const size_t mem_per_device_graphs_aux = n_cgraphs_per_device*backend_ctx->max_subgraphs*ggml_graph_overhead_custom(1, cgraph->grads);
+            const size_t mem_per_device_nodes_aux = n_nodes_per_device*backend_ctx->max_subgraphs*ggml_tensor_overhead();
+            ggml_init_params params = {
+                /*.mem_size   =*/ n_backends * (mem_per_device_graphs_main + mem_per_device_graphs_aux + mem_per_device_nodes_aux),
+                /*.mem_buffer =*/ nullptr,
+                /*.no_alloc   =*/ true,
+            };
+            backend_ctx->ctx.reset(ggml_init(params));
+            for (size_t j = 0; j < n_backends; j++) {
+                auto & bcj = backend_ctx->backend_configs[j];
+                for (size_t i = 0; i < n_subgraphs; i++) {
+                    bcj.cgraphs[i].cgraph_main = ggml_new_graph_custom(backend_ctx->ctx.get(), cgraph->n_nodes, /*grads =*/ false);
+                }
+            }
+            backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
+                backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
+            }
+            backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
+            for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
+                backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
             }
         }
-        backend_ctx->cgraphs_aux.resize(n_backends*n_cgraphs_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->cgraphs_aux.size(); k++) {
-            backend_ctx->cgraphs_aux[k] = ggml_new_graph_custom(backend_ctx->ctx.get(), 1, cgraph->grads);
-        }
-        backend_ctx->nodes_aux.resize(n_backends*n_nodes_per_device*backend_ctx->max_subgraphs);
-        for (size_t k = 0; k < backend_ctx->nodes_aux.size(); k++) {
-            backend_ctx->nodes_aux[k] = ggml_new_tensor_1d(backend_ctx->ctx.get(), GGML_TYPE_F32, 1);
-        }
-    }
 
-    for (size_t j = 0; j < n_backends; j++) {
-        auto & bcj = backend_ctx->backend_configs[j];
-        for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
-            ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
-            const size_t i_node_start = bcj.cgraphs[i_graph].offset;
-            const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
-            cgraph_ij->n_nodes = i_node_stop - i_node_start;
-            ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
-            for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
-                ggml_tensor * node_ij = bcj.nodes[i_node];
-                cgraph_ij->nodes[i_node - i_node_start] = node_ij;
-                const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
-                const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
-                cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            for (size_t i_graph = 0; i_graph < n_subgraphs; i_graph++) {
+                ggml_cgraph * cgraph_ij = bcj.cgraphs[i_graph].cgraph_main;
+                const size_t i_node_start = bcj.cgraphs[i_graph].offset;
+                const size_t i_node_stop = i_graph + 1 < n_subgraphs ? bcj.cgraphs[i_graph + 1].offset : cgraph->n_nodes;
+                cgraph_ij->n_nodes = i_node_stop - i_node_start;
+                ggml_hash_set_reset(&cgraph_ij->visited_hash_set);
+                for (size_t i_node = i_node_start; i_node < i_node_stop; i_node++) {
+                    ggml_tensor * node_ij = bcj.nodes[i_node];
+                    cgraph_ij->nodes[i_node - i_node_start] = node_ij;
+                    const size_t hash_pos_orig = ggml_hash_find(&cgraph->visited_hash_set, cgraph->nodes[i_node]);
+                    const size_t hash_pos_ij = ggml_hash_insert(&cgraph_ij->visited_hash_set, node_ij);
+                    cgraph_ij->use_counts[hash_pos_ij] = cgraph->use_counts[hash_pos_orig];
+                }
+                cgraph_ij->uid = ggml_graph_next_uid();
             }
         }
     }
@@ -1735,11 +1835,6 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
     size_t iga = 0; // i graph aux
     size_t ina = 0; // i node aux
 
-    // FIXME usage_counts
-    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
-        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
-        return ret;
-    };
     auto get_node_aux = [&](ggml_tensor * t) -> ggml_tensor * {
         ggml_tensor * ret = backend_ctx->nodes_aux[ina++];
         memset(ret, 0, sizeof(ggml_tensor));
@@ -1751,75 +1846,110 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
         }
         return ret;
     };
+    auto set_tmp_data = [&](ggml_tensor * tensor, const size_t j, const size_t i_buf) {
+        auto & bcj = backend_ctx->backend_configs[j];
+        ggml_backend_buffer_ptr & buf_ptr = bcj.bufs[i_buf];
+        if (!buf_ptr || ggml_backend_buffer_get_size(buf_ptr.get()) < backend_ctx->max_tmp_size) {
+            buf_ptr.reset(ggml_backend_alloc_buffer(bcj.backend, backend_ctx->max_tmp_size));
+        }
+        tensor->buffer = buf_ptr.get();
+        tensor->data   = ggml_backend_buffer_get_base(buf_ptr.get());
+    };
+    // FIXME usage_counts
+    auto get_cgraph_aux = [&]() -> ggml_cgraph * {
+        ggml_cgraph * ret = backend_ctx->cgraphs_aux[iga++];
+        return ret;
+    };
 
     // Preferentially use backend-specific allreduce_tensor_async (e.g. NCCL for CUDA), use a generic fallback if unavailable:
     auto allreduce_fallback = [&](size_t i) -> ggml_status {
         std::vector<ggml_cgraph *> step_cgraphs(n_backends, nullptr);
 
-        for (size_t offset_j = 1; offset_j < n_backends; offset_j *= 2) {
+        // Zero out nodes that were disabled due to having a zero-sized slice:
+        for (size_t j = 0; j < n_backends; j++) {
+            auto & bcj = backend_ctx->backend_configs[j];
+            ggml_tensor * node = bcj.cgraphs[i].cgraph_main->nodes[bcj.cgraphs[i].cgraph_main->n_nodes - 1];
+            if (node->flags & GGML_TENSOR_FLAG_COMPUTE) {
+                continue;
+            }
+            ggml_tensor * node_zero = get_node_aux(node);
+            node_zero->op = GGML_OP_SCALE; // FIXME 0.0f * NaN == NaN
+            node_zero->src[0] = node;
+            ggml_set_op_params_f32(node_zero, 0, 0.0f);
+            node_zero->data = node->data;
+            node_zero->flags |= GGML_TENSOR_FLAG_COMPUTE;
+
+            step_cgraphs[j] = get_cgraph_aux();
+            step_cgraphs[j]->nodes[0] = node_zero;
+            step_cgraphs[j]->n_nodes = 1;
+            const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, step_cgraphs[j]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+        }
+        std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
+
+        auto push_data = [&](const size_t j_src, const size_t j_dst, const size_t i_buf) {
+            assert(step_cgraphs[j_dst] == nullptr);
+            auto & bcj_src = backend_ctx->backend_configs[j_src];
+            auto & bcj_dst = backend_ctx->backend_configs[j_dst];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            GGML_ASSERT(ggml_is_contiguous(node_src));
+            GGML_ASSERT(ggml_is_contiguous(node_dst));
+
+            ggml_tensor * node_tmp = get_node_aux(node_dst);
+            set_tmp_data(node_tmp, j_dst, i_buf);
+
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_tmp);
+
+            ggml_tensor * node_red = get_node_aux(node_dst);
+            node_red->view_src = node_dst->view_src == nullptr ? node_dst : node_dst->view_src;
+            node_red->view_offs = node_dst->view_offs;
+            node_red->op = GGML_OP_ADD;
+            node_red->src[0] = node_dst;
+            node_red->src[1] = node_tmp;
+            node_red->flags |= GGML_TENSOR_FLAG_COMPUTE;
+            ggml_backend_view_init(node_red);
+
+            ggml_cgraph * cgraph_aux = get_cgraph_aux();
+            cgraph_aux->nodes[0] = node_red;
+            cgraph_aux->n_nodes = 1;
+            step_cgraphs[j_dst] = cgraph_aux;
+        };
+
+        size_t offset_j = n_backends/2;
+        while ((offset_j & (offset_j - 1)) != 0) {
+            offset_j--;
+        }
+        const size_t offset_j_max = offset_j;
+        size_t i_buf = 0;
+
+        // If n_backends is not a power of 2, fold in the excess prior to butterfly reduction:
+        for (size_t j_src = 2*offset_j_max; j_src < n_backends; j_src++) {
+            const size_t j_dst = j_src - 2*offset_j_max;
+            push_data(j_src, j_dst, i_buf);
+            const ggml_status status = ggml_backend_graph_compute_async(backend_ctx->backend_configs[j_dst].backend, step_cgraphs[j_dst]);
+            if (status != GGML_STATUS_SUCCESS) {
+                return status;
+            }
+            i_buf = 1;
+        }
+
+        // Butterfly reduction:
+        for (; offset_j >= 1; offset_j /= 2) {
             std::fill(step_cgraphs.begin(), step_cgraphs.end(), nullptr);
 
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 const size_t j_other = j ^ offset_j;
-                if (j_other > j) {
+                if (j_other >= n_backends) {
                     continue;
                 }
-
-                auto & bcj1 = backend_ctx->backend_configs[j];
-                auto & bcj2 = backend_ctx->backend_configs[j_other];
-
-                ggml_tensor * node1 = bcj1.cgraphs[i].cgraph_main->nodes[bcj1.cgraphs[i].cgraph_main->n_nodes - 1];
-                ggml_tensor * node2 = bcj2.cgraphs[i].cgraph_main->nodes[bcj2.cgraphs[i].cgraph_main->n_nodes - 1];
-                GGML_ASSERT(ggml_is_contiguous(node1));
-                GGML_ASSERT(ggml_is_contiguous(node2));
-
-                // Tmp tensors to receive P2P copies
-                ggml_tensor * node_tmp_1 = get_node_aux(node1);
-                node_tmp_1->buffer = bcj1.buf.get();
-                node_tmp_1->data = ggml_backend_buffer_get_base(bcj1.buf.get());
-
-                ggml_tensor * node_tmp_2 = get_node_aux(node2);
-                node_tmp_2->buffer = bcj2.buf.get();
-                node_tmp_2->data = ggml_backend_buffer_get_base(bcj2.buf.get());
-
-                // 2 P2P copies: exchange full buffers
-                ggml_backend_tensor_copy_async(bcj1.backend, bcj2.backend, node1, node_tmp_2);
-                ggml_backend_tensor_copy_async(bcj2.backend, bcj1.backend, node2, node_tmp_1);
-
-                // Local ADD: node1 += tmp1 (in-place via view)
-                ggml_tensor * node_red_1 = get_node_aux(node1);
-                node_red_1->view_src = node1->view_src == nullptr ? node1 : node1->view_src;
-                node_red_1->view_offs = node1->view_offs;
-                node_red_1->op = GGML_OP_ADD;
-                node_red_1->src[0] = node1;
-                node_red_1->src[1] = node_tmp_1;
-                node_red_1->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_1);
-
-                // Local ADD: node2 += tmp2 (in-place via view)
-                ggml_tensor * node_red_2 = get_node_aux(node2);
-                node_red_2->view_src = node2->view_src == nullptr ? node2 : node2->view_src;
-                node_red_2->view_offs = node2->view_offs;
-                node_red_2->op = GGML_OP_ADD;
-                node_red_2->src[0] = node2;
-                node_red_2->src[1] = node_tmp_2;
-                node_red_2->flags |= GGML_TENSOR_FLAG_COMPUTE;
-                ggml_backend_view_init(node_red_2);
-
-                // Build 1-node cgraphs for the ADD ops
-                ggml_cgraph * cgraph_aux_1 = get_cgraph_aux();
-                cgraph_aux_1->nodes[0] = node_red_1;
-                cgraph_aux_1->n_nodes = 1;
-                step_cgraphs[j] = cgraph_aux_1;
-
-                ggml_cgraph * cgraph_aux_2 = get_cgraph_aux();
-                cgraph_aux_2->nodes[0] = node_red_2;
-                cgraph_aux_2->n_nodes = 1;
-                step_cgraphs[j_other] = cgraph_aux_2;
+                push_data(j, j_other, i_buf);
             }
 
-            // Execute local ADDs for this step
-            for (size_t j = 0; j < n_backends; j++) {
+            for (size_t j = 0; j < 2*offset_j_max; j++) {
                 if (step_cgraphs[j] == nullptr) {
                     continue;
                 }
@@ -1829,12 +1959,25 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
                     return status;
                 }
             }
+            i_buf++;
         }
+        assert(i_buf == backend_ctx->n_reduce_steps);
+
+        // If n_backends is not a power of 2, copy back the reduced tensors to the excess:
+        for (size_t j = 2*offset_j_max; j < n_backends; j++) {
+            auto & bcj_src = backend_ctx->backend_configs[j - 2*offset_j_max];
+            auto & bcj_dst = backend_ctx->backend_configs[j];
+
+            ggml_tensor * node_src = bcj_src.cgraphs[i].cgraph_main->nodes[bcj_src.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_tensor * node_dst = bcj_dst.cgraphs[i].cgraph_main->nodes[bcj_dst.cgraphs[i].cgraph_main->n_nodes - 1];
+            ggml_backend_tensor_copy_async(bcj_src.backend, bcj_dst.backend, node_src, node_dst);
+        }
+
         return GGML_STATUS_SUCCESS;
     };
 
 
-    for (size_t i = 0; i < n_subgraphs; i++) {
+    for (size_t i = 0; i < backend_ctx->n_subgraphs; i++) {
         for (size_t j = 0; j < n_backends; j++) {
             auto & bcj = backend_ctx->backend_configs[j];
             const ggml_status status = ggml_backend_graph_compute_async(bcj.backend, bcj.cgraphs[i].cgraph_main);
@@ -1843,22 +1986,17 @@ static enum ggml_status ggml_backend_meta_graph_compute(ggml_backend_t backend,
             }
         }
 
-        if (n_backends > 1 && i < n_subgraphs - 1) {
+        if (n_backends > 1 && i < backend_ctx->n_subgraphs - 1) {
             bool backend_allreduce_success = false;
-            ggml_backend_allreduce_tensor_t allreduce_tensor = (ggml_backend_allreduce_tensor_t) ggml_backend_reg_get_proc_address(
-                ggml_backend_dev_backend_reg(ggml_backend_get_device(backend_ctx->backend_configs[0].backend)), "ggml_backend_allreduce_tensor");
-            if (allreduce_tensor) {
-                std::vector<ggml_backend_t> backends;
-                backends.reserve(n_backends);
+            if (backend_ctx->comm_ctx) {
                 std::vector<ggml_tensor *> nodes;
                 nodes.reserve(n_backends);
                 for (size_t j = 0; j < n_backends; j++) {
                     auto & bcj = backend_ctx->backend_configs[j];
-                    backends.push_back(bcj.backend);
                     ggml_cgraph * cgraph_ij = bcj.cgraphs[i].cgraph_main;
                     nodes.push_back(cgraph_ij->nodes[cgraph_ij->n_nodes-1]);
                 }
-                backend_allreduce_success = allreduce_tensor(backends.data(), nodes.data(), n_backends);
+                backend_allreduce_success = backend_ctx->comm_allreduce(backend_ctx->comm_ctx, nodes.data());
             }
 
             if (!backend_allreduce_success) {

ggml/src/ggml-backend.cpp

@@ -1030,6 +1030,8 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         GGML_ABORT("%s: failed to initialize context\n", __func__);
     }
 
+    graph->uid = ggml_graph_next_uid();
+
     // pass 1: assign backends to ops with pre-allocated inputs
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
@@ -1477,6 +1479,11 @@ void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgra
         assert(graph_copy->size > graph_copy->n_leafs);
         graph_copy->leafs[graph_copy->n_leafs++] = leaf;
     }
+
+    // set ids for all splits
+    for (int i = 0; i < sched->n_splits; ++i) {
+        sched->splits[i].graph.uid = ggml_graph_next_uid();
+    }
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {

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

@@ -83,7 +83,6 @@
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // quants.c
 #define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
-#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
 // repack.cpp
 #define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
 #define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4

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

@@ -151,8 +151,6 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
     const block_q1_0 * GGML_RESTRICT x = vx;
     const block_q8_0 * GGML_RESTRICT y = vy;
 
-    float sumf = 0.0f;
-
 #if defined(__ARM_NEON)
     float32x4_t sumv = vdupq_n_f32(0.0f);
 
@@ -212,31 +210,13 @@ void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const voi
         }
     }
 
-    sumf = vaddvq_f32(sumv);
+    *s = vaddvq_f32(sumv);
 #else
-    // Scalar fallback
-    for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
-
-        // Process 4 Q8_0 blocks
-        for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
-            int sumi = 0;
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
-
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi += xi * y[i*4 + k].qs[j];
-            }
-            sumf += d0 * d1 * sumi;
-        }
-    }
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
 #endif
-
-    *s = sumf;
 }
 
 

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

@@ -274,6 +274,18 @@ static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const
 }
 #endif
 #elif defined(__SSSE3__)
+static inline __m128i bytes_from_bits_16(const uint8_t * x) {
+    uint16_t x16;
+    memcpy(&x16, x, sizeof(uint16_t));
+
+    const __m128i shuf_mask = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+    __m128i bytes = _mm_shuffle_epi8(_mm_set1_epi16((short) x16), shuf_mask);
+    const __m128i bit_mask = _mm_set_epi64x(0x7fbfdfeff7fbfdfe, 0x7fbfdfeff7fbfdfe);
+    bytes = _mm_or_si128(bytes, bit_mask);
+
+    return _mm_cmpeq_epi8(bytes, _mm_set1_epi64x(-1));
+}
+
 // horizontally add 4x4 floats
 static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
     __m128 res_0 =_mm_hadd_ps(a, b);
@@ -540,6 +552,152 @@ static inline __m128i get_scale_shuffle(int i) {
 }
 #endif
 
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    const int qk = QK1_0;
+    const int nb = n / qk;
+
+    assert(n % qk == 0);
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const block_q1_0 * GGML_RESTRICT x = vx;
+    const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+    const __m256i ones_8 = _mm256_set1_epi8(1);
+    const __m256i ones_16 = _mm256_set1_epi16(1);
+    const __m256i byte_shuf = _mm256_setr_epi8(
+            0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+            2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
+    const __m256i bit_masks = _mm256_setr_epi8(
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128,
+            1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128);
+    const __m256i zero = _mm256_setzero_si256();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const uint32_t * GGML_RESTRICT qs32 = (const uint32_t *) x[ib].qs;
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+        __m256 acc_block;
+        {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[0].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[0]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), _mm256_cvtepi32_ps(s32));
+        }
+        for (int K = 1; K < 4; ++K) {
+            const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[K].qs);
+            const __m256i sm = _mm256_cmpeq_epi8(
+                    _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[K]), byte_shuf), bit_masks), zero);
+            const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+            const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+            acc_block = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[K].d)), _mm256_cvtepi32_ps(s32), acc_block);
+        }
+        acc = _mm256_fmadd_ps(_mm256_set1_ps(d0), acc_block, acc);
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m256 acc = _mm256_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+        __m256 acc_block;
+        {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[0]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), q);
+        }
+        for(int K = 1; K < 4; ++K) {
+            const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[(K) * 4]);
+            const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+            const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[0]);
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[16]);
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+            const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+            const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+            const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+            const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+            const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+            acc_block = _mm256_add_ps(acc_block, _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(K)].d)), q));
+        }
+#undef Q1_AVX_BLOCK
+
+        acc = _mm256_add_ps(acc, _mm256_mul_ps(_mm256_set1_ps(d0), acc_block));
+    }
+
+    *s = hsum_float_8(acc);
+#elif defined(__SSSE3__)
+    const __m128i ones_8 = _mm_set1_epi8(1);
+    const __m128i ones_16 = _mm_set1_epi16(1);
+    const __m128i zero = _mm_setzero_si128();
+    __m128 acc_0 = _mm_setzero_ps();
+    __m128 acc_1 = _mm_setzero_ps();
+    __m128 acc_2 = _mm_setzero_ps();
+    __m128 acc_3 = _mm_setzero_ps();
+
+    for (int ib = 0; ib < nb; ++ib) {
+        const __m128 d0 = _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+        const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+#define Q1_SSSE3_BLOCK(QS_OFF, Y_IDX, ACC) \
+        { \
+            const __m128i bit_mask_0 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 0]); \
+            const __m128i bit_mask_1 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 2]); \
+            const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[0]); \
+            const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[16]); \
+            const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero); \
+            const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero); \
+            const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0); \
+            const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1); \
+            const __m128i sum_0 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_0), ones_16); \
+            const __m128i sum_1 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_1), ones_16); \
+            const __m128 q = _mm_cvtepi32_ps(_mm_add_epi32(sum_0, sum_1)); \
+            (ACC) = _mm_add_ps((ACC), _mm_mul_ps(_mm_mul_ps(d0, _mm_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(Y_IDX)].d))), q)); \
+        }
+        Q1_SSSE3_BLOCK(0,  0, acc_0)
+        Q1_SSSE3_BLOCK(4,  1, acc_1)
+        Q1_SSSE3_BLOCK(8,  2, acc_2)
+        Q1_SSSE3_BLOCK(12, 3, acc_3)
+#undef Q1_SSSE3_BLOCK
+    }
+
+    *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#else
+    UNUSED(nb);
+    UNUSED(x);
+    UNUSED(y);
+    ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/quants.c

@@ -137,22 +137,28 @@ void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, c
     float sumf = 0.0;
 
     for (int i = 0; i < nb; i++) {
-        const float d0 = GGML_FP16_TO_FP32(x[i].d);
+        const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
 
         float sumi = 0.0f;
 
         for (int k = 0; k < 4; k++) {
-            const float d1 = GGML_FP16_TO_FP32(y[i*4 + k].d);
-
+            const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+            const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
             int sumi_block = 0;
 
-            for (int j = 0; j < QK8_0; j++) {
-                const int bit_index = k * QK8_0 + j;
-                const int byte_index = bit_index / 8;
-                const int bit_offset = bit_index % 8;
+            const uint8_t * GGML_RESTRICT bits = &x[i].qs[k * 4];
+            const int8_t  * GGML_RESTRICT qy   = yb->qs;
 
-                const int xi = ((x[i].qs[byte_index] >> bit_offset) & 1) ? 1 : -1;
-                sumi_block += xi * y[i*4 + k].qs[j];
+            for (int b = 0; b < 4; ++b, qy += 8) {
+                const unsigned mask = bits[b];
+                sumi_block += ((mask & 0x01) ? qy[0] : -qy[0])
+                           +  ((mask & 0x02) ? qy[1] : -qy[1])
+                           +  ((mask & 0x04) ? qy[2] : -qy[2])
+                           +  ((mask & 0x08) ? qy[3] : -qy[3])
+                           +  ((mask & 0x10) ? qy[4] : -qy[4])
+                           +  ((mask & 0x20) ? qy[5] : -qy[5])
+                           +  ((mask & 0x40) ? qy[6] : -qy[6])
+                           +  ((mask & 0x80) ? qy[7] : -qy[7]);
             }
 
             sumi += d1 * sumi_block;

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

@@ -109,6 +109,96 @@ static void simd_gemm(
         C += N;
     }
 }
+#elif defined(GGML_SIMD) && defined(__riscv_v_intrinsic)
+// RM accumulators + 1 B vector = RM + 1 <= 8  =>  RM <= 7
+// Microkernel: C[RM x vl] += A[RM x K] * B[K x N]
+template <int RM>
+static inline void rvv_simd_gemm_ukernel(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N, size_t vl)
+{
+    static_assert(RM >= 1 && RM <= 7, "RM must be 1..7 for LMUL=4");
+
+    vfloat32m4_t acc_0 = __riscv_vle32_v_f32m4(C + 0 * N, vl);
+    vfloat32m4_t acc_1, acc_2, acc_3, acc_4, acc_5, acc_6;
+    if constexpr (RM > 1) acc_1 = __riscv_vle32_v_f32m4(C + 1 * N, vl);
+    if constexpr (RM > 2) acc_2 = __riscv_vle32_v_f32m4(C + 2 * N, vl);
+    if constexpr (RM > 3) acc_3 = __riscv_vle32_v_f32m4(C + 3 * N, vl);
+    if constexpr (RM > 4) acc_4 = __riscv_vle32_v_f32m4(C + 4 * N, vl);
+    if constexpr (RM > 5) acc_5 = __riscv_vle32_v_f32m4(C + 5 * N, vl);
+    if constexpr (RM > 6) acc_6 = __riscv_vle32_v_f32m4(C + 6 * N, vl);
+
+    for (int kk = 0; kk < K; kk++) {
+        vfloat32m4_t b_0 = __riscv_vle32_v_f32m4(B + kk * N, vl);
+
+                              acc_0 = __riscv_vfmacc_vf_f32m4(acc_0, A[0 * K + kk], b_0, vl);
+        if constexpr (RM > 1) acc_1 = __riscv_vfmacc_vf_f32m4(acc_1, A[1 * K + kk], b_0, vl);
+        if constexpr (RM > 2) acc_2 = __riscv_vfmacc_vf_f32m4(acc_2, A[2 * K + kk], b_0, vl);
+        if constexpr (RM > 3) acc_3 = __riscv_vfmacc_vf_f32m4(acc_3, A[3 * K + kk], b_0, vl);
+        if constexpr (RM > 4) acc_4 = __riscv_vfmacc_vf_f32m4(acc_4, A[4 * K + kk], b_0, vl);
+        if constexpr (RM > 5) acc_5 = __riscv_vfmacc_vf_f32m4(acc_5, A[5 * K + kk], b_0, vl);
+        if constexpr (RM > 6) acc_6 = __riscv_vfmacc_vf_f32m4(acc_6, A[6 * K + kk], b_0, vl);
+    }
+
+                          __riscv_vse32_v_f32m4(C + 0 * N, acc_0, vl);
+    if constexpr (RM > 1) __riscv_vse32_v_f32m4(C + 1 * N, acc_1, vl);
+    if constexpr (RM > 2) __riscv_vse32_v_f32m4(C + 2 * N, acc_2, vl);
+    if constexpr (RM > 3) __riscv_vse32_v_f32m4(C + 3 * N, acc_3, vl);
+    if constexpr (RM > 4) __riscv_vse32_v_f32m4(C + 4 * N, acc_4, vl);
+    if constexpr (RM > 5) __riscv_vse32_v_f32m4(C + 5 * N, acc_5, vl);
+    if constexpr (RM > 6) __riscv_vse32_v_f32m4(C + 6 * N, acc_6, vl);
+}
+
+template <int RM>
+static inline void rvv_simd_gemm_dispatch_tail(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int K, int N, int KN, int remaining_rows)
+{
+    if constexpr (RM > 0) {
+        if (remaining_rows == RM) {
+            int64_t jj = 0;
+            for (; jj + KN <= N; jj += KN) {
+                rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, KN);
+            }
+            if (jj < N) {
+                rvv_simd_gemm_ukernel<RM>(C + jj, A, B + jj, K, N, N - jj);
+            }
+        } else {
+            rvv_simd_gemm_dispatch_tail<RM - 1>(C, A, B, K, N, KN, remaining_rows);
+        }
+    }
+}
+
+static constexpr int GEMM_RM = 7;
+
+// C[M x N] += A[M x K] * B[K x N]
+static void simd_gemm(
+    float       * GGML_RESTRICT C,
+    const float * GGML_RESTRICT A,
+    const float * GGML_RESTRICT B,
+    int M, int K, int N)
+{
+    const int KN = (int)__riscv_vlenb();
+    int64_t ii = 0;
+    for (; ii + GEMM_RM <= M; ii += GEMM_RM) {
+        int64_t jj = 0;
+        for (; jj + KN <= N; jj += KN) {
+            rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, KN);
+        }
+        if (jj < N) {
+            rvv_simd_gemm_ukernel<GEMM_RM>(C + jj, A, B + jj, K, N, N - jj);
+        }
+        A += GEMM_RM * K;
+        C += GEMM_RM * N;
+    }
+
+    int remaining_rows = M - ii;
+    rvv_simd_gemm_dispatch_tail<GEMM_RM - 1>(C, A, B, K, N, KN, remaining_rows);
+}
 
