llama : support GEGLU for jina-bert-v2 (#14090)

ggml-ci

.editorconfig

@@ -49,6 +49,6 @@ charset = unset
 trim_trailing_whitespace = unset
 insert_final_newline = unset
 
-[tools/mtmd/miniaudio.h]
+[vendor/miniaudio/miniaudio.h]
 trim_trailing_whitespace = unset
 insert_final_newline = unset

.github/labeler.yml

@@ -86,3 +86,10 @@ nix:
 embedding:
     - changed-files:
         - any-glob-to-any-file: examples/embedding/
+
+Ascend NPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-cann.h
+            - ggml/src/ggml-cann/**
+            - docs/backend/CANN.md

.github/workflows/build-linux-cross.yml

@@ -26,12 +26,12 @@ jobs:
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-riscv64-linux-gnu \
-                  g++-14-riscv64-linux-gnu \
-                  libcurl4-openssl-dev:riscv64
+                  g++-14-riscv64-linux-gnu
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
                          -DLLAMA_BUILD_TOOLS=ON \
@@ -72,12 +72,12 @@ jobs:
                   glslc \
                   gcc-14-riscv64-linux-gnu \
                   g++-14-riscv64-linux-gnu \
-                  libvulkan-dev:riscv64 \
-                  libcurl4-openssl-dev:riscv64
+                  libvulkan-dev:riscv64
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -118,12 +118,12 @@ jobs:
                   build-essential \
                   glslc \
                   crossbuild-essential-arm64 \
-                  libvulkan-dev:arm64 \
-                  libcurl4-openssl-dev:arm64
+                  libvulkan-dev:arm64
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -163,12 +163,12 @@ jobs:
           sudo apt-get install -y --no-install-recommends \
                   build-essential \
                   gcc-14-powerpc64le-linux-gnu \
-                  g++-14-powerpc64le-linux-gnu \
-                  libcurl4-openssl-dev:ppc64el
+                  g++-14-powerpc64le-linux-gnu
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
                          -DLLAMA_BUILD_TOOLS=ON \
@@ -209,12 +209,12 @@ jobs:
                   glslc \
                   gcc-14-powerpc64le-linux-gnu \
                   g++-14-powerpc64le-linux-gnu \
-                  libvulkan-dev:ppc64el \
-                  libcurl4-openssl-dev:ppc64el
+                  libvulkan-dev:ppc64el
 
       - name: Build
         run: |
-          cmake -B build -DCMAKE_BUILD_TYPE=Release \
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
                          -DGGML_VULKAN=ON \
                          -DGGML_OPENMP=OFF \
                          -DLLAMA_BUILD_EXAMPLES=ON \
@@ -231,3 +231,116 @@ jobs:
                          -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
 
           cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-cpu-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)
+
+  debian-13-loongarch64-vulkan-cross:
+    runs-on: ubuntu-24.04
+    container: debian@sha256:653dfb9f86c3782e8369d5f7d29bb8faba1f4bff9025db46e807fa4c22903671
+
+    steps:
+      - uses: actions/checkout@v4
+      - name: Setup LoongArch
+        run: |
+          rm -f /etc/apt/sources.list.d/*
+          cat << EOF | tee /etc/apt/sources.list.d/debian-ports.list
+          deb http://snapshot.debian.org/archive/debian/20250515T202920Z/ trixie main
+          EOF
+          ( echo 'quiet "true";'; \
+            echo 'APT::Get::Assume-Yes "true";'; \
+            echo 'APT::Install-Recommends "false";'; \
+            echo 'Acquire::Check-Valid-Until "false";'; \
+            echo 'Acquire::Retries "5";'; \
+          ) > /etc/apt/apt.conf.d/99snapshot-repos
+
+          apt-get update
+          apt-get install -y ca-certificates debian-ports-archive-keyring cmake git zip
+          dpkg --add-architecture loong64
+
+          # Add arch-specific repositories for non-amd64 architectures
+          cat << EOF | tee /etc/apt/sources.list.d/loong64-ports.list
+          deb [arch=loong64] http://snapshot.debian.org/archive/debian-ports/20250515T194251Z/ sid main
+          EOF
+
+          apt-get update || true    ;# Prevent failure due to missing URLs.
+
+          apt-get install -y --no-install-recommends \
+                  build-essential \
+                  glslc \
+                  gcc-14-loongarch64-linux-gnu \
+                  g++-14-loongarch64-linux-gnu \
+                  libvulkan-dev:loong64
+
+      - name: Build
+        run: |
+          cmake -B build -DLLAMA_CURL=OFF \
+                         -DCMAKE_BUILD_TYPE=Release \
+                         -DGGML_VULKAN=ON \
+                         -DGGML_OPENMP=OFF \
+                         -DLLAMA_BUILD_EXAMPLES=ON \
+                         -DLLAMA_BUILD_TOOLS=ON \
+                         -DLLAMA_BUILD_TESTS=OFF \
+                         -DCMAKE_SYSTEM_NAME=Linux \
+                         -DCMAKE_SYSTEM_PROCESSOR=loongarch64 \
+                         -DCMAKE_C_COMPILER=loongarch64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=loongarch64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/loongarch64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -306,6 +306,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
+          export GGML_VK_VISIBLE_DEVICES=0
           # This is using llvmpipe and runs slower than other backends
           ctest -L main --verbose --timeout 3600
 
@@ -687,8 +688,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'cpu-x64'
-            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF'
+          - build: 'cpu-x64 (static)'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF'
           - build: 'openblas-x64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/x64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON -DGGML_OPENMP=OFF -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'vulkan-x64'
@@ -839,12 +840,12 @@ jobs:
               -DGGML_CUDA=ON
             cmake --build build
 
-  windows-2019-cmake-cuda:
-    runs-on: windows-2019
+  windows-2022-cmake-cuda:
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -878,7 +879,7 @@ jobs:
         env:
           CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DLLAMA_BUILD_SERVER=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/release.yml

@@ -131,8 +131,9 @@ jobs:
         include:
           - build: 'x64'
             os: ubuntu-22.04
-          - build: 'arm64'
-            os: ubuntu-22.04-arm
+          # GGML_BACKEND_DL and GGML_CPU_ALL_VARIANTS are not currently supported on arm
+          # - build: 'arm64'
+          #   os: ubuntu-22.04-arm
 
     runs-on: ${{ matrix.os }}
 
@@ -159,6 +160,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -207,6 +211,9 @@ jobs:
         id: cmake_build
         run: |
           cmake -B build \
+            -DGGML_BACKEND_DL=ON \
+            -DGGML_NATIVE=OFF \
+            -DGGML_CPU_ALL_VARIANTS=ON \
             -DGGML_VULKAN=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
@@ -373,11 +380,11 @@ jobs:
           name: llama-bin-win-${{ matrix.backend }}-${{ matrix.arch }}.zip
 
   windows-cuda:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     strategy:
       matrix:
-        cuda: ['12.4', '11.7']
+        cuda: ['12.4']
 
     steps:
       - name: Clone
@@ -405,7 +412,7 @@ jobs:
         id: cmake_build
         shell: cmd
         run: |
-          call "C:\Program Files (x86)\Microsoft Visual Studio\2019\Enterprise\VC\Auxiliary\Build\vcvars64.bat"
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" x64
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_NATIVE=OFF ^

.github/workflows/server.yml

@@ -180,7 +180,7 @@ jobs:
 
 
   server-windows:
-    runs-on: windows-2019
+    runs-on: windows-2022
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -159,6 +159,11 @@ if (NOT TARGET ggml AND NOT LLAMA_USE_SYSTEM_GGML)
     # ... otherwise assume ggml is added by a parent CMakeLists.txt
 endif()
 
+if (MINGW)
+    # Target Windows 8 for PrefetchVirtualMemory
+    add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
+endif()
+
 #
 # build the library
 #

Makefile

@@ -367,7 +367,7 @@ ifdef LLAMA_SERVER_SSL
 endif
 
 ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_REPACK
 endif
 
 # warnings
@@ -970,7 +970,7 @@ OBJ_GGML = \
 	$(DIR_GGML)/src/ggml-threading.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu_cpp.o \
-	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-aarch64.o \
+	$(DIR_GGML)/src/ggml-cpu/repack.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-hbm.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-quants.o \
 	$(DIR_GGML)/src/ggml-cpu/ggml-cpu-traits.o \

README.md

@@ -3,6 +3,7 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![Release](https://img.shields.io/github/v/release/ggml-org/llama.cpp)](https://github.com/ggml-org/llama.cpp/releases)
 [![Server](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggml-org/llama.cpp/actions/workflows/server.yml)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggml-org/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggml-org/llama.cpp/discussions/205) / [ggml](https://github.com/ggml-org/ggml)
@@ -28,6 +29,30 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ----
 
+## Quick start
+
+Getting started with llama.cpp is straightforward. Here are several ways to install it on your machine:
+
+- Install `llama.cpp` using [brew, nix or winget](docs/install.md)
+- Run with Docker - see our [Docker documentation](docs/docker.md)
+- Download pre-built binaries from the [releases page](https://github.com/ggml-org/llama.cpp/releases)
+- Build from source by cloning this repository - check out [our build guide](docs/build.md)
+
+Once installed, you'll need a model to work with. Head to the [Obtaining and quantizing models](#obtaining-and-quantizing-models) section to learn more.
+
+Example command:
+
+```sh
+# Use a local model file
+llama-cli -m my_model.gguf
+
+# Or download and run a model directly from Hugging Face
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+
+# Launch OpenAI-compatible API server
+llama-server -hf ggml-org/gemma-3-1b-it-GGUF
+```
+
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -130,6 +155,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 <details>
 <summary>Bindings</summary>
 
+- Python: [ddh0/easy-llama](https://github.com/ddh0/easy-llama)
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
@@ -229,6 +255,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 
 </details>
 
+
 ## Supported backends
 
 | Backend | Target devices |
@@ -245,16 +272,6 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 | [OpenCL](docs/backend/OPENCL.md) | Adreno GPU |
 | [RPC](https://github.com/ggml-org/llama.cpp/tree/master/tools/rpc) | All |
 
-## Building the project
-
-The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
-The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server. Possible methods for obtaining the binaries:
-
-- Clone this repository and build locally, see [how to build](docs/build.md)
-- On MacOS or Linux, install `llama.cpp` via [brew, flox or nix](docs/install.md)
-- Use a Docker image, see [documentation for Docker](docs/docker.md)
-- Download pre-built binaries from [releases](https://github.com/ggml-org/llama.cpp/releases)
-
 ## Obtaining and quantizing models
 
 The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](https://huggingface.co/models?library=gguf&sort=trending) compatible with `llama.cpp`:
@@ -262,7 +279,11 @@ The [Hugging Face](https://huggingface.co) platform hosts a [number of LLMs](htt
 - [Trending](https://huggingface.co/models?library=gguf&sort=trending)
 - [LLaMA](https://huggingface.co/models?sort=trending&search=llama+gguf)
 
-You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`.
+You can either manually download the GGUF file or directly use any `llama.cpp`-compatible models from [Hugging Face](https://huggingface.co/) or other model hosting sites, such as [ModelScope](https://modelscope.cn/), by using this CLI argument: `-hf <user>/<model>[:quant]`. For example:
+
+```sh
+llama-cli -hf ggml-org/gemma-3-1b-it-GGUF
+```
 
 By default, the CLI would download from Hugging Face, you can switch to other options with the environment variable `MODEL_ENDPOINT`. For example, you may opt to downloading model checkpoints from ModelScope or other model sharing communities by setting the environment variable, e.g. `MODEL_ENDPOINT=https://www.modelscope.cn/`.
 

ci/run.sh

@@ -46,7 +46,20 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON"
+
+    if command -v nvidia-smi >/dev/null 2>&1; then
+        CUDA_ARCH=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader,nounits 2>/dev/null | head -1 | tr -d '.')
+        if [[ -n "$CUDA_ARCH" && "$CUDA_ARCH" =~ ^[0-9]+$ ]]; then
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCH}"
+        else
+            echo "Warning: Using fallback CUDA architectures"
+            CMAKE_EXTRA="${CMAKE_EXTRA} -DCMAKE_CUDA_ARCHITECTURES=61;70;75;80;86;89"
+        fi
+    else
+        echo "Error: nvidia-smi not found, cannot build with CUDA"
+        exit 1
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then

common/CMakeLists.txt

@@ -58,23 +58,20 @@ add_library(${TARGET} STATIC
     arg.cpp
     arg.h
     base64.hpp
-    chat.cpp
-    chat.h
     chat-parser.cpp
     chat-parser.h
+    chat.cpp
+    chat.h
     common.cpp
     common.h
     console.cpp
     console.h
-    json-schema-to-grammar.cpp
-    json.hpp
-    json-partial.h
     json-partial.cpp
+    json-partial.h
+    json-schema-to-grammar.cpp
     llguidance.cpp
     log.cpp
     log.h
-    minja/chat-template.hpp
-    minja/minja.hpp
     ngram-cache.cpp
     ngram-cache.h
     regex-partial.cpp
@@ -147,7 +144,7 @@ if (LLAMA_LLGUIDANCE)
     set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} llguidance ${LLGUIDANCE_PLATFORM_LIBS})
 endif ()
 
-target_include_directories(${TARGET} PUBLIC .)
+target_include_directories(${TARGET} PUBLIC . ../vendor)
 target_compile_features   (${TARGET} PUBLIC cxx_std_17)
 target_link_libraries     (${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
 

common/arg.cpp

@@ -1,10 +1,11 @@
-#include "gguf.h" // for reading GGUF splits
 #include "arg.h"
 
+#include "chat.h"
 #include "common.h"
+#include "gguf.h" // for reading GGUF splits
+#include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
-#include "chat.h"
 
 // fix problem with std::min and std::max
 #if defined(_WIN32)
@@ -15,6 +16,9 @@
 #include <windows.h>
 #endif
 
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
 #include <climits>
 #include <cstdarg>
@@ -34,8 +38,6 @@
 #include <future>
 #endif
 
-#include "json-schema-to-grammar.h"
-
 using json = nlohmann::ordered_json;
 
 std::initializer_list<enum llama_example> mmproj_examples = {
@@ -1346,9 +1348,9 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ));
     add_opt(common_arg(
         {"--prio"}, "N",
-        string_format("set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: %d)\n", params.cpuparams.priority),
+        string_format("set process/thread priority : low(-1), normal(0), medium(1), high(2), realtime(3) (default: %d)\n", params.cpuparams.priority),
         [](common_params & params, int prio) {
-            if (prio < 0 || prio > 3) {
+            if (prio < GGML_SCHED_PRIO_LOW || prio > GGML_SCHED_PRIO_REALTIME) {
                 throw std::invalid_argument("invalid value");
             }
             params.cpuparams.priority = (enum ggml_sched_priority) prio;
@@ -2867,6 +2869,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         "(default: deepseek)",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "deepseek") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK; }
+            else if (value == "deepseek-legacy") { params.reasoning_format = COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY; }
             else if (value == "none") {     params.reasoning_format = COMMON_REASONING_FORMAT_NONE; }
             else { throw std::invalid_argument("invalid value"); }
         }

common/chat-parser.cpp

@@ -154,9 +154,10 @@ bool common_chat_msg_parser::try_parse_reasoning(const std::string & start_think
             if (!rest.empty()) {
                 handle_reasoning(rest, /* closed */ !is_partial());
             }
-            if (!syntax_.thinking_forced_open) {
-                throw common_chat_msg_partial_exception(end_think);
-            }
+            // Allow unclosed thinking tags, for now (https://github.com/ggml-org/llama.cpp/issues/13812, https://github.com/ggml-org/llama.cpp/issues/13877)
+            // if (!syntax_.thinking_forced_open) {
+            //     throw common_chat_msg_partial_exception(end_think);
+            // }
             return true;
         }
     }

common/chat-parser.h

@@ -2,9 +2,10 @@
 
 #include "chat.h"
 #include "json-partial.h"
-#include "json.hpp"
 #include "regex-partial.h"
 
+#include <nlohmann/json.hpp>
+
 #include <optional>
 #include <string>
 #include <vector>

common/chat.cpp

@@ -1,13 +1,14 @@
 #include "chat.h"
 #include "chat-parser.h"
 #include "common.h"
+#include "json-partial.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "json-partial.h"
-#include "minja/chat-template.hpp"
-#include "minja/minja.hpp"
 #include "regex-partial.h"
 
+#include <minja/chat-template.hpp>
+#include <minja/minja.hpp>
+
 #include <cstdio>
 #include <exception>
 #include <iostream>
@@ -16,7 +17,6 @@
 #include <string>
 #include <vector>
 
-
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto time = std::chrono::system_clock::to_time_t(now);
     auto local_time = *std::localtime(&time);
@@ -82,10 +82,10 @@ json common_chat_msg::to_json_oaicompat() const
 
 std::vector<common_chat_msg_diff> common_chat_msg_diff::compute_diffs(const common_chat_msg & previous_msg, const common_chat_msg & new_msg) {
     std::vector<common_chat_msg_diff> diffs;
-    // if (previous_msg.reasoning_content != current.reasoning_content) {
-    //     auto & diff = diffs.emplace_back();
-    //     diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, current.reasoning_content);
-    // }
+    if (previous_msg.reasoning_content != new_msg.reasoning_content) {
+        auto & diff = diffs.emplace_back();
+        diff.reasoning_content_delta = string_diff(previous_msg.reasoning_content, new_msg.reasoning_content);
+    }
     if (previous_msg.content != new_msg.content) {
         auto & diff = diffs.emplace_back();
         diff.content_delta = string_diff(previous_msg.content, new_msg.content);
@@ -385,9 +385,9 @@ json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & t
 
 template <> json common_chat_msg_diff_to_json_oaicompat(const common_chat_msg_diff & diff) {
     json delta = json::object();
-    // if (!diff.reasoning_content_delta.empty()) {
-    //     delta["reasoning_content"] = msg.reasoning_content;
-    // }
+    if (!diff.reasoning_content_delta.empty()) {
+        delta["reasoning_content"] = diff.reasoning_content_delta;
+    }
     if (!diff.content_delta.empty()) {
         delta["content"] = diff.content_delta;
     }
@@ -598,6 +598,7 @@ const char * common_reasoning_format_name(common_reasoning_format format) {
     switch (format) {
         case COMMON_REASONING_FORMAT_NONE:     return "none";
         case COMMON_REASONING_FORMAT_DEEPSEEK: return "deepseek";
+        case COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY: return "deepseek-legacy";
         default:
             throw std::runtime_error("Unknown reasoning format");
     }

common/chat.h

@@ -70,7 +70,7 @@ struct common_chat_msg {
 };
 
 struct common_chat_msg_diff {
-    // std::string reasoning_content_delta;
+    std::string reasoning_content_delta;
     std::string content_delta;
     size_t tool_call_index = std::string::npos;
     common_chat_tool_call tool_call_delta;

common/common.cpp

@@ -203,6 +203,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     DWORD p = NORMAL_PRIORITY_CLASS;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = BELOW_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_NORMAL:   p = NORMAL_PRIORITY_CLASS;       break;
         case GGML_SCHED_PRIO_MEDIUM:   p = ABOVE_NORMAL_PRIORITY_CLASS; break;
         case GGML_SCHED_PRIO_HIGH:     p = HIGH_PRIORITY_CLASS;         break;
@@ -228,6 +229,7 @@ bool set_process_priority(enum ggml_sched_priority prio) {
 
     int p = 0;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p =  5;  break;
         case GGML_SCHED_PRIO_NORMAL:   p =  0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   p = -5;  break;
         case GGML_SCHED_PRIO_HIGH:     p = -10; break;
@@ -903,13 +905,16 @@ struct common_init_result common_init_from_params(common_params & params) {
             ok = false;
         }
 
-        if (llama_vocab_eos(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have an EOS token, reranking will not work\n", __func__);
-            ok = false;
-        }
+        bool has_eos = llama_vocab_eos(vocab) != LLAMA_TOKEN_NULL;
+        bool has_sep = llama_vocab_sep(vocab) != LLAMA_TOKEN_NULL;
 
-        if (llama_vocab_sep(vocab) == LLAMA_TOKEN_NULL) {
-            LOG_WRN("%s: warning: vocab does not have a  SEP token, reranking will not work\n", __func__);
+        if (!has_eos && !has_sep) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token or SEP token, reranking will not work\n", __func__);
+            ok = false;
+        } else if (!has_eos) {
+            LOG_WRN("%s: warning: vocab does not have an EOS token, using SEP token as fallback\n", __func__);
+        } else if (!has_sep) {
+            LOG_WRN("%s: warning: vocab does not have a SEP token, reranking will not work\n", __func__);
             ok = false;
         }
 
@@ -929,7 +934,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         return iparams;
     }
 
-    if (params.ctx_shift && !llama_kv_self_can_shift(lctx)) {
+    if (params.ctx_shift && !llama_memory_can_shift(llama_get_memory(lctx))) {
         LOG_WRN("%s: KV cache shifting is not supported for this context, disabling KV cache shifting\n", __func__);
         params.ctx_shift = false;
     }
@@ -1036,7 +1041,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         if (llama_model_has_decoder(model)) {
             llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
-        llama_kv_self_clear(lctx);
+        llama_memory_clear(llama_get_memory(lctx), true);
         llama_synchronize(lctx);
         llama_perf_context_reset(lctx);
         llama_set_warmup(lctx, false);

common/common.h

@@ -215,7 +215,8 @@ struct common_params_vocoder {
 
 enum common_reasoning_format {
     COMMON_REASONING_FORMAT_NONE,
-    COMMON_REASONING_FORMAT_DEEPSEEK, // Extract thinking tag contents and return as `message.reasoning_content`
+    COMMON_REASONING_FORMAT_DEEPSEEK_LEGACY, // Extract thinking tag contents and return as `message.reasoning_content`, or leave inline in <think> tags in stream mode
+    COMMON_REASONING_FORMAT_DEEPSEEK,        // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
 struct common_params {

common/json-partial.cpp

@@ -1,9 +1,10 @@
-#include <json-partial.h>
-#include "ggml.h"
-#include "log.h"
-#include <string>
+#include "json-partial.h"
 
-#include <json.hpp>
+#include "log.h"
+
+#include <nlohmann/json.hpp>
+
+#include <string>
 
 using json = nlohmann::ordered_json;
 

common/json-partial.h

@@ -1,5 +1,6 @@
 #pragma once
-#include <json.hpp>
+
+#include <nlohmann/json.hpp>
 
 // Healing marker (empty if the JSON was fully parsed / wasn't healed).
 struct common_healing_marker {

common/json-schema-to-grammar.cpp

@@ -1,8 +1,9 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
+#include <nlohmann/json.hpp>
+
 #include <algorithm>
-#include <fstream>
 #include <map>
 #include <regex>
 #include <sstream>

common/json-schema-to-grammar.h

@@ -1,9 +1,9 @@
 #pragma once
 
-#include "ggml.h"
-// Change JSON_ASSERT from assert() to GGML_ASSERT:
-#define JSON_ASSERT GGML_ASSERT
-#include "json.hpp"
+#include <nlohmann/json_fwd.hpp>
+
+#include <functional>
+#include <string>
 
 std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
                                    bool force_gbnf = false);

common/speculative.cpp

@@ -144,6 +144,8 @@ llama_tokens common_speculative_gen_draft(
     auto & smpl   = spec->smpl;
     auto & prompt = spec->prompt;
 
+    auto * mem = llama_get_memory(ctx);
+
     int reuse_i = 0;
     int reuse_n = 0;
 
@@ -173,7 +175,7 @@ llama_tokens common_speculative_gen_draft(
     result.reserve(params.n_draft);
 
     if (reuse_n == 0) {
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(mem, false);
 
         prompt.clear();
     } else {
@@ -192,14 +194,14 @@ llama_tokens common_speculative_gen_draft(
         }
 
         if (reuse_i > 0) {
-            llama_kv_self_seq_rm (ctx, 0, 0, reuse_i);
-            llama_kv_self_seq_add(ctx, 0, reuse_i, -1, -reuse_i);
+            llama_memory_seq_rm (mem, 0, 0, reuse_i);
+            llama_memory_seq_add(mem, 0, reuse_i, -1, -reuse_i);
 
             prompt.erase(prompt.begin(), prompt.begin() + reuse_i);
         }
 
         if (reuse_n < (int) prompt.size()) {
-            llama_kv_self_seq_rm (ctx, 0, reuse_n, -1);
+            llama_memory_seq_rm (mem, 0, reuse_n, -1);
 
             prompt.erase(prompt.begin() + reuse_n, prompt.end());
         }

convert_hf_to_gguf.py

@@ -423,19 +423,19 @@ class ModelBase:
         try:
             # for security reason, we don't allow loading remote code by default
             # if a model need remote code, we will fallback to config.json
-            return AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
+            config = AutoConfig.from_pretrained(dir_model, trust_remote_code=False).to_dict()
         except Exception as e:
             logger.warning(f"Failed to load model config from {dir_model}: {e}")
             logger.warning("Trying to load config.json instead")
             with open(dir_model / "config.json", "r", encoding="utf-8") as f:
                 config = json.load(f)
-                if "llm_config" in config:
-                    # rename for InternVL
-                    config["text_config"] = config["llm_config"]
-                if "thinker_config" in config:
-                    # rename for Qwen2.5-Omni
-                    config["text_config"] = config["thinker_config"]["text_config"]
-                return config
+        if "llm_config" in config:
+            # rename for InternVL
+            config["text_config"] = config["llm_config"]
+        if "thinker_config" in config:
+            # rename for Qwen2.5-Omni
+            config["text_config"] = config["thinker_config"]["text_config"]
+        return config
 
     @classmethod
     def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
@@ -523,15 +523,15 @@ class TextModel(ModelBase):
             self.gguf_writer.add_context_length(n_ctx)
             logger.info(f"gguf: context length = {n_ctx}")
 
-        if (n_embd := self.find_hparam(["hidden_size", "n_embd"], optional=True)) is not None:
+        if (n_embd := self.find_hparam(["hidden_size", "n_embd", "dim"], optional=True)) is not None:
             self.gguf_writer.add_embedding_length(n_embd)
             logger.info(f"gguf: embedding length = {n_embd}")
 
-        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner", "hidden_dim"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
             logger.info(f"gguf: feed forward length = {n_ff}")
 
-        if (n_head := self.find_hparam(["num_attention_heads", "n_head"], optional=True)) is not None:
+        if (n_head := self.find_hparam(["num_attention_heads", "n_head", "n_heads"], optional=True)) is not None:
             self.gguf_writer.add_head_count(n_head)
             logger.info(f"gguf: head count = {n_head}")
 
@@ -674,12 +674,12 @@ class TextModel(ModelBase):
         if chkhsh == "8aeee3860c56296a157a1fe2fad249ec40aa59b1bb5709f4ade11c4e6fe652ed":
             # ref: https://huggingface.co/tiiuae/falcon-7b
             res = "falcon"
-        if chkhsh == "9d032fcbd5501f4a38150912590928bfb36091efb5df11b8e2124b0390e3fb1e":
-            # ref: https://huggingface.co/tiiuae/Falcon3-7B-Base
-            res = "falcon3"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
+        if chkhsh == "9d032fcbd5501f4a38150912590928bfb36091efb5df11b8e2124b0390e3fb1e":
+            # ref: https://huggingface.co/tiiuae/Falcon3-7B-Base
+            res = "falcon3"
         if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
             # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
             res = "bert-bge-large"
@@ -731,9 +731,6 @@ class TextModel(ModelBase):
         if chkhsh == "7967bfa498ade6b757b064f31e964dddbb80f8f9a4d68d4ba7998fcf281c531a":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-code
             res = "jina-v2-code"
-        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b" or chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
-            res = "chatglm-bpe"
         if chkhsh == "7fc505bd3104ca1083b150b17d088b59534ede9bde81f0dd2090967d7fe52cee":
             # ref: https://huggingface.co/LumiOpen/Viking-7B
             res = "viking"
@@ -764,9 +761,6 @@ class TextModel(ModelBase):
         if chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
             # ref: https://huggingface.co/facebook/chameleon-7b
             res = "chameleon"
-        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
-            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
-            res = "minerva-7b"
         if chkhsh == "8b5a93ed704057481f240da0be7e7dca721d7f8f4755263b6807227a2cbeae65":
             # ref: https://huggingface.co/sentence-transformers/stsb-roberta-base
             res = "roberta-bpe"
@@ -797,15 +791,24 @@ class TextModel(ModelBase):
         if chkhsh == "d353350c764d8c3b39c763113960e4fb4919bea5fbf208a0e3b22e8469dc7406":
             # ref: https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct
             res = "llama4"
-        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
-            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
-            res = "glm4"
         if chkhsh == "0e9433cbbb161f89e264eb32e8e64bfe69e834973ffca5d41d3948a604a3e2a3":
             # ref: https://huggingface.co/mistral-community/pixtral-12b
             res = "pixtral"
         if chkhsh == "d5f1dd6f980fec569fb218a81a7658ac45fc56b38c5a0adeb1c232fbe04ef5ec":
             # ref: https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base
             res = "seed-coder"
+        if chkhsh == "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-chat
+            res = "chatglm-bpe"
+        if chkhsh == "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2":
+            # ref: https://huggingface.co/THUDM/glm-4-9b-hf
+            res = "glm4"
+        if chkhsh == "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35":
+            # ref: https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0
+            res = "minerva-7b"
 
         if res is None:
             logger.warning("\n")
@@ -1044,6 +1047,10 @@ class TextModel(ModelBase):
         special_vocab.chat_template = "rwkv-world"
         # hack: Add '\n\n' as the EOT token to make it chat normally
         special_vocab._set_special_token("eot", 261)
+        # hack: Override these as they have already been set (incorrectly)
+        special_vocab.special_token_ids["bos"] = 0
+        special_vocab.special_token_ids["eos"] = 0
+
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_builtin(self, model_name: Literal["gpt-neox", "llama-spm"], vocab_size: int):
@@ -1207,7 +1214,7 @@ class MmprojModel(ModelBase):
             self.gguf_writer.add_audio_block_count(self.find_aparam(self.n_block_keys))
             self.gguf_writer.add_audio_head_count(self.find_aparam(["num_attention_heads"]))
 
-        else:
+        if not self.has_vision_encoder and not self.has_audio_encoder:
             raise ValueError("MmprojModel must have either vision or audio encoder")
 
     def write_vocab(self):
@@ -1841,7 +1848,8 @@ class StableLMModel(TextModel):
     "MistralForCausalLM",
     "MixtralForCausalLM",
     "VLlama3ForCausalLM",
-    "LlavaForConditionalGeneration")
+    "LlavaForConditionalGeneration",
+    "LlamaModel")
 class LlamaModel(TextModel):
     model_arch = gguf.MODEL_ARCH.LLAMA
     undo_permute = True
@@ -1921,6 +1929,8 @@ class LlamaModel(TextModel):
 
         if is_vision_tensor:
             return [] # skip vision tensors
+        elif self.hf_arch == "LlamaModel":
+            name = "model." + name
         elif name.startswith("model.text_model"):
             name = name.replace("text_model.", "") # for SmolVLM
         elif name.startswith("language_model."):
@@ -2169,6 +2179,9 @@ class Llama4VisionModel(MmprojModel):
             # process vision tensors
             if "positional_embedding_vlm" in name and ".weight" not in name:
                 name += ".weight"
+            if "multi_modal_projector.linear_1" in name:
+                # despite the name with number postfix, this is a single fully connected layer
+                return [(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.V_MMPROJ_FC], data_torch)]
             return [(self.map_tensor_name(name), data_torch)]
         return []
 
@@ -3676,7 +3689,7 @@ class InternLM3Model(TextModel):
         return [(self.map_tensor_name(name), data_torch)]
 
 
-@ModelBase.register("BertModel", "BertForMaskedLM", "CamembertModel")
+@ModelBase.register("BertModel", "BertForMaskedLM", "CamembertModel", "BertForSequenceClassification")
 class BertModel(TextModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -3684,11 +3697,20 @@ class BertModel(TextModel):
         super().__init__(*args, **kwargs)
         self.vocab_size = None
 