 #if defined(__GNUC__) && !defined(__clang__)
 #pragma GCC diagnostic pop

ggml/src/ggml-cuda/common.cuh

@@ -269,10 +269,6 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
-#if defined(TURING_MMA_AVAILABLE)
-#define LDMATRIX_TRANS_AVAILABLE
-#endif // defined(TURING_MMA_AVAILABLE)
-
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);
@@ -924,6 +920,13 @@ struct ggml_cuda_type_traits<GGML_TYPE_F16> {
     static constexpr int qr = 1;
 };
 
+template<>
+struct ggml_cuda_type_traits<GGML_TYPE_Q1_0> {
+    static constexpr int qk = QK1_0;
+    static constexpr int qr = QR1_0;
+    static constexpr int qi = QI1_0;
+};
+
 template<>
 struct ggml_cuda_type_traits<GGML_TYPE_Q4_0> {
     static constexpr int qk = QK4_0;
@@ -1092,10 +1095,6 @@ struct ggml_cuda_device_info {
     cuda_device_info devices[GGML_CUDA_MAX_DEVICES] = {};
 
     std::array<float, GGML_CUDA_MAX_DEVICES> default_tensor_split = {};
-
-#ifdef GGML_USE_NCCL
-    ncclComm_t comms[GGML_CUDA_MAX_DEVICES];
-#endif // GGML_USE_NCCL
 };
 
 const ggml_cuda_device_info & ggml_cuda_info();
@@ -1183,6 +1182,8 @@ struct ggml_cuda_graph {
     std::vector<cudaGraphNode_t> nodes;
     bool disable_due_to_gpu_arch = false;
     bool warmup_complete = false;
+    uint64_t uid = 0;
+    int64_t last_used_time = 0;
     struct node_properties {
         ggml_tensor node;
         void *   node_src_data_ptrs[GGML_MAX_SRC];
@@ -1364,12 +1365,28 @@ struct ggml_backend_cuda_context {
     // when the computation is split across CPU/GPU (e.g., with --n-cpu-moe)
     std::unordered_map<const void *, std::unique_ptr<ggml_cuda_graph>> cuda_graphs;
 
+    int64_t last_graph_eviction_sweep = 0;
+
     ggml_cuda_graph * cuda_graph(const void * first_node_ptr) {
+        const int64_t time_now = ggml_time_us();
+
+        // sweep every 5s, evicting cuda graphs unused for >=10s
+        if (time_now - last_graph_eviction_sweep >= 5'000'000) {
+            last_graph_eviction_sweep = time_now;
+            for (auto it = cuda_graphs.begin(); it != cuda_graphs.end(); ) {
+                if (time_now - it->second->last_used_time >= 10'000'000) {
+                    it = cuda_graphs.erase(it);
+                } else {
+                    ++it;
+                }
+            }
+        }
+
         auto it = cuda_graphs.find(first_node_ptr);
         if (it == cuda_graphs.end()) {
-            cuda_graphs[first_node_ptr] = std::make_unique<ggml_cuda_graph>();
-            return cuda_graphs[first_node_ptr].get();
+            it = cuda_graphs.emplace(first_node_ptr, std::make_unique<ggml_cuda_graph>()).first;
         }
+        it->second->last_used_time = time_now;
         return it->second.get();
     }
 

ggml/src/ggml-cuda/convert.cu

@@ -711,6 +711,8 @@ to_bf16_cuda_t ggml_get_to_bf16_cuda(ggml_type type) {
 
 to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -767,6 +769,8 @@ to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
 
 to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cont_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_row_q4_0_cuda;
         case GGML_TYPE_Q4_1:
@@ -822,6 +826,8 @@ to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -843,6 +849,8 @@ to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F32:
             return convert_unary_cuda<float, nv_bfloat16>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:
@@ -864,6 +872,8 @@ to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type) {
     switch (type) {
         case GGML_TYPE_F16:
             return convert_unary_cuda<half, float>;
+        case GGML_TYPE_Q1_0:
+            return dequantize_block_cuda<QK1_0, QR1_0, dequantize_q1_0>;
         case GGML_TYPE_Q4_0:
             return dequantize_block_cuda<QK4_0, QR4_0, dequantize_q4_0>;
         case GGML_TYPE_Q4_1:

ggml/src/ggml-cuda/dequantize.cuh

@@ -1,5 +1,27 @@
 #include "common.cuh"
 
+static __device__ __forceinline__ void dequantize_q1_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
+    const block_q1_0 * x = (const block_q1_0 *) vx;
+
+    const float d = x[ib].d;
+
+    const int bit_index_0 = iqs;
+    const int bit_index_1 = iqs + 1;
+
+    const int byte_index_0 = bit_index_0 / 8;
+    const int bit_offset_0 = bit_index_0 % 8;
+
+    const int byte_index_1 = bit_index_1 / 8;
+    const int bit_offset_1 = bit_index_1 % 8;
+
+    // Extract bits: 1 = +d, 0 = -d (branchless)
+    const int bit_0 = (x[ib].qs[byte_index_0] >> bit_offset_0) & 1;
+    const int bit_1 = (x[ib].qs[byte_index_1] >> bit_offset_1) & 1;
+
+    v.x = (2*bit_0 - 1) * d;
+    v.y = (2*bit_1 - 1) * d;
+}
+
 static __device__ __forceinline__ void dequantize_q4_0(const void * vx, const int64_t ib, const int iqs, float2 & v){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -305,12 +305,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
     constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
-    // The minimum granularity with cp.async is 16 bytes, with synchronous data loading it's 4 bytes.
+    // The minimum granularity is 16 bytes.
+    constexpr int h2_per_chunk = 16/sizeof(half2);
+    const int chunks_per_row = D2 / h2_per_chunk;
     if constexpr (use_cp_async) {
+        static_assert(warp_size == 32, "bad warp_size");
         static_assert(!oob_check, "OOB check not compatible with cp_async");
         constexpr int preload = 64;
-        constexpr int h2_per_chunk = 16/sizeof(half2);
-        const int chunks_per_row = D2 / h2_per_chunk;
 
         const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
 
@@ -348,11 +349,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
         // 6: max  1*16= 16 bytes,   8 half
         ggml_cuda_unroll<6>{}(load);
     } else {
-        // TODO use ggml_cuda_memcpy_1
+        const half2 zero[4] = {{0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}, {0.0f, 0.0f}};
         auto load = [&] __device__ (const int n) {
-            const int stride_k = warp_size >> n;
-            const int k0_start = stride_k == warp_size ? 0 : D2 - D2 % (2*stride_k);
-            const int k0_stop  =                             D2 - D2 % (1*stride_k);
+            const int stride_k = 32 >> n;
+            const int k0_start = stride_k == 32 ? 0 : chunks_per_row - chunks_per_row % (2*stride_k);
+            const int k0_stop  =                      chunks_per_row - chunks_per_row % (1*stride_k);
             const int stride_i = warp_size / stride_k;
 
             if (k0_start == k0_stop) {
@@ -371,15 +372,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                 for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                     const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);
 
-                    tile_KV[i*stride_tile + k] = !oob_check || i < i_sup ? KV[i*stride_KV + k] : make_half2(0.0f, 0.0f);
+                    ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
+                        !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
                 }
             }
         };
-        // 1: max 32* 4=128 bytes,  64 half
-        // 2: max 16* 4= 64 bytes,  32 half
-        // 3: max  8* 4= 32 bytes,  16 half
-        // 4: max  4* 4= 16 bytes,   8 half
-        ggml_cuda_unroll<4>{}(load);
+        // 1: max 32*16=512 bytes, 256 half
+        // 2: max 16*16=256 bytes, 128 half
+        // 3: max  8*16=128 bytes,  64 half
+        // 4: max  4*16= 64 bytes,  32 half
+        // 5: max  2*16= 32 bytes,  16 half
+        // 6: max  1*16= 16 bytes,   8 half
+        ggml_cuda_unroll<6>{}(load);
     }
 }
 
@@ -862,11 +866,6 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     }
 
 
-#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-    T_A_VKQ A_identity;
-    make_identity_mat(A_identity);
-#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
-
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
     for (int i0_start = 0; i0_start < DV; i0_start += 2*nbatch_V2) {
@@ -897,29 +896,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
-#if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-#elif defined(AMD_MFMA_AVAILABLE)
-                // MFMA A register layout: A_mat[i=lane%16][k=4*(lane/16)+reg].
-                // Normal load gives A_mat[seq][dv] but we need A_mat[dv][seq] = V^T.
-                // Load with transposed addressing: 4 strided half loads.
-                {
-                    const half2 * xs0 = tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2;
-                    const half * xs0_h = (const half *) xs0;
-                    const int stride_h = stride_tile_V * 2; // stride in half units
-                    half * A_h = (half *) A.x;
-#pragma unroll
-                    for (int l = 0; l < 4; ++l) {
-                        A_h[l] = xs0_h[(4*(threadIdx.x / 16) + l) * stride_h + threadIdx.x % 16];
-                    }
-                }
-#else
-                // TODO: Try to transpose tile_V when loading gmem to smem.
-                // Use mma to transpose T_A_VKQ for RDNA.
-                T_A_VKQ A_trans;
-                load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
-                mma(A, A_trans, A_identity);
-#endif // defined(LDMATRIX_TRANS_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {

ggml/src/ggml-cuda/getrows.cu

@@ -179,6 +179,10 @@ static void ggml_cuda_get_rows_switch_src0_type(
             get_rows_cuda_float((const nv_bfloat16 *) src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
             break;
+        case GGML_TYPE_Q1_0:
+            get_rows_cuda_q<QK1_0, QR1_0, dequantize_q1_0>(src0_d, src1_d, dst_d,
+                ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);
+            break;
         case GGML_TYPE_Q4_0:
             get_rows_cuda_q<QK4_0, QR4_0, dequantize_q4_0>(src0_d, src1_d, dst_d,
                 ne00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb1, nb2, nb3, stream);

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

@@ -324,28 +324,22 @@ static ggml_cuda_device_info ggml_cuda_init() {
     // configure logging to stdout
     // CUBLAS_CHECK(cublasLoggerConfigure(1, 1, 0, nullptr));
 
-    for (int id = 0; id < info.device_count; ++id) {
-        ggml_cuda_set_device(id);
-        for (int id_other = 0; id_other < info.device_count; ++id_other) {
-            if (id == id_other) {
-                continue;
-            }
-            int can_access_peer;
-            CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
-            if (can_access_peer) {
-                CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+    if (getenv("GGML_CUDA_P2P") != nullptr) {
+        for (int id = 0; id < info.device_count; ++id) {
+            ggml_cuda_set_device(id);
+            for (int id_other = 0; id_other < info.device_count; ++id_other) {
+                if (id == id_other) {
+                    continue;
+                }
+                int can_access_peer;
+                CUDA_CHECK(cudaDeviceCanAccessPeer(&can_access_peer, id, id_other));
+                if (can_access_peer) {
+                    CUDA_CHECK(cudaDeviceEnablePeerAccess(id_other, 0));
+                }
             }
         }
     }
 
-#ifdef GGML_USE_NCCL
-    int dev_ids[GGML_CUDA_MAX_DEVICES];
-    for (int id = 0; id < info.device_count; ++id) {
-        dev_ids[id] = id;
-    }
-    NCCL_CHECK(ncclCommInitAll(info.comms, info.device_count, dev_ids));
-#endif // GGML_USE_NCCL
-
     return info;
 }
 
@@ -1125,66 +1119,51 @@ static const ggml_backend_buffer_type_i ggml_backend_cuda_split_buffer_type_inte
     /* .is_host          = */ ggml_backend_cuda_split_buffer_type_is_host,
 };
 
-bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_tensor ** tensors, size_t n_backends) {
 #ifdef GGML_USE_NCCL
-    const int64_t ne = ggml_nelements(tensors[0]);
-    // FIXME the input of llm_graph_context::build_in_out_ids can produce a tensor with 0 elements if n_outputs == 0
-    // This then causes a crash in this function
-    if (ne == 0) {
-        return true;
-    }
-    for (size_t i = 0; i < n_backends; ++i) {
-        GGML_ASSERT(tensors[i] != nullptr);
-        GGML_ASSERT(ggml_nelements(tensors[i]) == ne);
-        GGML_ASSERT(ggml_is_contiguously_allocated(tensors[i]));
-    }
+struct ggml_backend_cuda_comm_context {
+    std::vector<ggml_backend_t> backends;
+    std::vector<ncclComm_t> comms;
 
-    const ggml_cuda_device_info info = ggml_cuda_info();
-
-    // For small tensors, simply reduce them as FP32.
-    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
-    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
-        NCCL_CHECK(ncclGroupStart());
-        for (size_t i = 0; i < n_backends; ++i) {
-            ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-            NCCL_CHECK(ncclAllReduce(tensors[i]->data, tensors[i]->data, ne, ncclFloat, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
+    ~ggml_backend_cuda_comm_context() {
+        for (ncclComm_t comm : comms) {
+            NCCL_CHECK(ncclCommDestroy(comm));
         }
-        NCCL_CHECK(ncclGroupEnd());
-
-        return true;
     }
+};
+#endif // GGML_USE_NCCL
 
-    // For large tensors it's faster to compress them to BF16 for the reduction:
-    to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
-    to_fp32_cuda_t to_fp32 = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
+static void ggml_backend_cuda_comm_free(void * comm_ctx_v) {
+#ifdef GGML_USE_NCCL
+    if (comm_ctx_v == nullptr) {
+        return;
+    }
+    ggml_backend_cuda_comm_context * comm_ctx = (ggml_backend_cuda_comm_context *) comm_ctx_v;
+    delete comm_ctx;
+#else
+    GGML_UNUSED(comm_ctx_v);
+#endif // GGML_USE_NCCL
+}
 
-    ggml_cuda_pool_alloc<nv_bfloat16> tmp[GGML_CUDA_MAX_DEVICES];
-    for (size_t i = 0; i < n_backends; ++i) {
+static void * ggml_backend_cuda_comm_init(ggml_backend_t * backends, size_t n_backends) {
+#ifdef GGML_USE_NCCL
+    for (size_t i = 0; i < n_backends; i++) {
+        if (!ggml_backend_is_cuda(backends[i])) {
+            return nullptr;
+        }
+    }
+    ggml_backend_cuda_comm_context * ret = new ggml_backend_cuda_comm_context;
+    std::vector<int> dev_ids;
+    ret->backends.reserve(n_backends);
+    dev_ids.reserve(n_backends);
+    for (size_t i = 0; i < n_backends; i++) {
+        ret->backends.push_back(backends[i]);
         ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-        tmp[i].pool = &cuda_ctx->pool();
-        tmp[i].alloc(ne);
-
-        ggml_cuda_set_device(i);
-        to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
-        CUDA_CHECK(cudaGetLastError());
+        dev_ids.push_back(cuda_ctx->device);
     }
 
-    NCCL_CHECK(ncclGroupStart());
-    for (size_t i = 0; i < n_backends; ++i) {
-        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-        NCCL_CHECK(ncclAllReduce(tmp[i].get(), tmp[i].get(), ne, ncclBfloat16, ncclSum, info.comms[cuda_ctx->device], cuda_ctx->stream()));
-    }
-    NCCL_CHECK(ncclGroupEnd());
-
-    for (size_t i = 0; i < n_backends; ++i) {
-        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backends[i]->context;
-
-        ggml_cuda_set_device(i);
-        to_fp32(tmp[i].get(), (float *) tensors[i]->data, ne, cuda_ctx->stream());
-        CUDA_CHECK(cudaGetLastError());
-    }
-
-    return true;
+    ret->comms.resize(n_backends);
+    NCCL_CHECK(ncclCommInitAll(ret->comms.data(), n_backends, dev_ids.data()));
+    return ret;
 #else
     // If NCCL is installed it is used by default for optimal performance.
     // However, NVIDIA does not distribute NCCL with CUDA so users may be unwittingly missing this package.
@@ -1197,7 +1176,87 @@ bool ggml_backend_cuda_allreduce_tensor(ggml_backend_t * backends, struct ggml_t
         warning_printed = true;
     }
 #endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
-    GGML_UNUSED_VARS(backends, tensors, n_backends);
+    GGML_UNUSED_VARS(backends, n_backends);
+    return nullptr;
+#endif // GGML_USE_NCCL
+}
+
+static bool ggml_backend_cuda_comm_allreduce_tensor(void * comm_ctx_v, struct ggml_tensor ** tensors) {
+#ifdef GGML_USE_NCCL
+    const int64_t ne = ggml_nelements(tensors[0]);
+    // FIXME the input of llm_graph_context::build_in_out_ids can produce a tensor with 0 elements if n_outputs == 0
+    // This then causes a crash in this function
+    if (ne == 0) {
+        return true;
+    }
+
+    GGML_ASSERT(comm_ctx_v != nullptr);
+    ggml_backend_cuda_comm_context * comm_ctx = (ggml_backend_cuda_comm_context *) comm_ctx_v;
+    const size_t n_backends = comm_ctx->backends.size();
+
+    for (size_t i = 0; i < n_backends; ++i) {
+        GGML_ASSERT(tensors[i] != nullptr);
+        GGML_ASSERT(ggml_nelements(tensors[i]) == ne);
+        GGML_ASSERT(ggml_is_contiguously_allocated(tensors[i]));
+    }
+
+    // For small tensors, simply reduce them as FP32.
+    // The following heuristic for how "small" a tensor should be is based on RTX 4090s connected via 16x PCIe 4.0.
+    if ((n_backends <= 2 && ne < 32768) || (n_backends == 3 && ne < 131072) || (n_backends >= 4 && ne < 262144)) {
+        for (size_t i = 0; i < n_backends; ++i) {
+            if ((tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) == 0) {
+                ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+                ggml_cuda_set_device(cuda_ctx->device);
+                CUDA_CHECK(cudaMemsetAsync(tensors[i]->data, 0, ggml_nbytes(tensors[i]), cuda_ctx->stream()));
+            }
+        }
+        NCCL_CHECK(ncclGroupStart());
+        for (size_t i = 0; i < n_backends; ++i) {
+            ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+            NCCL_CHECK(ncclAllReduce(tensors[i]->data, tensors[i]->data, ne, ncclFloat, ncclSum, comm_ctx->comms[i], cuda_ctx->stream()));
+        }
+        NCCL_CHECK(ncclGroupEnd());
+
+        return true;
+    }
+
+    // For large tensors it's faster to compress them to BF16 for the reduction:
+    to_bf16_cuda_t to_bf16 = ggml_get_to_bf16_cuda(GGML_TYPE_F32);
+    to_fp32_cuda_t to_fp32 = ggml_get_to_fp32_cuda(GGML_TYPE_BF16);
+
+    ggml_cuda_pool_alloc<nv_bfloat16> tmp[GGML_CUDA_MAX_DEVICES];
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+        tmp[i].pool = &cuda_ctx->pool();
+        tmp[i].alloc(ne);
+
+        ggml_cuda_set_device(cuda_ctx->device);
+        if (tensors[i]->flags & GGML_TENSOR_FLAG_COMPUTE) {
+            to_bf16(tensors[i]->data, tmp[i].get(), ne, cuda_ctx->stream());
+        } else {
+            CUDA_CHECK(cudaMemsetAsync(tmp[i].get(), 0, ne * sizeof(nv_bfloat16), cuda_ctx->stream()));
+        }
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    NCCL_CHECK(ncclGroupStart());
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+        NCCL_CHECK(ncclAllReduce(tmp[i].get(), tmp[i].get(), ne, ncclBfloat16, ncclSum, comm_ctx->comms[i], cuda_ctx->stream()));
+    }
+    NCCL_CHECK(ncclGroupEnd());
+
+    for (size_t i = 0; i < n_backends; ++i) {
+        ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) comm_ctx->backends[i]->context;
+
+        ggml_cuda_set_device(cuda_ctx->device);
+        to_fp32(tmp[i].get(), (float *) tensors[i]->data, ne, cuda_ctx->stream());
+        CUDA_CHECK(cudaGetLastError());
+    }
+
+    return true;
+#else
+    GGML_UNUSED_VARS(comm_ctx_v, tensors);
     return false;
 #endif // GGML_USE_NCCL
 }
@@ -3060,6 +3119,15 @@ static bool ggml_cuda_graph_update_required(ggml_backend_cuda_context * cuda_ctx
     const void * graph_key = ggml_cuda_graph_get_key(cgraph);
     ggml_cuda_graph * graph = cuda_ctx->cuda_graph(graph_key);
 
+    if (cgraph->uid != 0 &&
+        cgraph->uid == graph->uid) {
+        GGML_LOG_DEBUG("CUDA Graph id %zu reused\n", cgraph->uid);
+        GGML_ASSERT((int)graph->node_props.size() == cgraph->n_nodes);
+        return false;
+    }
+
+    graph->uid = cgraph->uid;
+
     // Check if the graph size has changed
     if ((int)graph->node_props.size() != cgraph->n_nodes) {
         res = true;
@@ -4783,6 +4851,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                 switch (a->type) {
                     case GGML_TYPE_F32:
                     case GGML_TYPE_F16:
+                    case GGML_TYPE_Q1_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -4820,6 +4889,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
                     case GGML_TYPE_F32:
                     case GGML_TYPE_BF16:
                     case GGML_TYPE_I32:
+                    case GGML_TYPE_Q1_0:
                     case GGML_TYPE_Q4_0:
                     case GGML_TYPE_Q4_1:
                     case GGML_TYPE_Q5_0:
@@ -5220,8 +5290,14 @@ static ggml_backend_feature * ggml_backend_cuda_get_features(ggml_backend_reg_t
 
 static void * ggml_backend_cuda_reg_get_proc_address(ggml_backend_reg_t reg, const char * name) {
     GGML_UNUSED(reg);
-    if (strcmp(name, "ggml_backend_allreduce_tensor") == 0) {
-        return (void *)ggml_backend_cuda_allreduce_tensor;
+    if (strcmp(name, "ggml_backend_comm_init") == 0) {
+        return (void *)ggml_backend_cuda_comm_init;
+    }
+    if (strcmp(name, "ggml_backend_comm_free") == 0) {
+        return (void *)ggml_backend_cuda_comm_free;
+    }
+    if (strcmp(name, "ggml_backend_comm_allreduce_tensor") == 0) {
+        return (void *)ggml_backend_cuda_comm_allreduce_tensor;
     }
     if (strcmp(name, "ggml_backend_split_buffer_type") == 0) {
         return (void *)ggml_backend_cuda_split_buffer_type;

ggml/src/ggml-cuda/mma.cuh

@@ -86,17 +86,12 @@ namespace ggml_cuda_mma {
     //   - (I_MAJOR, I_MAJOR_MIRRORED) -> I_MAJOR
     //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
 
-    static constexpr bool is_i_major(const data_layout dl) {
-        return dl == DATA_LAYOUT_I_MAJOR ||
-               dl == DATA_LAYOUT_I_MAJOR_MIRRORED;
-    }
-
     static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
         return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
         return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -113,7 +108,6 @@ namespace ggml_cuda_mma {
         T x[ne] = {0};
 
         static constexpr __device__ bool supported() {
-            if (I == 64 && J ==  2) return true;
             if (I == 16 && J ==  8) return true;
             if (I == 32 && J ==  4) return true;
             if (I == 16 && J == 16) return true;
@@ -122,7 +116,7 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_i(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
+            if constexpr (I == 16 && J == 4) {
                 return threadIdx.x % 16;
             } else if constexpr (I == 16 && J == 8) {
                 return threadIdx.x % 16;
@@ -139,8 +133,8 @@ namespace ggml_cuda_mma {
         }
 
         static __device__ __forceinline__ int get_j(const int l) {
-            if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-                return (2 * ((threadIdx.x / 16) % 2) + l);
+            if constexpr (I == 16 && J == 4) {
+                return threadIdx.x / 16;
             } else if constexpr (I == 16 && J == 8) {
                 return 2 * (threadIdx.x / 16) + l;
             } else if constexpr (I == 32 && J == 4) {
@@ -154,7 +148,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#elif __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#elif defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / 32;
         T x[ne] = {0};
 
@@ -283,7 +277,7 @@ namespace ggml_cuda_mma {
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR;
 
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         static constexpr int ne = I * J / WARP_SIZE;
         half2 x[ne] = {{0.0f, 0.0f}};
 
@@ -407,7 +401,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     };
 
     template <int I_, int J_>
@@ -701,57 +695,12 @@ namespace ggml_cuda_mma {
     }
 #endif // defined(TURING_MMA_AVAILABLE)
 
-    static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
-#if defined(RDNA4)
-        const int row = t.get_i(0);
-        const int left_right = t.get_j(0) / 4;
-        const int up_down = row / 8;
-        const int idx = row % 8;
-        reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
-#else
-        GGML_UNUSED_VARS(t);
-        NO_DEVICE_CODE;
-#endif // defined(RDNA4)
-    }
-
     template <int I, int J, typename T, data_layout dl>
     static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
-#if defined(AMD_MFMA_AVAILABLE)
-        if constexpr (I == 64 && J == 2) { // Special tile size to load <16, 4> as <16, 8>
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        } else {
-            ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-        }
-#elif defined(AMD_WMMA_AVAILABLE)
-        // All wmma layout has contiguous data when i-major.
-        if constexpr (is_i_major(dl)) {
-            // the data must be aligned to 16 bytes when bigger than ggml_cuda_get_max_cpy_bytes()
-            constexpr int aligned_copy_bytes = ggml_cuda_get_max_cpy_bytes();
-            if constexpr (sizeof(t.x) > aligned_copy_bytes) {
-                static_assert(sizeof(t.x) % aligned_copy_bytes == 0, "bad type size");
-                constexpr int aligned_copy_count = sizeof(t.x)/aligned_copy_bytes;
-#pragma unroll
-                for (int i = 0; i < aligned_copy_count; ++i) {
-                    ggml_cuda_memcpy_1<aligned_copy_bytes>(t.x + t.ne/aligned_copy_count*i, xs0 + t.get_i(0) * stride + t.get_j(t.ne/aligned_copy_count*i));
-                }
-            } else {
-                ggml_cuda_memcpy_1<sizeof(t.x)>(t.x, xs0 + t.get_i(0) * stride + t.get_j(0));
-            }
-        } else {
-#pragma unroll
-            for (int l = 0; l < t.ne; ++l) {
-                t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
-            }
-        }
-#else
 #pragma unroll
         for (int l = 0; l < t.ne; ++l) {
             t.x[l] = xs0[t.get_i(l)*stride + t.get_j(l)];
         }
-#endif // defined(AMD_MFMA_AVAILABLE)
     }
 
     template <typename T>
@@ -764,26 +713,37 @@ namespace ggml_cuda_mma {
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
 #else
-        load_generic(t, xs0, stride);
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
-    template <typename T>
+    template <typename T, data_layout dl>
     static __device__ __forceinline__ void load_ldmatrix(
-            tile<16, 4, T> & t, const T * __restrict__ xs0, const int stride) {
+            tile<16, 4, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
         int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride;
         asm volatile("ldmatrix.sync.aligned.m8n8.x2.b16 {%0, %1}, [%2];"
             : "=r"(xi[0]), "=r"(xi[1])
             : "l"(xs));
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<8>(t.x + 2, xs0 + t.get_i(0)*stride + 2);
+#else
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 4, "bad ne");
+        ggml_cuda_memcpy_1<4>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -796,19 +756,26 @@ namespace ggml_cuda_mma {
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[1]), "=r"(xi[2]), "=r"(xi[3])
             : "l"(xs));
-#else
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
-#if 1
-        // TODO: more generic handling
-        static_assert(sizeof(T) == 4, "bad type size");
+#elif defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
         ggml_cuda_memcpy_1<4*sizeof(T)>(t.x + 4, xs0 + t.get_i(4)*stride + 4);
+#elif defined(AMD_WMMA_AVAILABLE)
+#ifdef RDNA3
+        static_assert(dl == DATA_LAYOUT_I_MAJOR_MIRRORED, "bad data layout");
+        static_assert(sizeof(t.x) == 32, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x + 0, xs0 + t.get_i(0)*stride + 0);
+        ggml_cuda_memcpy_1<16>(t.x + 4, xs0 + t.get_i(0)*stride + 4);
 #else
-        load_generic(t, xs0, stride);
-#endif // 1
+        static_assert(dl == DATA_LAYOUT_I_MAJOR, "bad data layout");
+        static_assert(sizeof(t.x) == 16, "bad ne");
+        ggml_cuda_memcpy_1<16>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
+#endif // RDNA3
+#elif defined(AMD_MFMA_AVAILABLE)
+        static_assert(sizeof(t.x) == 8, "bad ne");
+        ggml_cuda_memcpy_1<8>(t.x, xs0 + t.get_i(0)*stride + t.get_j(0));
 #else
-        load_generic(t, xs0, stride);
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+        GGML_UNUSED_VARS(t, xs0, stride);
+        NO_DEVICE_CODE;
 #endif // TURING_MMA_AVAILABLE
     }
 