+        if cls_out_labels := self.hparams.get("id2label"):
+            if len(cls_out_labels) == 2 and cls_out_labels[0] == "LABEL_0":
+                # Remove dummy labels added by AutoConfig
+                cls_out_labels = None
+        self.cls_out_labels = cls_out_labels
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
         self._try_set_pooling_type()
 
+        if self.cls_out_labels:
+            self.gguf_writer.add_classifier_output_labels([v for k, v in sorted(self.cls_out_labels.items())])
+
     def set_vocab(self):
         tokens, toktypes, tokpre = self.get_vocab_base()
         self.vocab_size = len(tokens)
@@ -3739,6 +3761,14 @@ class BertModel(TextModel):
         if name.startswith("cls.seq_relationship"):
             return []
 
+        if self.cls_out_labels:
+            # For BertForSequenceClassification (direct projection layer)
+            if name == "classifier.weight":
+                name = "classifier.out_proj.weight"
+
+            if name == "classifier.bias":
+                name = "classifier.out_proj.bias"
+
         return [(self.map_tensor_name(name), data_torch)]
 
     def _xlmroberta_tokenizer_init(self) -> None:
@@ -3758,62 +3788,111 @@ class BertModel(TextModel):
         from sentencepiece import sentencepiece_model_pb2 as model
 
         tokenizer_path = self.dir_model / 'sentencepiece.bpe.model'
+
+        tokenizer_json = {}
+        tokenizer_config_json = {}
         if not tokenizer_path.is_file():
-            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+            tokenizer_path = self.dir_model / 'tokenizer.json'
+            tokenizer_config_path = self.dir_model / 'tokenizer_config.json'
 
-        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
-        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
-        assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+            if not tokenizer_path.is_file():
+                raise FileNotFoundError(f"File not found: {tokenizer_path}")
 
-        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
-        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
-        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+            from base64 import b64decode
+            from transformers import AutoTokenizer
+            tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
 
-        tokenizer = SentencePieceProcessor()
-        tokenizer.LoadFromFile(str(tokenizer_path))
+            with open(tokenizer_path, "r", encoding="utf-8") as fp:
+                tokenizer_json = json.load(fp)
 
-        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+            if tokenizer_config_path.is_file():
+                with open(tokenizer_config_path, "r", encoding="utf-8") as fp:
+                    tokenizer_config_json = json.load(fp)
+
+            add_prefix = tokenizer.add_prefix_space
+            remove_whitespaces = tokenizer.clean_up_tokenization_spaces
+            precompiled_charsmap = b64decode(tokenizer_json["normalizer"]["precompiled_charsmap"])
+
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size)
+        else:
+            sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+            sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+            add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+            remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+            precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+            tokenizer = SentencePieceProcessor()
+            tokenizer.LoadFromFile(str(tokenizer_path))
+
+            vocab_size = max(self.hparams.get("vocab_size", 0), tokenizer.vocab_size())
 
         tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
         scores: list[float] = [-10000.0] * vocab_size
         toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
 
-        for token_id in range(tokenizer.vocab_size()):
-            piece = tokenizer.IdToPiece(token_id)
-            text = piece.encode("utf-8")
-            score = tokenizer.GetScore(token_id)
+        if isinstance(tokenizer, SentencePieceProcessor):
+            for token_id in range(tokenizer.vocab_size()):
+                piece = tokenizer.IdToPiece(token_id)
+                text = piece.encode("utf-8")
+                score = tokenizer.GetScore(token_id)
 
-            toktype = SentencePieceTokenTypes.NORMAL
-            if tokenizer.IsUnknown(token_id):
-                toktype = SentencePieceTokenTypes.UNKNOWN
-            elif tokenizer.IsControl(token_id):
-                toktype = SentencePieceTokenTypes.CONTROL
-            elif tokenizer.IsUnused(token_id):
-                toktype = SentencePieceTokenTypes.UNUSED
-            elif tokenizer.IsByte(token_id):
-                toktype = SentencePieceTokenTypes.BYTE
+                toktype = SentencePieceTokenTypes.NORMAL
+                if tokenizer.IsUnknown(token_id):
+                    toktype = SentencePieceTokenTypes.UNKNOWN
+                elif tokenizer.IsControl(token_id):
+                    toktype = SentencePieceTokenTypes.CONTROL
+                elif tokenizer.IsUnused(token_id):
+                    toktype = SentencePieceTokenTypes.UNUSED
+                elif tokenizer.IsByte(token_id):
+                    toktype = SentencePieceTokenTypes.BYTE
 
-            tokens[token_id] = text
-            scores[token_id] = score
-            toktypes[token_id] = toktype
+                tokens[token_id] = text
+                scores[token_id] = score
+                toktypes[token_id] = toktype
+        else:
+            added_vocab = tokenizer.get_added_vocab()
+            unk_token = tokenizer_config_json.get("unk_token")
+            unk_token_id = added_vocab.get(unk_token, tokenizer_json["model"].get("unk_id", 3))
 
-        if vocab_size > len(tokens):
-            pad_count = vocab_size - len(tokens)
-            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
-            for i in range(1, pad_count + 1):
-                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
-                scores.append(-1000.0)
-                toktypes.append(SentencePieceTokenTypes.UNUSED)
+            for token_id in range(tokenizer.vocab_size):
+                piece = tokenizer._convert_id_to_token(token_id)
+                if (piece := tokenizer._convert_id_to_token(token_id)) is not None:
+                    text = piece.encode("utf-8")
+                    score = tokenizer_json["model"]["vocab"][token_id][1]
 
-        # realign tokens (see HF tokenizer code)
-        tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
-        scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
-        toktypes = [
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.CONTROL,
-            SentencePieceTokenTypes.UNKNOWN,
-        ] + toktypes[3:-1]
+                    toktype = SentencePieceTokenTypes.NORMAL
+                    if token_id == unk_token_id:
+                        toktype = SentencePieceTokenTypes.UNKNOWN
+                    elif token_id in tokenizer.all_special_ids:
+                        toktype = SentencePieceTokenTypes.CONTROL
+                    elif token_id in added_vocab.values():
+                        toktype = SentencePieceTokenTypes.USER_DEFINED
+                    # No reliable way to detect this, but jina doesn't have any
+                    # elif tokenizer.IsByte(token_id):
+                    #     toktype = SentencePieceTokenTypes.BYTE
+
+                    tokens[token_id] = text
+                    scores[token_id] = score
+                    toktypes[token_id] = toktype
+
+        if isinstance(tokenizer, SentencePieceProcessor):
+            # realign tokens (see HF tokenizer code)
+            tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
+            scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
+            toktypes = [
+                SentencePieceTokenTypes.CONTROL,
+                SentencePieceTokenTypes.CONTROL,
+                SentencePieceTokenTypes.CONTROL,
+                SentencePieceTokenTypes.UNKNOWN,
+            ] + toktypes[3:-1]
+
+            if self.model_arch == gguf.MODEL_ARCH.NOMIC_BERT_MOE:
+                # Add mask token missing from sentencepiece.bpe.model
+                tokens[250001] = b'<mask>'
+                scores[250001] = 0.0
+                toktypes[250001] = SentencePieceTokenTypes.CONTROL
 
         self.gguf_writer.add_tokenizer_model("t5")
         self.gguf_writer.add_tokenizer_pre("default")
@@ -3833,7 +3912,27 @@ class BertModel(TextModel):
         self.gguf_writer.add_add_eos_token(True)
 
 
-@ModelBase.register("RobertaModel")
+@ModelBase.register("DistilBertModel", "DistilBertForMaskedLM", "DistilBertForSequenceClassification")
+class DistilBertModel(BertModel):
+    model_arch = gguf.MODEL_ARCH.BERT
+
+    def set_gguf_parameters(self):
+        self.gguf_writer.add_layer_norm_eps(1e-12)
+        logger.info("gguf: layer norm epsilon = 1e-12")
+        super().set_gguf_parameters()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("distilbert."):
+            name = name[11:]
+
+        # These layers act as MLM head, so we don't need them
+        if name.startswith("vocab_"):
+            return []
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@ModelBase.register("RobertaModel", "RobertaForSequenceClassification")
 class RobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -4699,25 +4798,6 @@ class OlmoeModel(TextModel):
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.intermediate_size = self.hparams["intermediate_size"]
-
-    def get_tensors(self):
-        for name, data in super().get_tensors():
-            if 'gated_layer' in name:
-                d1 = data[:self.intermediate_size, :]
-                name1 = name.replace('gated_layers', 'gated_layers_w')
-                name1 = name1.replace('up_gated_layer', 'gated_layers_v')
-                d2 = data[self.intermediate_size:, :]
-                name2 = name.replace('gated_layers', 'gated_layers_v')
-                name2 = name2.replace('up_gated_layer', 'gated_layers_w')
-                yield name1, d1
-                yield name2, d2
-                continue
-
-            yield name, data
-
     def set_vocab(self):
         tokenizer_class = 'BertTokenizer'
         with open(self.dir_model / "tokenizer_config.json", "r", encoding="utf-8") as f:
@@ -4733,14 +4813,6 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        # if name starts with "bert.", remove the prefix
-        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
-        if name.startswith("bert."):
-            name = name[5:]
-
-        return super().modify_tensors(data_torch, name, bid)
-
 
 @ModelBase.register("OpenELMForCausalLM")
 class OpenELMModel(TextModel):

convert_hf_to_gguf_update.py

@@ -1,28 +1,6 @@
 #!/usr/bin/env python3
 # -*- coding: utf-8 -*-
 
-# This script downloads the tokenizer models of the specified models from Huggingface and
-# generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
-#
-# This is necessary in order to analyze the type of pre-tokenizer used by the model and
-# provide the necessary information to llama.cpp via the GGUF header in order to implement
-# the same pre-tokenizer.
-#
-# ref: https://github.com/ggml-org/llama.cpp/pull/6920
-#
-# Instructions:
-#
-# - Add a new model to the "models" list
-# - Run the script with your huggingface token:
-#
-#   python3 convert_hf_to_gguf_update.py <huggingface_token>
-#
-# - The convert_hf_to_gguf.py script will have had its get_vocab_base_pre() function updated
-# - Update llama.cpp with the new pre-tokenizer if necessary
-#
-# TODO: generate tokenizer tests for llama.cpp
-#
-
 import logging
 import os
 import pathlib
@@ -32,6 +10,7 @@ import requests
 import sys
 import json
 import shutil
+import argparse
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -41,6 +20,11 @@ logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger("convert_hf_to_gguf_update")
 sess = requests.Session()
 
+convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
+convert_py = convert_py_pth.read_text(encoding="utf-8")
+hf_token_pth = pathlib.Path.home() / ".cache" / "huggingface" / "token"
+hf_token = hf_token_pth.read_text(encoding="utf-8").strip() if hf_token_pth.exists() else None
+
 
 class TOKENIZER_TYPE(IntEnum):
     SPM = auto()
@@ -49,20 +33,49 @@ class TOKENIZER_TYPE(IntEnum):
     UGM = auto()
 
 
+DOC_STRING = """
+This script downloads the tokenizer models of the specified models from Huggingface and
+generates the get_vocab_base_pre() function for convert_hf_to_gguf.py
+
+/!\\ It is intended to be used by contributors and is not meant to be run by end users
+
+This is necessary in order to analyze the type of pre-tokenizer used by the model and
+provide the necessary information to llama.cpp via the GGUF header in order to implement
+the same pre-tokenizer.
+
+ref: https://github.com/ggml-org/llama.cpp/pull/6920
+
+Instructions:
+
+- Add a new model to the "models" list
+- Run the script with your huggingface token
+    By default, token will be read from ~/.cache/huggingface/token
+- The convert_hf_to_gguf.py script will have had its get_vocab_base_pre() function updated
+- Update llama.cpp with the new pre-tokenizer if necessary
+"""
+# TODO: generate tokenizer tests for llama.cpp
+
+parser = argparse.ArgumentParser(description=DOC_STRING, formatter_class=argparse.RawTextHelpFormatter)
+parser.add_argument(
+    "--full", action="store_true",
+    help="download full list of models - make sure you have access to all of them",
+)
+parser.add_argument(
+    "hf_token",
+    help="optional HF token",
+    nargs="?",
+)
+args = parser.parse_args()
+hf_token = args.hf_token if args.hf_token is not None else hf_token
+
+if hf_token is None:
+    logger.error("HF token is required. Please provide it as an argument or set it in ~/.cache/huggingface/token")
+    sys.exit(1)
+
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
 CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
 
-if len(sys.argv) == 2:
-    token = sys.argv[1]
-    if not token.startswith("hf_"):
-        logger.info("Huggingface token seems invalid")
-        logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
-        sys.exit(1)
-else:
-    logger.info("Usage: python convert_hf_to_gguf_update.py <huggingface_token>")
-    sys.exit(1)
-
 # TODO: add models here, base models preferred
 models = [
     {"name": "llama-spm",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
@@ -103,7 +116,6 @@ models = [
     {"name": "exaone",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
     {"name": "phi-2",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
     {"name": "chameleon",        "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
-    {"name": "minerva-7b",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", },
     {"name": "roberta-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sentence-transformers/stsb-roberta-base"},
     {"name": "gigachat",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ai-sage/GigaChat-20B-A3B-instruct"},
     {"name": "megrez",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Infinigence/Megrez-3B-Instruct"},
@@ -114,11 +126,19 @@ models = [
     {"name": "trillion",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/trillionlabs/Trillion-7B-preview", },
     {"name": "bailingmoe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/inclusionAI/Ling-lite", },
     {"name": "llama4",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct", },
-    {"name": "glm4",             "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", },
     {"name": "pixtral",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistral-community/pixtral-12b", },
     {"name": "seed-coder",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/ByteDance-Seed/Seed-Coder-8B-Base", },
 ]
 
+# some models are known to be broken upstream, so we will skip them as exceptions
+pre_computed_hashes = [
+    # chatglm-bpe has 2 hashes, why?
+    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "b6e8e1518dc4305be2fe39c313ed643381c4da5db34a98f6a04c093f8afbe99b"},
+    {"name": "chatglm-bpe", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-chat", "chkhsh": "81d72c7348a9f0ebe86f23298d37debe0a5e71149e29bd283904c02262b27516"},
+    {"name": "glm4", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/THUDM/glm-4-9b-hf", "chkhsh": "a1336059768a55c99a734006ffb02203cd450fed003e9a71886c88acf24fdbc2"},
+    {"name": "minerva-7b", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/sapienzanlp/Minerva-7B-base-v1.0", "chkhsh": "1431a23e583c97432bc230bff598d103ddb5a1f89960c8f1d1051aaa944d0b35"},
+]
+
 
 def download_file_with_auth(url, token, save_path):
     headers = {"Authorization": f"Bearer {token}"}
@@ -169,9 +189,29 @@ def download_model(model):
             if os.path.isfile(save_path):
                 logger.info(f"{name}: File {save_path} already exists - skipping")
                 continue
-            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+            download_file_with_auth(f"{repo}/resolve/main/{file}", hf_token, save_path)
 
 
+# get list of existing models and chkhsh from the convert_hf_to_gguf.py file
+# returns mapping res --> chkhsh
+def get_existing_models(convert_py):
+    pattern = r'if chkhsh == "([a-f0-9]{64})":\s*\n\s*.*\s*res = "([^"]+)"'
+    matches = re.findall(pattern, convert_py)
+    output = {}
+    for chkhsh, res in matches:
+        output[res] = chkhsh
+    return output
+
+
+existing_models = {}
+all_models = models.copy()
+if not args.full:
+    # Filter out models that already exist in convert_hf_to_gguf.py
+    existing_models = get_existing_models(convert_py)
+    all_models = models.copy()
+    models = [model for model in all_models if model["name"] not in existing_models]
+
+logging.info(f"Downloading {len(models)} models...")
 for model in models:
     try:
         download_model(model)
@@ -182,9 +222,10 @@ for model in models:
 # generate the source code for the convert_hf_to_gguf.py:get_vocab_base_pre() function:
 
 src_ifs = ""
-for model in models:
+for model in [*all_models, *pre_computed_hashes]:
     name = model["name"]
     tokt = model["tokt"]
+    chkhsh = model.get("chkhsh")
 
     if tokt == TOKENIZER_TYPE.SPM or tokt == TOKENIZER_TYPE.UGM:
         continue
@@ -195,35 +236,44 @@ for model in models:
         continue
 
     # create the tokenizer
-    try:
-        if name == "t5":
-            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
-        else:
-            tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
-    except OSError as e:
-        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
-        continue  # Skip to the next model if the tokenizer can't be loaded
+    if chkhsh is not None:
+        # if the model has a pre-computed hash, use it
+        logger.info(f"Using pre-computed hash for model {name}: {chkhsh}")
+    elif name in existing_models:
+        # if the model already exists in convert_hf_to_gguf.py, skip compute hash
+        chkhsh = existing_models[name]
+    else:
+        # otherwise, compute the hash of the tokenizer
+        try:
+            logger.info(f"Loading tokenizer from {f'models/tokenizers/{name}'}...")
+            if name == "t5":
+                tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}", use_fast=False)
+            else:
+                tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+        except OSError as e:
+            logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+            continue  # Skip to the next model if the tokenizer can't be loaded
 
-    chktok = tokenizer.encode(CHK_TXT)
-    chkhsh = sha256(str(chktok).encode()).hexdigest()
+        chktok = tokenizer.encode(CHK_TXT)
+        chkhsh = sha256(str(chktok).encode()).hexdigest()
 
-    logger.info(f"model: {name}")
-    logger.info(f"tokt: {tokt}")
-    logger.info(f"repo: {model['repo']}")
-    logger.info(f"chktok: {chktok}")
-    logger.info(f"chkhsh: {chkhsh}")
+        logger.info(f"model: {name}")
+        logger.info(f"tokt: {tokt}")
+        logger.info(f"repo: {model['repo']}")
+        logger.info(f"chktok: {chktok}")
+        logger.info(f"chkhsh: {chkhsh}")
 
-    # print the "pre_tokenizer" content from the tokenizer.json
-    with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
-        cfg = json.load(f)
-        normalizer = cfg["normalizer"]
-        logger.info("normalizer: " + json.dumps(normalizer, indent=4))
-        pre_tokenizer = cfg["pre_tokenizer"]
-        logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
-        if "ignore_merges" in cfg["model"]:
-            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
+        # print the "pre_tokenizer" content from the tokenizer.json
+        with open(f"models/tokenizers/{name}/tokenizer.json", "r", encoding="utf-8") as f:
+            cfg = json.load(f)
+            normalizer = cfg["normalizer"]
+            logger.info("normalizer: " + json.dumps(normalizer, indent=4))
+            pre_tokenizer = cfg["pre_tokenizer"]
+            logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+            if "ignore_merges" in cfg["model"]:
+                logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))
 
-    logger.info("")
+        logger.info("")
 
     src_ifs += f"        if chkhsh == \"{chkhsh}\":\n"
     src_ifs += f"            # ref: {model['repo']}\n"
@@ -271,8 +321,6 @@ src_func = f"""
         return res
 """
 
-convert_py_pth = pathlib.Path("convert_hf_to_gguf.py")
-convert_py = convert_py_pth.read_text(encoding="utf-8")
 convert_py = re.sub(
     r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
     lambda m: m.group(1) + src_func + m.group(3),
@@ -288,7 +336,7 @@ logger.info("+++ convert_hf_to_gguf.py was updated")
 
 tests = [
     "ied 4 ½ months",
-    "Führer",
+    "Äpfel",
     "",
     " ",
     "  ",
@@ -367,6 +415,10 @@ for model in models:
         logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
         continue  # Skip this model and continue with the next one in the loop
 
+    if not os.path.exists(f"models/ggml-vocab-{name}.gguf"):
+        logger.info(f"Skip vocab files for model {name}, no GGUF file found")
+        continue
+
     with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
         for text in tests:
             f.write(f"{text}")

docs/backend/CANN.md

@@ -8,6 +8,7 @@
  - [DataType Supports](#datatype-supports)
  - [Docker](#docker)
  - [Linux](#linux)
+ - [Environment variable setup](#environment-variable-setup)
  - [TODO](#todo)
 
 
@@ -290,5 +291,24 @@ Authors from Peking University: Bizhao Shi (bshi@pku.edu.cn), Yuxin Yang (yxyang
 
 We would like to thank Tuo Dai, Shanni Li, and all of the project maintainers from Huawei Technologies Co., Ltd for their help during the code development and pull request.
 
+## Environment variable setup
+
+### GGML_CANN_ASYNC_MODE
+
+Enables asynchronous operator submission. Disabled by default.
+
+### GGML_CANN_MEM_POOL
+
+Specifies the memory pool management strategy:
+
+- vmm: Utilizes a virtual memory manager pool. If hardware support for VMM is unavailable, falls back to the legacy (leg) memory pool.
+
+- prio: Employs a priority queue-based memory pool management.
+- leg: Uses a fixed-size buffer pool.
+
+### GGML_CANN_DISABLE_BUF_POOL_CLEAN
+
+Controls automatic cleanup of the memory pool. This option is only effective when using the prio or leg memory pool strategies.
+
 ## TODO
 - Support more models and data types.

docs/build.md

@@ -1,5 +1,9 @@
 # Build llama.cpp locally
 
+The main product of this project is the `llama` library. Its C-style interface can be found in [include/llama.h](include/llama.h).
+
+The project also includes many example programs and tools using the `llama` library. The examples range from simple, minimal code snippets to sophisticated sub-projects such as an OpenAI-compatible HTTP server.
+
 **To get the Code:**
 
 ```bash
@@ -63,6 +67,7 @@ cmake --build build --config Release
       cmake --preset x64-windows-llvm-release
       cmake --build build-x64-windows-llvm-release
       ```
+- Curl usage is enabled by default and can be turned off with `-DLLAMA_CURL=OFF`. Otherwise you need to install development libraries for libcurl.
 
 ## BLAS Build
 

docs/install.md

@@ -1,28 +1,42 @@
 # Install pre-built version of llama.cpp
 
-## Homebrew
+| Install via | Windows | Mac | Linux |
+|-------------|---------|-----|-------|
+| Winget      | ✅      |      |      |
+| Homebrew    |         | ✅   | ✅   |
+| MacPorts    |         | ✅   |      |
+| Nix         |         | ✅   | ✅   |
 
-On Mac and Linux, the homebrew package manager can be used via
+## Winget (Windows)
+
+```sh
+winget install llama.cpp
+```
+
+The package is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/issues/8188
+
+## Homebrew (Mac and Linux)
 
 ```sh
 brew install llama.cpp
 ```
+
 The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggml-org/llama.cpp/discussions/7668
 
-## MacPorts
+## MacPorts (Mac)
 
 ```sh
 sudo port install llama.cpp
 ```
-see also: https://ports.macports.org/port/llama.cpp/details/
 
-## Nix
+See also: https://ports.macports.org/port/llama.cpp/details/
 
-On Mac and Linux, the Nix package manager can be used via
+## Nix (Mac and Linux)
 
 ```sh
 nix profile install nixpkgs#llama-cpp
 ```
+
 For flake enabled installs.
 
 Or
@@ -34,13 +48,3 @@ nix-env --file '<nixpkgs>' --install --attr llama-cpp
 For non-flake enabled installs.
 
 This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-## Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-
-```sh
-flox install llama-cpp
-```
-
-Flox follows the nixpkgs build of llama.cpp.

examples/batched.swift/Sources/main.swift

@@ -116,7 +116,7 @@ if llama_decode(context, batch) != 0 {
 }
 
 for i in 1 ..< n_parallel {
-    llama_kv_self_seq_cp(context, 0, Int32(i), 0, batch.n_tokens)
+    llama_memory_seq_cp(llama_get_memory(context), 0, Int32(i), 0, batch.n_tokens)
 }
 
 if n_parallel > 1 {

examples/embedding/embedding.cpp

@@ -37,7 +37,7 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
 
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
@@ -236,9 +236,24 @@ int main(int argc, char ** argv) {
                 LOG("\n");
             }
         } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            const uint32_t n_cls_out = llama_model_n_cls_out(model);
+            std::vector<std::string> cls_out_labels;
+
+            for (uint32_t i = 0; i < n_cls_out; i++) {
+                const char * label = llama_model_cls_label(model, i);
+                const std::string label_i(label == nullptr ? "" : label);
+                cls_out_labels.emplace_back(label_i.empty() ? std::to_string(i) : label_i);
+            }
+
             for (int j = 0; j < n_embd_count; j++) {
-                // NOTE: if you change this log - update the tests in ci/run.sh
-                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                for (uint32_t i = 0; i < n_cls_out; i++) {
+                    // NOTE: if you change this log - update the tests in ci/run.sh
+                    if (n_cls_out == 1) {
+                        LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+                    } else {
+                        LOG("rerank score %d: %8.3f [%s]\n", j, emb[j * n_embd + i], cls_out_labels[i].c_str());
+                    }
+                }
             }
         } else {
             // print the first part of the embeddings or for a single prompt, the full embedding

examples/gritlm/gritlm.cpp

@@ -45,7 +45,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_self_clear(ctx);
+        llama_memory_clear(llama_get_memory(ctx), true);
         llama_set_embeddings(ctx, true);
         llama_set_causal_attn(ctx, false);
 
@@ -102,7 +102,7 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_token eos_token = llama_vocab_eos(vocab);
 
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), true);
     llama_set_embeddings(ctx, false);
     llama_set_causal_attn(ctx, true);
 

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -194,7 +194,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         }
 
         batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp_start = ggml_time_us();
         if (llama_decode(context, *batch) != 0) {
@@ -206,7 +206,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         LOGi("Benchmark text generation (tg)");
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
@@ -223,7 +223,7 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
 
         const auto t_tg_end = ggml_time_us();
 
-        llama_kv_self_clear(context);
+        llama_memory_clear(llama_get_memory(context), false);
 
         const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
         const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
@@ -448,5 +448,5 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_self_clear(reinterpret_cast<llama_context *>(context));
+    llama_memory_clear(llama_get_memory(reinterpret_cast<llama_context *>(context)), true);
 }

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -210,7 +210,7 @@ actor LlamaContext {
             }
             batch.logits[Int(batch.n_tokens) - 1] = 1 // true
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -223,7 +223,7 @@ actor LlamaContext {
 
             // bench text generation
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_tg_start = DispatchTime.now().uptimeNanoseconds / 1000;
 
@@ -242,7 +242,7 @@ actor LlamaContext {
 
             let t_tg_end = DispatchTime.now().uptimeNanoseconds / 1000;
 
-            llama_kv_self_clear(context)
+            llama_memory_clear(llama_get_memory(context), false)
 
             let t_pp = Double(t_pp_end - t_pp_start) / 1000000.0
             let t_tg = Double(t_tg_end - t_tg_start) / 1000000.0
@@ -292,7 +292,7 @@ actor LlamaContext {
     func clear() {
         tokens_list.removeAll()
         temporary_invalid_cchars.removeAll()
-        llama_kv_self_clear(context)
+        llama_memory_clear(llama_get_memory(context), true)
     }
 
     private func tokenize(text: String, add_bos: Bool) -> [llama_token] {

examples/lookahead/lookahead.cpp

@@ -60,6 +60,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // Tokenize the prompt
@@ -94,7 +96,7 @@ int main(int argc, char ** argv) {
     llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
-        llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+        llama_memory_seq_cp(mem, 0, s, -1, -1);
     }
 
     const auto t_enc_end = ggml_time_us();
@@ -427,17 +429,17 @@ int main(int argc, char ** argv) {
 
         // KV cache management
         // if no verification token matched, we simply remove all cells from this batch -> no fragmentation
-        llama_kv_self_seq_rm(ctx, -1, n_past, -1);
+        llama_memory_seq_rm(mem, -1, n_past, -1);
 
         if (seq_id_best != 0) {
             // if a verification token matched, we keep the best sequence and remove the rest
             // this leads to some KV cache fragmentation
-            llama_kv_self_seq_keep(ctx, seq_id_best);
-            llama_kv_self_seq_cp  (ctx, seq_id_best, 0, -1, -1);
-            llama_kv_self_seq_rm  (ctx, seq_id_best,    -1, -1);
+            llama_memory_seq_keep(mem, seq_id_best);
+            llama_memory_seq_cp  (mem, seq_id_best, 0, -1, -1);
+            llama_memory_seq_rm  (mem, seq_id_best,    -1, -1);
 
             for (int s = 1; s < W + G + 1; ++s) {
-                llama_kv_self_seq_cp(ctx, 0, s, -1, -1);
+                llama_memory_seq_cp(mem, 0, s, -1, -1);
             }
         }
     }

examples/lookup/lookup.cpp

@@ -181,7 +181,7 @@ int main(int argc, char ** argv){
 
         // KV cache management
         // clean the cache of draft tokens that weren't accepted
-        llama_kv_self_seq_rm(ctx, 0, n_past, -1);
+        llama_memory_seq_rm(llama_get_memory(ctx), 0, n_past, -1);
 
         common_batch_clear(batch_tgt);
         common_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);

examples/parallel/README.md

@@ -4,7 +4,7 @@ Simplified simulation of serving incoming requests in parallel
 
 ## Example
 
-Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of 10 junk questions (`-j 10`) followed by the actual question.
+Generate 128 client requests (`-ns 128`), simulating 8 concurrent clients (`-np 8`). The system prompt is shared (`-pps`), meaning that it is computed once at the start. The client requests consist of up to 10 junk questions (`--junk 10`) followed by the actual question.
 
 ```bash
 llama-parallel -m model.gguf -np 8 -ns 128 --top-k 1 -pps --junk 10 -c 16384

examples/parallel/parallel.cpp

@@ -158,7 +158,7 @@ int main(int argc, char ** argv) {
     common_params params;
 
     params.n_predict = 128;
-    params.n_junk = 0;
+    params.n_junk = 1;
 
     if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
@@ -182,7 +182,7 @@ int main(int argc, char ** argv) {
     const bool is_sp_shared = params.is_pp_shared;
 
     // extra text to insert in each client's prompt in order to make it larger
-    const int32_t n_junk = params.n_junk;
+    const int32_t n_junk = std::max(1, params.n_junk);
 
     // init llama.cpp
     llama_backend_init();
@@ -194,6 +194,8 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_init.model.get();
     llama_context * ctx = llama_init.context.get();
 
+    auto * mem = llama_get_memory(ctx);
+
     const llama_vocab * vocab = llama_model_get_vocab(model);
 
     // load the prompts from an external file if there are any
@@ -259,7 +261,7 @@ int main(int argc, char ** argv) {
 
         // assign the system KV cache to all parallel sequences
         for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+            llama_memory_seq_cp(mem, 0, i, -1, -1);
         }
 
         LOG_INF("\n");
@@ -286,9 +288,9 @@ int main(int argc, char ** argv) {
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
             for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_self_seq_rm(ctx, i, -1, -1);
+                llama_memory_seq_rm(mem, i, -1, -1);
                 // but keep the system prompt
-                llama_kv_self_seq_cp(ctx, 0, i, -1, -1);
+                llama_memory_seq_cp(mem, 0, i, -1, -1);
             }
 
             LOG_INF("%s: clearing the KV cache\n", __func__);
@@ -315,7 +317,10 @@ int main(int argc, char ** argv) {
                     } else {
                         client.prompt += k_system;
                     }
-                    for (int i = 0; i < n_junk; ++i) {
+
+                    const int n_junk_cur = rand() % n_junk;
+
+                    for (int i = 0; i < n_junk_cur; ++i) {
                         const int r = rand() % k_questions.size();
                         client.prompt += "User:\n" + k_questions[r] + "\nAssistant:\n " + k_answers[r] + "\n";
                     }
@@ -340,7 +345,7 @@ int main(int argc, char ** argv) {
                     client.n_decoded = 0;
                     client.i_batch   = batch.n_tokens - 1;
 
-                    LOG_INF("\033[31mClient %3d, seq %4d, started decoding ...\033[0m\n", client.id, client.seq_id);
+                    LOG_INF("\033[31mClient %3d, seq %4d, junk = %4d, started decoding ...\033[0m\n", client.id, client.seq_id, n_junk_cur);
 
                     g_seq_id += 1;
 