@@ -827,23 +794,30 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void load_ldmatrix(
             tile<32, 4, half2> & t, const half2 * __restrict__ xs0, const int stride) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         ggml_cuda_memcpy_1<4*sizeof(half2)>(t.x, xs0 + t.get_i(0)*stride);
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <typename T>
     static __device__ __forceinline__ void load_ldmatrix_trans(
             tile<16, 8, T> & t, const T * __restrict__ xs0, const int stride) {
 #ifdef TURING_MMA_AVAILABLE
-        int * xi = (int * ) t.x;
+        int * xi = (int *) t.x;
         const int * xs = (const int *) xs0 + (threadIdx.x % t.I) * stride + (threadIdx.x / t.I) * (t.J / 2);
         asm volatile("ldmatrix.sync.aligned.m8n8.x4.trans.b16 {%0, %1, %2, %3}, [%4];"
             : "=r"(xi[0]), "=r"(xi[2]), "=r"(xi[1]), "=r"(xi[3])
             : "l"(xs));
+#elif defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        half * xh = (half *) t.x;
+#pragma unroll
+        for (int l = 0; l < t.ne; ++l) {
+            xh[2*l + 0] = ((const half *) xs0)[(2*t.get_j(l) + 0)*(2*stride) + t.get_i(l)];
+            xh[2*l + 1] = ((const half *) xs0)[(2*t.get_j(l) + 1)*(2*stride) + t.get_i(l)];
+        }
 #else
         GGML_UNUSED_VARS(t, xs0, stride);
         NO_DEVICE_CODE;
@@ -1218,73 +1192,27 @@ namespace ggml_cuda_mma {
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * acc = (int32x4_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0],
-                                                      B.x[0],
-                                                      acc[0],
-                                                      0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1],
-                                                      B.x[1],
-                                                      acc[0],
-                                                      0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
-
 #elif defined(AMD_WMMA_AVAILABLE)
-
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
-
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            true
-        );
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[1],
-            true,
-            b_vec[1],
-            acc[0],
-            true
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
 #endif // RDNA4
-
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1297,19 +1225,10 @@ namespace ggml_cuda_mma {
         using int32x16_t = __attribute__((__vector_size__(16 * sizeof(int)))) int;
         int32x16_t * acc = (int32x16_t *) D.x;
 #if defined(CDNA4) || defined(CDNA3)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0],
-                                                       ((int64_t *) B.x)[0],
-                                                       acc[0],
-                                                       0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x16_i8(((int64_t *) A.x)[0], ((int64_t *) B.x)[0], acc[0], 0, 0, 0);
 #elif defined(CDNA2) || defined(CDNA1)
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0],
-                                                     B.x[0],
-                                                     acc[0],
-                                                     0, 0, 0);
-        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1],
-                                                     B.x[1],
-                                                     acc[0],
-                                                     0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+        acc[0] = __builtin_amdgcn_mfma_i32_32x32x8i8(A.x[1], B.x[1], acc[0], 0, 0, 0);
 #endif // defined(CDNA4) || defined(CDNA3)
 
 #else
@@ -1329,7 +1248,7 @@ namespace ggml_cuda_mma {
 
     static __device__ __forceinline__ void mma(
             tile<32, 8, float> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1344,12 +1263,12 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     static __device__ __forceinline__ void mma(
             tile<32, 4, half2> & D, const tile<32, 4, half2> & A, const tile<8, 4, half2, DATA_LAYOUT_J_MAJOR_MIRRORED> & B) {
-#if __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
+#if defined(VOLTA_MMA_AVAILABLE)
         const int * Axi = (const int *) A.x;
         const int * Bxi = (const int *) B.x;
         int       * Dxi = (int       *) D.x;
@@ -1364,41 +1283,35 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // __CUDA_ARCH__ >= GGML_CUDA_CC_VOLTA
+#endif // defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <data_layout dl_d, data_layout dl_ab>
     static __device__ __forceinline__ void mma(
             tile<16, 16, int, dl_d> & D, const tile<16, 4, int, dl_ab> & A, const tile<16, 4, int, dl_ab> & B) {
-#if defined(AMD_WMMA_AVAILABLE)
+#if defined(AMD_MFMA_AVAILABLE)
+        using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
+        int32x4_t * acc = (int32x4_t *) D.x;
+#if defined(CDNA4) || defined(CDNA3)
+        const int64_t xA = uint32_t(A.x[0]);
+        const int64_t xB = uint32_t(B.x[0]);
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x32_i8(xA, xB, acc[0], 0, 0, 0);
+#elif defined(CDNA2) || defined(CDNA1)
+        acc[0] = __builtin_amdgcn_mfma_i32_16x16x16i8(A.x[0], B.x[0], acc[0], 0, 0, 0);
+#endif // defined(CDNA4) || defined(CDNA3)
+#elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
 #if defined(RDNA4)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
 #elif defined(RDNA3)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
-
-        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(
-            true,
-            a_vec[0],
-            true,
-            b_vec[0],
-            acc[0],
-            false
-        );
+        acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
 #endif // RDNA4
 #else
         GGML_UNUSED(D);

ggml/src/ggml-cuda/mmq.cu

@@ -5,6 +5,9 @@
 
 static void ggml_cuda_mul_mat_q_switch_type(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
     switch (args.type_x) {
+        case GGML_TYPE_Q1_0:
+            mul_mat_q_case<GGML_TYPE_Q1_0>(ctx, args, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_q_case<GGML_TYPE_Q4_0>(ctx, args, stream);
             break;
@@ -270,6 +273,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     bool mmq_supported;
 
     switch (type) {
+        case GGML_TYPE_Q1_0:
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q5_0:

ggml/src/ggml-cuda/mmq.cuh

@@ -57,6 +57,8 @@ static_assert(sizeof(block_fp4_mmq)  == sizeof(block_q8_1_mmq),    "Unexpected b
 
 static mmq_q8_1_ds_layout mmq_get_q8_1_ds_layout(const ggml_type type_x) {
     switch (type_x) {
+        case GGML_TYPE_Q1_0:
+            return MMQ_Q8_1_DS_LAYOUT_D4;
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
             return MMQ_Q8_1_DS_LAYOUT_DS4;
@@ -102,7 +104,7 @@ struct tile_x_sizes {
 };
 
 static int get_mmq_x_max_host(const int cc) {
-    return (amd_mfma_available(cc) || turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
+    return (turing_mma_available(cc) || amd_wmma_available(cc)) ? 128 :
         GGML_CUDA_CC_IS_NVIDIA(cc) && ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_VOLTA ?
 #ifdef GGML_CUDA_FORCE_MMQ
             128                     : 64;
@@ -112,9 +114,9 @@ static int get_mmq_x_max_host(const int cc) {
 }
 
 static constexpr __device__ int get_mmq_x_max_device() {
-#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     return 128;
-#else // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#else // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
 #if defined(GGML_USE_HIP)
     return 64;
@@ -185,6 +187,7 @@ static constexpr __device__ int get_mmq_y_device() {
 
 static constexpr __host__ __device__ tile_x_sizes mmq_get_dp4a_tile_x_sizes(ggml_type type, int mmq_y) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return MMQ_DP4A_TXS_Q8_0;
         case GGML_TYPE_Q4_0:    return MMQ_DP4A_TXS_Q4_0;
         case GGML_TYPE_Q4_1:    return MMQ_DP4A_TXS_Q4_1;
         case GGML_TYPE_Q5_0:    return MMQ_DP4A_TXS_Q8_0;
@@ -229,6 +232,7 @@ static_assert(MMQ_MMA_TILE_X_K_NVFP4 % 8 == 4, "Wrong padding.");
 
 static constexpr __host__ __device__ int mmq_get_mma_tile_x_k(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q4_0:    return MMQ_MMA_TILE_X_K_Q8_0;
         case GGML_TYPE_Q4_1:    return MMQ_MMA_TILE_X_K_Q8_1;
         case GGML_TYPE_Q5_0:    return MMQ_MMA_TILE_X_K_Q8_0;
@@ -302,6 +306,87 @@ static constexpr __device__ int mmq_get_nwarps_device() {
 
 // ------------------------------------------------------------
 
+template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q1_0(
+    const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
+    constexpr int nwarps = mmq_get_nwarps_device();
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + 2*MMQ_TILE_NE_K);
+#else
+    constexpr tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(GGML_TYPE_Q8_0, mmq_y);
+    int   * x_qs = (int   *)  x_tile;
+    float * x_df = (float *) (x_qs + txs.qs);
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+
+    constexpr int blocks_per_iter = MMQ_ITER_K / QK1_0;
+    constexpr int threads_per_row = blocks_per_iter * QI1_0;
+    constexpr int nrows = warp_size / threads_per_row;
+    constexpr int scale_entries_per_block = QK1_0 / QK8_1;
+    constexpr int scale_entries_per_row = blocks_per_iter * scale_entries_per_block;
+
+    const int txi  = threadIdx.x % threads_per_row;
+    const int kbx  = txi / QI1_0;
+    const int kqsx = txi % QI1_0;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nrows*nwarps) {
+        int i = i0 + threadIdx.y*nrows + threadIdx.x/threads_per_row;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + kbx;
+        const int qs_offset = 4*kqsx;
+        const int qs0 = bxi->qs[qs_offset + 0] | (bxi->qs[qs_offset + 1] << 8) |
+                        (bxi->qs[qs_offset + 2] << 16) | (bxi->qs[qs_offset + 3] << 24);
+
+        int unpacked_bytes[8];
+#pragma unroll
+        for (int j = 0; j < 8; ++j) {
+            const int shift = j * 4;
+            const int bits4 = (qs0 >> shift) & 0x0F;
+            const int b0 = (bits4 & 0x01) ? 1 : -1;
+            const int b1 = (bits4 & 0x02) ? 1 : -1;
+            const int b2 = (bits4 & 0x04) ? 1 : -1;
+            const int b3 = (bits4 & 0x08) ? 1 : -1;
+            unpacked_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+        }
+
+        const int dst_offset = kbx*(scale_entries_per_block*QI8_0) + kqsx*QI8_0;
+#pragma unroll
+        for (int j = 0; j < 8; ++j) {
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+            x_qs[i*MMQ_MMA_TILE_X_K_Q8_0 + dst_offset + j] = unpacked_bytes[j];
+#else
+            x_qs[i*(2*MMQ_TILE_NE_K + 1) + dst_offset + j] = unpacked_bytes[j];
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        }
+    }
+
+    const int ksx = threadIdx.x % scale_entries_per_row;
+    const int scale_block = ksx / scale_entries_per_block;
+
+#pragma unroll
+    for (int i0 = 0; i0 < mmq_y; i0 += nwarps) {
+        int i = i0 + threadIdx.y;
+
+        if (need_check) {
+            i = min(i, i_max);
+        }
+
+        const block_q1_0 * bxi = (const block_q1_0 *) x + kbx0 + i*stride + scale_block;
+
+#if defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+        x_df[i*MMQ_MMA_TILE_X_K_Q8_0 + ksx] = bxi->d;
+#else
+        x_df[i*(2*MMQ_TILE_NE_K/QI8_0) + i/(QI8_0/2) + ksx] = bxi->d;
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
+    }
+}
+
 template <int mmq_y, bool need_check> static __device__ __forceinline__ void load_tiles_q4_0(
     const char * __restrict__ x, int * __restrict__ x_tile, const int kbx0, const int i_max, const int stride) {
     constexpr int nwarps = mmq_get_nwarps_device();
@@ -969,13 +1054,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_0 + k0, MMQ_MMA_TILE_X_K_Q8_0);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             float dB;
             const int j = j0 + tile_C::get_j(0);
@@ -1210,13 +1295,13 @@ static __device__ __forceinline__ void vec_dot_q8_1_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q8_1 + k0, MMQ_MMA_TILE_X_K_Q8_1);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float2 dsB = __half22float2(y_dm[j*MMQ_TILE_Y_K + k01/QI8_1]);
@@ -1350,57 +1435,7 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * x_df[i*MMQ_MMA_TILE_X_K_Q3_K + k0/4] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1425,13 +1460,13 @@ static __device__ __forceinline__ void vec_dot_q8_0_16_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q3_K + k0, MMQ_MMA_TILE_X_K_Q3_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -1657,74 +1692,7 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const half2 * x_dm = (const half2 *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * y_qs = (const int   *) y + 4;
-    const half2 * y_ds = (const half2 *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x/2 : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y/2;
-            const float sB = (k01 >= MMQ_TILE_NE_K * 3/4) ? 0
-                                              : (((k01/4)%2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).y
-                                                             : __half22float2(y_ds[j*MMQ_TILE_Y_K + (1 + k01/QI8_1)]).x);
-
-            tile_C Cm;
-            if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                tile_A A1;
-                A1.x[0] = 0x01010101;
-                A1.x[1] = 0x01010101;
-                mma(Cm, A1, B[0]);
-            }
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C Cd;
-                mma(Cd, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const float2 dm = __half22float2(x_dm[i*MMQ_MMA_TILE_X_K_Q2_K + k0/4]);
-                    float tmp = Cd.x[l]*dm.x;
-                    if (k01 >= MMQ_TILE_NE_K * 3/4) {
-                        tmp -= Cm.x[l]*dm.y;
-                    }
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += tmp*dB;
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] -= dm.y*sB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -1749,13 +1717,13 @@ static __device__ __forceinline__ void vec_dot_q2_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q2_K + k0, MMQ_MMA_TILE_X_K_Q2_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = (k01 < MMQ_TILE_NE_K/2) ? __half22float2(y_ds[j*MMQ_TILE_Y_K]).x : __half22float2(y_ds[j*MMQ_TILE_Y_K]).y;
@@ -2488,59 +2456,7 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_dp4a(
 template <int mmq_x, int mmq_y>
 static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
     const int * __restrict__ x, const int * __restrict__ y, float * __restrict__ sum, const int k00) {
-#if defined(AMD_MFMA_AVAILABLE)
-    constexpr data_layout input_layout = get_input_data_layout();
-    typedef tile<16,  8, int, input_layout>        tile_A;
-    typedef tile<16,  8, int, input_layout>        tile_B;
-    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;
-    typedef tile<64,  2, int, input_layout>        tile_load;
-
-    constexpr int granularity = mmq_get_granularity_device(mmq_x);
-    constexpr int rows_per_warp = granularity;
-    constexpr int ntx = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
-
-    y += (threadIdx.y % ntx) * (tile_C::J*MMQ_TILE_Y_K);
-
-    const int   * x_qs = (const int   *) x;
-    const float * x_df = (const float *) x_qs + MMQ_TILE_NE_K*2;
-    const int   * x_sc = (const int   *) x_df + MMQ_TILE_NE_K/QI6_K;
-    const int   * y_qs = (const int   *) y + 4;
-    const float * y_df = (const float *) y;
-
-    const int i0 = (threadIdx.y / ntx) * rows_per_warp;
-
-    for (int k01 = 0; k01 < MMQ_TILE_NE_K; k01 += 4) {
-        const int k0 = k00 + k01;
-
-        tile_A A[ntx];
-#pragma unroll
-        for (int n = 0; n < ntx; ++n) {
-            load_generic(((tile_load *) A)[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
-        }
-
-#pragma unroll
-        for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
-            tile_B B[1];
-            load_generic(((tile_load *) B)[0], y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
-
-            const int j = j0 + tile_C::get_j(0);
-            const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1] / 2;
-
-#pragma unroll
-            for (int n = 0; n < ntx; ++n) {
-                tile_C C;
-                mma(C, A[n], B[0]);
-
-#pragma unroll
-                for (int l = 0; l < tile_C::ne; ++l) {
-                    const int i = i0 + n*tile_C::I + tile_C::get_i(l);
-                    const int8_t * sc = (const int8_t *) (x_sc + i*MMQ_MMA_TILE_X_K_Q6_K + k00/16);
-                    sum[(j0/tile_C::J + n)*tile_C::ne + l] += C.x[l] * sc[k01/4] * x_df[i*MMQ_MMA_TILE_X_K_Q6_K] * dB;
-                }
-            }
-        }
-    }
-#elif defined(AMD_WMMA_AVAILABLE) //wmma instructions can handle 16x4 tiles, does not require loading 64x2 tiles
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
     constexpr data_layout input_layout = get_input_data_layout();
     typedef tile<16,  4, int, input_layout>        tile_A;
     typedef tile<16,  4, int, input_layout>        tile_B;
@@ -2566,13 +2482,13 @@ static __device__ __forceinline__ void vec_dot_q6_K_q8_1_mma(
         tile_A A[ntx];
 #pragma unroll
         for (int n = 0; n < ntx; ++n) {
-            load_generic(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
+            load_ldmatrix(A[n], x_qs + (i0 + n*tile_A::I)*MMQ_MMA_TILE_X_K_Q6_K + k0, MMQ_MMA_TILE_X_K_Q6_K);
         }
 
 #pragma unroll
         for (int j0 = 0; j0 < mmq_x; j0 += ntx*tile_C::J) {
             tile_B B;
-            load_generic(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
+            load_ldmatrix(B, y_qs + j0*MMQ_TILE_Y_K + k01, MMQ_TILE_Y_K);
 
             const int j = j0 + tile_C::get_j(0);
             const float dB = y_df[j*MMQ_TILE_Y_K + k01/QI8_1];
@@ -3290,6 +3206,14 @@ static __device__ __forceinline__ void mmq_write_back_mma(
 template <int mmq_x, int mmq_y, bool need_check, ggml_type type>
 struct mmq_type_traits;
 
+template <int mmq_x, int mmq_y, bool need_check>
+struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q1_0> {
+    static constexpr int              vdr          = VDR_Q1_0_Q8_1_MMQ;
+    static constexpr load_tiles_mmq_t load_tiles   = load_tiles_q1_0<mmq_y, need_check>;
+    static constexpr vec_dot_mmq_t    vec_dot_mma  = vec_dot_q8_0_q8_1_mma<mmq_x, mmq_y, MMQ_Q8_1_DS_LAYOUT_D4>;
+    static constexpr vec_dot_mmq_t    vec_dot_dp4a = vec_dot_q8_0_q8_1_dp4a<mmq_x, mmq_y>;
+};
+
 template <int mmq_x, int mmq_y, bool need_check>
 struct mmq_type_traits<mmq_x, mmq_y, need_check, GGML_TYPE_Q4_0> {
     static constexpr int              vdr          = VDR_Q4_0_Q8_1_MMQ;

ggml/src/ggml-cuda/mmvq.cu

@@ -9,6 +9,7 @@ typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_
 
 static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return vec_dot_q1_0_q8_1;
         case GGML_TYPE_Q4_0:    return vec_dot_q4_0_q8_1;
         case GGML_TYPE_Q4_1:    return vec_dot_q4_1_q8_1;
         case GGML_TYPE_Q5_0:    return vec_dot_q5_0_q8_1;
@@ -36,6 +37,7 @@ static constexpr __device__ vec_dot_q_cuda_t get_vec_dot_q_cuda(ggml_type type)
 
 static constexpr __host__ __device__ int get_vdr_mmvq(ggml_type type) {
     switch (type) {
+        case GGML_TYPE_Q1_0:    return VDR_Q1_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_0:    return VDR_Q4_0_Q8_1_MMVQ;
         case GGML_TYPE_Q4_1:    return VDR_Q4_1_Q8_1_MMVQ;
         case GGML_TYPE_Q5_0:    return VDR_Q5_0_Q8_1_MMVQ;
@@ -886,6 +888,12 @@ static void mul_mat_vec_q_switch_type(
         const int nsamples_x, const int nsamples_dst, const int stride_sample_x, const int stride_sample_y, const int stride_sample_dst,
         const int ids_stride, cudaStream_t stream) {
     switch (type_x) {
+        case GGML_TYPE_Q1_0:
+            mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q1_0>
+                (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,
+                 nchannels_x, nchannels_y, nchannels_dst, stride_channel_x, stride_channel_y, stride_channel_dst,
+                 nsamples_x, nsamples_dst, stride_sample_x, stride_sample_y, stride_sample_dst, ids_stride, stream);
+            break;
         case GGML_TYPE_Q4_0:
             mul_mat_vec_q_switch_ncols_dst<GGML_TYPE_Q4_0>
                 (vx, vy, ids, fusion, dst, ncols_x, nrows_x, ncols_dst, stride_row_x, stride_col_y, stride_col_dst,

ggml/src/ggml-cuda/template-instances/generate_cu_files.py

@@ -32,6 +32,7 @@ SOURCE_FATTN_MMA_START = """// This file has been autogenerated by generate_cu_f
 SOURCE_FATTN_MMA_CASE = "DECL_FATTN_MMA_F16_CASE({head_size_kq}, {head_size_v}, {ncols1}, {ncols2});\n"
 
 TYPES_MMQ = [
+    "GGML_TYPE_Q1_0",
     "GGML_TYPE_Q4_0", "GGML_TYPE_Q4_1", "GGML_TYPE_Q5_0", "GGML_TYPE_Q5_1", "GGML_TYPE_Q8_0",
     "GGML_TYPE_Q2_K", "GGML_TYPE_Q3_K", "GGML_TYPE_Q4_K", "GGML_TYPE_Q5_K", "GGML_TYPE_Q6_K",
     "GGML_TYPE_IQ2_XXS", "GGML_TYPE_IQ2_XS", "GGML_TYPE_IQ2_S", "GGML_TYPE_IQ3_XXS", "GGML_TYPE_IQ3_S",

ggml/src/ggml-cuda/template-instances/mmq-instance-q1_0.cu

@@ -0,0 +1,5 @@
+// This file has been autogenerated by generate_cu_files.py, do not edit manually.
+
+#include "../mmq.cuh"
+
+DECL_MMQ_CASE(GGML_TYPE_Q1_0);

ggml/src/ggml-cuda/vecdotq.cuh

@@ -106,6 +106,9 @@ static __device__ __forceinline__ uint32_t unpack_ksigns(const uint8_t v) {
 // VDR = vec dot ratio, how many contiguous integers each thread processes when the vec dot kernel is called
 // MMVQ = mul_mat_vec_q, MMQ = mul_mat_q
 
+#define VDR_Q1_0_Q8_1_MMVQ 1  // Process one 32-element chunk at a time for parallelism
+#define VDR_Q1_0_Q8_1_MMQ  4  // Q1_0 has 128 bits (4 ints) per block
+
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4
 
@@ -669,6 +672,51 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_impl_mmq(
     return d6 * sumf_d;
 }
 
+static __device__ __forceinline__ float vec_dot_q1_0_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
+
+    const block_q1_0 * bq1_0 = (const block_q1_0 *) vbq + kbx;
+
+    // Q1_0: 128 elements with ONE scale
+    // Q8_1: 32 elements per block with individual scales
+    // iqs selects which of the 4 chunks of 32 elements to process (0-3)
+
+    const float d1 = bq1_0->d;
+
+    // Process only the chunk specified by iqs
+    const block_q8_1 * bq8_1_chunk = bq8_1 + iqs;
+
+    // Load 32 bits (4 bytes) for this chunk from Q1_0
+    const int offset = iqs * 4;
+    const int v = bq1_0->qs[offset + 0] | (bq1_0->qs[offset + 1] << 8) |
+                  (bq1_0->qs[offset + 2] << 16) | (bq1_0->qs[offset + 3] << 24);
+
+    // Unpack 32 bits into 32 signed values (-1 or +1)
+    int vi_bytes[8];
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int shift = j * 4;
+        const int bits4 = (v >> shift) & 0x0F;
+        const int b0 = (bits4 & 0x01) ? 1 : -1;
+        const int b1 = (bits4 & 0x02) ? 1 : -1;
+        const int b2 = (bits4 & 0x04) ? 1 : -1;
+        const int b3 = (bits4 & 0x08) ? 1 : -1;
+        vi_bytes[j] = (b0 & 0xFF) | ((b1 & 0xFF) << 8) | ((b2 & 0xFF) << 16) | ((b3 & 0xFF) << 24);
+    }
+
+    // Compute dot product for this 32-element chunk
+    int sumi = 0;
+#pragma unroll
+    for (int j = 0; j < 8; ++j) {
+        const int u = get_int_b4(bq8_1_chunk->qs, j);
+        sumi = ggml_cuda_dp4a(vi_bytes[j], u, sumi);
+    }
+
+    // Apply Q1_0's single scale and this chunk's Q8_1 scale
+    const float d8 = __low2float(bq8_1_chunk->ds);
+    return d1 * d8 * sumi;
+}
+
 static __device__ __forceinline__ float vec_dot_q4_0_q8_1(
     const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & kbx, const int & iqs) {
 

ggml/src/ggml-cuda/vendors/hip.h

@@ -33,7 +33,6 @@
 #define CU_MEM_LOCATION_TYPE_DEVICE hipMemLocationTypeDevice
 #define CU_MEM_ACCESS_FLAGS_PROT_READWRITE hipMemAccessFlagsProtReadWrite
 #define CU_CHECK(fn) {hipError_t err = fn; if(err != hipSuccess) { GGML_ABORT("HipVMM Failure: %s\n", hipGetErrorString(err)); }}
-#define NCCL_CHECK(fn) {ncclResult_t err = fn; if(err != ncclSuccess) { GGML_ABORT("RCCL Failure RCCL returned: %i\n", err); }}
 #define __shfl_sync(mask, var, laneMask, width) __shfl(var, laneMask, width)
 #define __shfl_up_sync(mask, var, laneMask, width) __shfl_up(var, laneMask, width)
 #define __shfl_xor_sync(mask, var, laneMask, width) __shfl_xor(var, laneMask, width)

ggml/src/ggml-hexagon/htp/hmx-matmul-ops.c

@@ -648,9 +648,9 @@ static void dequantize_x4x2_weight_chunk_to_fp16_tiles(
     assert(n_cols  % HMX_FP16_TILE_N_COLS == 0);
     assert(k_block % HMX_FP16_TILE_N_COLS == 0);
 
-    int n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
-    int n_k_tiles   = k_block / HMX_FP16_TILE_N_COLS;
-    int n_tot_tiles = n_col_tiles * n_k_tiles;
+    size_t n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
+    size_t n_k_tiles   = k_block / HMX_FP16_TILE_N_COLS;
+    size_t n_tot_tiles = n_col_tiles * n_k_tiles;
 
     size_t n_tiles_per_task = hmx_ceil_div(n_tot_tiles, ctx->n_threads);
 
@@ -678,9 +678,8 @@ static void core_dot_chunk_fp16(__fp16 *restrict output, const __fp16 *restrict
     __builtin_assume(n_dot_tiles > 0);
 
     Q6_bias_mxmem2_A((void *)scales);
-
     for (int r = 0; r < n_row_tiles; ++r) {
-        for (int c = 0; c < n_col_tiles; ++c) {
+        for (size_t c = 0; c < n_col_tiles; ++c) {
             Q6_mxclracc_hf();
 
             const __fp16 *row_tiles = activation + r * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
@@ -738,25 +737,25 @@ static inline void hmx_matmul_job_init(hmx_matmul_job_t * job,
 
 static void transfer_output_chunk_fp16_to_fp32(float *restrict dst, const __fp16 *restrict vtcm_src, int n_rows, int n_cols, int n) {
     assert(n_cols % HMX_FP16_TILE_N_COLS == 0);
-    const int n_col_tiles = n_cols / HMX_FP16_TILE_N_COLS;
+    const size_t tile_row_stride = (n_cols / HMX_FP16_TILE_N_COLS) * HMX_FP16_TILE_N_ELMS;
 
     const HVX_Vector one = hvx_vec_splat_f16(1.0);
 