@@ -359,7 +364,9 @@ int main(int argc, char ** argv) {
         // process in chunks of params.n_batch
         int32_t n_batch = params.n_batch;
 
-        for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
+        int32_t i_next = 0;
+
+        for (int32_t i = 0; i < batch.n_tokens; i = i_next) {
             // experiment: process in powers of 2
             //if (i + n_batch > (int32_t) batch.n_tokens && n_batch > 32) {
             //    n_batch /= 2;
@@ -367,7 +374,7 @@ int main(int argc, char ** argv) {
             //    continue;
             //}
 
-            const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
+            const int32_t n_tokens = std::min(n_batch, batch.n_tokens - i);
 
             llama_batch batch_view = {
                 n_tokens,
@@ -387,19 +394,24 @@ int main(int argc, char ** argv) {
                     return 1;
                 }
 
-                LOG_ERR("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
+                LOG_WRN("%s : failed to decode the batch, retrying with n_batch = %d\n", __func__, n_batch / 2);
 
                 n_cache_miss += 1;
 
                 // retry with half the batch size to try to find a free slot in the KV cache
                 n_batch /= 2;
-                i -= n_batch;
 
                 continue;
             }
 
             LOG_DBG("%s : decoded batch of %d tokens\n", __func__, n_tokens);
 
+            // move the head of the batch forward with the number of tokens we just processed
+            i_next = i + n_tokens;
+
+            // on successful decode, restore the original batch size
+            n_batch = params.n_batch;
+
             for (auto & client : clients) {
                 if (client.i_batch < (int) i || client.i_batch >= (int) (i + n_tokens)) {
                     continue;
@@ -437,8 +449,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_self_seq_rm(ctx,    client.id + 1, -1, -1);
-                    llama_kv_self_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_memory_seq_rm(mem,    client.id + 1, -1, -1);
+                    llama_memory_seq_cp(mem, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,6 +126,8 @@ int main(int argc, char ** argv) {
 
     int n_past = 0;
 
+    auto * mem = llama_get_memory(ctx);
+
     // fill the KV cache
     for (int i = 0; i < n_ctx; i += n_batch) {
         if (i > 0 && n_grp > 1) {
@@ -133,11 +135,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_self_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_self_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_self_update  (ctx);
+            llama_memory_seq_add(mem, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_memory_seq_div(mem, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
 
         common_batch_clear(batch);
@@ -167,12 +168,10 @@ int main(int argc, char ** argv) {
 
         LOG_INF("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_self_defrag (ctx);
-        llama_kv_self_update (ctx);
+        llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+        llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+        n_past = llama_memory_seq_pos_max(mem, 0) + 1;
 
         common_batch_clear(batch);
 
@@ -198,12 +197,10 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_INF("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_self_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_self_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_self_defrag (ctx);
-            llama_kv_self_update (ctx);
+            llama_memory_seq_rm (mem, 0, n_keep            , n_keep + n_discard);
+            llama_memory_seq_add(mem, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_self_seq_pos_max(ctx, 0) + 1;
+            n_past = llama_memory_seq_pos_max(mem, 0) + 1;
         }
     }
 

examples/retrieval/retrieval.cpp

@@ -83,7 +83,7 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 
 static void batch_process(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
     // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_self_clear(ctx);
+    llama_memory_clear(llama_get_memory(ctx), false);
 
     // run model
     LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);

examples/save-load-state/save-load-state.cpp

@@ -196,7 +196,7 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
 
         // erase whole kv
-        llama_kv_self_clear(ctx3);
+        llama_memory_clear(llama_get_memory(ctx3), true);
         fprintf(stderr, "%s : kv cache cleared\n", __func__);
 
         // restore kv into seq 1

examples/simple-chat/simple-chat.cpp

@@ -98,7 +98,7 @@ int main(int argc, char ** argv) {
     auto generate = [&](const std::string & prompt) {
         std::string response;
 
-        const bool is_first = llama_kv_self_seq_pos_max(ctx, 0) == 0;
+        const bool is_first = llama_memory_seq_pos_max(llama_get_memory(ctx), 0) == 0;
 
         // tokenize the prompt
         const int n_prompt_tokens = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, is_first, true);
@@ -113,7 +113,7 @@ int main(int argc, char ** argv) {
         while (true) {
             // check if we have enough space in the context to evaluate this batch
             int n_ctx = llama_n_ctx(ctx);
-            int n_ctx_used = llama_kv_self_seq_pos_max(ctx, 0);
+            int n_ctx_used = llama_memory_seq_pos_max(llama_get_memory(ctx), 0);
             if (n_ctx_used + batch.n_tokens > n_ctx) {
                 printf("\033[0m\n");
                 fprintf(stderr, "context size exceeded\n");

examples/speculative-simple/speculative-simple.cpp

@@ -217,7 +217,7 @@ int main(int argc, char ** argv) {
         {
             LOG_DBG("clear kv cache from any extra tokens, n_past = %d\n", n_past);
 
-            llama_kv_self_seq_rm(ctx_tgt, 0, n_past, -1);
+            llama_memory_seq_rm(llama_get_memory(ctx_tgt), 0, n_past, -1);
         }
 
         if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {

examples/speculative/speculative.cpp

@@ -142,6 +142,8 @@ int main(int argc, char ** argv) {
         }
     }
 
+    auto * mem_tgt = llama_get_memory(ctx_tgt);
+    auto * mem_dft = llama_get_memory(ctx_dft);
 
     // Tokenize the prompt
     std::vector<llama_token> inp;
@@ -420,14 +422,14 @@ int main(int argc, char ** argv) {
             {
                 LOG_DBG("keeping sequence %d, n_past_tgt = %d, n_past_dft = %d\n", s_keep, n_past_tgt, n_past_dft);
 
-                llama_kv_self_seq_keep(ctx_dft, s_keep);
-                llama_kv_self_seq_cp  (ctx_dft, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_dft, 0);
+                llama_memory_seq_keep(mem_dft, s_keep);
+                llama_memory_seq_cp  (mem_dft, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_dft, 0);
 
-                llama_kv_self_seq_rm  (ctx_tgt, s_keep, n_past_tgt, -1);
-                llama_kv_self_seq_keep(ctx_tgt, s_keep);
-                llama_kv_self_seq_cp  (ctx_tgt, s_keep, 0, -1, -1);
-                llama_kv_self_seq_keep(ctx_tgt, 0);
+                llama_memory_seq_rm  (mem_tgt, s_keep, n_past_tgt, -1);
+                llama_memory_seq_keep(mem_tgt, s_keep);
+                llama_memory_seq_cp  (mem_tgt, s_keep, 0, -1, -1);
+                llama_memory_seq_keep(mem_tgt, 0);
             }
 
             for (int s = 0; s < n_seq_dft; ++s) {
@@ -444,7 +446,7 @@ int main(int argc, char ** argv) {
             common_batch_clear(batch_dft);
             common_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
 
-            llama_kv_self_seq_rm(ctx_dft, 0, n_past_dft, -1);
+            llama_memory_seq_rm(mem_dft, 0, n_past_dft, -1);
             // LOG_DBG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
             llama_decode(ctx_dft, batch_dft);
 
@@ -503,8 +505,8 @@ int main(int argc, char ** argv) {
                     if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_draft_split) {
                         LOG_DBG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
-                        llama_kv_self_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
-                        llama_kv_self_seq_cp(ctx_dft, s, n_seq_cur, -1, -1);
+                        llama_memory_seq_rm(mem_dft,    n_seq_cur, -1, -1);
+                        llama_memory_seq_cp(mem_dft, s, n_seq_cur, -1, -1);
 
                         // all previous tokens from this branch are now also part of the new branch
                         for (int t = 0; t < batch_tgt.n_tokens; ++t) {
@@ -585,9 +587,9 @@ int main(int argc, char ** argv) {
 
         // evaluate the target model on the drafted tokens
         {
-            llama_kv_self_seq_keep(ctx_tgt, 0);
+            llama_memory_seq_keep(mem_tgt, 0);
             for (int s = 1; s < n_seq_dft; ++s) {
-                llama_kv_self_seq_cp(ctx_tgt, 0, s, -1, -1);
+                llama_memory_seq_cp(mem_tgt, 0, s, -1, -1);
             }
 
             // LOG_DBG("target batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_tgt, batch_tgt).c_str());

ggml/CMakeLists.txt

@@ -105,7 +105,7 @@ message(DEBUG "GGML_NATIVE_DEFAULT : ${GGML_NATIVE_DEFAULT}")
 message(DEBUG "INS_ENB             : ${INS_ENB}")
 
 option(GGML_CPU_HBM          "ggml: use memkind for CPU HBM" OFF)
-option(GGML_CPU_AARCH64      "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
+option(GGML_CPU_REPACK       "ggml: use runtime weight conversion of Q4_0 to Q4_X_X" ON)
 option(GGML_CPU_KLEIDIAI     "ggml: use KleidiAI optimized kernels if applicable" OFF)
 option(GGML_SSE42            "ggml: enable SSE 4.2"          ${INS_ENB})
 option(GGML_AVX              "ggml: enable AVX"              ${INS_ENB})
@@ -137,7 +137,7 @@ set(GGML_CPU_ARM_ARCH        "" CACHE STRING "ggml: CPU architecture for ARM")
 set(GGML_CPU_POWERPC_CPUTYPE "" CACHE STRING "ggml: CPU type for PowerPC")
 
 
-if (WIN32)
+if (MINGW)
     set(GGML_WIN_VER "0x602" CACHE STRING   "ggml: Windows version")
 endif()
 
@@ -177,7 +177,6 @@ option(GGML_VULKAN_CHECK_RESULTS            "ggml: run Vulkan op checks"
 option(GGML_VULKAN_DEBUG                    "ggml: enable Vulkan debug output"                OFF)
 option(GGML_VULKAN_MEMORY_DEBUG             "ggml: enable Vulkan memory debug output"         OFF)
 option(GGML_VULKAN_SHADER_DEBUG_INFO        "ggml: enable Vulkan shader debug info"           OFF)
-option(GGML_VULKAN_PERF                     "ggml: enable Vulkan perf output"                 OFF)
 option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"                  OFF)
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)

ggml/cmake/common.cmake

@@ -24,3 +24,28 @@ function(ggml_get_flags CCID CCVER)
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
+
+function(ggml_get_system_arch)
+    if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
+        CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
+        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
+        set(GGML_SYSTEM_ARCH "ARM" PARENT_SCOPE)
+    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR
+            CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
+            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
+            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
+        set(GGML_SYSTEM_ARCH "x86" PARENT_SCOPE)
+    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR
+            "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+        set(GGML_SYSTEM_ARCH "PowerPC" PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+        set(GGML_SYSTEM_ARCH "loongarch64"  PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
+        set(GGML_SYSTEM_ARCH "riscv64" PARENT_SCOPE)
+    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+        set(GGML_SYSTEM_ARCH "s390x" PARENT_SCOPE)
+    else()
+        set(GGML_SYSTEM_ARCH "UNKNOWN" PARENT_SCOPE)
+    endif()
+endfunction()

ggml/include/ggml.h

@@ -2095,9 +2095,6 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_graph_get_grad    (const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
     GGML_API struct ggml_tensor * ggml_graph_get_grad_acc(const struct ggml_cgraph * cgraph, const struct ggml_tensor * node);
 
-    GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
-    GGML_API struct ggml_cgraph * ggml_graph_import(const char * fname, struct ggml_context ** ctx_data, struct ggml_context ** ctx_eval);
-
     // print info and performance information for the graph
     GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);
 
@@ -2181,6 +2178,7 @@ extern "C" {
 
     // scheduling priorities
     enum ggml_sched_priority {
+        GGML_SCHED_PRIO_LOW = -1,
         GGML_SCHED_PRIO_NORMAL,
         GGML_SCHED_PRIO_MEDIUM,
         GGML_SCHED_PRIO_HIGH,

ggml/src/CMakeLists.txt

@@ -109,6 +109,8 @@ if (MSVC)
 else ()
     set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
+ggml_get_system_arch()
+message(STATUS "GGML_SYSTEM_ARCH: ${GGML_SYSTEM_ARCH}")
 
 if (NOT MSVC)
     if (GGML_STATIC)
@@ -123,7 +125,6 @@ if (NOT MSVC)
 endif()
 
 if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
     add_compile_definitions(_WIN32_WINNT=${GGML_WIN_VER})
 endif()
 
@@ -194,6 +195,7 @@ add_library(ggml-base
             ../include/ggml-opt.h
             ../include/gguf.h
             ggml.c
+            ggml.cpp
             ggml-alloc.c
             ggml-backend.cpp
             ggml-opt.cpp
@@ -210,6 +212,7 @@ endif()
 
 add_library(ggml
             ggml-backend-reg.cpp)
+add_library(ggml::ggml ALIAS ggml)
 
 target_link_libraries(ggml PUBLIC ggml-base)
 
@@ -224,6 +227,7 @@ function(ggml_add_backend_library backend)
         set_target_properties(${backend} PROPERTIES LIBRARY_OUTPUT_DIRECTORY ${CMAKE_RUNTIME_OUTPUT_DIRECTORY})
         target_compile_definitions(${backend} PRIVATE GGML_BACKEND_DL)
         add_dependencies(ggml ${backend})
+        install(TARGETS ${backend} LIBRARY DESTINATION ${CMAKE_INSTALL_BINDIR})
     else()
         add_library(${backend} ${ARGN})
         target_link_libraries(ggml PUBLIC ${backend})
@@ -287,16 +291,20 @@ if (GGML_CPU_ALL_VARIANTS)
     if (NOT GGML_BACKEND_DL)
         message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS requires GGML_BACKEND_DL")
     endif()
-    ggml_add_cpu_backend_variant(x64)
-    ggml_add_cpu_backend_variant(sse42        SSE42)
-    ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
-    ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
-    ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
-    ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
-    ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
-    if (NOT MSVC)
-        # MSVC doesn't support AMX
-        ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+    if (GGML_SYSTEM_ARCH STREQUAL "x86")
+        ggml_add_cpu_backend_variant(x64)
+        ggml_add_cpu_backend_variant(sse42        SSE42)
+        ggml_add_cpu_backend_variant(sandybridge  SSE42 AVX)
+        ggml_add_cpu_backend_variant(haswell      SSE42 AVX F16C AVX2 BMI2 FMA)
+        ggml_add_cpu_backend_variant(skylakex     SSE42 AVX F16C AVX2 BMI2 FMA AVX512)
+        ggml_add_cpu_backend_variant(icelake      SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI)
+        ggml_add_cpu_backend_variant(alderlake    SSE42 AVX F16C AVX2 BMI2 FMA AVX_VNNI)
+        if (NOT MSVC)
+            # MSVC doesn't support AMX
+            ggml_add_cpu_backend_variant(sapphirerapids SSE42 AVX F16C AVX2 BMI2 FMA AVX512 AVX512_VBMI AVX512_VNNI AVX512_BF16 AMX_TILE AMX_INT8)
+        endif()
+    else()
+        message(FATAL_ERROR "GGML_CPU_ALL_VARIANTS not yet supported on ${GGML_SYSTEM_ARCH}")
     endif()
 elseif (GGML_CPU)
     ggml_add_cpu_backend_variant_impl("")

ggml/src/ggml-backend.cpp

@@ -1340,7 +1340,10 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     // allocate graph
     if (backend_ids_changed || !ggml_gallocr_alloc_graph(sched->galloc, &sched->graph)) {
         // the re-allocation may cause the split inputs to be moved to a different address
-        ggml_backend_sched_synchronize(sched);
+        // synchronize without ggml_backend_sched_synchronize to avoid changing cur_copy
+        for (int i = 0; i < sched->n_backends; i++) {
+            ggml_backend_synchronize(sched->backends[i]);
+        }
 #ifndef NDEBUG
         GGML_LOG_DEBUG("%s: failed to allocate graph, reserving (backend_ids_changed = %d)\n", __func__, backend_ids_changed);
 #endif
@@ -1564,7 +1567,6 @@ bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgra
 
     ggml_backend_sched_split_graph(sched, graph);
 
-
     if (!ggml_backend_sched_alloc_splits(sched)) {
         return false;
     }
@@ -1598,9 +1600,12 @@ void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
     for (int i = 0; i < sched->n_backends; i++) {
         ggml_backend_synchronize(sched->backends[i]);
     }
-    // reset the current copy to 0 so that the graphs will be similar during generation
-    // necessary for CUDA graphs
-    sched->cur_copy = 0;
+    if (!sched->is_alloc) {
+        // if the graph is not already allocated, always use copy 0 after a synchronization
+        // this ensures that during generation the same copy is used every time,
+        // which avoids changes in the graph that could cause CUDA or other graphs to be disabled
+        sched->cur_copy = 0;
+    }
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {

ggml/src/ggml-blas/CMakeLists.txt

@@ -81,7 +81,7 @@ if (BLAS_FOUND)
     target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
     target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
 else()
-    message(ERROR "BLAS not found, please refer to "
-                  "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
-                  " to set correct GGML_BLAS_VENDOR")
+    message(FATAL_ERROR "BLAS not found, please refer to "
+                        "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
+                        " to set correct GGML_BLAS_VENDOR")
 endif()

ggml/src/ggml-cann/CMakeLists.txt

@@ -30,6 +30,7 @@ string(TOLOWER ${SOC_TYPE} SOC_VERSION) # SOC_VERSION need lower
 string(REGEX MATCH "[0-9]+[a-zA-Z]" SOC_TYPE_MAJOR_SN "${SOC_VERSION}")
 set(SOC_TYPE_COMPILE_OPTION "ASCEND_${SOC_TYPE_MAJOR_SN}")
 string(TOUPPER ${SOC_TYPE_COMPILE_OPTION} SOC_TYPE_COMPILE_OPTION)
+message(STATUS "CANN: SOC_VERSION =  ${SOC_VERSION}")
 
 if (CANN_INSTALL_DIR)
     # Only Support Linux.

ggml/src/ggml-cann/common.h

@@ -37,6 +37,7 @@
 #include <thread>
 #include <unistd.h>
 #include <functional>
+#include <optional>
 
 #include "../include/ggml-cann.h"
 #include "../include/ggml.h"
@@ -103,6 +104,9 @@ const ggml_cann_device_info& ggml_cann_info();
 void ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
+std::optional<std::string> get_env(const std::string& name);
+bool parse_bool(const std::string& value);
+
 /**
  * @brief Abstract base class for memory pools used by CANN.
  */
@@ -354,7 +358,8 @@ struct ggml_backend_cann_context {
         : device(device), name("CANN" + std::to_string(device)), task_queue(1024, device) {
         ggml_cann_set_device(device);
         description = aclrtGetSocName();
-        async_mode = (getenv("GGML_CANN_ASYNC_MODE") != nullptr);
+
+        bool async_mode = parse_bool(get_env("GGML_CANN_ASYNC_MODE").value_or(""));
         GGML_LOG_INFO("%s: device %d async operator submission is %s\n", __func__,
             device, async_mode ? "ON" : "OFF");
     }

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

@@ -31,6 +31,8 @@
 #include <mutex>
 #include <queue>
 #include <chrono>
+#include <unordered_set>
+#include <optional>
 
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
@@ -93,6 +95,26 @@ int32_t ggml_cann_get_device() {
     return id;
 }
 
+/**
+ * @brief Get the value of the specified environment variable (name).
+ *        if not empty, return a std::string object
+ */
+std::optional<std::string> get_env(const std::string& name) {
+    const char* val = std::getenv(name.c_str());
+    if (!val) return std::nullopt;
+    std::string res = std::string(val);
+    std::transform(res.begin(), res.end(), res.begin(), ::tolower);
+    return res;
+}
+
+/**
+ * @brief Verify whether the environment variable is a valid value.
+ */
+bool parse_bool(const std::string& value) {
+    std::unordered_set<std::string> valid_values = {"on", "1", "yes", "y", "enable", "true"};
+    return valid_values.find(value) != valid_values.end();
+}
+
 /**
  * @brief Initialize the CANN device information.
  *
@@ -214,7 +236,7 @@ struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf_prio(int device) : device(device) {
-        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -410,7 +432,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf(int device) : device(device) {
-        disable_clean = getenv("GGML_CANN_DISABLE_BUF_POOL_CLEAN") != nullptr;
+        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -731,16 +753,18 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(
     int device) {
-    bool disable_vmm = (getenv("GGML_CANN_DISABLE_VMM_POOL") != nullptr);
-    if (!disable_vmm && ggml_cann_info().devices[device].vmm) {
-        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
-        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
-    }
-    bool enable_buf_prio = (getenv("GGML_CANN_ENABLE_BUF_PRIO_POOL") != nullptr);
-    if (enable_buf_prio) {
+    std::string mem_pool_type = get_env("GGML_CANN_MEM_POOL").value_or("");
+
+    if (mem_pool_type == "prio") {
         GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
         return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf_prio(device));
     }
+
+    if (ggml_cann_info().devices[device].vmm && mem_pool_type != "leg") {
+        GGML_LOG_INFO("%s: device %d use vmm pool\n", __func__, device);
+        return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_vmm(device));
+    }
+
     GGML_LOG_INFO("%s: device %d use buffer pool\n", __func__, device);
     return std::unique_ptr<ggml_cann_pool>(new ggml_cann_pool_buf(device));
 }

ggml/src/ggml-common.h

@@ -1074,6 +1074,10 @@ GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
     0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
 GGML_TABLE_END()
 
+GGML_TABLE_BEGIN(int8_t, kvalues_iq4nl, 16)
+    -127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113,
+GGML_TABLE_END()
+
 #define NGRID_IQ1S 2048
 #define IQ1S_DELTA 0.125f
 #define IQ1M_DELTA 0.125f

ggml/src/ggml-cpu/CMakeLists.txt

@@ -10,14 +10,14 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
     list (APPEND GGML_CPU_SOURCES
         ggml-cpu/ggml-cpu.c
         ggml-cpu/ggml-cpu.cpp
-        ggml-cpu/ggml-cpu-aarch64.cpp
-        ggml-cpu/ggml-cpu-aarch64.h
-        ggml-cpu/ggml-cpu-hbm.cpp
-        ggml-cpu/ggml-cpu-hbm.h
-        ggml-cpu/ggml-cpu-quants.c
-        ggml-cpu/ggml-cpu-quants.h
-        ggml-cpu/ggml-cpu-traits.cpp
-        ggml-cpu/ggml-cpu-traits.h
+        ggml-cpu/repack.cpp
+        ggml-cpu/repack.h
+        ggml-cpu/hbm.cpp
+        ggml-cpu/hbm.h
+        ggml-cpu/quants.c
+        ggml-cpu/quants.h
+        ggml-cpu/traits.cpp
+        ggml-cpu/traits.h
         ggml-cpu/amx/amx.cpp
         ggml-cpu/amx/amx.h
         ggml-cpu/amx/mmq.cpp
@@ -82,12 +82,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         target_link_libraries(${GGML_CPU_NAME} PUBLIC memkind)
     endif()
 
-    if (CMAKE_OSX_ARCHITECTURES      STREQUAL "arm64" OR
-        CMAKE_GENERATOR_PLATFORM_LWR STREQUAL "arm64" OR
-        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-            CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
-
+    if (GGML_SYSTEM_ARCH STREQUAL "ARM")
         message(STATUS "ARM detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/arm/quants.c
+            ggml-cpu/arch/arm/repack.cpp
+            )
 
         if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
             message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
@@ -170,11 +170,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 endforeach()
             endif()
         endif()
-    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
-
+    elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
         message(STATUS "x86 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/x86/quants.c
+            ggml-cpu/arch/x86/repack.cpp
+            )
 
         if (MSVC)
             # instruction set detection for MSVC only
@@ -311,15 +312,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             # Since multiple variants of the CPU backend may be included in the same
             # build, using set_source_files_properties() to set the arch flags is not possible
             set(GGML_CPU_FEATS_NAME ${GGML_CPU_NAME}-feats)
-            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/cpu-feats-x86.cpp)
+            add_library(${GGML_CPU_FEATS_NAME} OBJECT ggml-cpu/arch/x86/cpu-feats.cpp)
             target_include_directories(${GGML_CPU_FEATS_NAME} PRIVATE . .. ../include)
             target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE ${ARCH_DEFINITIONS})
             target_compile_definitions(${GGML_CPU_FEATS_NAME} PRIVATE GGML_BACKEND_DL GGML_BACKEND_BUILD GGML_BACKEND_SHARED)
             set_target_properties(${GGML_CPU_FEATS_NAME} PROPERTIES POSITION_INDEPENDENT_CODE ON)
             target_link_libraries(${GGML_CPU_NAME} PRIVATE ${GGML_CPU_FEATS_NAME})
         endif()
-    elseif ("${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "ppc64le " OR "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
         message(STATUS "PowerPC detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/powerpc/quants.c)
         if (GGML_NATIVE)
             if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
                 file(READ "/proc/cpuinfo" POWER10_M)
@@ -327,7 +329,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 execute_process(COMMAND bash -c "prtconf |grep 'Implementation' | head -n 1" OUTPUT_VARIABLE POWER10_M)
             endif()
 
-            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M}")
+            string(TOUPPER "${POWER10_M}" POWER10_M_UPPER)
+            string(REGEX MATCHALL "POWER *([0-9]+)" MATCHED_STRING "${POWER10_M_UPPER}")
             string(REGEX REPLACE "POWER *([0-9]+)" "\\1" EXTRACTED_NUMBER "${MATCHED_STRING}")
 
             if (EXTRACTED_NUMBER GREATER_EQUAL 10)
@@ -344,8 +347,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 list(APPEND ARCH_FLAGS -mcpu=${GGML_CPU_POWERPC_CPUTYPE})
             endif()
         endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "loongarch64")
         message(STATUS "loongarch64 detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/loongarch/quants.c)
 
         list(APPEND ARCH_FLAGS -march=loongarch64)
         if (GGML_LASX)
@@ -354,8 +358,12 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_LSX)
             list(APPEND ARCH_FLAGS -mlsx)
         endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
-        message(STATUS "RISC-V detected")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        message(STATUS "riscv64 detected")
+        list(APPEND GGML_CPU_SOURCES
+            ggml-cpu/arch/riscv/quants.c
+            ggml-cpu/arch/riscv/repack.cpp
+            )
         if (GGML_RVV)
             if (GGML_XTHEADVECTOR)
                 list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -365,8 +373,9 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                 list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
             endif()
         endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
         message(STATUS "s390x detected")
+        list(APPEND GGML_CPU_SOURCES ggml-cpu/arch/s390/quants.c)
         file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
         string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
 
@@ -390,12 +399,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
         if (GGML_VXE)
             list(APPEND ARCH_FLAGS -mvx -mzvector)
         endif()
+    elseif (CMAKE_SYSTEM_PROCESSOR MATCHES "wasm")
+        message(STATUS "Wasm detected")
+        list (APPEND GGML_CPU_SOURCES ggml-cpu/arch/wasm/quants.c)
     else()
-        message(STATUS "Unknown architecture")
+        message(WARNING "Unknown CPU architecture. Falling back to generic implementations.")
+        list(APPEND ARCH_FLAGS -DGGML_CPU_GENERIC)
     endif()
 
-    if (GGML_CPU_AARCH64)
-        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_AARCH64)
+    if (GGML_CPU_REPACK)
+        target_compile_definitions(${GGML_CPU_NAME} PRIVATE GGML_USE_CPU_REPACK)
     endif()
 
     if (GGML_CPU_KLEIDIAI)

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

@@ -5,7 +5,7 @@
 #include "ggml-backend.h"
 #include "ggml-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #if defined(__gnu_linux__)
 #include <sys/syscall.h>

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

@@ -8,7 +8,7 @@
 #include "mmq.h"
 #include "ggml-impl.h"
 #include "ggml-cpu-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-quants.h"
 #include <algorithm>
 #include <type_traits>