-    for (int r = 0; r < n_rows; r += 2) {
-        int r0 = r / HMX_FP16_TILE_N_ROWS;
-        int r1 = r % HMX_FP16_TILE_N_ROWS;
+    for (size_t r = 0; r < n_rows; r += 2) {
+        const size_t r0 = r / HMX_FP16_TILE_N_ROWS;
+        const size_t r1 = (r % HMX_FP16_TILE_N_ROWS) / 2;  // index of the row pair within the tile
+        const __fp16 *row_base = vtcm_src + r0 * tile_row_stride;
+        float *output_row_base = dst + r * n;  // global memory row base for row r (and r+1)
 
         #pragma unroll(4)
-        for (int c = 0; c < n_cols; c += HMX_FP16_TILE_N_COLS) {
-            int c0 = c / HMX_FP16_TILE_N_COLS;
-
-            const __fp16 *tile = vtcm_src + (r0 * n_col_tiles + c0) * HMX_FP16_TILE_N_ELMS;
-
-            HVX_Vector v = ((const HVX_Vector *) tile)[r1 / 2];
+        for (size_t c = 0; c < n_cols; c += HMX_FP16_TILE_N_COLS) {
+            const size_t c0 = c / HMX_FP16_TILE_N_COLS;
+            const __fp16 *tile = row_base + c0 * HMX_FP16_TILE_N_ELMS;
+            HVX_Vector v = ((const HVX_Vector *) tile)[r1];
             HVX_VectorPair vp = Q6_Wqf32_vmpy_VhfVhf(v, one);
 
-            volatile HVX_Vector *pv_out0 = (volatile HVX_Vector *) (dst + (r * n + c + 0));
-            volatile HVX_Vector *pv_out1 = (volatile HVX_Vector *) (dst + (r * n + c + n));  // next row in global memory
+            volatile HVX_Vector *pv_out0 = (volatile HVX_Vector *) (output_row_base + c + 0);
+            volatile HVX_Vector *pv_out1 = (volatile HVX_Vector *) (output_row_base + c + n);  // next row in global memory
 
             *pv_out0 = Q6_Vsf_equals_Vqf32(Q6_V_lo_W(vp));
             if (r + 1 < n_rows) {
@@ -794,7 +793,7 @@ static void transfer_output_chunk_threaded(struct htp_context *ctx, float *dst,
     assert(n_cols % HMX_FP16_TILE_N_COLS == 0);
 
     size_t n_tot_chunks      = n_rows;
-    size_t n_chunks_per_task = 32;  // must be multiple of HMX_FP16_TILE_N_ROWS (32)
+    size_t n_chunks_per_task = HMX_FP16_TILE_N_ROWS;  // must be multiple of HMX_FP16_TILE_N_ROWS (32)
 
     output_transfer_task_state_t state;
     state.n_tasks           = (n_tot_chunks + n_chunks_per_task - 1) / n_chunks_per_task;
@@ -926,7 +925,7 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
         return hmx_mat_mul_permuted_w16a32_batched_legacy(ctx, params);
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: grouped path m=%d k=%d n=%d group=%d streams=%d mc=%zu nc=%zu vtcm=%zu/%zu",
             __func__, params->m, params->k, params->n, group_size, params->ne13,
@@ -944,12 +943,15 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
     const size_t fp16_row_bytes   = (size_t) params->k * sizeof(__fp16);
     const size_t weight_row_bytes = (size_t) params->weight_stride * sizeof(__fp16);
 
+    HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
+
     for (int b3 = 0; b3 < params->ne13; ++b3) {
         for (int b2_base = 0; b2_base < params->ne12; b2_base += group_size) {
             const __fp16 *weight_group = hmx_matmul_weight_batch_ptr(params, b2_base, b3);
 
             for (size_t mr = 0; mr < (size_t) params->m; mr += m_chunk_n_rows) {
                 const size_t n_rows = hex_smin((size_t) params->m - mr, m_chunk_n_rows);
+                const size_t n_row_tiles = hmx_ceil_div((int) n_rows, HMX_FP16_TILE_N_ROWS);
 
                 // Pre-load activations for all heads in the group (once per m_chunk).
                 // When the source is strided (permuted Q), use 2D DMA to gather
@@ -987,10 +989,9 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                                       fp16_row_bytes, weight_row_bytes, fp16_row_bytes, n_cols_first);
                 }
 
-                HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
-
                 for (size_t nc = 0; nc < (size_t) params->n; nc += n_chunk_n_cols) {
                     const size_t n_cols = hex_smin((size_t) params->n - nc, n_chunk_n_cols);
+                    const size_t n_col_tiles = hmx_ceil_div((int) n_cols, HMX_FP16_TILE_N_COLS);
 
                     TIMER_START(weight_load);
                     {
@@ -1014,11 +1015,9 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                     for (int g = 0; g < group_size; ++g) {
                         TIMER_START(hmx_core);
                         {
-                            const __fp16 *vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
-                            const int n_row_tiles = hmx_ceil_div((int) n_rows, HMX_FP16_TILE_N_ROWS);
-                            const int n_col_tiles = hmx_ceil_div((int) n_cols, HMX_FP16_TILE_N_COLS);
-                            core_dot_chunk_fp16(vtcm_output, vtcm_act_g, vtcm_weight, vtcm_scales,
-                                                n_row_tiles, n_col_tiles, params->k / 32);
+                            const __fp16 * vtcm_act_g = vtcm_activation + (size_t) g * act_head_stride;
+                            core_dot_chunk_fp16(vtcm_output, vtcm_act_g, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles,
+                                                params->k / 32);
                         }
                         TIMER_STOP(hmx_core);
 
@@ -1030,12 +1029,12 @@ int hmx_mat_mul_permuted_w16a32_batched(struct htp_context *ctx, const hmx_matmu
                         TIMER_STOP(output_store);
                     }
                 }
-
-                HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
             }
         }
     }
 
+    HAP_compute_res_hmx_unlock(ctx->vtcm_rctx);
+
     TIMER_STOP(total);
 
 #if defined(ENABLE_PROFILE_TIMERS)
@@ -1103,7 +1102,7 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
         return -1;
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: m=%d k=%d n=%d mc=%zu nc=%zu vtcm=%zu/%zu",
          __func__, m, k, n, m_chunk_n_rows, n_chunk_n_cols,
@@ -1121,7 +1120,8 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
 
     for (size_t mr = 0; mr < m; mr += m_chunk_n_rows) {
         // transfer activation matrix chunk into VTCM
-        size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+        const size_t n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
 
         TIMER_START(activation_load);
         {
@@ -1159,7 +1159,8 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
         }
 
         for (size_t nc = 0; nc < n; nc += n_chunk_n_cols) {
-            size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+            const size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+            const size_t n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
 
             TIMER_START(weight_load);
             {
@@ -1184,8 +1185,6 @@ int hmx_mat_mul_permuted_w16a32(struct htp_context *ctx, float *restrict dst, co
 
             TIMER_START(hmx_core);
             {
-                const int n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
-                const int n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
                 core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles, k / 32);
             }
             TIMER_STOP(hmx_core);
@@ -1307,7 +1306,7 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
         return -1;
     }
 
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     FARF(MEDIUM, "%s: m=%d k=%d n=%d wtype=%d pipe=%d mc=%zu nc=%zu vtcm=%zu/%zu",
          __func__, m, k, n, weight_type, use_pipeline,
@@ -1330,7 +1329,8 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
         HAP_compute_res_hmx_lock(ctx->vtcm_rctx);
         for (size_t mr = 0; mr < m; mr += m_chunk_n_rows) {
             // transfer activation matrix chunk into VTCM
-            size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+            const size_t n_rows = hex_smin(m - mr, m_chunk_n_rows);
+            const size_t n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
 
             TIMER_START(activation_load);
             {
@@ -1348,7 +1348,8 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
             }
 
             for (size_t nc = 0; nc < n; nc += n_chunk_n_cols) {
-                size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+                const size_t n_cols = hex_smin(n - nc, n_chunk_n_cols);
+                const size_t n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
 
                 TIMER_START(weight_load);
                 {
@@ -1373,8 +1374,6 @@ int hmx_mat_mul_permuted_qk_0_d16a32(struct htp_context *ctx, float *restrict ds
 
                 TIMER_START(hmx_core);
                 {
-                    const int n_row_tiles = hmx_ceil_div(n_rows, HMX_FP16_TILE_N_ROWS);
-                    const int n_col_tiles = hmx_ceil_div(n_cols, HMX_FP16_TILE_N_COLS);
                     core_dot_chunk_fp16(vtcm_output, vtcm_activation, vtcm_weight, vtcm_scales, n_row_tiles, n_col_tiles, k / 32);
                 }
                 TIMER_STOP(hmx_core);
@@ -1521,14 +1520,16 @@ void core_mma_chunk_fp16(__fp16 *restrict c, const __fp16 *restrict a, const __f
 
     Q6_bias_mxmem2_A((void *)col_scales);
 
-    for (int i = 0; i < n_row_tiles; ++i) {
-        for (int j = 0; j < n_col_tiles; ++j) {
+    const size_t dot_tile_stride = n_dot_tiles * HMX_FP16_TILE_N_ELMS;
+    for (size_t i = 0; i < n_row_tiles; ++i) {
+        const __fp16 *row_base = a + i * dot_tile_stride;
+        __fp16 *res_base = c + i * n_col_tiles * HMX_FP16_TILE_N_ELMS;
+        for (size_t j = 0; j < n_col_tiles; ++j) {
             Q6_mxclracc_hf();
 
-            const __fp16 *row_tiles = a + i * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
-            const __fp16 *col_tiles = b + j * n_dot_tiles * HMX_FP16_TILE_N_ELMS;
-
-            __fp16 *accum_tile = c + (i * n_col_tiles + j) * HMX_FP16_TILE_N_ELMS;
+            const __fp16 *col_tiles = b + j * dot_tile_stride;
+            const __fp16 *row_tiles = row_base;
+            __fp16 *accum_tile = res_base + j * HMX_FP16_TILE_N_ELMS;
             if (!zero_init) {
                 Q6_activation_hf_mxmem_RR((unsigned int)accum_tile, 2047);
                 Q6_weight_hf_mxmem_RR((unsigned int)eye_tile, 2047);
@@ -1697,7 +1698,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
             v = Q6_V_vror_VR(v, VLEN - 8);
         }
     }
-    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // fp16: 1.0
+    hmx_init_column_scales(vtcm_scales, Q6_V_vsplat_R(0x3c00));  // scale: 1.0, bias: 0.0 in FP16
 
     TIMER_DEFINE(fetch);
     TIMER_DEFINE(act_load);
@@ -1715,7 +1716,7 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
             const int n_col_tiles = hmx_ceil_div(n_blk_sz, HMX_FP16_TILE_N_COLS);
 
             for (size_t kk = 0; kk < k; kk += K_BLOCK_SIZE) {
-                size_t k_blk_sz = hex_smin(k - kk, K_BLOCK_SIZE);
+                const size_t k_blk_sz = hex_smin(k - kk, K_BLOCK_SIZE);
 
                 TIMER_START(fetch);
                 // fetch activation block into VTCM
@@ -1731,13 +1732,13 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                 }
 
                 // fetch weight block into VTCM (x4x2 sub-block: quants + scales)
+                const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                 {
                     qweight_fetch_task_state_t s;
 
                     const int blk_start = kk / QK_Q4_0x4x2;
                     const int nb_sub = (k_blk_sz + QK_Q4_0x4x2 - 1) / QK_Q4_0x4x2;
                     const int    full_qrow      = (weight_type == HTP_TYPE_Q8_0) ? k : (k / 2);
-                    const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                     const int    scale_blk_size =
                         (weight_type == HTP_TYPE_MXFP4) ? HMX_X4X2_MXFP4_EBLK_SIZE : HMX_X4X2_DBLK_SIZE;
 
@@ -1777,7 +1778,6 @@ int mat_mul_qk_0_d16a32_out_stationary(struct htp_context *ctx, float *restrict
                     dma_queue_pop(ctx->dma[0]);
                     // vtcm_scratch0 is used to store the qweight chunk
                     // worker_pool_run_func already returned, so fetch is done
-                    const size_t sub_row_stride = get_x4x2_row_stride(weight_type, k_blk_sz);
                     dequantize_x4x2_weight_chunk_to_fp16_tiles(ctx, vtcm_weight, vtcm_scratch0,
                                                                 n_blk_sz, k_blk_sz, sub_row_stride, weight_type);
                 }

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

@@ -98,6 +98,8 @@ enum htp_op_code {
 #define HTP_OP_MAX_VMEM    (3221225472u)
 #endif
 
+#define HTP_MMAP_MAX_VMEM  (2147483648u)
+
 enum htp_tensor_flags {
     HTP_TENSOR_COMPUTE = (1U << 0), // Tensor buffer temporal compute data (not weights)
     HTP_TENSOR_FLUSHED = (1U << 1)  // Tensor buffer has been flushed (set by the NPU)

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

@@ -118,7 +118,11 @@ AEEResult htp_iface_close(remote_handle64 handle) {
     // release the mmaps (if any)
     for (uint32_t i=0; i<HTP_MAX_MMAPS; i++) {
         if (ctx->mmap[i].size) {
+#if __HVX_ARCH__ > 73
             HAP_munmap2((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#else
+            HAP_munmap((void *) ctx->mmap[i].base, ctx->mmap[i].size);
+#endif
             ctx->mmap[i].size = 0;
             ctx->mmap[i].base = NULL;
             ctx->mmap[i].fd   = -1;
@@ -173,8 +177,16 @@ AEEResult htp_iface_mmap(remote_handle64 handle, int fd, uint32_t size, uint32_t
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
             FARF(HIGH, "mmap : fd %u size %u pinned %u", fd, size, pinned);
-
+#if __HVX_ARCH__ > 73
             void *va = HAP_mmap2(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#else
+            if (size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, size, HAP_PROT_READ | HAP_PROT_WRITE, 0, fd, 0);
+#endif
             if (va == (void*)-1) {
                 FARF(ERROR, "mmap failed : va %p fd %u size %u", va, fd, (uint32_t) size);
                 return AEE_EFAILED;
@@ -202,7 +214,11 @@ AEEResult htp_iface_munmap(remote_handle64 handle, int fd) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (fd < 0 || m->fd == fd) {
             FARF(HIGH, "unmmap : base %p fd %u size %u", (void*) m->base, m->fd, (uint32_t) m->size);
+#if __HVX_ARCH__ > 73
             HAP_munmap2((void *) m->base, m->size);
+#else
+            HAP_munmap((void *) m->base, m->size);
+#endif
             m->size   = 0;
             m->base   = NULL;
             m->fd     = -1;
@@ -526,7 +542,11 @@ static inline bool reuse_buf(struct htp_context *ctx, uint32_t *m_reuse, struct
 static inline void drop_mmap(struct htp_context *ctx, struct htp_mmap *m) {
     if (m->size && !m->pinned) {
         FARF(HIGH, "unmap : fd %u base %p size %u pinned %u", m->fd, (void*) m->base, (uint32_t) m->size, m->pinned);
+#if __HVX_ARCH__ > 73
         HAP_munmap2((void *) m->base, m->size);
+#else
+        HAP_munmap((void *) m->base, m->size);
+#endif
         m->size = 0;
         m->base = 0;
         m->fd   = -1;
@@ -540,7 +560,16 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
     for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) {
         struct htp_mmap *m = &ctx->mmap[i];
         if (!m->size) {
+#if __HVX_ARCH__ > 73
             void *va = HAP_mmap2(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#else
+            if (b->size > HTP_MMAP_MAX_VMEM) { // HAP_mmap has a size limit of 2GB
+                FARF(ERROR, "mmap failed : size %u exceeds 2GB limit for HAP_mmap", (uint32_t) b->size);
+                abort(); // can't do much else at this point
+            }
+
+            void *va = HAP_mmap(NULL, b->size, HAP_PROT_READ | HAP_PROT_WRITE, 0, b->fd, 0);
+#endif
             if (va == (void*)-1) {
                 FARF(ERROR, "mmap failed : va %p fd %u size %u", va, b->fd, (uint32_t) b->size);
                 abort(); // can't do much else at this point

ggml/src/ggml-impl.h

@@ -30,6 +30,8 @@ extern "C" {
 
 void ggml_print_backtrace(void);
 
+uint64_t ggml_graph_next_uid(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif
@@ -338,6 +340,10 @@ struct ggml_cgraph {
     struct ggml_hash_set visited_hash_set;
 
     enum ggml_cgraph_eval_order order;
+
+    // an optional identifier that can be utilized to recognize same graphs if two non-zero values match
+    // a value of 0 means it is not set and should be ignored
+    uint64_t uid;
 };
 
 // returns a slice of cgraph with nodes [i0, i1)

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

@@ -1819,6 +1819,23 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale(ggml_met
     return res;
 }
 
+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_roll(ggml_metal_library_t lib, const ggml_tensor * op) {
+    assert(op->op == GGML_OP_ROLL);
+
+    char base[256];
+    char name[256];
+
+    snprintf(base, 256, "kernel_roll_%s", ggml_type_name(op->src[0]->type));
+    snprintf(name, 256, "%s", base);
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, base, name, nullptr);
+    }
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad(ggml_metal_library_t lib, const ggml_tensor * op) {
     assert(op->op == GGML_OP_PAD);
 

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

@@ -152,6 +152,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_conv_3d
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_upscale           (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad               (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_pad_reflect_1d    (ggml_metal_library_t lib, const struct ggml_tensor * op);
+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_roll              (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_arange            (ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_timestep_embedding(ggml_metal_library_t lib, const struct ggml_tensor * op);
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_opt_step_adamw    (ggml_metal_library_t lib, const struct ggml_tensor * op);

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

@@ -1138,6 +1138,7 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
         case GGML_OP_ARGSORT:
         case GGML_OP_TOP_K:
         case GGML_OP_ARANGE:
+        case GGML_OP_ROLL:
             return true;
         case GGML_OP_FLASH_ATTN_EXT:
             // for new head sizes, add checks here

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

@@ -1017,6 +1017,29 @@ typedef struct {
     int32_t  p1;
 } ggml_metal_kargs_pad_reflect_1d;
 
+typedef struct {
+    int64_t  ne00;
+    int64_t  ne01;
+    int64_t  ne02;
+    int64_t  ne03;
+    uint64_t nb00;
+    uint64_t nb01;
+    uint64_t nb02;
+    uint64_t nb03;
+    int64_t  ne0;
+    int64_t  ne1;
+    int64_t  ne2;
+    int64_t  ne3;
+    uint64_t nb0;
+    uint64_t nb1;
+    uint64_t nb2;
+    uint64_t nb3;
+    int32_t  s0;
+    int32_t  s1;
+    int32_t  s2;
+    int32_t  s3;
+} ggml_metal_kargs_roll;
+
 typedef struct {
     uint64_t nb1;
     int      dim;

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

@@ -410,6 +410,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
             {
                 n_fuse = ggml_metal_op_pad_reflect_1d(ctx, idx);
             } break;
+        case GGML_OP_ROLL:
+            {
+                n_fuse = ggml_metal_op_roll(ctx, idx);
+            } break;
         case GGML_OP_ARANGE:
             {
                 n_fuse = ggml_metal_op_arange(ctx, idx);
@@ -3945,6 +3949,59 @@ int ggml_metal_op_upscale(ggml_metal_op_t ctx, int idx) {
     return 1;
 }
 
+int ggml_metal_op_roll(ggml_metal_op_t ctx, int idx) {
+    ggml_tensor * op = ctx->node(idx);
+
+    ggml_metal_library_t lib = ctx->lib;
+    ggml_metal_encoder_t enc = ctx->enc;
+
+    GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
+    GGML_TENSOR_LOCALS( int32_t, ne,  op,         ne);
+    GGML_TENSOR_LOCALS(uint64_t, nb,  op,         nb);
+
+    const int32_t s0 = ggml_get_op_params_i32(op, 0);
+    const int32_t s1 = ggml_get_op_params_i32(op, 1);
+    const int32_t s2 = ggml_get_op_params_i32(op, 2);
+    const int32_t s3 = ggml_get_op_params_i32(op, 3);
+
+    ggml_metal_kargs_roll args = {
+        /*.ne00 =*/ ne00,
+        /*.ne01 =*/ ne01,
+        /*.ne02 =*/ ne02,
+        /*.ne03 =*/ ne03,
+        /*.nb00 =*/ nb00,
+        /*.nb01 =*/ nb01,
+        /*.nb02 =*/ nb02,
+        /*.nb03 =*/ nb03,
+        /*.ne0  =*/ ne0,
+        /*.ne1  =*/ ne1,
+        /*.ne2  =*/ ne2,
+        /*.ne3  =*/ ne3,
+        /*.nb0  =*/ nb0,
+        /*.nb1  =*/ nb1,
+        /*.nb2  =*/ nb2,
+        /*.nb3  =*/ nb3,
+        /*.s0   =*/ s0,
+        /*.s1   =*/ s1,
+        /*.s2   =*/ s2,
+        /*.s3   =*/ s3
+    };
+
+    auto pipeline = ggml_metal_library_get_pipeline_roll(lib, op);
+
+    const int nth = std::min(1024, ne0);
+
+    ggml_metal_encoder_set_pipeline(enc, pipeline);
+    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op->src[0]), 1);
+    ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         2);
+
+    ggml_metal_encoder_dispatch_threadgroups(enc, ne1, ne2, ne3, nth, 1, 1);
+
+    return 1;
+}
+
 int ggml_metal_op_pad(ggml_metal_op_t ctx, int idx) {
     ggml_tensor * op = ctx->node(idx);
 

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

@@ -81,6 +81,7 @@ int ggml_metal_op_conv_transpose_2d (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_upscale           (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pad               (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_pad_reflect_1d    (ggml_metal_op_t ctx, int idx);
+int ggml_metal_op_roll              (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_arange            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_timestep_embedding(ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_argmax            (ggml_metal_op_t ctx, int idx);

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

@@ -5247,6 +5247,40 @@ kernel void kernel_upscale_bicubic_f32(
     }
 }
 
+kernel void kernel_roll_f32(
+    constant ggml_metal_kargs_roll & args,
+    device  const char * src0,
+    device        char * dst,
+    uint3 tgpig[[threadgroup_position_in_grid]],
+    uint3 tpitg[[thread_position_in_threadgroup]],
+    uint3   ntg[[threads_per_threadgroup]]) {
+
+    const int64_t i3 = tgpig.z;
+    const int64_t i2 = tgpig.y;
+    const int64_t i1 = tgpig.x;
+
+    device const float * src0_ptr = (device const float *) src0;
+    device       float * dst_ptr  = (device       float *) dst;
+
+    for (int i0 = tpitg.x; i0 < args.ne0; i0 += ntg.x) {
+        // apply shifts and wrap around
+        int64_t i00 = i0 - args.s0;
+        int64_t i01 = i1 - args.s1;
+        int64_t i02 = i2 - args.s2;
+        int64_t i03 = i3 - args.s3;
+
+        if (i00 < 0) { i00 += args.ne00; } else if (i00 >= args.ne00) { i00 -= args.ne00; }
+        if (i01 < 0) { i01 += args.ne01; } else if (i01 >= args.ne01) { i01 -= args.ne01; }
+        if (i02 < 0) { i02 += args.ne02; } else if (i02 >= args.ne02) { i02 -= args.ne02; }
+        if (i03 < 0) { i03 += args.ne03; } else if (i03 >= args.ne03) { i03 -= args.ne03; }
+
+        int64_t src_idx = i03*args.ne02*args.ne01*args.ne00 + i02*args.ne01*args.ne00 + i01*args.ne00 + i00;
+        int64_t dst_idx = i3 *args.ne2 *args.ne1 *args.ne0  + i2 *args.ne1 *args.ne0  + i1 *args.ne0  + i0;
+
+        dst_ptr[dst_idx] = src0_ptr[src_idx];
+    }
+}
+
 kernel void kernel_pad_f32(
     constant ggml_metal_kargs_pad & args,
     device  const char * src0,

ggml/src/ggml-opencl/CMakeLists.txt

@@ -121,6 +121,8 @@ set(GGML_OPENCL_KERNELS
     gemm_noshuffle_q4_k_f32
     gemv_noshuffle_q6_k_f32
     gemm_noshuffle_q6_k_f32
+    gemv_noshuffle_q5_k_f32
+    gemm_noshuffle_q5_k_f32
     mul
     neg
     norm

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

@@ -542,6 +542,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_restore_block_q4_K_noshuffle;
     cl_kernel kernel_convert_block_q4_K, kernel_restore_block_q4_K;
     cl_kernel kernel_convert_block_q5_K, kernel_restore_block_q5_K;
+    cl_kernel kernel_convert_block_q5_K_noshuffle;
+    cl_kernel kernel_restore_block_q5_K_noshuffle;
     cl_kernel kernel_convert_block_q6_K, kernel_restore_block_q6_K;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q4_1_f32;
@@ -730,6 +732,8 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gemm_noshuffle_q4_k_f32;
     cl_kernel kernel_gemv_noshuffle_q6_K_f32;
     cl_kernel kernel_gemm_noshuffle_q6_K_f32;
+    cl_kernel kernel_gemv_noshuffle_q5_k_f32;
+    cl_kernel kernel_gemm_noshuffle_q5_k_f32;
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
     void free() {
@@ -944,6 +948,8 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_restore_block_q4_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q4_K_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q5_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_K", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q5_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_K", &err), err));
+        CL_CHECK((backend_ctx->kernel_convert_block_q5_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q5_K_noshuffle", &err), err));
+        CL_CHECK((backend_ctx->kernel_restore_block_q5_K_noshuffle = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q5_K_noshuffle", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K", &err), err));
         CL_CHECK((backend_ctx->kernel_restore_block_q6_K  = clCreateKernel(backend_ctx->program_cvt, "kernel_restore_block_q6_K", &err), err));
         CL_CHECK((backend_ctx->kernel_convert_block_q6_K_noshuffle  = clCreateKernel(backend_ctx->program_cvt, "kernel_convert_block_q6_K_noshuffle", &err), err));
@@ -2794,6 +2800,45 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q6_K_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q6_K_f32", &err), err));
         GGML_LOG_CONT(".");
     }
+
+    // gemv_noshuffle_q5_k_f32
+    {
+        std::string CL_gemv_compile_opts = std::string("-cl-std=") + opencl_c_std +
+                                       " -cl-mad-enable ";
+        if (backend_ctx->has_vector_subgroup_broadcast) {
+            CL_gemv_compile_opts += " -DVECTOR_SUB_GROUP_BROADCAST ";
+        }
+
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemv_noshuffle_q5_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemv_noshuffle_q5_k_f32.cl");
+#endif
+
+        cl_program prog = build_program_from_source(
+            backend_ctx->context, backend_ctx->device, kernel_src.c_str(), CL_gemv_compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_gemv_noshuffle_q5_k_f32 = clCreateKernel(prog, "kernel_gemv_noshuffle_q5_k_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
+
+    // gemm_noshuffle_q5_k_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "gemm_noshuffle_q5_k_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("gemm_noshuffle_q5_k_f32.cl");
+#endif
+        cl_program prog = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+        CL_CHECK((backend_ctx->kernel_gemm_noshuffle_q5_k_f32 = clCreateKernel(prog, "kernel_gemm_noshuffle_q5_k_f32", &err), err));
+        CL_CHECK(clReleaseProgram(prog));
+        GGML_LOG_CONT(".");
+    }
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
     GGML_LOG_CONT("\n");
 }
@@ -5071,115 +5116,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
             GGML_ASSERT(tensor->ne[2] == 1);
             GGML_ASSERT(tensor->ne[3] == 1);
 