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

@@ -0,0 +1,396 @@
+#define GGML_COMMON_IMPL_CPP
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+#include "ggml-backend-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "ggml-cpu-impl.h"
+#include "traits.h"
+
+#include <cmath>
+#include <cstring>
+#include <cassert>
+#include <cstdlib> // for qsort
+#include <cstdio>  // for GGML_ASSERT
+
+#define GGML_CPU_CLANG_WORKAROUND
+#include "../../repack.h"
+
+#if defined(__GNUC__)
+#pragma GCC diagnostic ignored "-Woverlength-strings"
+#endif
+
+#define UNUSED GGML_UNUSED
+
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            vfloat32m1_t sumf = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+            for (int l = 0; l < nb; l++) {
+                const int64_t a0 = *(const int64_t *)&a_ptr[l].qs[0];
+                const int64_t a1 = *(const int64_t *)&a_ptr[l].qs[8];
+                const int64_t a2 = *(const int64_t *)&a_ptr[l].qs[16];
+                const int64_t a3 = *(const int64_t *)&a_ptr[l].qs[24];
+                __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment constraints
+                const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a0, vl / 4));
+                const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a1, vl / 4));
+                const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a2, vl / 4));
+                const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(a3, vl / 4));
+
+                const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_hi_m));
+                const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                // vector version needs Zvfhmin extension
+                const float a_scale = GGML_FP16_TO_FP32(a_ptr[l].d);
+                const float b_scales[8] = {
+                    GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                    GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                };
+                const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+                const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scale, vl / 4);
+                sumf = __riscv_vfmacc_vv_f32m1(sumf, tmp1, b_scales_vec, vl / 4);
+            }
+            __riscv_vse32_v_f32m1(s + x * ncols_interleaved, sumf, vl / 4);
+        }
+        return;
+    }
+
+#endif
+    {
+        float sumf[8];
+        int sumi;
+
+        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+
+            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int j = 0; j < ncols_interleaved; j++) {
+                        sumi = 0;
+                        for (int i = 0; i < blocklen; ++i) {
+                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
+                        }
+                        sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
+                    }
+                }
+            }
+            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+        }
+    }
+}
+
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert (n % qk == 0);
+    assert (nr % 4 == 0);
+    assert (nc % ncols_interleaved == 0);
+
+    UNUSED(s);
+    UNUSED(bs);
+    UNUSED(vx);
+    UNUSED(vy);
+    UNUSED(nr);
+    UNUSED(nc);
+    UNUSED(nb);
+    UNUSED(ncols_interleaved);
+    UNUSED(blocklen);
+
+#if defined __riscv_v
+    if (__riscv_vlenb() >= QK4_0) {
+        const size_t vl = QK4_0;
+
+        for (int y = 0; y < nr / 4; y++) {
+            const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+            for (int x = 0; x < nc / ncols_interleaved; x++) {
+                const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+                vfloat32m1_t sumf0 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf1 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf2 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                vfloat32m1_t sumf3 = __riscv_vfmv_v_f_f32m1(0.0, vl / 4);
+                for (int l = 0; l < nb; l++) {
+                    const vint8m4_t rhs_raw_vec = __riscv_vle8_v_i8m4((const int8_t *)b_ptr[l].qs, vl * 4);
+                    const vint8m4_t rhs_vec_lo = __riscv_vsra_vx_i8m4(__riscv_vsll_vx_i8m4(rhs_raw_vec, 4, vl * 4), 4, vl * 4);
+                    const vint8m4_t rhs_vec_hi = __riscv_vsra_vx_i8m4(rhs_raw_vec, 4, vl * 4);
+                    const vint8m2_t rhs_vec_lo_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 0);
+                    const vint8m2_t rhs_vec_lo_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_lo, 1);
+                    const vint8m2_t rhs_vec_hi_0 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 0);
+                    const vint8m2_t rhs_vec_hi_1 = __riscv_vget_v_i8m4_i8m2(rhs_vec_hi, 1);
+
+                    // vector version needs Zvfhmin extension
+                    const float a_scales[4] = {
+                        GGML_FP16_TO_FP32(a_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(a_ptr[l].d[3])
+                    };
+                    const float b_scales[8] = {
+                        GGML_FP16_TO_FP32(b_ptr[l].d[0]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[1]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[2]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[3]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[4]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[5]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[6]),
+                        GGML_FP16_TO_FP32(b_ptr[l].d[7])
+                    };
+                    const vfloat32m1_t b_scales_vec = __riscv_vle32_v_f32m1(b_scales, vl / 4);
+
+                    const int64_t A0 = *(const int64_t *)&a_ptr[l].qs[0];
+                    const int64_t A4 = *(const int64_t *)&a_ptr[l].qs[32];
+                    const int64_t A8 = *(const int64_t *)&a_ptr[l].qs[64];
+                    const int64_t Ac = *(const int64_t *)&a_ptr[l].qs[96];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l0;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A0, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A4, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A8, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ac, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l0 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l0));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[0], vl / 4);
+                        sumf0 = __riscv_vfmacc_vv_f32m1(sumf0, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A1 = *(const int64_t *)&a_ptr[l].qs[8];
+                    const int64_t A5 = *(const int64_t *)&a_ptr[l].qs[40];
+                    const int64_t A9 = *(const int64_t *)&a_ptr[l].qs[72];
+                    const int64_t Ad = *(const int64_t *)&a_ptr[l].qs[104];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l1;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A1, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A5, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A9, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ad, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l1 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l1));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[1], vl / 4);
+                        sumf1 = __riscv_vfmacc_vv_f32m1(sumf1, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A2 = *(const int64_t *)&a_ptr[l].qs[16];
+                    const int64_t A6 = *(const int64_t *)&a_ptr[l].qs[48];
+                    const int64_t Aa = *(const int64_t *)&a_ptr[l].qs[80];
+                    const int64_t Ae = *(const int64_t *)&a_ptr[l].qs[112];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l2;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A2, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A6, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Aa, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ae, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l2 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l2));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[2], vl / 4);
+                        sumf2 = __riscv_vfmacc_vv_f32m1(sumf2, tmp1, b_scales_vec, vl / 4);
+                    }
+
+                    const int64_t A3 = *(const int64_t *)&a_ptr[l].qs[24];
+                    const int64_t A7 = *(const int64_t *)&a_ptr[l].qs[56];
+                    const int64_t Ab = *(const int64_t *)&a_ptr[l].qs[88];
+                    const int64_t Af = *(const int64_t *)&a_ptr[l].qs[120];
+                    __asm__ __volatile__("" ::: "memory"); // prevent gcc from emitting fused vlse64, violating alignment
+                    vint16m4_t sumi_l3;
+                    {
+                        const vint8m2_t lhs_0_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A3, vl / 4));
+                        const vint8m2_t lhs_1_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(A7, vl / 4));
+                        const vint8m2_t lhs_2_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Ab, vl / 4));
+                        const vint8m2_t lhs_3_8 =__riscv_vreinterpret_v_i64m2_i8m2(__riscv_vmv_v_x_i64m2(Af, vl / 4));
+                        const vint16m4_t sumi_lo_0 = __riscv_vwmul_vv_i16m4(rhs_vec_lo_0, lhs_0_8, vl * 2);
+                        const vint16m4_t sumi_lo_1 = __riscv_vwmacc_vv_i16m4(sumi_lo_0, rhs_vec_lo_1, lhs_1_8, vl * 2);
+                        const vint16m4_t sumi_hi_0 = __riscv_vwmacc_vv_i16m4(sumi_lo_1, rhs_vec_hi_0, lhs_2_8, vl * 2);
+                        const vint16m4_t sumi_hi_m = __riscv_vwmacc_vv_i16m4(sumi_hi_0, rhs_vec_hi_1, lhs_3_8, vl * 2);
+
+                        sumi_l3 = sumi_hi_m;
+                    }
+
+                    {
+                        const vuint32m4_t sumi_i32 = __riscv_vreinterpret_v_i32m4_u32m4(__riscv_vreinterpret_v_i16m4_i32m4(sumi_l3));
+                        const vuint16m2_t sumi_h2_0 = __riscv_vnsrl_wx_u16m2(sumi_i32, 0, vl);
+                        const vuint16m2_t sumi_h2_1 = __riscv_vnsrl_wx_u16m2(sumi_i32, 16, vl);
+                        const vuint16m2_t sumi_h2 = __riscv_vadd_vv_u16m2(sumi_h2_0, sumi_h2_1, vl);
+                        const vuint32m2_t sumi_h2_i32 = __riscv_vreinterpret_v_u16m2_u32m2(sumi_h2);
+                        const vuint16m1_t sumi_h4_0 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 0, vl / 2);
+                        const vuint16m1_t sumi_h4_1 = __riscv_vnsrl_wx_u16m1(sumi_h2_i32, 16, vl / 2);
+                        const vuint16m1_t sumi_h4 = __riscv_vadd_vv_u16m1(sumi_h4_0, sumi_h4_1, vl / 2);
+                        const vuint32m1_t sumi_h4_i32 = __riscv_vreinterpret_v_u16m1_u32m1(sumi_h4);
+                        const vint16mf2_t sumi_h8_0 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 0, vl / 4));
+                        const vint16mf2_t sumi_h8_1 = __riscv_vreinterpret_v_u16mf2_i16mf2(__riscv_vnsrl_wx_u16mf2(sumi_h4_i32, 16, vl / 4));
+                        const vint32m1_t sumi_h8 = __riscv_vwadd_vv_i32m1(sumi_h8_0, sumi_h8_1, vl / 4);
+                        const vfloat32m1_t facc = __riscv_vfcvt_f_x_v_f32m1(sumi_h8, vl / 4);
+
+                        const vfloat32m1_t tmp1 = __riscv_vfmul_vf_f32m1(facc, a_scales[3], vl / 4);
+                        sumf3 = __riscv_vfmacc_vv_f32m1(sumf3, tmp1, b_scales_vec, vl / 4);
+                    }
+                }
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 0) * bs + x * ncols_interleaved], sumf0, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 1) * bs + x * ncols_interleaved], sumf1, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 2) * bs + x * ncols_interleaved], sumf2, vl / 4);
+                __riscv_vse32_v_f32m1(&s[(y * 4 + 3) * bs + x * ncols_interleaved], sumf3, vl / 4);
+            }
+        }
+
+        return;
+    }
+
+#endif // #if ! ((defined(_MSC_VER)) && ! defined(__clang__)) && defined(__aarch64__)
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                                const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
+                            }
+                            sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}

ggml/src/ggml-cpu/common.h

@@ -1,7 +1,7 @@
 #pragma once
 
 #include "ggml.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-impl.h"
 

ggml/src/ggml-cpu/ggml-cpu-aarch64.h

@@ -1,8 +0,0 @@
-#pragma once
-
-#include "ggml-cpu-traits.h"
-#include "ggml.h"
-
-// GGML internal header
-
-ggml_backend_buffer_type_t ggml_backend_cpu_aarch64_buffer_type(void);

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

@@ -506,3 +506,28 @@ void ggml_barrier(struct ggml_threadpool * tp);
 #ifdef __cplusplus
 }
 #endif
+
+#define GGML_DO_PRAGMA_(x) _Pragma (#x)
+#define GGML_DO_PRAGMA(x) GGML_DO_PRAGMA_(x)
+#if defined(GGML_CPU_GENERIC) || defined(__HIPCC__)
+// Note for Apple targets:
+// - clang: aliases are not supported on darwin
+// - all native kernels need to be implemented in both x86 and arm files
+// - on iOS, tvOS, and visionOS, if cmake cannot determine the target architecture, all `_generic` names are replaced by defines
+# define GGML_WEAK_ALIAS(name, alias)
+#elif defined(__GNUC__)
+// GCC/Clang on *nix
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(weak name = alias) // NOLINT
+#elif defined(_MSC_VER) && defined(_WIN64)
+// MSVC
+// Note: C name mangling varies across different calling conventions
+// see https://learn.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:" #name "=" #alias))
+#elif defined(_MSC_VER) && defined(WIN32)
+// ref: https://github.com/ggml-org/whisper.cpp/pull/3239#issuecomment-2958224591
+# define GGML_WEAK_ALIAS(name, alias) GGML_DO_PRAGMA(comment(linker, "/alternatename:_" #name "=_" #alias))
+#else
+# error "Unsupported compiler for GGML_WEAK_ALIAS"
+#endif
+
+#define GGML_CPU_NATIVE_IMPL(name) GGML_WEAK_ALIAS(name, name ## _generic)

ggml/src/ggml-cpu/ggml-cpu-quants.h

@@ -1,63 +0,0 @@
-#pragma once
-
-#define GGML_COMMON_DECL_C
-#include "ggml-common.h"
-
-#include "ggml.h"
-
-// GGML CPU internal header
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-// Quantization
-void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
-
-// Dot product
-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);
-void ggml_vec_dot_q4_1_q8_1(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);
-void ggml_vec_dot_q5_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);
-void ggml_vec_dot_q5_1_q8_1(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);
-void ggml_vec_dot_q8_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);
-
-void ggml_vec_dot_q2_K_q8_K(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);
-void ggml_vec_dot_q3_K_q8_K(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);
-void ggml_vec_dot_q4_K_q8_K(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);
-void ggml_vec_dot_q5_K_q8_K(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);
-void ggml_vec_dot_q6_K_q8_K(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);
-
-void ggml_vec_dot_tq1_0_q8_K(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);
-void ggml_vec_dot_tq2_0_q8_K(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);
-
-void ggml_vec_dot_iq2_xxs_q8_K(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);
-void ggml_vec_dot_iq2_xs_q8_K (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);
-void ggml_vec_dot_iq2_s_q8_K  (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);
-void ggml_vec_dot_iq3_xxs_q8_K(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);
-void ggml_vec_dot_iq1_s_q8_K  (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);
-void ggml_vec_dot_iq1_m_q8_K  (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);
-void ggml_vec_dot_iq4_nl_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);
-void ggml_vec_dot_iq4_xs_q8_K (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);
-void ggml_vec_dot_iq3_s_q8_K  (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);
-
-#ifdef __cplusplus
-}
-#endif

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

@@ -3,11 +3,11 @@
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-quants.h"
+#include "quants.h"
 #include "ggml-threading.h"
 #include "unary-ops.h"
 #include "binary-ops.h"
@@ -270,7 +270,11 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .from_float               = quantize_row_q4_K,
         .vec_dot                  = ggml_vec_dot_q4_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
+#if defined (__ARM_FEATURE_MATMUL_INT8)
+        .nrows                    = 2,
+#else
         .nrows                    = 1,
+#endif
     },
     [GGML_TYPE_Q5_K] = {
         .from_float               = quantize_row_q5_K,
@@ -2414,12 +2418,32 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     // This is up to the applications.
     DWORD p = THREAD_PRIORITY_NORMAL;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      p = THREAD_PRIORITY_BELOW_NORMAL;  break;
         case GGML_SCHED_PRIO_NORMAL:   p = THREAD_PRIORITY_NORMAL;        break;
         case GGML_SCHED_PRIO_MEDIUM:   p = THREAD_PRIORITY_ABOVE_NORMAL;  break;
         case GGML_SCHED_PRIO_HIGH:     p = THREAD_PRIORITY_HIGHEST;       break;
         case GGML_SCHED_PRIO_REALTIME: p = THREAD_PRIORITY_TIME_CRITICAL; break;
     }
 
+    if (prio != GGML_SCHED_PRIO_LOW) {
+        // Tell Windows that this thread should not be throttled (needs its own CPU core).
+        // Newer Windows 11 versions aggresively park (offline) CPU cores and often place
+        // all our threads onto the first 4 cores which results in terrible performance with
+        // n_threads > 4
+        #if _WIN32_WINNT >= 0x0602
+        THREAD_POWER_THROTTLING_STATE t;
+        ZeroMemory(&t, sizeof(t));
+        t.Version     = THREAD_POWER_THROTTLING_CURRENT_VERSION;
+        t.ControlMask = THREAD_POWER_THROTTLING_EXECUTION_SPEED;
+        t.StateMask   = 0;
+
+        if (!SetThreadInformation(GetCurrentThread(), ThreadPowerThrottling, &t, sizeof(t))) {
+            GGML_LOG_DEBUG("failed to disable thread power throttling %d : (%d)\n", prio, (int) GetLastError());
+            return false;
+        }
+        #endif
+    }
+
     if (prio == GGML_SCHED_PRIO_NORMAL) {
         // Keep inherited policy/priority
         return true;
@@ -2447,6 +2471,8 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        // TODO: there seems to be no way to set lower prio on Apple platforms
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;
@@ -2503,6 +2529,7 @@ static bool ggml_thread_apply_priority(int32_t prio) {
     struct sched_param p;
     int32_t policy = SCHED_OTHER;
     switch (prio) {
+        case GGML_SCHED_PRIO_LOW:      policy = SCHED_BATCH; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_NORMAL:   policy = SCHED_OTHER; p.sched_priority = 0;  break;
         case GGML_SCHED_PRIO_MEDIUM:   policy = SCHED_FIFO;  p.sched_priority = 40; break;
         case GGML_SCHED_PRIO_HIGH:     policy = SCHED_FIFO;  p.sched_priority = 80; break;

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

@@ -1,8 +1,8 @@
 #include "ggml-backend.h"
 #include "ggml-backend-impl.h"
 #include "ggml-cpu.h"
-#include "ggml-cpu-aarch64.h"
-#include "ggml-cpu-traits.h"
+#include "repack.h"
+#include "traits.h"
 #include "ggml-impl.h"
 #include "amx/amx.h"
 
@@ -11,7 +11,7 @@
 #include <vector>
 
 #ifdef GGML_USE_CPU_HBM
-#    include "ggml-cpu-hbm.h"
+#    include "hbm.h"
 #endif
 
 #ifdef GGML_USE_CPU_KLEIDIAI
@@ -51,9 +51,9 @@ std::vector<ggml_backend_buffer_type_t>& ggml_backend_cpu_get_extra_buffers_type
         }
 #endif
 
-#ifdef GGML_USE_CPU_AARCH64
-        if (ggml_backend_cpu_aarch64_buffer_type()) {
-            bufts.push_back(ggml_backend_cpu_aarch64_buffer_type());
+#ifdef GGML_USE_CPU_REPACK
+        if (ggml_backend_cpu_repack_buffer_type()) {
+            bufts.push_back(ggml_backend_cpu_repack_buffer_type());
         }
 #endif
 
@@ -596,8 +596,8 @@ static ggml_backend_feature * ggml_backend_cpu_get_features(ggml_backend_reg_t r
     #ifdef GGML_USE_CPU_KLEIDIAI
         features.push_back({ "KLEIDIAI", "1" });
     #endif
-    #ifdef GGML_USE_CPU_AARCH64
-        features.push_back({ "AARCH64_REPACK", "1" });
+    #ifdef GGML_USE_CPU_REPACK
+        features.push_back({ "REPACK", "1" });
     #endif
 
         features.push_back({ nullptr, nullptr });

ggml/src/ggml-cpu/hbm.cpp

@@ -5,7 +5,7 @@
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 
-#include "ggml-cpu-hbm.h"
+#include "hbm.h"
 
 // buffer type HBM
 

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

@@ -26,7 +26,7 @@
 #include "ggml-impl.h"
 #include "ggml-backend-impl.h"
 #include "ggml-threading.h"
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "kernels.h"
 

ggml/src/ggml-cpu/ops.cpp

@@ -7633,39 +7633,83 @@ static void ggml_compute_forward_ssm_scan_f32(
     const int ir1 = MIN(ir0 + dr, nr);
     const int ir  = ir1 - ir0;
 
-    for (int i3 = 0; i3 < n_s; ++i3) {
-        for (int i2 = 0; i2 < n_t; ++i2) {
-            const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
-            const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-            const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
-            const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
-            const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
-            const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
-                  float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
-                  float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
+    #ifdef __ARM_FEATURE_SVE
+        for (int i3 = 0; i3 < n_s; ++i3) {
+            for (int i2 = 0; i2 < n_t; ++i2) {
+                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
+                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
+                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
+                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
+                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
+                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
 
-            // use the output as the source for the next token-wise iterations
-            if (i2 > 0) { s0 = s; }
+                // use the output as the source for the next token-wise iterations
+                if (i2 > 0) { s0 = s; }
 