-            // Transpose weights
-            size_t q_size_bytes = K * M / 4 * sizeof(float);
-            cl_buffer_region region;
-            region.origin = 0;
-            region.size = q_size_bytes;
-            cl_mem qT_d = clCreateSubBuffer(
-                backend_ctx->prealloc_quant_trans.buffer,
-                0,
-                CL_BUFFER_CREATE_TYPE_REGION,
-                &region,
-                &err);
-            CL_CHECK(err);
-
-            cl_mem q_d_image1D;
-            cl_mem qT_d_image1D;
-
-            cl_image_format img_fmt_1d;
-            cl_image_desc img_desc_1d;
-
-            img_fmt_1d = { CL_RGBA, CL_FLOAT };
-            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-            img_desc_1d.image_width = M * K / 4 / 4;
-            img_desc_1d.buffer = extra->q;
-            q_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
-            CL_CHECK(err);
-
-            img_fmt_1d = { CL_RGBA, CL_FLOAT };
-            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-            img_desc_1d.image_width = M * K / 4 / 4;
-            img_desc_1d.buffer = qT_d;
-            qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
-            CL_CHECK(err);
-
-            int height_q = M / 4;
-            int width_q = K / 4 / 4;
-            kernel = backend_ctx->kernel_transpose_32;
-
-            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &q_d_image1D));
-            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &qT_d_image1D));
-            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_q));
-            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_q));
-
-            size_t local_size_q[3] = {4, 16, 1};
-            size_t global_size_q[3] = {static_cast<size_t>(width_q), static_cast<size_t>(height_q), 1};
-            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_q, local_size_q, 0, NULL, &evt));
-            CL_CHECK(clWaitForEvents(1, &evt));
-
-            // Transpose scales
-            size_t d_size_bytes = M * (K / 32) * 2;
-            region.origin = 0;
-            region.size = d_size_bytes;
-            cl_mem dT_d = clCreateSubBuffer(
-                backend_ctx->prealloc_scales_trans.buffer,
-                0,
-                CL_BUFFER_CREATE_TYPE_REGION,
-                &region,
-                &err);
-            CL_CHECK(err);
-
-            cl_mem d_d_image1D;
-            cl_mem dT_d_image1D;
-
-            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
-            img_desc_1d.image_width = M * K / 32;
-            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-            img_desc_1d.buffer = extra->d;
-            d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
-            CL_CHECK(err);
-
-            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
-            memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-            img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-            img_desc_1d.image_width = M * K / 32 / 4;
-            img_desc_1d.buffer = dT_d;
-            dT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
-            CL_CHECK(err);
-
-            int height_s = M / 4;
-            int width_s = K / 32;
-
-            kernel = backend_ctx->kernel_transpose_16_4x1;
-
-            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
-            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
-            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
-            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int), &width_s));
-
-            size_t local_size_s[3] = {4, 16, 1};
-            size_t global_size_s[3] = {static_cast<size_t>(width_s), static_cast<size_t>(height_s), 1};
-            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_size_s, local_size_s, 0, NULL, &evt));
-            CL_CHECK(clWaitForEvents(1, &evt));
-
-            // copy transposed buffer contents to original buffers
-            CL_CHECK(clEnqueueCopyBuffer(queue, qT_d, extra->q, 0, 0, q_size_bytes, 0, NULL, &evt));
-            CL_CHECK(clWaitForEvents(1, &evt));
-
-            CL_CHECK(clEnqueueCopyBuffer(queue, dT_d, extra->d, 0, 0, d_size_bytes, 0, NULL, &evt));
-            CL_CHECK(clWaitForEvents(1, &evt));
-
-            CL_CHECK(clReleaseMemObject(qT_d));
-            CL_CHECK(clReleaseMemObject(dT_d));
-
-            CL_CHECK(clReleaseMemObject(q_d_image1D));
-            CL_CHECK(clReleaseMemObject(d_d_image1D));
-            CL_CHECK(clReleaseMemObject(qT_d_image1D));
-            CL_CHECK(clReleaseMemObject(dT_d_image1D));
+            transpose_2d_as_32b(backend_ctx, extra->q, extra->q, size_q, K/4,  M);
+            transpose_2d_as_16b(backend_ctx, extra->d, extra->d, size_d, K/32, M);
         } // end transpose
 #endif // GGML_OPENCL_USE_ADRENO_KERNELS
 
@@ -5354,7 +5292,17 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         CL_CHECK((extra->qh = clCreateSubBuffer(extra_orig->data_device, CL_MEM_READ_WRITE, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
         CL_CHECK(err);
 
+        #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
         cl_kernel kernel = backend_ctx->kernel_convert_block_q5_K;
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            kernel = backend_ctx->kernel_convert_block_q5_K_noshuffle;
+        }
+        #else
+        cl_kernel kernel = backend_ctx->kernel_convert_block_q5_K;
+        #endif
+
+        cl_uchar mask_0F = 0x0F;
+        cl_uchar mask_F0 = 0xF0;
 
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &data_device));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra->q));
@@ -5362,6 +5310,8 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra->s));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra->d));
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &extra->dm));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_0F));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_uchar), &mask_F0));
 
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {64, 1, 1};
@@ -5378,6 +5328,21 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         extra->size_dm = size_dm;
 
         tensor->extra = extra;
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+
+            GGML_ASSERT(K % 32 == 0);
+
+            // Transpose q, d, dm as ushort, qh as uchar
+            transpose_2d_as_16b(backend_ctx, extra->q,  extra->q,  size_q,  K/4,   M);
+            transpose_2d_as_8b (backend_ctx, extra->qh, extra->qh, size_qh, K/8,   M);
+            transpose_2d_as_16b(backend_ctx, extra->d,  extra->d,  size_d,  K/256, M);
+            transpose_2d_as_16b(backend_ctx, extra->dm, extra->dm, size_dm, K/256, M);
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
         return;
     }
     if (tensor->type == GGML_TYPE_Q6_K) {
@@ -5894,6 +5859,57 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
             ggml_nbytes(tensor), NULL, &err);
         CL_CHECK(err);
 
+        cl_uchar mask_0F = 0x0F;
+        cl_uchar mask_F0 = 0xF0;
+
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+        if (use_adreno_kernels(backend_ctx, tensor)) {
+            int M = tensor->ne[1];
+            int K = tensor->ne[0];
+
+            size_t size_q  = extra->size_q;
+            size_t size_qh = extra->size_qh;
+            size_t size_d  = extra->size_d;
+            size_t size_dm = extra->size_dm;
+
+            static ggml_cl_buffer buf_trans_q;
+            static ggml_cl_buffer buf_trans_qh;
+            static ggml_cl_buffer buf_trans_d;
+            static ggml_cl_buffer buf_trans_dm;
+
+            buf_trans_q.allocate(backend_ctx->context, size_q);
+            buf_trans_qh.allocate(backend_ctx->context, size_qh);
+            buf_trans_d.allocate(backend_ctx->context, size_d);
+            buf_trans_dm.allocate(backend_ctx->context, size_dm);
+
+            // Reverse transpose q, qh, d, dm
+            transpose_2d_as_16b(backend_ctx, extra->q,  buf_trans_q.buffer,  size_q,  M, K/4);
+            transpose_2d_as_8b (backend_ctx, extra->qh, buf_trans_qh.buffer, size_qh, M, K/8);
+            transpose_2d_as_16b(backend_ctx, extra->d,  buf_trans_d.buffer,  size_d,  M, K/256);
+            transpose_2d_as_16b(backend_ctx, extra->dm, buf_trans_dm.buffer, size_dm, M, K/256);
+
+            cl_kernel kernel = backend_ctx->kernel_restore_block_q5_K_noshuffle;
+            CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &buf_trans_q.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &buf_trans_qh.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra->s));
+            CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &buf_trans_d.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &buf_trans_dm.buffer));
+            CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &data_device));
+            CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_0F));
+            CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_uchar), &mask_F0));
+
+            size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
+            size_t local_work_size[] = {1, 1, 1};
+
+            CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL,
+                global_work_size, local_work_size, 0, NULL, NULL));
+            CL_CHECK(clEnqueueReadBuffer(queue, data_device, CL_TRUE, offset,
+                size, data, 0, NULL, NULL));
+            CL_CHECK(clReleaseMemObject(data_device));
+            return;
+        }
+#endif // GGML_OPENCL_USE_ADRENO_KERNELS
+
         cl_kernel kernel = backend_ctx->kernel_restore_block_q5_K;
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra->q));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),   &extra->qh));
@@ -5901,6 +5917,8 @@ static void ggml_backend_opencl_buffer_get_tensor(ggml_backend_buffer_t buffer,
         CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_mem),   &extra->d));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra->dm));
         CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_mem),   &data_device));
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_uchar), &mask_0F));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_uchar), &mask_F0));
 
         size_t global_work_size[] = {(size_t)ggml_nelements(tensor)/ggml_blck_size(tensor->type), 1, 1};
         size_t local_work_size[] = {1, 1, 1};
@@ -9831,19 +9849,18 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
     GGML_ASSERT(dst);
     GGML_ASSERT(dst->extra);
 
-    const enum ggml_type src0t = src0->type;
-    const enum ggml_type src1t = src1->type;
-
-    GGML_ASSERT(src0t == GGML_TYPE_Q8_0);
-    GGML_ASSERT(src1t == GGML_TYPE_F32);
+    GGML_ASSERT(src0->type == GGML_TYPE_Q8_0);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
-
     ggml_tensor_extra_cl_q8_0 * extra0_q8_0 = (ggml_tensor_extra_cl_q8_0 *)src0->extra;
 
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
     GGML_ASSERT(src1->view_offs == 0);
     GGML_ASSERT(dst->view_offs == 0);
 
@@ -9864,148 +9881,112 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
     cl_context context = backend_ctx->context;
     cl_kernel kernel;
 
-    // init CL objects
-    cl_int              status;
-    cl_image_format     img_fmt_1d;
-    cl_image_desc       img_desc_1d;
+    cl_int              err;
+    cl_image_format     img_fmt;
+    cl_image_desc       img_desc;
     cl_buffer_region    region;
-    cl_mem              A_image1d;
-    cl_mem              B_image1d;
-    cl_mem              B_sub_buffer;
-    cl_mem              S_image1d;
-    // for B transpose
-    cl_mem              B_image1d_trans = nullptr;
-    cl_mem              B_d = nullptr;
-
-    cl_mem              D_image1d;
-    cl_mem              D_sub_buffer;
 
     int M = ne01;
     int N = ne1;
     int K = ne00;
 
-    // create an image for A
-    img_fmt_1d = { CL_R, CL_FLOAT};
-    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-    img_desc_1d.image_width = M * K / 4;    // Divide by 4 for char -> float
-    img_desc_1d.buffer = extra0_q8_0->q;
-    A_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
-    CL_CHECK(status);
+    if (ne1 == 1) {
+        cl_mem q_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
 
-    // create an image for Scale
-    img_fmt_1d = { CL_R, CL_HALF_FLOAT};
-    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-    img_desc_1d.image_width = M * K / 32;    // Block size is 32
-    img_desc_1d.buffer = extra0_q8_0->d;
-    S_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
-    CL_CHECK(status);
+        // image for q
+        img_fmt = { CL_R, CL_UNSIGNED_INT32};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 4;
+        img_desc.buffer = extra0_q8_0->q;
+        CL_CHECK((q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
 
-    // create a sub_buffer for B
-    region.origin = (extra1->offset); // + src1->view_offs);
-    region.size = K * N * sizeof(float);
-    B_sub_buffer = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
-    CL_CHECK(status);
+        // create a sub_buffer for B
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer((extra1->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
 
-    // create an image for B from sub_buffer: RGBA (OCL)
-    img_fmt_1d = {CL_RGBA, CL_FLOAT};
-    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-    img_desc_1d.image_width = K * N / 4;
-    img_desc_1d.buffer = B_sub_buffer;
-    B_image1d = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
-    CL_CHECK(status);
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
 
-    // Create subbuffer and image1d_buffer for dst
-    region.origin = (extrad->offset); // + dst->view_offs;
-    region.size = M * N * sizeof(float);
-    D_sub_buffer = clCreateSubBuffer((extrad->data_device), 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &status);
-    CL_CHECK(status);
-
-    img_fmt_1d = {CL_R, CL_FLOAT};
-    memset(&img_desc_1d, 0, sizeof(img_desc_1d));
-    img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-    img_desc_1d.image_width = M * N;
-    img_desc_1d.buffer = D_sub_buffer;
-    D_image1d = clCreateImage(context, CL_MEM_WRITE_ONLY, &img_fmt_1d, &img_desc_1d, NULL, &status);
-    CL_CHECK(status);
-
-    size_t local_work_size[3] = {1, 1, 1};
-    size_t global_work_size[3] = {1, 1, 1};
-
-    if (N == 1) {
         kernel = backend_ctx->CL_mul_mat_vec_q8_0_f32;
 
         int r2 = 1;
         int r3 = 1;
-        cl_uint k_arg = 0;
 
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &A_image1d));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extra0_q8_0->d));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &B_image1d));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extra1->offset));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_mem),   &extrad->data_device));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(cl_ulong), &extrad->offset));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne00));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne01));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne02));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne10));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne12));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne0));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &ne1));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r2));
-        CL_CHECK(clSetKernelArg(kernel,  k_arg++, sizeof(int),      &r3));
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q_img));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &extra1->offset));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &extrad->offset));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(int),      &ne10));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne0));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
 
         size_t wavesize = backend_ctx->adreno_wave_size;
-        local_work_size[0] = wavesize;
-        local_work_size[1] = 4; // reduce factor
-        local_work_size[2] = 1;
+        size_t local_work_size[]  = { wavesize, 4, 1 };
+        size_t global_work_size[] = { CEIL_DIV(M, wavesize)*wavesize, 4, 1 };
 
-        global_work_size[0] = ((M + wavesize - 1) / wavesize) * wavesize;
-        global_work_size[1] = 4; // reduce factor
-        global_work_size[2] = 1;
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(q_img));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
     } else {
-        cl_ulong offsetd = extrad->offset + dst->view_offs;
-        int padding;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img = nullptr;
+        cl_mem b_img_trans = nullptr;
 
-        //how many extra elements beyond multiple of 8
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
         int extra_elements = N % 8;
-
-        //how much padding to add
-        padding = 0;
+        int padding = 0;
         if (extra_elements > 0){
             padding = 8 - extra_elements;
         }
 
-        // Specify the starting offset (in bytes)
+        // subbuffer for transposed activations
         region.origin = 0;
-        // Specify the size of the sub-buffer (divide by 2 for FP16)
         region.size = K * (N + padding) * sizeof(float)/2;
         backend_ctx->prealloc_act_trans.allocate(context, region.size);
-        B_d = clCreateSubBuffer(
-            backend_ctx->prealloc_act_trans.buffer,
-            0,
-            CL_BUFFER_CREATE_TYPE_REGION,
-            &region,
-            &status);
-        CL_CHECK(status);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
 
-        cl_image_format image_format_B_d_output = { CL_RGBA, CL_HALF_FLOAT }; //(CL_HALF_FLOAT for FP16)
-        cl_image_desc image_desc_B_d_output = {
-            CL_MEM_OBJECT_IMAGE1D_BUFFER,
-            static_cast<size_t>(K * (N + padding)/4),
-            0, 0, 0, 0, 0, 0, 0, { B_d }
-        };
-        B_image1d_trans = clCreateImage(
-            context,
-            0,
-            &image_format_B_d_output,
-            &image_desc_B_d_output,
-            NULL,
-            &status);
-        CL_CHECK(status);
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
 
+        // transpose activations
         int height_B = N/4;
         if (height_B == 0) {
             height_B = 1;
@@ -10014,58 +9995,39 @@ static void ggml_cl_mul_mat_q8_0_f32_adreno(ggml_backend_t backend, const ggml_t
         int padded_height_B = (N + padding)/4;
 
         kernel = backend_ctx->kernel_transpose_32_16;
-        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &B_image1d));
-        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
         CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
         CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
 
-        size_t local_size_t[2] = { 1, 16 };
-        size_t global_size_t[2] = {
-            static_cast<size_t>(width_B),
-            static_cast<size_t>(padded_height_B)
-        };
-
-        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_size_t, local_size_t, dst);
+        size_t local_work_size_t[2] = { 1, 16 };
+        size_t global_work_size_t[2] = { (size_t)width_B, (size_t)padded_height_B };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
 
+        // gemm
         kernel = backend_ctx->kernel_mul_mm_q8_0_f32_8x4;
-
-        int N_with_padding = N + padding;
+        int padded_N = N + padding;
 
         CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q8_0->q));
         CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q8_0->d));
-        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &B_image1d_trans));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &b_img_trans));
         CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extrad->data_device));
         CL_CHECK(clSetKernelArg(kernel,  4, sizeof(int),      &K));
         CL_CHECK(clSetKernelArg(kernel,  5, sizeof(int),      &M));
-        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &N_with_padding));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &padded_N));
         CL_CHECK(clSetKernelArg(kernel,  7, sizeof(int),      &N));
         CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &offsetd));
 
-        global_work_size[0] = (size_t)(N + 7) / 8;
-        global_work_size[1] = (size_t)(M + 3) / 4;
-        global_work_size[2] = 1;
+        size_t global_work_size[] = { (size_t)CEIL_DIV(N, 8), (size_t)CEIL_DIV(M, 4), 1 };
+        size_t local_work_size[]  = { 2, 128, 1 };
 
-        local_work_size[0] = 2;
-        local_work_size[1] = 128;
-        local_work_size[2] = 1;
-    }
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
 
-    // enqueue kernel with profiling
-    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
-
-    // deallocate sub buffers and images
-    CL_CHECK(clReleaseMemObject(A_image1d));
-    CL_CHECK(clReleaseMemObject(B_sub_buffer));
-    CL_CHECK(clReleaseMemObject(B_image1d));
-    CL_CHECK(clReleaseMemObject(S_image1d));
-    CL_CHECK(clReleaseMemObject(D_sub_buffer));
-    CL_CHECK(clReleaseMemObject(D_image1d));
-    if (B_image1d_trans) {
-        CL_CHECK(clReleaseMemObject(B_image1d_trans));
-    }
-    if (B_d) {
-        CL_CHECK(clReleaseMemObject(B_d));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
     }
 #else
     GGML_UNUSED(backend);
@@ -10451,6 +10413,201 @@ static void ggml_cl_mul_mat_q6_K_f32_adreno(ggml_backend_t backend, const ggml_t
 #endif
 }
 
+static void ggml_cl_mul_mat_q5_K_f32_adreno(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+#ifdef GGML_OPENCL_USE_ADRENO_KERNELS
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl_q5_K * extra0_q5_k = (ggml_tensor_extra_cl_q5_K *)src0->extra;
+
+    cl_ulong offset1 = extra1->offset + src1->view_offs;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne1  = dst->ne[1];
+
+    GGML_ASSERT(ne00 % ggml_blck_size(src0->type) == 0);
+
+    cl_context context = backend_ctx->context;
+    cl_kernel kernel;
+
+    cl_int           err;
+    cl_image_format  img_fmt;
+    cl_image_desc    img_desc;
+    cl_buffer_region region;
+
+    int M = ne01;
+    int N = ne1;
+    int K = ne00;
+
+    cl_uchar mask_d6  = 0x3F;
+    cl_uchar mask_d4  = 0x0F;
+    cl_uchar mask_hi2 = 0xC0;
+
+    if (ne1 == 1) {
+        cl_mem q_img  = nullptr;
+        cl_mem qh_img = nullptr;
+        cl_mem b_sub_buf = nullptr;
+        cl_mem b_img = nullptr;
+
+        // image for q (CL_R, CL_UNSIGNED_INT32): width = M*K/2/4
+        img_fmt = {CL_R, CL_UNSIGNED_INT32};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 2 / 4;
+        img_desc.buffer      = extra0_q5_k->q;
+        CL_CHECK((q_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // image for qh (CL_R, CL_HALF_FLOAT): width = M*K/16
+        img_fmt = {CL_R, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = M * K / 16;
+        img_desc.buffer      = extra0_q5_k->qh;
+        CL_CHECK((qh_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size   = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations (CL_RGBA, CL_FLOAT): width = K*N/4
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer      = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        kernel = backend_ctx->kernel_gemv_noshuffle_q5_k_f32;
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &q_img));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &qh_img));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q5_k->d));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q5_k->dm));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra0_q5_k->s));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_mem),   &b_img));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_uchar), &mask_d6));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_uchar), &mask_d4));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_uchar), &mask_hi2));
+
+        size_t local_work_size[3]  = {64, 4, 1};
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne01/2, 64)*64, 4, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(q_img));
+        CL_CHECK(clReleaseMemObject(qh_img));
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_img));
+    } else {
+        cl_mem b_sub_buf      = nullptr;
+        cl_mem b_sub_buf_trans = nullptr;
+        cl_mem b_img          = nullptr;
+        cl_mem b_img_trans    = nullptr;
+
+        // subbuffer for activations
+        region.origin = offset1;
+        region.size   = K * N * sizeof(float);
+        CL_CHECK((b_sub_buf = clCreateSubBuffer(extra1->data_device, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for activations
+        img_fmt = {CL_RGBA, CL_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * N / 4;
+        img_desc.buffer      = b_sub_buf;
+        CL_CHECK((b_img = clCreateImage(context, CL_MEM_READ_ONLY, &img_fmt, &img_desc, NULL, &err), err));
+
+        // pad N to multiple of 8
+        int extra_elements = N % 8;
+        int padding = 0;
+        if (extra_elements > 0) {
+            padding = 8 - extra_elements;
+        }
+
+        // subbuffer for transposed activations
+        region.origin = 0;
+        region.size   = K * (N + padding) * sizeof(float) / 2;
+        backend_ctx->prealloc_act_trans.allocate(context, region.size);
+        CL_CHECK((b_sub_buf_trans = clCreateSubBuffer(backend_ctx->prealloc_act_trans.buffer, 0, CL_BUFFER_CREATE_TYPE_REGION, &region, &err), err));
+
+        // image for transposed activations
+        img_fmt = {CL_RGBA, CL_HALF_FLOAT};
+        memset(&img_desc, 0, sizeof(img_desc));
+        img_desc.image_type  = CL_MEM_OBJECT_IMAGE1D_BUFFER;
+        img_desc.image_width = K * (N + padding) / 4;
+        img_desc.buffer      = b_sub_buf_trans;
+        CL_CHECK((b_img_trans = clCreateImage(context, 0, &img_fmt, &img_desc, NULL, &err), err));
+
+        // transpose activations
+        int height_B       = N / 4;
+        if (height_B == 0) height_B = 1;
+        int width_B        = K / 4;
+        int padded_height_B = (N + padding) / 4;
+
+        kernel = backend_ctx->kernel_transpose_32_16;
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &b_img));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int),    &height_B));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(int),    &width_B));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),    &padded_height_B));
+
+        size_t local_work_size_t[2]  = {1, 16};
+        size_t global_work_size_t[2] = {(size_t)width_B, (size_t)padded_height_B};
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size_t, local_work_size_t, dst);
+
+        // gemm
+        kernel = backend_ctx->kernel_gemm_noshuffle_q5_k_f32;
+        int padded_N = N + padding;
+
+        CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0_q5_k->q));
+        CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_mem),   &extra0_q5_k->qh));
+        CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra0_q5_k->s));
+        CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_mem),   &extra0_q5_k->d));
+        CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra0_q5_k->dm));
+        CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_mem),   &b_img_trans));
+        CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+        CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+        CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_int),   &ne01));
+        CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_int),   &padded_N));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_int),   &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_int),   &ne1));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_uchar), &mask_d6));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_uchar), &mask_d4));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_uchar), &mask_hi2));
+
+        size_t global_work_size[3] = {(size_t)CEIL_DIV(ne1, 8), (size_t)CEIL_DIV(ne01, 4), 1};
+        size_t local_work_size[3]  = {1, 128, 1};
+
+        backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size, dst);
+
+        CL_CHECK(clReleaseMemObject(b_sub_buf));
+        CL_CHECK(clReleaseMemObject(b_sub_buf_trans));
+        CL_CHECK(clReleaseMemObject(b_img));
+        CL_CHECK(clReleaseMemObject(b_img_trans));
+    }
+#else
+    GGML_UNUSED(backend);
+    GGML_UNUSED(src0);
+    GGML_UNUSED(src1);
+    GGML_UNUSED(dst);
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -10600,6 +10757,12 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
         return;
     }
 
+    // q5_K x fp32
+    if (src0t == GGML_TYPE_Q5_K && src1t == GGML_TYPE_F32) {
+        ggml_cl_mul_mat_q5_K_f32_adreno(backend, src0, src1, dst);
+        return;
+    }
+
     // q4_0 x fp32
     if(src0t == GGML_TYPE_Q4_0 && src1t == GGML_TYPE_F32) {
         // TODO: remove duplicate definitions of image description + format -- move to top

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

@@ -568,7 +568,9 @@ kernel void kernel_convert_block_q5_K(
     global uchar * dst_qh,
     global uchar * dst_s,
     global half  * dst_d,
-    global half  * dst_dm
+    global half  * dst_dm,
+    uchar mask_0F,
+    uchar mask_F0
 ) {
     global struct block_q5_K * b  = (global struct block_q5_K *) src0 + get_global_id(0);
     global uchar * q  = (global uchar *) dst_q  + QK_K/2*get_global_id(0);
@@ -599,7 +601,9 @@ kernel void kernel_restore_block_q5_K(
     global uchar * src_s,
     global half  * src_d,
     global half  * src_dm,
-    global struct block_q5_K * dst
+    global struct block_q5_K * dst,
+    uchar mask_0F,
+    uchar mask_F0
 ) {
     global struct block_q5_K * b  = (global struct block_q5_K *) dst + get_global_id(0);
     global uchar * q  = (global uchar *) src_q  + QK_K/2*get_global_id(0);
@@ -622,6 +626,92 @@ kernel void kernel_restore_block_q5_K(
     }
 }
 
+kernel void kernel_convert_block_q5_K_noshuffle(
+    global struct block_q5_K * src0,
+    global uchar * dst_q,
+    global uchar * dst_qh,
+    global uchar * dst_s,
+    global half  * dst_d,
+    global half  * dst_dm,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q5_K * b  = (global struct block_q5_K *) src0 + get_global_id(0);
+    global uchar * q  = (global uchar *) dst_q  + QK_K/2       * get_global_id(0);
+    global uchar * qh = (global uchar *) dst_qh + QK_K/8       * get_global_id(0);
+    global uchar * s  = (global uchar *) dst_s  + K_SCALE_SIZE * get_global_id(0);
+    global half  * d  = (global half  *) dst_d  + get_global_id(0);
+    global half  * dm = (global half  *) dst_dm + get_global_id(0);
+
+    *d  = b->d;
+    *dm = b->dm;
+
+    for (int i = 0; i < QK_K / 64; ++i) {
+        for (int j = 0; j < 16; ++j) {
+            uchar x0 = b->qs[i*32 + 2*j];
+            uchar x1 = b->qs[i*32 + 2*j + 1];
+            q[i*32 + j]      = convert_uchar(x0 & mask_0F) | convert_uchar((x1 & mask_0F) << 4);
+            q[i*32 + j + 16] = convert_uchar((x0 & mask_F0) >> 4) | convert_uchar(x1 & mask_F0);
+        }
+    }
+
+    for (int l = 0; l < QK_K/8; ++l) {
+        uchar x0 = 0;
+        for (int i = 0; i < 8; ++i) {
+            x0 |= ((b->qh[(l%4)*8+i] >> (l/4)) & 0x01) << i;
+        }
+        qh[l] = x0;
+    }
+
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        s[i] = b->s[i];
+    }
+}
+
+kernel void kernel_restore_block_q5_K_noshuffle(
+    global uchar * src_q,
+    global uchar * src_qh,
+    global uchar * src_s,
+    global half  * src_d,
+    global half  * src_dm,
+    global struct block_q5_K * dst,
+    uchar mask_0F,
+    uchar mask_F0
+) {
+    global struct block_q5_K * b  = (global struct block_q5_K *) dst + get_global_id(0);
+    global uchar * q  = (global uchar *) src_q  + QK_K/2       * get_global_id(0);
+    global uchar * qh = (global uchar *) src_qh + QK_K/8       * get_global_id(0);
+    global uchar * s  = (global uchar *) src_s  + K_SCALE_SIZE * get_global_id(0);
+    global half  * d  = (global half  *) src_d  + get_global_id(0);
+    global half  * dm = (global half  *) src_dm + get_global_id(0);
+
+    b->d  = *d;
+    b->dm = *dm;
+
+    for (int i = 0; i < QK_K / 64; ++i) {
+        for (int j = 0; j < 16; ++j) {
+            uchar lo = q[i*32 + j];
+            uchar hi = q[i*32 + j + 16];
+            b->qs[i*32 + 2*j]     = convert_uchar((lo & mask_0F) | ((hi & mask_0F) << 4));
+            b->qs[i*32 + 2*j + 1] = convert_uchar(((lo & mask_F0) >> 4) | (hi & mask_F0));
+        }
+    }
+
+    for (int g = 0; g < 4; ++g) {
+        for (int i = 0; i < 8; ++i) {
+            uchar x0 = 0;
+            for (int k = 0; k < 8; ++k) {
+                x0 |= ((qh[4*k+g] >> i) & 0x01) << k;
+            }
+            b->qh[g*8+i] = x0;
+        }
+    }
+
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        b->s[i] = s[i];
+    }
+}
+
 //------------------------------------------------------------------------------
 // kernel_convert_block_q6_K
 // Convert the block_q6_K format to 3 separate arrays (AOS -> SOA).

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q5_k_f32.cl

@@ -0,0 +1,176 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+#define QK_K         256
+#define K_SCALE_SIZE 12
+
+inline void get_scale_min_k4(
+    int j,
+    global const uchar * q,
+    uchar * d,
+    uchar * m,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    if (j < 4) {
+        *d = q[j]   & mask_d6;
+        *m = q[j+4] & mask_d6;
+    } else {
+        *d = (q[j+4] & mask_d4) | ((q[j-4] & mask_hi2) >> 2);
+        *m = ((q[j+4] >> 4) & mask_d4) | ((q[j]   & mask_hi2) >> 2);
+    }
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+kernel void kernel_gemm_noshuffle_q5_k_f32(
+    global const ushort * src0_q,
+    global const uchar  * src0_qh,
+    global const uchar  * src0_s,
+    global const half   * src0_d,
+    global const half   * src0_dm,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    int n_4 = n >> 2;
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 2;
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    half8 B;
+    half4 dequantized_weights;
+
+    int num_blocks_K = k / QK_K;
+
+    global const ushort * weight_ptr = src0_q  + gx_2;
+    global const uchar  * qh_ptr     = src0_qh + gx_2;
+    global const half   * d_ptr      = src0_d  + gx_2;
+    global const half   * dm_ptr     = src0_dm + gx_2;
+
+    for (int i = 0; i < k; i += 32) {
+        int sb_idx  = i / QK_K;
+        int sub_idx = (i / 32) % 8;
+
+        half4 d  = vload4(0, d_ptr  + sb_idx * m);
+        half4 dm = vload4(0, dm_ptr + sb_idx * m);
+
+        global const uchar * sc0 = src0_s + (gx_2+0) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc1 = src0_s + (gx_2+1) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc2 = src0_s + (gx_2+2) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+        global const uchar * sc3 = src0_s + (gx_2+3) * num_blocks_K * K_SCALE_SIZE + sb_idx * K_SCALE_SIZE;
+
+        uchar sv0, mn0, sv1, mn1, sv2, mn2, sv3, mn3;
+        get_scale_min_k4(sub_idx, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc2, &sv2, &mn2, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc3, &sv3, &mn3, mask_d6, mask_d4, mask_hi2);
+
+        half4 scale = convert_half4(convert_float4(d)  * convert_float4((uchar4)(sv0, sv1, sv2, sv3)));
+        half4 mval  = convert_half4(convert_float4(dm) * convert_float4((uchar4)(mn0, mn1, mn2, mn3)));
+
+        for (int l = 0; l < 32; l += 4) {
+            int ki = i + l;
+            ushort4 bits4   = vload4(0, weight_ptr + (ki/4) * m);
+            uchar4  qh_bits = vload4(0, qh_ptr     + (ki/8) * m);
+            int     qh_shift = ki % 8;
+
+            // j=0
+            B.s0123 = read_imageh(src1, gy*2   + (ki+0) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+0) * n_4);
+            dequantized_weights.s0 = ((bits4.s0 & 0x000F) | (((qh_bits.s0 >> (qh_shift+0)) & 1) << 4)) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits4.s1 & 0x000F) | (((qh_bits.s1 >> (qh_shift+0)) & 1) << 4)) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = ((bits4.s2 & 0x000F) | (((qh_bits.s2 >> (qh_shift+0)) & 1) << 4)) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = ((bits4.s3 & 0x000F) | (((qh_bits.s3 >> (qh_shift+0)) & 1) << 4)) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=1
+            B.s0123 = read_imageh(src1, gy*2   + (ki+1) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+1) * n_4);
+            dequantized_weights.s0 = (((bits4.s0 & 0x00F0) >> 4) | (((qh_bits.s0 >> (qh_shift+1)) & 1) << 4)) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = (((bits4.s1 & 0x00F0) >> 4) | (((qh_bits.s1 >> (qh_shift+1)) & 1) << 4)) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = (((bits4.s2 & 0x00F0) >> 4) | (((qh_bits.s2 >> (qh_shift+1)) & 1) << 4)) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = (((bits4.s3 & 0x00F0) >> 4) | (((qh_bits.s3 >> (qh_shift+1)) & 1) << 4)) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=2
+            B.s0123 = read_imageh(src1, gy*2   + (ki+2) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+2) * n_4);
+            dequantized_weights.s0 = (((bits4.s0 & 0x0F00) >> 8) | (((qh_bits.s0 >> (qh_shift+2)) & 1) << 4)) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = (((bits4.s1 & 0x0F00) >> 8) | (((qh_bits.s1 >> (qh_shift+2)) & 1) << 4)) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = (((bits4.s2 & 0x0F00) >> 8) | (((qh_bits.s2 >> (qh_shift+2)) & 1) << 4)) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = (((bits4.s3 & 0x0F00) >> 8) | (((qh_bits.s3 >> (qh_shift+2)) & 1) << 4)) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+
+            // j=3
+            B.s0123 = read_imageh(src1, gy*2   + (ki+3) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+3) * n_4);
+            dequantized_weights.s0 = (((bits4.s0 & 0xF000) >> 12) | (((qh_bits.s0 >> (qh_shift+3)) & 1) << 4)) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = (((bits4.s1 & 0xF000) >> 12) | (((qh_bits.s1 >> (qh_shift+3)) & 1) << 4)) * scale.s1 - mval.s1;
+            dequantized_weights.s2 = (((bits4.s2 & 0xF000) >> 12) | (((qh_bits.s2 >> (qh_shift+3)) & 1) << 4)) * scale.s2 - mval.s2;
+            dequantized_weights.s3 = (((bits4.s3 & 0xF000) >> 12) | (((qh_bits.s3 >> (qh_shift+3)) & 1) << 4)) * scale.s3 - mval.s3;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+            c2 += B * dequantized_weights.s2;
+            c3 += B * dequantized_weights.s3;
+        }
+    }
+
+    int idx = (gy<<3)*m + (gx<<2);
+
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx+3 < m*n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q5_k_f32.cl

@@ -0,0 +1,326 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK_K  256
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+inline void get_scale_min_k4(
+    int j,
+    global const uchar * q,
+    uchar * d,
+    uchar * m,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    if (j < 4) {
+        *d = q[j]   & mask_d6;
+        *m = q[j+4] & mask_d6;
+    } else {
+        *d = (q[j+4] & mask_d4) | ((q[j-4] & mask_hi2) >> 2);
+        *m = ((q[j+4] >> 4) & mask_d4) | ((q[j]   & mask_hi2) >> 2);
+    }
+}
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, bits1, scale, minv, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | ((bits1.s0 & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | ((bits1.s1 & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s0 >> 1) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s1 >> 1) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s0 >> 2) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s1 >> 2) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s0 >> 4) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s1 >> 4) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s0 >> 5) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s1 >> 5) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s0 >> 6) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s1 >> 6) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | ((bits1.s2 & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | ((bits1.s3 & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s2 >> 1) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s3 >> 1) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s2 >> 2) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s3 >> 2) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s2 >> 4) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s3 >> 4) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s2 >> 5) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s3 >> 5) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s2 >> 6) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s3 >> 6) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, bits1, scale, minv, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F) | ((bits1.s4 & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x000F) | ((bits1.s5 & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4) | (((bits1.s4 >> 1) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4) | (((bits1.s5 >> 1) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8) | (((bits1.s4 >> 2) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8) | (((bits1.s5 >> 2) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x000F) | (((bits1.s4 >> 4) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x000F) | (((bits1.s5 >> 4) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4) | (((bits1.s4 >> 5) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4) | (((bits1.s5 >> 5) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8) | (((bits1.s4 >> 6) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8) | (((bits1.s5 >> 6) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F) | ((bits1.s6 & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x000F) | ((bits1.s7 & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4) | (((bits1.s6 >> 1) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4) | (((bits1.s7 >> 1) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8) | (((bits1.s6 >> 2) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8) | (((bits1.s7 >> 2) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x000F) | (((bits1.s6 >> 4) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x000F) | (((bits1.s7 >> 4) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4) | (((bits1.s6 >> 5) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4) | (((bits1.s7 >> 5) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8) | (((bits1.s6 >> 6) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8) | (((bits1.s7 >> 6) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4)) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4)) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, bits1, scale, minv, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | ((bits1.s0 & 0x01) << 4))         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s0 >> 1) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s0 >> 2) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s0 >> 3) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s0 >> 4) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s0 >> 5) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s0 >> 6) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s0 >> 7) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | ((bits1.s1 & 0x01) << 4))         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s1 >> 1) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s1 >> 2) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s1 >> 3) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s1 >> 4) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s1 >> 5) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s1 >> 6) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s1 >> 7) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | ((bits1.s2 & 0x01) << 4))         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s2 >> 1) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s2 >> 2) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s2 >> 3) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s2 >> 4) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s2 >> 5) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s2 >> 6) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s2 >> 7) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | ((bits1.s3 & 0x01) << 4))         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s3 >> 1) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s3 >> 2) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s3 >> 3) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s3 >> 4) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s3 >> 5) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s3 >> 6) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s3 >> 7) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, bits1, scale, minv, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x000F)         | ((bits1.s4 & 0x01) << 4))         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s0 & 0x00F0) >> 4)  | (((bits1.s4 >> 1) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s0 & 0x0F00) >> 8)  | (((bits1.s4 >> 2) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s0 & 0xF000) >> 12) | (((bits1.s4 >> 3) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s2 & 0x000F)         | (((bits1.s4 >> 4) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s2 & 0x00F0) >> 4)  | (((bits1.s4 >> 5) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s2 & 0x0F00) >> 8)  | (((bits1.s4 >> 6) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s2 & 0xF000) >> 12) | (((bits1.s4 >> 7) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s1 & 0x000F)         | ((bits1.s5 & 0x01) << 4))         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s1 & 0x00F0) >> 4)  | (((bits1.s5 >> 1) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s1 & 0x0F00) >> 8)  | (((bits1.s5 >> 2) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s1 & 0xF000) >> 12) | (((bits1.s5 >> 3) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s3 & 0x000F)         | (((bits1.s5 >> 4) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s3 & 0x00F0) >> 4)  | (((bits1.s5 >> 5) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s3 & 0x0F00) >> 8)  | (((bits1.s5 >> 6) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s3 & 0xF000) >> 12) | (((bits1.s5 >> 7) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x000F)         | ((bits1.s6 & 0x01) << 4))         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += ((((bits4.s4 & 0x00F0) >> 4)  | (((bits1.s6 >> 1) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += ((((bits4.s4 & 0x0F00) >> 8)  | (((bits1.s6 >> 2) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += ((((bits4.s4 & 0xF000) >> 12) | (((bits1.s6 >> 3) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += (((bits4.s6 & 0x000F)         | (((bits1.s6 >> 4) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += ((((bits4.s6 & 0x00F0) >> 4)  | (((bits1.s6 >> 5) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += ((((bits4.s6 & 0x0F00) >> 8)  | (((bits1.s6 >> 6) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += ((((bits4.s6 & 0xF000) >> 12) | (((bits1.s6 >> 7) & 0x01) << 4))  * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += (((bits4.s5 & 0x000F)         | ((bits1.s7 & 0x01) << 4))         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += ((((bits4.s5 & 0x00F0) >> 4)  | (((bits1.s7 >> 1) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += ((((bits4.s5 & 0x0F00) >> 8)  | (((bits1.s7 >> 2) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += ((((bits4.s5 & 0xF000) >> 12) | (((bits1.s7 >> 3) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += (((bits4.s7 & 0x000F)         | (((bits1.s7 >> 4) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += ((((bits4.s7 & 0x00F0) >> 4)  | (((bits1.s7 >> 5) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += ((((bits4.s7 & 0x0F00) >> 8)  | (((bits1.s7 >> 6) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += ((((bits4.s7 & 0xF000) >> 12) | (((bits1.s7 >> 7) & 0x01) << 4))  * scale.s1 - minv.s1) * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q5_k_f32(
+        read_only  image1d_buffer_t src0_q,
+        read_only  image1d_buffer_t src0_qh,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A     = M / 2;
+    uint BLOCK_STRIDE_A    = NSUBGROUPS * M;
+
+    uint LINE_STRIDE_A_QH  = M / 2;
+    uint BLOCK_STRIDE_A_QH = NSUBGROUPS * M / 2;
+    uint scales_per_row    = (K / QK_K) * 12;
+
+    private uint4     regA;
+    private ushort4   regH;
+    private half2     regS;
+    private half2     regM;
+    private float8    regB;
+
+    private float2 totalSum = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / 32); k += NSUBGROUPS) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+
+        half2 d   = src0_d[gid + sb * LINE_STRIDE_A];
+        half2 dm  = src0_m[gid + sb * LINE_STRIDE_A];
+
+        global const uchar * sc0 = src0_s + 2 * gid * scales_per_row + sb * 12;
+        global const uchar * sc1 = src0_s + (2 * gid + 1) * scales_per_row + sb * 12;
+
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(j, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        regH.s0 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 0)).x);
+        regH.s1 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 1)).x);
+        regH.s2 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 2)).x);
+        regH.s3 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 3)).x);
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), as_uchar8(regH), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), as_uchar8(regH), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), as_uchar8(regH), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), as_uchar8(regH), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // reduction in local memory, assumes #wave=4
+    local float2 reduceLM[SUBGROUP_SIZE * 3];
+    if (groupId == 1) {
+        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
+    }
+    if (groupId == 2) {
+        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
+    }
+    if (groupId == 3) {
+        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
+    }
+    if (groupId == 0) {
+        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        vstore2(totalSum, 0, &(dst[gid * 2]));
+    }
+}

ggml/src/ggml-rpc/CMakeLists.txt

@@ -2,8 +2,32 @@ message(STATUS "Using RPC backend")
 
 ggml_add_backend_library(ggml-rpc
                          ggml-rpc.cpp
+                         transport.cpp
                         )
 
 if (WIN32)
     target_link_libraries(ggml-rpc PRIVATE ws2_32)
 endif()
+
+# RDMA auto-detection (Linux only, requires libibverbs)
+if (NOT WIN32 AND NOT APPLE)
+    find_library(IBVERBS_LIB ibverbs)
+    if (IBVERBS_LIB)
+        option(GGML_RPC_RDMA "ggml: enable RDMA transport for RPC" ON)
+    else()
+        option(GGML_RPC_RDMA "ggml: enable RDMA transport for RPC" OFF)
+    endif()
+else()
+    set(GGML_RPC_RDMA OFF CACHE BOOL "RDMA not available on this platform" FORCE)
+endif()
+
+if (GGML_RPC_RDMA)
+    if (NOT IBVERBS_LIB)
+        find_library(IBVERBS_LIB ibverbs REQUIRED)
+    endif()
+    target_compile_definitions(ggml-rpc PRIVATE GGML_RPC_RDMA)
+    target_link_libraries(ggml-rpc PRIVATE ${IBVERBS_LIB})
+    message(STATUS "  RDMA transport enabled (auto-detected)")
+else()
+    message(STATUS "  RDMA transport disabled")
+endif()

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

@@ -2,30 +2,17 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpp.h"
+#include "transport.h"
 
+#include <array>
 #include <cinttypes>
+#include <optional>
 #include <string>
 #include <vector>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
 #include <unordered_set>
-#ifdef _WIN32
-#  define WIN32_LEAN_AND_MEAN
-#  ifndef NOMINMAX
-#     define NOMINMAX
-#  endif
-#  include <windows.h>
-#  include <winsock2.h>
-#else
-#  include <arpa/inet.h>
-#  include <sys/socket.h>
-#  include <sys/types.h>
-#  include <netinet/in.h>
-#  include <netinet/tcp.h>
-#  include <netdb.h>
-#  include <unistd.h>
-#endif
 #include <cstring>
 #include <fstream>
 #include <filesystem>
@@ -39,29 +26,6 @@ static const char * RPC_DEBUG = std::getenv("GGML_RPC_DEBUG");
 
 namespace fs = std::filesystem;
 
-static constexpr size_t MAX_CHUNK_SIZE = 1024ull * 1024ull * 1024ull; // 1 GiB
-
-#ifdef _WIN32
-typedef SOCKET sockfd_t;
-using ssize_t = __int64;
-#else
-typedef int sockfd_t;
-#endif
-
-// cross-platform socket
-struct socket_t {
-    sockfd_t fd;
-    socket_t(sockfd_t fd) : fd(fd) {}
-    ~socket_t() {
-        LOG_DBG("[%s] closing socket %d\n", __func__, this->fd);
-#ifdef _WIN32
-        closesocket(this->fd);
-#else
-        close(this->fd);
-#endif
-    }
-};
-
 // macro for nicer error messages on server crash
 #define RPC_STATUS_ASSERT(x) if (!(x)) GGML_ABORT("Remote RPC server crashed or returned malformed response")
 
@@ -115,10 +79,16 @@ static_assert(RPC_CMD_HELLO == 14, "RPC_CMD_HELLO must be always 14");
 // Try RPC_CMD_SET_TENSOR_HASH first when data size is larger than this threshold
 const size_t HASH_THRESHOLD = 10 * 1024 * 1024;
 
+struct rpc_msg_hello_req {
+    uint8_t conn_caps[RPC_CONN_CAPS_SIZE];
+};
+
 struct rpc_msg_hello_rsp {
     uint8_t major;
     uint8_t minor;
     uint8_t patch;
+    uint8_t padding;
+    uint8_t conn_caps[RPC_CONN_CAPS_SIZE];
 };
 
 struct rpc_msg_device_count_rsp {
@@ -288,153 +258,27 @@ static uint64_t fnv_hash(const uint8_t * data, size_t len) {
     return hash;
 }
 
-static std::shared_ptr<socket_t> make_socket(sockfd_t fd) {
-#ifdef _WIN32
-    if (fd == INVALID_SOCKET) {
-        return nullptr;
-    }
-#else
-    if (fd < 0) {
-        return nullptr;
-    }
-#endif
-    return std::make_shared<socket_t>(fd);
-}
-
-static bool set_no_delay(sockfd_t sockfd) {
-    int flag = 1;
-    // set TCP_NODELAY to disable Nagle's algorithm
-    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
-    return ret == 0;
-}
-
-static bool set_reuse_addr(sockfd_t sockfd) {
-    int flag = 1;
-    int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
-    return ret == 0;
-}
-
-static std::shared_ptr<socket_t> socket_connect(const char * host, int port) {
-    struct sockaddr_in addr;
-    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
-    auto sock_ptr = make_socket(sockfd);
-    if (sock_ptr == nullptr) {
-        return nullptr;
-    }
-    if (!set_no_delay(sockfd)) {
-        GGML_LOG_ERROR("Failed to set TCP_NODELAY\n");
-        return nullptr;
-    }
-    addr.sin_family = AF_INET;
-    addr.sin_port = htons(port);
-    struct hostent * server = gethostbyname(host);
-    if (server == NULL) {
-        GGML_LOG_ERROR("Cannot resolve host '%s'\n", host);
-        return nullptr;
-    }
-    memcpy(&addr.sin_addr.s_addr, server->h_addr, server->h_length);
-    if (connect(sock_ptr->fd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
-        return nullptr;
-    }
-    return sock_ptr;
-}
-
-static std::shared_ptr<socket_t> socket_accept(sockfd_t srv_sockfd) {
-    auto client_socket_fd = accept(srv_sockfd, NULL, NULL);
-    auto client_socket = make_socket(client_socket_fd);
-    if (client_socket == nullptr) {
-        return nullptr;
-    }
-    if (!set_no_delay(client_socket_fd)) {
-        GGML_LOG_ERROR("Failed to set TCP_NODELAY\n");
-        return nullptr;
-    }
-    return client_socket;
-}
-
-static std::shared_ptr<socket_t> create_server_socket(const char * host, int port) {
-    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
-    auto sock = make_socket(sockfd);
-    if (sock == nullptr) {
-        return nullptr;
-    }
-    if (!set_reuse_addr(sockfd)) {
-        GGML_LOG_ERROR("Failed to set SO_REUSEADDR\n");
-        return nullptr;
-    }
-    if (inet_addr(host) == INADDR_NONE) {
-        GGML_LOG_ERROR("Invalid host address: %s\n", host);
-        return nullptr;
-    }
-    struct sockaddr_in serv_addr;
-    serv_addr.sin_family = AF_INET;
-    serv_addr.sin_addr.s_addr = inet_addr(host);
-    serv_addr.sin_port = htons(port);
-
-    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
-        return nullptr;
-    }
-    if (listen(sockfd, 1) < 0) {
-        return nullptr;
-    }
-    return sock;
-}
-
-static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
-    size_t bytes_sent = 0;
-    while (bytes_sent < size) {
-        size_t size_to_send = std::min(size - bytes_sent, MAX_CHUNK_SIZE);
-        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size_to_send, 0);
-        if (n < 0) {
-            GGML_LOG_ERROR("send failed (bytes_sent=%zu, size_to_send=%zu)\n",
-                           bytes_sent, size_to_send);
-            return false;
-        }
-        bytes_sent += (size_t)n;
-    }
-    return true;
-}
-
-static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
-    size_t bytes_recv = 0;
-    while (bytes_recv < size) {
-        size_t size_to_recv = std::min(size - bytes_recv, MAX_CHUNK_SIZE);
-        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size_to_recv, 0);
-        if (n < 0) {
-            GGML_LOG_ERROR("recv failed (bytes_recv=%zu, size_to_recv=%zu)\n",
-                           bytes_recv, size_to_recv);
-            return false;
-        }
-        if (n == 0) {
-            LOG_DBG("recv returned 0 (peer closed?)\n");
-            return false;
-        }
-        bytes_recv += (size_t)n;
-    }
-    return true;
-}
-
-static bool send_msg(sockfd_t sockfd, const void * msg, size_t msg_size) {
-    if (!send_data(sockfd, &msg_size, sizeof(msg_size))) {
+static bool send_msg(socket_ptr sock, const void * msg, size_t msg_size) {
+    if (!sock->send_data(&msg_size, sizeof(msg_size))) {
         return false;
     }
-    return send_data(sockfd, msg, msg_size);
+    return sock->send_data(msg, msg_size);
 }
 
-static bool recv_msg(sockfd_t sockfd, void * msg, size_t msg_size) {
+static bool recv_msg(socket_ptr sock, void * msg, size_t msg_size) {
     uint64_t size;
-    if (!recv_data(sockfd, &size, sizeof(size))) {
+    if (!sock->recv_data(&size, sizeof(size))) {
         return false;
     }
     if (size != msg_size) {
         return false;
     }
-    return recv_data(sockfd, msg, msg_size);
+    return sock->recv_data(msg, msg_size);
 }
 
-static bool recv_msg(sockfd_t sockfd, std::vector<uint8_t> & input) {
+static bool recv_msg(socket_ptr sock, std::vector<uint8_t> & input) {
     uint64_t size;
-    if (!recv_data(sockfd, &size, sizeof(size))) {
+    if (!sock->recv_data(&size, sizeof(size))) {
         return false;
     }
     try {
@@ -443,7 +287,7 @@ static bool recv_msg(sockfd_t sockfd, std::vector<uint8_t> & input) {
         GGML_LOG_ERROR("Failed to allocate input buffer of size %" PRIu64 "\n", size);
         return false;
     }
-    return recv_data(sockfd, input.data(), size);
+    return sock->recv_data(input.data(), size);
 }
 
 static bool parse_endpoint(const std::string & endpoint, std::string & host, int & port) {
@@ -452,21 +296,25 @@ static bool parse_endpoint(const std::string & endpoint, std::string & host, int
         return false;
     }
     host = endpoint.substr(0, pos);
-    port = std::stoi(endpoint.substr(pos + 1));
+    try {
+        port = std::stoi(endpoint.substr(pos + 1));
+    } catch (...) {
+        return false;
+    }
     return true;
 }
 
 // RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
 // No response
-static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size) {
+static bool send_rpc_cmd(socket_ptr sock, enum rpc_cmd cmd, const void * input, size_t input_size) {
     uint8_t cmd_byte = cmd;
-    if (!send_data(sock->fd, &cmd_byte, sizeof(cmd_byte))) {
+    if (!sock->send_data(&cmd_byte, sizeof(cmd_byte))) {
         return false;
     }
-    if (!send_data(sock->fd, &input_size, sizeof(input_size))) {
+    if (!sock->send_data(&input_size, sizeof(input_size))) {
         return false;
     }
-    if (!send_data(sock->fd, input, input_size)) {
+    if (!sock->send_data(input, input_size)) {
         return false;
     }
     return true;
@@ -474,20 +322,18 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 
 // RPC request : | rpc_cmd (1 byte) | request_size (8 bytes) | request_data (request_size bytes) |
 // RPC response: | response_size (8 bytes) | response_data (response_size bytes) |
-static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
+static bool send_rpc_cmd(socket_ptr sock, enum rpc_cmd cmd, const void * input, size_t input_size, void * output, size_t output_size) {
     if (!send_rpc_cmd(sock, cmd, input, input_size)) {
         return false;
     }
-    // TODO: currently the output_size is always known, do we need support for commands with variable output size?
-    // even if we do, we can skip sending output_size from the server for commands with known output size
     uint64_t out_size;
-    if (!recv_data(sock->fd, &out_size, sizeof(out_size))) {
+    if (!sock->recv_data(&out_size, sizeof(out_size))) {
         return false;
     }
     if (out_size != output_size) {
         return false;
     }
-    if (!recv_data(sock->fd, output, output_size)) {
+    if (!sock->recv_data(output, output_size)) {
         return false;
     }
     return true;
@@ -495,17 +341,25 @@ static bool send_rpc_cmd(const std::shared_ptr<socket_t> & sock, enum rpc_cmd cm
 
 // RPC client-side implementation
 
-static bool check_server_version(const std::shared_ptr<socket_t> & sock) {
-    rpc_msg_hello_rsp response;
-    bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, nullptr, 0, &response, sizeof(response));
+// Performs HELLO handshake with transport auto-negotiation.
+// Advertises local capabilities via conn_caps; if the server responds with
+// matching capabilities, the socket is upgraded transparently.
+static bool negotiate_hello(const std::shared_ptr<socket_t> & sock) {
+    rpc_msg_hello_req request = {};
+    rpc_msg_hello_rsp response = {};
+
+    sock->get_caps(request.conn_caps);
+
+    bool status = send_rpc_cmd(sock, RPC_CMD_HELLO, &request, sizeof(request), &response, sizeof(response));
     RPC_STATUS_ASSERT(status);
+
     if (response.major != RPC_PROTO_MAJOR_VERSION || response.minor > RPC_PROTO_MINOR_VERSION) {
-        GGML_LOG_ERROR("RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
+        GGML_LOG_ERROR("RPC server version mismatch: %d.%d.%d\n",
+                       response.major, response.minor, response.patch);
         return false;
     }
-    if (response.minor != RPC_PROTO_MINOR_VERSION || response.patch != RPC_PROTO_PATCH_VERSION) {
-        GGML_LOG_INFO("WARNING: RPC server version mismatch: %d.%d.%d\n", response.major, response.minor, response.patch);
-    }
+
+    sock->update_caps(response.conn_caps);
     return true;
 }
 
@@ -513,7 +367,6 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
     static std::mutex mutex;
     std::lock_guard<std::mutex> lock(mutex);
     static std::unordered_map<std::string, std::weak_ptr<socket_t>> sockets;
-    static bool initialized = false;
 
     auto it = sockets.find(endpoint);
     if (it != sockets.end()) {
@@ -527,26 +380,18 @@ static std::shared_ptr<socket_t> get_socket(const std::string & endpoint) {
         GGML_LOG_ERROR("Failed to parse endpoint: %s\n", endpoint.c_str());
         return nullptr;
     }
-#ifdef _WIN32
-    if (!initialized) {
-        WSADATA wsaData;
-        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
-        if (res != 0) {
-            return nullptr;
-        }
-        initialized = true;
+
+    if (!rpc_transport_init()) {
+        return nullptr;
     }
-#else
-    GGML_UNUSED(initialized);
-#endif
-    auto sock = socket_connect(host.c_str(), port);
+    auto sock = socket_t::connect(host.c_str(), port);
     if (sock == nullptr) {
         return nullptr;
     }
-    if (!check_server_version(sock)) {
+    if (!negotiate_hello(sock)) {
         return nullptr;
     }
-    LOG_DBG("[%s] connected to %s, sockfd=%d\n", __func__, endpoint.c_str(), sock->fd);
+    LOG_DBG("[%s] connected to %s\n", __func__, endpoint.c_str());
     sockets[endpoint] = sock;
     return sock;
 }
@@ -1597,27 +1442,46 @@ rpc_server::~rpc_server() {
 }
 
 static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const char * cache_dir,
-                             sockfd_t sockfd) {
+                             socket_ptr sock) {
     rpc_server server(backends, cache_dir);
     uint8_t cmd;
-    if (!recv_data(sockfd, &cmd, 1)) {
+    if (!sock->recv_data(&cmd, 1)) {
         return;
     }
-    // the first command sent by the client must be HELLO
     if (cmd != RPC_CMD_HELLO) {
         GGML_LOG_ERROR("Expected HELLO command, update client\n");
         return;
     }
-    if (!recv_msg(sockfd, nullptr, 0)) {
+
+    // Read input_size and validate protocol version
+    uint64_t hello_input_size;
+    if (!sock->recv_data(&hello_input_size, sizeof(hello_input_size))) {
         return;
     }
-    rpc_msg_hello_rsp response;
-    server.hello(response);
-    if (!send_msg(sockfd, &response, sizeof(response))) {
+
+    if (hello_input_size != sizeof(rpc_msg_hello_req)) {
+        GGML_LOG_ERROR("HELLO request size mismatch (%zu vs %zu) — client needs upgrade to protocol v%d.x\n",
+                       (size_t)hello_input_size, sizeof(rpc_msg_hello_req), RPC_PROTO_MAJOR_VERSION);
         return;
     }
+
+    rpc_msg_hello_req req = {};
+    if (!sock->recv_data(&req, sizeof(req))) {
+        return;
+    }
+
+    rpc_msg_hello_rsp rsp = {};
+    server.hello(rsp);
+    // Advertise server transport capabilities based on client's caps
+    sock->get_caps(rsp.conn_caps);
+    if (!send_msg(sock, &rsp, sizeof(rsp))) {
+        return;
+    }
+
+    // Activate transport upgrade using client's caps
+    sock->update_caps(req.conn_caps);
     while (true) {
-        if (!recv_data(sockfd, &cmd, 1)) {
+        if (!sock->recv_data(&cmd, 1)) {
             break;
         }
         if (cmd >= RPC_CMD_COUNT) {
@@ -1631,115 +1495,115 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
                 return;
             }
             case RPC_CMD_DEVICE_COUNT: {
-                if (!recv_msg(sockfd, nullptr, 0)) {
+                if (!recv_msg(sock, nullptr, 0)) {
                     return;
                 }
                 rpc_msg_device_count_rsp response;
                 response.device_count = backends.size();
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_ALLOC_BUFFER: {
                 rpc_msg_alloc_buffer_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_alloc_buffer_rsp response;
                 if (!server.alloc_buffer(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_GET_ALLOC_SIZE: {
                 rpc_msg_get_alloc_size_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_get_alloc_size_rsp response;
                 if (!server.get_alloc_size(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_GET_ALIGNMENT: {
                 rpc_msg_get_alignment_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_get_alignment_rsp response;
                 if (!server.get_alignment(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_GET_MAX_SIZE: {
                 rpc_msg_get_max_size_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_get_max_size_rsp response;
                 if (!server.get_max_size(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_BUFFER_GET_BASE: {
                 rpc_msg_buffer_get_base_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_buffer_get_base_rsp response;
                 if (!server.buffer_get_base(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_FREE_BUFFER: {
                 rpc_msg_free_buffer_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 if (!server.free_buffer(request)) {
                     return;
                 }
-                if (!send_msg(sockfd, nullptr, 0)) {
+                if (!send_msg(sock, nullptr, 0)) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_BUFFER_CLEAR: {
                 rpc_msg_buffer_clear_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 if (!server.buffer_clear(request)) {
                     return;
                 }
-                if (!send_msg(sockfd, nullptr, 0)) {
+                if (!send_msg(sock, nullptr, 0)) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_SET_TENSOR: {
                 std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
+                if (!recv_msg(sock, input)) {
                     return;
                 }
                 if (!server.set_tensor(input)) {
@@ -1749,62 +1613,62 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
             }
             case RPC_CMD_SET_TENSOR_HASH: {
                 rpc_msg_set_tensor_hash_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_set_tensor_hash_rsp response;
                 if (!server.set_tensor_hash(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_INIT_TENSOR: {
                 rpc_msg_init_tensor_req request;
-                if (!recv_msg(sockfd, &request,sizeof(request))) {
+                if (!recv_msg(sock, &request,sizeof(request))) {
                     return;
                 }
                 if (!server.init_tensor(request)) {
                     return;
                 }
-                if (!send_msg(sockfd, nullptr, 0)) {
+                if (!send_msg(sock, nullptr, 0)) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_GET_TENSOR: {
                 rpc_msg_get_tensor_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 std::vector<uint8_t> response;
                 if (!server.get_tensor(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, response.data(), response.size())) {
+                if (!send_msg(sock, response.data(), response.size())) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_COPY_TENSOR: {
                 rpc_msg_copy_tensor_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_copy_tensor_rsp response;
                 if (!server.copy_tensor(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
             }
             case RPC_CMD_GRAPH_COMPUTE: {
                 std::vector<uint8_t> input;
-                if (!recv_msg(sockfd, input)) {
+                if (!recv_msg(sock, input)) {
                     return;
                 }
                 if (!server.graph_compute(input)) {
@@ -1814,7 +1678,7 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
             }
             case RPC_CMD_GRAPH_RECOMPUTE: {
                 rpc_msg_graph_recompute_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 if (!server.graph_recompute(request)) {
@@ -1824,14 +1688,14 @@ static void rpc_serve_client(const std::vector<ggml_backend_t> & backends, const
             }
             case RPC_CMD_GET_DEVICE_MEMORY: {
                 rpc_msg_get_device_memory_req request;
-                if (!recv_msg(sockfd, &request, sizeof(request))) {
+                if (!recv_msg(sock, &request, sizeof(request))) {
                     return;
                 }
                 rpc_msg_get_device_memory_rsp response;
                 if (!server.get_device_memory(request, response)) {
                     return;
                 }
-                if (!send_msg(sockfd, &response, sizeof(response))) {
+                if (!send_msg(sock, &response, sizeof(response))) {
                     return;
                 }
                 break;
@@ -1884,36 +1748,34 @@ void ggml_backend_rpc_start_server(const char * endpoint, const char * cache_dir
     if (!parse_endpoint(endpoint, host, port)) {
         return;
     }
-#ifdef _WIN32
-    {
-        WSADATA wsaData;
-        int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
-        if (res != 0) {
-            fprintf(stderr, "WSAStartup failed: %d\n", res);
-            return;
-        }
+
+#ifdef GGML_RPC_RDMA
+    printf("  transport      : TCP (RDMA auto-negotiate enabled)\n");
+#else
+    printf("  transport      : TCP\n");
+#endif // GGML_RPC_RDMA
+    if (!rpc_transport_init()) {
+        fprintf(stderr, "Failed to initialize RPC transport\n");
+        return;
     }
-#endif
-    auto server_socket = create_server_socket(host.c_str(), port);
+    auto server_socket = socket_t::create_server(host.c_str(), port);
     if (server_socket == nullptr) {
         fprintf(stderr, "Failed to create server socket\n");
         return;
     }
     while (true) {
-        auto client_socket = socket_accept(server_socket->fd);
+        auto client_socket = server_socket->accept();
         if (client_socket == nullptr) {
             fprintf(stderr, "Failed to accept client connection\n");
             return;
         }
         printf("Accepted client connection\n");
         fflush(stdout);
-        rpc_serve_client(backends, cache_dir, client_socket->fd);
+        rpc_serve_client(backends, cache_dir, client_socket);
         printf("Client connection closed\n");
         fflush(stdout);
     }
-#ifdef _WIN32
-    WSACleanup();
-#endif
+    rpc_transport_shutdown();
     for (auto backend : backends) {
         ggml_backend_free(backend);
     }

ggml/src/ggml-rpc/transport.cpp

@@ -0,0 +1,683 @@
+#include "transport.h"
+#include "ggml-impl.h"
+
+#ifdef _WIN32
+#  define WIN32_LEAN_AND_MEAN
+#  ifndef NOMINMAX
+#     define NOMINMAX
+#  endif
+#  include <windows.h>
+#  include <winsock2.h>
+#else
+#  include <arpa/inet.h>
+#  include <sys/socket.h>
+#  include <sys/types.h>
+#  include <netinet/in.h>
+#  include <netinet/tcp.h>
+#  include <netdb.h>
+#  include <unistd.h>
+#endif
+#include <cstdlib>
+#include <mutex>
+#include <optional>
+
+#ifdef GGML_RPC_RDMA
+#  include <infiniband/verbs.h>
+#  include <time.h>
+#  ifndef _WIN32
+#    include <poll.h>
+#  endif
+#endif // GGML_RPC_RDMA
+
+#ifdef _WIN32
+typedef SOCKET sockfd_t;
+using ssize_t = __int64;
+#else
+typedef int sockfd_t;
+#endif
+
+static const char * RPC_DEBUG = std::getenv("GGML_RPC_DEBUG");
+
+#define LOG_DBG(...) \
+    do { if (RPC_DEBUG) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)
+
+#ifdef GGML_RPC_RDMA
+static constexpr size_t RDMA_CHUNK    = 256 * 1024;   // 256 KiB per send/recv (fits default 8 MiB memlock)
+static constexpr int    RDMA_RX_DEPTH = 24;            // pre-posted recv ring: 24 × 256 KiB = 6 MiB
+static constexpr size_t RDMA_GID_SIZE = 16;            // RoCE GID / IB GID is always 16 bytes
+using rdma_gid_t = std::array<uint8_t, RDMA_GID_SIZE>;
+
+struct rdma_conn {
+    struct ibv_context * ctx = nullptr;
+    struct ibv_pd * pd  = nullptr;
+    struct ibv_cq * scq = nullptr;   // send completions
+    struct ibv_cq * rcq = nullptr;   // recv completions
+    struct ibv_qp * qp  = nullptr;
+
+    void          * tx_buf = nullptr;
+    struct ibv_mr * tx_mr  = nullptr;
+
+    void          * rx_buf = nullptr; // RDMA_RX_DEPTH × RDMA_CHUNK contiguous
+    struct ibv_mr * rx_mr  = nullptr;
+    int             rx_head = 0;
+
+    uint32_t        max_inline = 0;
+
+    uint8_t * rx_slot(int i) const {
+        return static_cast<uint8_t *>(rx_buf) + static_cast<size_t>(i) * RDMA_CHUNK;
+    }
+
+    bool post_rx(int i) {
+        struct ibv_sge sge = {};
+        sge.addr   = (uintptr_t)rx_slot(i);
+        sge.length = RDMA_CHUNK;
+        sge.lkey   = rx_mr->lkey;
+        struct ibv_recv_wr wr = {}, * bad = nullptr;
+        wr.wr_id   = (uint64_t)i;
+        wr.sg_list = &sge;
+        wr.num_sge = 1;
+        return ibv_post_recv(qp, &wr, &bad) == 0;
+    }
+
+    ~rdma_conn() {
+        if (tx_mr) ibv_dereg_mr(tx_mr);
+        if (rx_mr) ibv_dereg_mr(rx_mr);
+        free(tx_buf);
+        free(rx_buf);
+        if (qp)  ibv_destroy_qp(qp);
+        if (scq) ibv_destroy_cq(scq);
+        if (rcq) ibv_destroy_cq(rcq);
+        if (pd)  ibv_dealloc_pd(pd);
+        if (ctx) ibv_close_device(ctx);
+    }
+};
+
+// Local RDMA parameters captured during the probe phase and later consumed
+// by rdma_activate() after the remote side's caps arrive via HELLO.
+struct rdma_local_info {
+    uint32_t qpn     = 0;
+    uint32_t psn     = 0;
+    uint8_t  gid[RDMA_GID_SIZE] = {};
+    uint8_t  ib_port = 0;
+    int      gid_idx = 0;
+    enum ibv_mtu path_mtu = IBV_MTU_1024;
+};
+
+struct rdma_caps {
+    uint32_t qpn;
+    uint32_t psn;
+    uint8_t  gid[RDMA_GID_SIZE];
+};
+
+static_assert(sizeof(rdma_caps) == RPC_CONN_CAPS_SIZE, "rdma_caps must match conn_caps size");
+
+#endif // GGML_RPC_RDMA
+
+struct socket_t::impl {
+    impl(sockfd_t fd) : use_rdma(false), fd(fd) {}
+    ~impl();
+    bool send_data(const void * data, size_t size);
+    bool recv_data(void * data, size_t size);
+    void get_caps(uint8_t * local_caps);
+    void update_caps(const uint8_t * remote_caps);
+
+#ifdef GGML_RPC_RDMA
+    bool tcp_peer_closed();
+    std::optional<rdma_gid_t> rdma_build_target_gid();
+    bool rdma_probe();
+    bool rdma_activate(uint32_t remote_qpn, uint32_t remote_psn, const uint8_t * remote_gid);
+    bool rdma_poll(struct ibv_cq * cq, struct ibv_wc * wc);
+    bool rdma_send(const void * data, size_t size);
+    bool rdma_recv(void * data, size_t size);
+
+    std::unique_ptr<rdma_conn> rdma;
+    rdma_local_info            rdma_local = {};
+#endif // GGML_RPC_RDMA
+    bool     use_rdma;
+    sockfd_t fd;
+};
+
+socket_t::impl::~impl() {
+#ifdef GGML_RPC_RDMA
+    rdma.reset();
+#endif // GGML_RPC_RDMA
+    LOG_DBG("[%s] closing socket %d\n", __func__, this->fd);
+#ifdef _WIN32
+    if (fd != INVALID_SOCKET) closesocket(this->fd);
+#else
+    if (fd >= 0) close(this->fd);
+#endif
+}
+
+#ifdef GGML_RPC_RDMA
+
+bool socket_t::impl::tcp_peer_closed() {
+    if (fd < 0) return false;
+#ifndef _WIN32
+    struct pollfd pfd = { fd, POLLIN | POLLRDHUP, 0 };
+    int r = poll(&pfd, 1, 0);
+    return r > 0 && (pfd.revents & (POLLHUP | POLLERR | POLLRDHUP));
+#else
+    return false;
+#endif
+}
+
+// Build a RoCE GID-shaped 16-byte target from a TCP socket's local address.
+// Used to match the socket's local IP against the kernel's GID table so that
+// a single memcmp handles IPv4, IPv4-mapped IPv6, and native IPv6 uniformly:
+//   AF_INET                -> ::ffff:a.b.c.d  (bytes 10-11 = 0xff, last 4 = IPv4)
+//   AF_INET6 (IPv4-mapped) -> ::ffff:a.b.c.d  (already in GID shape)
+//   AF_INET6 (native v6)   -> the 16-byte IPv6 address as-is
+// Returns std::nullopt on unsupported family or getsockname failure.
+std::optional<rdma_gid_t> socket_t::impl::rdma_build_target_gid() {
+    sockaddr_storage addr = {};
+    socklen_t addr_len = sizeof(addr);
+    if (getsockname(fd, reinterpret_cast<sockaddr *>(&addr), &addr_len) != 0) {
+        return std::nullopt;
+    }
+    rdma_gid_t target = {};
+    if (addr.ss_family == AF_INET) {
+        const auto * a = reinterpret_cast<const sockaddr_in *>(&addr);
+        target[10] = 0xff;
+        target[11] = 0xff;
+        memcpy(&target[12], &a->sin_addr, 4);
+        return target;
+    }
+    if (addr.ss_family == AF_INET6) {
+        const auto * a = reinterpret_cast<const sockaddr_in6 *>(&addr);
+        memcpy(target.data(), &a->sin6_addr, RDMA_GID_SIZE);
+        return target;
+    }
+    return std::nullopt;
+}
+
+bool socket_t::impl::rdma_probe() {
+    const char * dev_env = std::getenv("GGML_RDMA_DEV");
+    const char * gid_env = std::getenv("GGML_RDMA_GID");
+
+    auto target_gid = rdma_build_target_gid();
+    if (!target_gid) {
+        return false;
+    }
+
+    const uint8_t ib_port = 1;
+    int num_devs = 0;
+    ibv_device ** devs = ibv_get_device_list(&num_devs);
+    if (!devs || num_devs == 0) return false;
+
+    ibv_context * ibctx = nullptr;
+    const char * matched_dev = nullptr;
+    int gid_idx = gid_env ? atoi(gid_env) : -1;
+    int gid_version = IBV_GID_TYPE_IB;  // 0 = unknown/IB
+
+    for (int d = 0; d < num_devs; d++) {
+        const char * dn = ibv_get_device_name(devs[d]);
+        if (dev_env && strcmp(dev_env, dn) != 0) continue;
+
+        ibv_context * ctx = ibv_open_device(devs[d]);
+        if (!ctx) continue;
+
+        ibv_port_attr pa;
+        if (ibv_query_port(ctx, ib_port, &pa) != 0) { ibv_close_device(ctx); continue; }
+
+        int found_gid = gid_idx;
+        int found_version = IBV_GID_TYPE_IB;
+        if (found_gid < 0) {
+            // Find a GID on this port whose bytes equal the local TCP address
+            // (IPv4 or IPv6). Prefer RoCE v2 (UDP/IP, L3-routable) over v1
+            // (raw Ethernet, same-L2 only) so silent hangs on L3-routed paths
+            // are avoided. ibv_query_gid_ex returns gid+type in one call.
+            int v2_idx = -1;
+            int v1_idx = -1;
+            for (int i = 0; i < pa.gid_tbl_len; i++) {
+                ibv_gid_entry entry = {};
+                if (ibv_query_gid_ex(ctx, ib_port, i, &entry, 0) != 0) continue;
+                if (memcmp(entry.gid.raw, target_gid->data(), RDMA_GID_SIZE) != 0) continue;
+                if (entry.gid_type == IBV_GID_TYPE_ROCE_V2 && v2_idx < 0) {
+                    v2_idx = i;
+                } else if (entry.gid_type == IBV_GID_TYPE_ROCE_V1 && v1_idx < 0) {
+                    v1_idx = i;
+                }
+            }
+            if (v2_idx >= 0) {
+                found_gid = v2_idx;
+                found_version = IBV_GID_TYPE_ROCE_V2;
+            } else if (v1_idx >= 0) {
+                found_gid = v1_idx;
+                found_version = IBV_GID_TYPE_ROCE_V1;
+            }
+        } else {
+            // Explicit GID index from GGML_RDMA_GID — fetch its type for logging.
+            ibv_gid_entry entry = {};
+            if (ibv_query_gid_ex(ctx, ib_port, found_gid, &entry, 0) == 0) {
+                found_version = entry.gid_type;
+            }
+        }
+        if (found_gid >= 0) {
+            ibctx = ctx;
+            gid_idx = found_gid;
+            gid_version = found_version;
+            matched_dev = dn;
+            rdma_local.path_mtu = pa.active_mtu;
+            break;
+        }
+        ibv_close_device(ctx);
+    }
+    ibv_free_device_list(devs);
+    if (!ibctx) return false;
+
+    rdma_local.ib_port = ib_port;
+    rdma_local.gid_idx = gid_idx;
+
+    rdma = std::make_unique<rdma_conn>();
+    rdma->ctx = ibctx;
+
+    rdma->pd = ibv_alloc_pd(ibctx);
+    if (!rdma->pd) return false;
+
+    rdma->scq = ibv_create_cq(ibctx, 16, nullptr, nullptr, 0);
+    rdma->rcq = ibv_create_cq(ibctx, RDMA_RX_DEPTH + 4, nullptr, nullptr, 0);
+    if (!rdma->scq || !rdma->rcq) return false;
+
+    ibv_qp_init_attr qia = {};
+    qia.send_cq = rdma->scq;
+    qia.recv_cq = rdma->rcq;
+    qia.qp_type = IBV_QPT_RC;
+    qia.cap.max_send_wr     = 4;
+    qia.cap.max_recv_wr     = RDMA_RX_DEPTH + 4;
+    qia.cap.max_send_sge    = 1;
+    qia.cap.max_recv_sge    = 1;
+    qia.cap.max_inline_data = 256;
+
+    rdma->qp = ibv_create_qp(rdma->pd, &qia);
+    if (!rdma->qp) return false;
+    rdma->max_inline = qia.cap.max_inline_data;
+
+    rdma->tx_buf = aligned_alloc(4096, RDMA_CHUNK);
+    rdma->rx_buf = aligned_alloc(4096, static_cast<size_t>(RDMA_RX_DEPTH) * RDMA_CHUNK);
+    if (!rdma->tx_buf || !rdma->rx_buf) return false;
+
+    rdma->tx_mr = ibv_reg_mr(rdma->pd, rdma->tx_buf, RDMA_CHUNK, IBV_ACCESS_LOCAL_WRITE);
+    rdma->rx_mr = ibv_reg_mr(rdma->pd, rdma->rx_buf, static_cast<size_t>(RDMA_RX_DEPTH) * RDMA_CHUNK,
+                           IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);
+    if (!rdma->tx_mr || !rdma->rx_mr) return false;
+
+    ibv_gid local_gid;
+    if (ibv_query_gid(ibctx, ib_port, gid_idx, &local_gid) != 0) return false;
+
+    rdma_local.qpn = rdma->qp->qp_num;
+    rdma_local.psn = rdma->qp->qp_num & 0xffffff;
+    memcpy(&rdma_local.gid, &local_gid, RDMA_GID_SIZE);
+
+    const char * ver_str = "";
+    if (gid_version == IBV_GID_TYPE_ROCE_V2) {
+        ver_str = " RoCEv2";
+    } else if (gid_version == IBV_GID_TYPE_ROCE_V1) {
+        ver_str = " RoCEv1";
+    }
+    GGML_LOG_INFO("RDMA probed: dev=%s gid=%d%s qpn=%u inline=%u\n",
+                  matched_dev, gid_idx, ver_str, rdma_local.qpn, rdma->max_inline);
+    return true;
+}
+
+// Phase 2: Given remote QPN/PSN/GID, transition QP: RESET->INIT->pre-post->RTR->RTS.
+// On success, the connection is live and ready for rdma_send/rdma_recv.
+bool socket_t::impl::rdma_activate(uint32_t remote_qpn, uint32_t remote_psn, const uint8_t * remote_gid) {
+    // RESET -> INIT
+    {
+        struct ibv_qp_attr a = {};
+        a.qp_state        = IBV_QPS_INIT;
+        a.port_num        = rdma_local.ib_port;
+        a.pkey_index      = 0;
+        a.qp_access_flags = IBV_ACCESS_REMOTE_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_LOCAL_WRITE;
+        if (ibv_modify_qp(rdma->qp, &a,
+                IBV_QP_STATE | IBV_QP_PKEY_INDEX | IBV_QP_PORT | IBV_QP_ACCESS_FLAGS) != 0) {
+            return false;
+        }
+    }
+
+    for (int i = 0; i < RDMA_RX_DEPTH; i++) {
+        if (!rdma->post_rx(i)) return false;
+    }
+
+    // INIT -> RTR
+    {
+        struct ibv_qp_attr a = {};
+        a.qp_state           = IBV_QPS_RTR;
+        a.path_mtu           = rdma_local.path_mtu;
+        a.dest_qp_num        = remote_qpn;
+        a.rq_psn             = remote_psn;
+        a.max_dest_rd_atomic = 1;
+        a.min_rnr_timer      = 1;
+        a.ah_attr.is_global  = 1;
+        memcpy(&a.ah_attr.grh.dgid, remote_gid, RDMA_GID_SIZE);
+        a.ah_attr.grh.hop_limit  = 1;
+        a.ah_attr.grh.sgid_index = rdma_local.gid_idx;
+        a.ah_attr.dlid       = 0;
+        a.ah_attr.port_num   = rdma_local.ib_port;
+        if (ibv_modify_qp(rdma->qp, &a,
+                IBV_QP_STATE | IBV_QP_AV | IBV_QP_PATH_MTU | IBV_QP_DEST_QPN |
+                IBV_QP_RQ_PSN | IBV_QP_MAX_DEST_RD_ATOMIC | IBV_QP_MIN_RNR_TIMER) != 0) {
+            return false;
+        }
+    }
+
+    // RTR -> RTS
+    {
+        struct ibv_qp_attr a = {};
+        a.qp_state     = IBV_QPS_RTS;
+        a.timeout      = 14;
+        a.retry_cnt    = 7;
+        a.rnr_retry    = 7;
+        a.sq_psn       = rdma_local.psn;
+        a.max_rd_atomic = 1;
+        if (ibv_modify_qp(rdma->qp, &a,
+                IBV_QP_STATE | IBV_QP_TIMEOUT | IBV_QP_RETRY_CNT | IBV_QP_RNR_RETRY |
+                IBV_QP_SQ_PSN | IBV_QP_MAX_QP_RD_ATOMIC) != 0) {
+            return false;
+        }
+    }
+
+    GGML_LOG_INFO("RDMA activated: qpn=%u->%u mtu=%d rx_depth=%d\n",
+                  rdma_local.qpn, remote_qpn, 128 << rdma_local.path_mtu, RDMA_RX_DEPTH);
+    return true;
+}
+
+bool socket_t::impl::rdma_poll(struct ibv_cq * cq, struct ibv_wc * wc) {
+    for (uint64_t s = 0; ; s++) {
+        int n = ibv_poll_cq(cq, 1, wc);
+        if (n > 0) {
+            if (wc->status != IBV_WC_SUCCESS) {
+                GGML_LOG_ERROR("RDMA CQ wc error: status=%d (%s) vendor_err=0x%x\n",
+                    wc->status, ibv_wc_status_str(wc->status), wc->vendor_err);
+            }
+            return wc->status == IBV_WC_SUCCESS;
+        }
+        if (n < 0) return false;
+        if ((s & 0xFFFFF) == 0 && s > 0) {
+            if (tcp_peer_closed()) {
+                return false;
+            }
+        }
+    }
+}
+
+bool socket_t::impl::rdma_send(const void * data, size_t size) {
+    rdma_conn * c = rdma.get();
+    const uint8_t * src = (const uint8_t *)data;
+    size_t rem = size;
+    while (rem > 0) {
+        size_t chunk = std::min(rem, RDMA_CHUNK);
+
+        struct ibv_sge sge = {};
+        struct ibv_send_wr wr = {}, * bad = nullptr;
+        wr.opcode  = IBV_WR_SEND;
+        wr.sg_list = &sge;
+        wr.num_sge = 1;
+
+        if (chunk <= c->max_inline) {
+            sge.addr   = (uintptr_t)src;
+            sge.length = chunk;
+            wr.send_flags = IBV_SEND_SIGNALED | IBV_SEND_INLINE;
+        } else {
+            memcpy(c->tx_buf, src, chunk);
+            sge.addr   = (uintptr_t)c->tx_buf;
+            sge.length = chunk;
+            sge.lkey   = c->tx_mr->lkey;
+            wr.send_flags = IBV_SEND_SIGNALED;
+        }
+
+        if (ibv_post_send(c->qp, &wr, &bad) != 0) return false;
+        struct ibv_wc wc;
+        if (!rdma_poll(c->scq, &wc)) return false;
+
+        src += chunk;
+        rem -= chunk;
+    }
+    return true;
+}
+
+bool socket_t::impl::rdma_recv(void * data, size_t size) {
+    rdma_conn * c = rdma.get();
+    uint8_t * dst = (uint8_t *)data;
+    size_t rem = size;
+    while (rem > 0) {
+        struct ibv_wc wc;
+        if (!rdma_poll(c->rcq, &wc)) return false;
+
+        int slot = (int)wc.wr_id;
+        size_t got = wc.byte_len;
+        memcpy(dst, c->rx_slot(slot), got);
+
+        if (!c->post_rx(slot)) return false;
+
+        dst += got;
+        rem -= got;
+    }
+    return true;
+}
+
+#endif // GGML_RPC_RDMA
+
+bool socket_t::impl::send_data(const void * data, size_t size) {
+#ifdef GGML_RPC_RDMA
+    if (use_rdma) {
+        return rdma_send(data, size);
+    }
+#endif
+    size_t bytes_sent = 0;
+    while (bytes_sent < size) {
+        size_t size_to_send = std::min(size - bytes_sent, MAX_CHUNK_SIZE);
+        ssize_t n = send(fd, (const char *)data + bytes_sent, size_to_send, 0);
+        if (n < 0) {
+            GGML_LOG_ERROR("send failed (bytes_sent=%zu, size_to_send=%zu)\n",
+                           bytes_sent, size_to_send);
+            return false;
+        }
+        bytes_sent += (size_t)n;
+    }
+    return true;
+}
+
+bool socket_t::impl::recv_data(void * data, size_t size) {
+#ifdef GGML_RPC_RDMA
+    if (use_rdma) {
+        return rdma_recv(data, size);
+    }
+#endif
+    size_t bytes_recv = 0;
+    while (bytes_recv < size) {
+        size_t size_to_recv = std::min(size - bytes_recv, MAX_CHUNK_SIZE);
+        ssize_t n = recv(fd, (char *)data + bytes_recv, size_to_recv, 0);
+        if (n < 0) {
+            GGML_LOG_ERROR("recv failed (bytes_recv=%zu, size_to_recv=%zu)\n",
+                           bytes_recv, size_to_recv);
+            return false;
+        }
+        if (n == 0) {
+            LOG_DBG("recv returned 0 (peer closed?)\n");
+            return false;
+        }
+        bytes_recv += (size_t)n;
+    }
+    return true;
+}
+
+void socket_t::impl::get_caps(uint8_t * local_caps) {
+    memset(local_caps, 0, RPC_CONN_CAPS_SIZE);
+#ifdef GGML_RPC_RDMA
+    rdma_local = {};
+    if (rdma_probe()) {
+        rdma_caps rc = {};
+        rc.qpn = rdma_local.qpn;
+        rc.psn = rdma_local.psn;
+        memcpy(rc.gid, rdma_local.gid, RDMA_GID_SIZE);
+        memcpy(local_caps, &rc, sizeof(rc));
+    } else {
+        rdma.reset();
+    }
+#endif // GGML_RPC_RDMA
+}
+
+void socket_t::impl::update_caps(const uint8_t * remote_caps) {
+#ifdef GGML_RPC_RDMA
+    if (!rdma) {
+        return;
+    }
+    rdma_caps rc = {};
+    memcpy(&rc, remote_caps, sizeof(rc));
+    if (rc.qpn == 0) {
+        rdma.reset();
+        return;
+    }
+    if (rdma_activate(rc.qpn, rc.psn, rc.gid)) {
+        use_rdma = true;
+    } else {
+        GGML_LOG_ERROR("RDMA activate failed, staying on TCP\n");
+        rdma.reset();
+    }
+#else
+    (void)remote_caps;
+#endif // GGML_RPC_RDMA
+}
+
+
+/////////////////////////////////////////////////////////////////////////////
+
+socket_t::socket_t(std::unique_ptr<impl> p) : pimpl(std::move(p)) {}
+
+socket_t::~socket_t() = default;
+
+bool socket_t::send_data(const void * data, size_t size) {
+    return pimpl->send_data(data, size);
+}
+
+bool socket_t::recv_data(void * data, size_t size) {
+    return pimpl->recv_data(data, size);
+}
+
+void socket_t::get_caps(uint8_t * local_caps) {
+    return pimpl->get_caps(local_caps);
+}
+
+void socket_t::update_caps(const uint8_t * remote_caps) {
+    return pimpl->update_caps(remote_caps);
+}
+
+static bool is_valid_fd(sockfd_t sockfd) {
+#ifdef _WIN32
+    return sockfd != INVALID_SOCKET;
+#else
+    return sockfd >= 0;
+#endif
+}
+
+static bool set_no_delay(sockfd_t sockfd) {
+    int flag = 1;
+    // set TCP_NODELAY to disable Nagle's algorithm
+    int ret = setsockopt(sockfd, IPPROTO_TCP, TCP_NODELAY, (char *)&flag, sizeof(int));
+    return ret == 0;
+}
+
+static bool set_reuse_addr(sockfd_t sockfd) {
+    int flag = 1;
+    int ret = setsockopt(sockfd, SOL_SOCKET, SO_REUSEADDR, (char *)&flag, sizeof(int));
+    return ret == 0;
+}
+
+socket_ptr socket_t::accept() {
+    auto client_socket_fd = ::accept(pimpl->fd, NULL, NULL);
+    if (!is_valid_fd(client_socket_fd)) {
+        return nullptr;
+    }
+    if (!set_no_delay(client_socket_fd)) {
+        GGML_LOG_ERROR("Failed to set TCP_NODELAY\n");
+        return nullptr;
+    }
+    return socket_ptr(new socket_t(std::make_unique<impl>(client_socket_fd)));
+}
+
+socket_ptr socket_t::create_server(const char * host, int port) {
+    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    if (!is_valid_fd(sockfd)) {
+        return nullptr;
+    }
+    if (!set_reuse_addr(sockfd)) {
+        GGML_LOG_ERROR("Failed to set SO_REUSEADDR\n");
+        return nullptr;
+    }
+    if (inet_addr(host) == INADDR_NONE) {
+        GGML_LOG_ERROR("Invalid host address: %s\n", host);
+        return nullptr;
+    }
+    struct sockaddr_in serv_addr;
+    serv_addr.sin_family = AF_INET;
+    serv_addr.sin_addr.s_addr = inet_addr(host);
+    serv_addr.sin_port = htons(port);
+
+    if (bind(sockfd, (struct sockaddr *) &serv_addr, sizeof(serv_addr)) < 0) {
+        return nullptr;
+    }
+    if (listen(sockfd, 1) < 0) {
+        return nullptr;
+    }
+    return socket_ptr(new socket_t(std::make_unique<impl>(sockfd)));
+}
+
+socket_ptr socket_t::connect(const char * host, int port) {
+    auto sockfd = socket(AF_INET, SOCK_STREAM, 0);
+    if (!is_valid_fd(sockfd)) {
+        return nullptr;
+    }
+    if (!set_no_delay(sockfd)) {
+        GGML_LOG_ERROR("Failed to set TCP_NODELAY\n");
+        return nullptr;
+    }
+    struct sockaddr_in addr;
+    addr.sin_family = AF_INET;
+    addr.sin_port = htons(port);
+    struct hostent * server = gethostbyname(host);
+    if (server == NULL) {
+        GGML_LOG_ERROR("Cannot resolve host '%s'\n", host);
+        return nullptr;
+    }
+    memcpy(&addr.sin_addr.s_addr, server->h_addr, server->h_length);
+    if (::connect(sockfd, (struct sockaddr *)&addr, sizeof(addr)) < 0) {
+        return nullptr;
+    }
+    return socket_ptr(new socket_t(std::make_unique<impl>(sockfd)));
+}
+
+#ifdef _WIN32
+static std::mutex g_rpc_transport_mu;
+static bool g_rpc_transport_wsa_started = false;
+#endif
+
+bool rpc_transport_init() {
+#ifdef _WIN32
+    std::lock_guard<std::mutex> lock(g_rpc_transport_mu);
+    if (g_rpc_transport_wsa_started) {
+        return true;
+    }
+    WSADATA wsaData;
+    int res = WSAStartup(MAKEWORD(2, 2), &wsaData);
+    if (res != 0) {
+        return false;
+    }
+    g_rpc_transport_wsa_started = true;
+    return true;
+#else
+    return true;
+#endif
+}
+
+void rpc_transport_shutdown() {
+#ifdef _WIN32
+    std::lock_guard<std::mutex> lock(g_rpc_transport_mu);
+    if (!g_rpc_transport_wsa_started) {
+        return;
+    }
+    WSACleanup();
+    g_rpc_transport_wsa_started = false;
+#endif
+}

ggml/src/ggml-rpc/transport.h

@@ -0,0 +1,34 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <memory>
+
+struct socket_t;
+typedef std::shared_ptr<socket_t> socket_ptr;
+
+static constexpr size_t MAX_CHUNK_SIZE = 1024ull * 1024ull * 1024ull; // 1 GiB
+static constexpr size_t RPC_CONN_CAPS_SIZE = 24;
+
+struct socket_t {
+    ~socket_t();
+
+    bool send_data(const void * data, size_t size);
+    bool recv_data(void * data, size_t size);
+
+    socket_ptr accept();
+
+    void get_caps(uint8_t * local_caps);
+    void update_caps(const uint8_t * remote_caps);
+
+    static socket_ptr create_server(const char * host, int port);
+    static socket_ptr connect(const char * host, int port);
+
+private:
+    struct impl;
+    explicit socket_t(std::unique_ptr<impl> p);
+    std::unique_ptr<impl> pimpl;
+};
+
+bool rpc_transport_init();
+void rpc_transport_shutdown();

ggml/src/ggml-sycl/CMakeLists.txt

@@ -154,6 +154,11 @@ if (GGML_SYCL_GRAPH)
     target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_GRAPH)
 endif()
 
+if (GGML_SYCL_HOST_MEM_FALLBACK)
+    message(STATUS "find GGML_SYCL_HOST_MEM_FALLBACK")
+    target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_HOST_MEM_FALLBACK)
+endif()
+
 if (GGML_SYCL_DEVICE_ARCH)
     target_compile_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})
     target_link_options(ggml-sycl PRIVATE -Xsycl-target-backend --offload-arch=${GGML_SYCL_DEVICE_ARCH})

ggml/src/ggml-sycl/convert.cpp

@@ -151,6 +151,25 @@ static void dequantize_row_q4_0_sycl_reorder(const void *vx, dst_t *y, const int
 
 }
 
+template <typename dst_t>
+static void dequantize_row_q8_0_sycl_reorder(const void *vx, dst_t *y, const int64_t k,
+                                     dpct::queue_ptr stream) {
+
+    dpct::has_capability_or_fail(stream->get_device(),
+                                    {sycl::aspect::fp16});
+
+    int constexpr WARP_K = WARP_SIZE * QK8_0;
+    const int n_warp = (k + WARP_K - 1) / WARP_K;
+    GGML_ASSERT(k % QK8_0 == 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_q8_0_reorder(vx, y, k, item_ct1);
+        });
+
+}
+
 template <typename dst_t>
 static void dequantize_row_q4_1_sycl(const void *vx, dst_t *y, const int64_t k,
                                      dpct::queue_ptr stream) {
@@ -614,7 +633,12 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q8_0_sycl_reorder;
+            } else {
+                return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            }
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_sycl;
         case GGML_TYPE_Q3_K:
@@ -683,7 +707,12 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
         case GGML_TYPE_Q5_1:
             return dequantize_block_sycl<QK5_1, QR5_1, dequantize_q5_1>;
         case GGML_TYPE_Q8_0:
-            return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            if (dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu*)dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q8_0_sycl_reorder;
+            } else {
+                return dequantize_block_sycl<QK8_0, QR8_0, dequantize_q8_0>;
+            }
         case GGML_TYPE_Q2_K:
             return dequantize_row_q2_K_sycl;
         case GGML_TYPE_Q3_K:

ggml/src/ggml-sycl/dequantize.hpp

@@ -239,6 +239,34 @@ static void dequantize_block_q4_0_reorder(const void * __restrict__ vx, dst_t *
 
 }
 
+// Dequantize Q8_0 from reorder layout: [all qs (k bytes)][all d values]
+// Each thread handles one block of QK8_0 elements.
+template<typename dst_t>
+static void dequantize_block_q8_0_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t k,
+                                  const sycl::nd_item<3> &item_ct1) {
+
+    const int64_t i = item_ct1.get_group(2);
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int lane_ib = i * WARP_SIZE + tid;
+
+    if (lane_ib >= k / QK8_0) {
+        return;
+    }
+
+    dst_t * y_ptr = yy + lane_ib * QK8_0;
+
+    auto qs = (const int8_t*)vx + lane_ib * QK8_0;
+    auto s_ptr = (const sycl::half*)((const uint8_t*)vx + k) + lane_ib;
+
+    const float d = float(*s_ptr);
+
+#pragma unroll
+    for (int l = 0; l < QK8_0; ++l) {
+        y_ptr[l] = d * qs[l];
+    }
+
+}
+
 template<typename dst_t>
 static void dequantize_block_q4_1(const void * __restrict__ vx, dst_t * __restrict__ yy, int64_t nb32,
                                   const sycl::nd_item<3> &item_ct1) {

ggml/src/ggml-sycl/dmmv.cpp

@@ -615,6 +615,162 @@ static void dequantize_mul_mat_vec_q4_k(const void *__restrict__ vx,
     }
 }
 
+static void dequantize_mul_mat_vec_q4_k_reorder(const void *__restrict__ vx,
+                                                const float *__restrict__ yy,
+                                                float *__restrict__ dst,
+                                                const int ncols, int nrows,
+                                                const sycl::nd_item<3> &item_ct1) {
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    // SOA base pointers for the reordered layout:
+    //   [qs: nb * QK_K/2] [scales: nb * K_SCALE_SIZE] [dm: nb * sizeof(half2)]
+    const int nb = nrows * num_blocks_per_row;
+    const uint8_t     * qs_base     = (const uint8_t *)vx;
+    const uint8_t     * scales_base = qs_base + (size_t)nb * (QK_K / 2);
+    const sycl::half2 * dm_base     = (const sycl::half2 *)(scales_base + (size_t)nb * K_SCALE_SIZE);
+
+#if QK_K == 256
+    const uint16_t kmask1 = 0x3f3f;
+    const uint16_t kmask2 = 0x0f0f;
+    const uint16_t kmask3 = 0xc0c0;
+
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0,1
+
+    const int step = 8/K_QUANTS_PER_ITERATION;           // 8 or 4
+
+    const int il  = tid/step;                            // 0...3
+    const int ir  = tid - step*il;                       // 0...7 or 0...3
+    const int n   = 2 * K_QUANTS_PER_ITERATION;          // 2 or 4
+
+    const int im = il/2;  // 0 or 1. 0 computes 0,32 + 128,160, 1 computes 64,96 + 192,224
+    const int in = il%2;
+
+    const int l0 = n*(2*ir + in);
+    const int q_offset = 32*im + l0;
+    const int y_offset = 64*im + l0;
+
+    uint16_t aux[4];
+    const uint8_t * sc = (const uint8_t *)aux;
+
+#if K_QUANTS_PER_ITERATION == 2
+    uint32_t q32[4];
+    const uint8_t * q4 = (const uint8_t *)q32;
+#else
+    uint16_t q16[4];
+    const uint8_t * q4 = (const uint8_t *)q16;
+#endif
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const int bi = ib0 + i;
+
+        const float   * y1 = yy + i*QK_K + y_offset;
+        const float   * y2 = y1 + 128;
+
+        const sycl::half2 dm_val = dm_base[bi];
+        const float dall = dm_val[0];
+        const float dmin = dm_val[1];
+
+        const uint16_t * a = (const uint16_t *)(scales_base + bi * K_SCALE_SIZE);
+        aux[0] = a[im+0] & kmask1;
+        aux[1] = a[im+2] & kmask1;
+        aux[2] = ((a[im+4] >> 0) & kmask2) | ((a[im+0] & kmask3) >> 2);
+        aux[3] = ((a[im+4] >> 4) & kmask2) | ((a[im+2] & kmask3) >> 2);
+
+#if K_QUANTS_PER_ITERATION == 2
+        const uint32_t * q1 = (const uint32_t *)(qs_base + bi * (QK_K / 2) + q_offset);
+        const uint32_t * q2 = q1 + 16;
+
+        q32[0] = q1[0] & 0x0f0f0f0f;
+        q32[1] = q1[0] & 0xf0f0f0f0;
+        q32[2] = q2[0] & 0x0f0f0f0f;
+        q32[3] = q2[0] & 0xf0f0f0f0;
+
+        sycl::float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 4; ++l) {
+            s.x() += y1[l] * q4[l + 0]; s.y() += y1[l + 32] * q4[l + 4];
+            s.z() += y2[l] * q4[l + 8]; s.w() += y2[l + 32] * q4[l + 12];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x() * sc[0] + s.y() * sc[1] * 1.f / 16.f +
+                       s.z() * sc[4] + s.w() * sc[5] * 1.f / 16.f) -
+               dmin * smin;
+#else
+        const uint16_t * q1 = (const uint16_t *)(qs_base + bi * (QK_K / 2) + q_offset);
+        const uint16_t * q2 = q1 + 32;
+
+        q16[0] = q1[0] & 0x0f0f;
+        q16[1] = q1[0] & 0xf0f0;
+        q16[2] = q2[0] & 0x0f0f;
+        q16[3] = q2[0] & 0xf0f0;
+
+        float4 s = {0.f, 0.f, 0.f, 0.f};
+        float smin = 0;
+        for (int l = 0; l < 2; ++l) {
+            s.x += y1[l] * q4[l+0]; s.y += y1[l+32] * q4[l+2];
+            s.z += y2[l] * q4[l+4]; s.w += y2[l+32] * q4[l+6];
+            smin += y1[l] * sc[2] + y1[l+32] * sc[3] + y2[l] * sc[6] + y2[l+32] * sc[7];
+        }
+        tmp += dall * (s.x * sc[0] + s.y * sc[1] * 1.f/16.f + s.z * sc[4] + s.w * sc[5] * 1.f/16.f) - dmin * smin;
+#endif
+
+    }
+#else
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...15
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    uint16_t aux16[2];
+    const uint8_t * s = (const uint8_t *)aux16;
+
+    float tmp = 0;
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const int bi = ib0 + i;
+
+        const uint8_t * q = qs_base + bi * (QK_K / 2) + step;
+        const float   * y = yy + i*QK_K + step;
+        const uint16_t * a = (const uint16_t *)(scales_base + bi * K_SCALE_SIZE);
+        aux16[0] = a[0] & 0x0f0f;
+        aux16[1] = (a[0] >> 4) & 0x0f0f;
+        const sycl::half2 dm_val = dm_base[bi];
+        const float d = (float)dm_val[0];
+        const float m = (float)dm_val[1];
+        float sum = 0.f;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
+                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
+                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
+                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
+        }
+        tmp += sum;
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
 /*
 DPCT1110:7: The total declared local variable size in device function
 dequantize_mul_mat_vec_q5_k exceeds 128 bytes and may cause high register
@@ -864,6 +1020,129 @@ static void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx, const floa
     }
 }
 
+static void dequantize_mul_mat_vec_q6_k_reorder(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows,
+                                                const sycl::nd_item<3> &item_ct1) {
+
+    static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
+
+    const int row = item_ct1.get_group(2) * item_ct1.get_local_range(1) +
+                    item_ct1.get_local_id(1);
+    if (row > nrows) return;
+
+    const int num_blocks_per_row = ncols / QK_K;
+    const int ib0 = row*num_blocks_per_row;
+
+    // SOA base pointers for the reordered layout:
+    //   [ql: nb * QK_K/2] [qh: nb * QK_K/4] [scales: nb * QK_K/16] [d: nb * sizeof(half)]
+    const int nb = nrows * num_blocks_per_row;
+    const uint8_t   * ql_base     = (const uint8_t *)vx;
+    const uint8_t   * qh_base     = ql_base + (size_t)nb * (QK_K / 2);
+    const int8_t    * scales_base = (const int8_t *)(qh_base + (size_t)nb * (QK_K / 4));
+    const sycl::half * d_base     = (const sycl::half *)((const uint8_t *)scales_base + (size_t)nb * (QK_K / 16));
+
+#if QK_K == 256
+
+    const int tid =
+        item_ct1.get_local_id(2) / K_QUANTS_PER_ITERATION; // 0...31 or 0...16
+    const int ix =
+        item_ct1.get_local_id(2) % K_QUANTS_PER_ITERATION; // 0 or 0, 1
+
+    const int step = 16/K_QUANTS_PER_ITERATION;          // 16 or 8
+
+    const int im = tid/step;                             // 0 or 1. 0 computes 0..., 1 computes 128...
+    const int in = tid - step*im;                        // 0...15 or 0...7
+
+#if K_QUANTS_PER_ITERATION == 1
+    const int l0 = K_QUANTS_PER_ITERATION*in;            // 0...15
+    const int is = 0;
+#else
+    const int l0 = 4 * in;                               // 0, 4, 8, ..., 28
+    const int is = in / 4;
+#endif
+    const int ql_offset = 64*im + l0;
+    const int qh_offset = 32*im + l0;
+    const int s_offset  =  8*im + is;
+    const int y_offset = 128*im + l0;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += K_QUANTS_PER_ITERATION) {
+        const int bi = ib0 + i;
+
+        const float   * y  = yy + i * QK_K + y_offset;
+        const uint8_t * ql = ql_base + bi * (QK_K / 2) + ql_offset;
+        const uint8_t * qh = qh_base + bi * (QK_K / 4) + qh_offset;
+        const int8_t  * s  = scales_base + bi * (QK_K / 16) + s_offset;
+
+        const float d = d_base[bi];
+
+#if K_QUANTS_PER_ITERATION == 1
+        float sum = y[ 0] * s[0] * d * ((int8_t)((ql[ 0] & 0xF) | ((qh[ 0] & 0x03) << 4)) - 32)
+                  + y[16] * s[1] * d * ((int8_t)((ql[16] & 0xF) | ((qh[16] & 0x03) << 4)) - 32)
+                  + y[32] * s[2] * d * ((int8_t)((ql[32] & 0xF) | ((qh[ 0] & 0x0c) << 2)) - 32)
+                  + y[48] * s[3] * d * ((int8_t)((ql[48] & 0xF) | ((qh[16] & 0x0c) << 2)) - 32)
+                  + y[64] * s[4] * d * ((int8_t)((ql[ 0]  >> 4) | ((qh[ 0] & 0x30) >> 0)) - 32)
+                  + y[80] * s[5] * d * ((int8_t)((ql[16]  >> 4) | ((qh[16] & 0x30) >> 0)) - 32)
+                  + y[96] * s[6] * d * ((int8_t)((ql[32]  >> 4) | ((qh[ 0] & 0xc0) >> 2)) - 32)
+                  +y[112] * s[7] * d * ((int8_t)((ql[48]  >> 4) | ((qh[16] & 0xc0) >> 2)) - 32);
+        tmp += sum;
+#else
+        float sum = 0;
+        for (int l = 0; l < 4; ++l) {
+            sum += y[l+ 0] * s[0] * d * ((int8_t)((ql[l+ 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32)
+                 + y[l+32] * s[2] * d * ((int8_t)((ql[l+32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32)
+                 + y[l+64] * s[4] * d * ((int8_t)((ql[l+ 0]  >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32)
+                 + y[l+96] * s[6] * d * ((int8_t)((ql[l+32]  >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32);
+        }
+        tmp += sum;
+#endif
+
+    }
+
+#else
+
+    const int tid = item_ct1.get_local_id(2)/(2*K_QUANTS_PER_ITERATION);  // 0...7
+    const int ix  = item_ct1.get_local_id(2)%(2*K_QUANTS_PER_ITERATION);  // 0...3
+
+    const int step = tid * K_QUANTS_PER_ITERATION;
+
+    float tmp = 0; // partial sum for thread in warp
+
+    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
+        const int bi = ib0 + i;
+
+        const float   * y  = yy + i * QK_K + step;
+        const uint8_t * ql = ql_base + bi * (QK_K / 2) + step;
+        const uint8_t * qh = qh_base + bi * (QK_K / 4) + step;
+        const int8_t  * s  = scales_base + bi * (QK_K / 16);
+
+        const float d = d_base[bi];
+
+        float sum = 0;
+        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
+            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
+                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
+                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
+                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
+        }
+        tmp += sum;
+
+    }
+
+#endif
+
+    // sum up partial sums and write back result
+#pragma unroll
+    for (int mask = QK_WARP_SIZE / 2; mask > 0; mask >>= 1) {
+        tmp +=
+            dpct::permute_sub_group_by_xor(item_ct1.get_sub_group(), tmp, mask);
+    }
+
+    if (tid == 0) {
+        dst[row] = tmp;
+    }
+}
+
 static void dequantize_mul_mat_vec_q4_0_sycl_reorder(const void *vx, const dfloat *y,
                                              float *dst, const int ncols,
                                              const int nrows,
@@ -1167,6 +1446,38 @@ static void dequantize_mul_mat_vec_q6_K_sycl(const void *vx, const float *y,
         });
 }
 
+static void dequantize_mul_mat_vec_q4_K_sycl_reorder(const void *vx, const float *y,
+                                                     float *dst, const int ncols,
+                                                     const int nrows,
+                                                     dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    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_reorder(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
+static void dequantize_mul_mat_vec_q6_K_sycl_reorder(const void *vx, const float *y,
+                                                     float *dst, const int ncols,
+                                                     const int nrows,
+                                                     dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int ny = 2 / K_QUANTS_PER_ITERATION;
+    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_reorder(vx, y, dst, ncols, nrows, item_ct1);
+        });
+}
+
 void ggml_sycl_op_dequantize_mul_mat_vec(
     ggml_backend_sycl_context & ctx,
     const ggml_tensor *src0, const ggml_tensor *src1, ggml_tensor *dst,
@@ -1235,8 +1546,7 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
         case GGML_TYPE_Q4_K:
             if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
                 ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
-                // reorder is currently not supported for dmmv
-                GGML_ABORT("Unimplemented dequantize case case for q4_k reorder");
+                dequantize_mul_mat_vec_q4_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             } else {
                 dequantize_mul_mat_vec_q4_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             }
@@ -1245,7 +1555,12 @@ void ggml_sycl_op_dequantize_mul_mat_vec(
             dequantize_mul_mat_vec_q5_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
             break;
         case GGML_TYPE_Q6_K:
-            dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                dequantize_mul_mat_vec_q6_K_sycl_reorder(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            } else {
+                dequantize_mul_mat_vec_q6_K_sycl(src0_dd_i, src1_ddf_i, dst_dd_i, ne00, row_diff, stream);
+            }
             break;
         case GGML_TYPE_F16:
             convert_mul_mat_vec_f16_sycl(src0_dd_i, src1_dfloat, dst_dd_i, ne00, row_diff, stream);