-            // d_inner
-            for (int i1 = 0; i1 < ir; ++i1) {
-                // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
-                float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
-                float x_dt = x[i1] * dt_soft_plus;
-                float sumf = 0.0f;
-                // d_state
-                for (int i0 = 0; i0 < nc; ++i0) {
-                    int i = i0 + i1*nc;
-                    // state = prev_state * dA + dB * x
-                    float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
-                    // y = rowwise_dotprod(state, C)
-                    sumf += state * C[i0];
-                    s[i] = state;
+                // d_inner
+                for (int i1 = 0; i1 < ir; ++i1) {
+                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
+                    float x_dt = x[i1] * dt_soft_plus;
+                    svfloat32_t vx_dt = GGML_F32_VEC_SET1(x_dt);
+                    svfloat32_t vdt_soft_plus = GGML_F32_VEC_SET1(dt_soft_plus);
+                    svfloat32_t r1_vector = GGML_F32_VEC_ZERO;
+
+                    for (int64_t k = 0; k < nc; k += svcntw()) {
+                        svfloat32_t vA = GGML_F32_VEC_LOAD(&A[i1*nc + k]);
+                        svfloat32_t vB = GGML_F32_VEC_LOAD(&B[k]);
+                        svfloat32_t vC = GGML_F32_VEC_LOAD(&C[k]);
+                        svfloat32_t vs0 = GGML_F32_VEC_LOAD(&s0[i1*nc + k]);
+
+                        svfloat32_t t1 = GGML_F32_VEC_MUL(vdt_soft_plus, vA);
+                        t1 = exp_ps_sve(svptrue_b32(), t1);
+                        svfloat32_t t2 = GGML_F32_VEC_MUL(vx_dt, vB);
+
+                        vs0 = GGML_F32_VEC_FMA(vs0, t1, t2);
+                        r1_vector = GGML_F32_VEC_ADD(GGML_F32_VEC_MUL(vs0, vC), r1_vector);
+
+                        GGML_F32_VEC_STORE(&s[i1*nc + k], vs0);
+                    }
+                    y[i1] = GGML_F32xt_REDUCE_ONE(r1_vector);
                 }
-                y[i1] = sumf;
             }
         }
-    }
+    #else
+        for (int i3 = 0; i3 < n_s; ++i3) {
+            for (int i2 = 0; i2 < n_t; ++i2) {
+                const float * s0 = (const float *) ((const char *) src0->data + ir0*(src0->nb[1]) + i3*(src0->nb[2])); // {d_state, d_inner, n_s}
+                const float * x  = (const float *) ((const char *) src1->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+                const float * dt = (const float *) ((const char *) src2->data + ir0*(src2->nb[0]) + i2*(src2->nb[1]) + i3*(src2->nb[2])); // {d_inner, n_t, n_s}
+                const float * A  = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
+                const float * B  = (const float *) ((const char *) src4->data +  i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
+                const float * C  = (const float *) ((const char *) src5->data +  i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
+                    float * y  = (      float *) ((      char *) dst->data  + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
+                    float * s  = (      float *) ((      char *) dst->data  + ir0*(src0->nb[1]) + i3*(src0->nb[2]) +     src1->nb[3]);  // {d_state, d_inner, n_s}
+
+                // use the output as the source for the next token-wise iterations
+                if (i2 > 0) { s0 = s; }
+
+                // d_inner
+                for (int i1 = 0; i1 < ir; ++i1) {
+                    // ref: https://github.com/state-spaces/mamba/blob/34076d664838588a3c97727b263478ab9f621a07/mamba_ssm/ops/triton/selective_state_update.py#L78
+                    float dt_soft_plus = dt[i1] <= 20.0f ? log1pf(expf(dt[i1])) : dt[i1];
+                    float x_dt = x[i1] * dt_soft_plus;
+                    float sumf = 0.0f;
+                    // d_state
+                    for (int i0 = 0; i0 < nc; ++i0) {
+                        int i = i0 + i1*nc;
+                        // state = prev_state * dA + dB * x
+                        float state = (s0[i] * expf(dt_soft_plus * A[i])) + (B[i0] * x_dt);
+                        // y = rowwise_dotprod(state, C)
+                        sumf += state * C[i0];
+                        s[i] = state;
+                    }
+                    y[i1] = sumf;
+                }
+            }
+        }
+    #endif
 }
 
 void ggml_compute_forward_ssm_scan(
@@ -8070,6 +8114,14 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
         #define GGML_F32X_MUL GGML_F32x16_MUL
         #define GGML_F32X_FMA GGML_F32x16_FMA
         #define WKV_VECTOR_SIZE 16
+    #elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+        #define GGML_F32X GGML_F32xt
+        #define GGML_F32X_SET1 GGML_F32xt_SET1
+        #define GGML_F32X_LOAD GGML_F32xt_LOAD
+        #define GGML_F32X_STORE GGML_F32xt_STORE
+        #define GGML_F32X_MUL GGML_F32xt_MUL
+        #define GGML_F32X_FMA GGML_F32xt_FMA
+        #define WKV_VECTOR_SIZE 8
     #elif defined(__ARM_NEON) && defined(__aarch64__)
         #define GGML_F32X GGML_F32x4
         #define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -8081,7 +8133,13 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
     #endif
 
     #ifdef WKV_VECTOR_SIZE
-        const int64_t vec_count = head_size / WKV_VECTOR_SIZE;
+        int wkv_vector_size;
+        #if defined(__ARM_FEATURE_SVE)
+            wkv_vector_size = svcntw();
+        #else
+            wkv_vector_size = WKV_VECTOR_SIZE;
+        #endif
+        const int64_t vec_count = head_size / wkv_vector_size;
 
         for (int64_t t = 0; t < T; t++) {
             size_t t_offset = t * t_stride;
@@ -8111,7 +8169,7 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
                     GGML_F32X time_decay_vec = GGML_F32X_SET1(time_decay_val);
 
                     for (int64_t j = 0; j < vec_count; j++) {
-                        size_t base_j = j * WKV_VECTOR_SIZE;
+                        size_t base_j = j * wkv_vector_size;
                         size_t t_h_j_offset = t_h_offset + base_j;
                         size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
 
@@ -8136,7 +8194,7 @@ static void ggml_compute_forward_rwkv_wkv6_f32(
                     }
 
                     // Handle remaining elements, this will not be used.
-                    for (int64_t j = vec_count * WKV_VECTOR_SIZE; j < head_size; j++) {
+                    for (int64_t j = vec_count * wkv_vector_size; j < head_size; j++) {
                         size_t t_h_j_offset = t_h_offset + j;
                         size_t h_2d_i_j_offset = h_2d_i_offset + j;
                         float v_val = v[t_h_j_offset];
@@ -8272,6 +8330,14 @@ static void ggml_compute_forward_gla_f32(
         #define GGML_F32X_MUL GGML_F32x16_MUL
         #define GGML_F32X_FMA GGML_F32x16_FMA
         #define GLA_VECTOR_SIZE 16
+    #elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+        #define GGML_F32X GGML_F32xt
+        #define GGML_F32X_SET1 GGML_F32xt_SET1
+        #define GGML_F32X_LOAD GGML_F32xt_LOAD
+        #define GGML_F32X_STORE GGML_F32xt_STORE
+        #define GGML_F32X_MUL GGML_F32xt_MUL
+        #define GGML_F32X_FMA GGML_F32xt_FMA
+        #define GLA_VECTOR_SIZE 8
     #elif defined(__ARM_NEON) && defined(__aarch64__)
         #define GGML_F32X GGML_F32x4
         #define GGML_F32X_SET1 GGML_F32x4_SET1
@@ -8283,7 +8349,13 @@ static void ggml_compute_forward_gla_f32(
     #endif
 
     #ifdef GLA_VECTOR_SIZE
-        const int64_t vec_count = head_size / GLA_VECTOR_SIZE;
+        int gla_vector_size;
+        #if defined(__ARM_FEATURE_SVE)
+            gla_vector_size = svcntw();
+        #else
+            gla_vector_size = GLA_VECTOR_SIZE;
+        #endif
+        const int64_t vec_count = head_size / gla_vector_size;
 
         for (int64_t t = 0; t < T; t++) {
             size_t t_offset = t * t_stride;
@@ -8310,7 +8382,7 @@ static void ggml_compute_forward_gla_f32(
                     GGML_F32X g_vec = GGML_F32X_SET1(g_val);
 
                     for (int64_t j = 0; j < vec_count; j++) {
-                        size_t base_j = j * GLA_VECTOR_SIZE;
+                        size_t base_j = j * gla_vector_size;
                         size_t t_h_j_offset = t_h_offset + base_j;
                         size_t h_2d_i_j_offset = h_2d_i_offset + base_j;
 
@@ -8334,7 +8406,7 @@ static void ggml_compute_forward_gla_f32(
                     }
 
                     // Handle remaining elements, this will not be used.
-                    for (int64_t j = vec_count * GLA_VECTOR_SIZE; j < head_size; j++) {
+                    for (int64_t j = vec_count * gla_vector_size; j < head_size; j++) {
                         size_t t_h_j_offset = t_h_offset + j;
                         size_t h_2d_i_j_offset = h_2d_i_offset + j;
                         float v_val = v[t_h_j_offset];
@@ -8443,83 +8515,126 @@ static void ggml_compute_forward_rwkv_wkv7_f32(
     int64_t h_stride_2d = head_size * head_size;
 
     #if defined(GGML_SIMD)
-        for (int64_t t = 0; t < T; t++) {
-            int64_t t_offset = t * t_stride;
-            int64_t state_offset = head_size * C * (t / (T / n_seqs));
-            float * state_cur = state + state_offset;
-            float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+        #if defined(__ARM_FEATURE_SVE)
+            // scalar Route to scalar implementation       //TODO: Write SVE code
+            for (int64_t t = 0; t < T; t++) {
+                int64_t t_offset = t * t_stride;
+                int64_t state_offset = head_size * C * (t / (T / n_seqs));
+                float * state_cur = state + state_offset;
+                float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
 
-            for (int64_t h = h_start; h < h_end; h++) {
-                int64_t h_offset = h * h_stride;
-                int64_t t_h_offset = t_offset + h_offset;
-                int64_t h_2d_offset = h * h_stride_2d;
+                for (int64_t h = h_start; h < h_end; h++) {
+                    int64_t h_offset = h * h_stride;
+                    int64_t t_h_offset = t_offset + h_offset;
+                    int64_t h_2d_offset = h * h_stride_2d;
 
-                for (int64_t ii = 0; ii < head_size; ii++) {
-                    int64_t t_h_i_offset = t_h_offset + ii;
-                    int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
+                    for (int64_t i = 0; i < head_size; i++) {
+                        int64_t t_h_i_offset = t_h_offset + i;
+                        int64_t h_2d_i_offset = h_2d_offset + i * h_stride;
 
-                    GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
+                        float v_val = v[t_h_i_offset];
 
-                    float sa = 0;
-                    {
-                        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
-                        GGML_F32_VEC ax[GGML_F32_ARR];
-                        GGML_F32_VEC ay[GGML_F32_ARR];
-                        for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
-                            for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
-                                ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
-                                ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
-                                sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
-                            }
+                        float sa = 0, result = 0;
+                        for (int64_t j = 0; j < head_size; j++) {
+                            sa += a[t_h_offset + j] * state_prev[h_2d_i_offset + j];
                         }
-                        GGML_F32_VEC_REDUCE(sa, sum);
-                    }
 
-                    GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
+                        for (int64_t j = 0; j < head_size; j++) {
+                            int64_t t_h_j_offset = t_h_offset + j;
+                            int64_t h_2d_i_j_offset = h_2d_i_offset + j;
 
-                    int64_t j = 0;
-                    GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
-                    for (; j < head_size; j += GGML_F32_STEP) {
-                        for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
-                            int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
-                            int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
-
-                            GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
-                            GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
-                            GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
-                            GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
-
-                            k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
-
-                            GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
-                            // kv + s * decay + sa * b
-                            state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
-                            state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
-                            GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
-
-                            result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
+                            float r_val = r[t_h_j_offset];
+                            float w_val = w[t_h_j_offset];
+                            float k_val = k[t_h_j_offset];
+                            float b_val = b[t_h_j_offset];
+                            float kv_val = v_val * k_val;
+                            float prev_state_val = state_prev[h_2d_i_j_offset];
+                            state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                            result += state_cur[h_2d_i_j_offset] * r_val;
                         }
-                    }
-                    GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
-
-                    // There shouldn't be left-overs though.
-                    for (; j < head_size; j++) {
-                        int64_t t_h_j_offset = t_h_offset + j;
-                        int64_t h_2d_i_j_offset = h_2d_i_offset + j;
-
-                        float r_val = r[t_h_j_offset];
-                        float w_val = w[t_h_j_offset];
-                        float k_val = k[t_h_j_offset];
-                        float b_val = b[t_h_j_offset];
-                        float kv_val = v[t_h_i_offset] * k_val;
-
-                        float prev_state_val = state_prev[h_2d_i_j_offset];
-                        state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
-                        dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                        dst_data[t_h_i_offset] = result;
                     }
                 }
             }
-        }
+        #else
+            for (int64_t t = 0; t < T; t++) {
+                int64_t t_offset = t * t_stride;
+                int64_t state_offset = head_size * C * (t / (T / n_seqs));
+                float * state_cur = state + state_offset;
+                float * state_prev = t % (T / n_seqs) ? state_cur : (float*)dst->src[6]->data + state_offset;
+
+                for (int64_t h = h_start; h < h_end; h++) {
+                    int64_t h_offset = h * h_stride;
+                    int64_t t_h_offset = t_offset + h_offset;
+                    int64_t h_2d_offset = h * h_stride_2d;
+
+                    for (int64_t ii = 0; ii < head_size; ii++) {
+                        int64_t t_h_i_offset = t_h_offset + ii;
+                        int64_t h_2d_i_offset = h_2d_offset + ii * h_stride;
+
+                        GGML_F32_VEC v_vec = GGML_F32_VEC_SET1(v[t_h_i_offset]);
+
+                        float sa = 0;
+                        {
+                            GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                            GGML_F32_VEC ax[GGML_F32_ARR];
+                            GGML_F32_VEC ay[GGML_F32_ARR];
+                            for (int64_t j = 0; j < head_size; j += GGML_F32_STEP) {
+                                for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                                    ax[kk] = GGML_F32_VEC_LOAD(&a[t_h_offset + j + kk * GGML_F32_EPR]);
+                                    ay[kk] = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_offset + j + kk * GGML_F32_EPR]);
+                                    sum[kk] = GGML_F32_VEC_FMA(sum[kk], ax[kk], ay[kk]);
+                                }
+                            }
+                            GGML_F32_VEC_REDUCE(sa, sum);
+                        }
+
+                        GGML_F32_VEC sa_vec = GGML_F32_VEC_SET1(sa);
+
+                        int64_t j = 0;
+                        GGML_F32_VEC result_vec[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+                        for (; j < head_size; j += GGML_F32_STEP) {
+                            for (int64_t kk = 0; kk < GGML_F32_ARR; kk++) {
+                                int64_t t_h_j_offset = t_h_offset + j + kk * GGML_F32_EPR;
+                                int64_t h_2d_i_j_offset = h_2d_i_offset + j + kk * GGML_F32_EPR;
+
+                                GGML_F32_VEC r_vec = GGML_F32_VEC_LOAD(&r[t_h_j_offset]);
+                                GGML_F32_VEC w_vec = GGML_F32_VEC_LOAD(&w[t_h_j_offset]);
+                                GGML_F32_VEC k_vec = GGML_F32_VEC_LOAD(&k[t_h_j_offset]);
+                                GGML_F32_VEC b_vec = GGML_F32_VEC_LOAD(&b[t_h_j_offset]);
+
+                                k_vec = GGML_F32_VEC_MUL(v_vec, k_vec);
+
+                                GGML_F32_VEC state_vec = GGML_F32_VEC_LOAD(&state_prev[h_2d_i_j_offset]);
+                                // kv + s * decay + sa * b
+                                state_vec = GGML_F32_VEC_FMA(k_vec, state_vec, w_vec);
+                                state_vec = GGML_F32_VEC_FMA(state_vec, sa_vec, b_vec);
+                                GGML_F32_VEC_STORE(&state_cur[h_2d_i_j_offset], state_vec);
+
+                                result_vec[kk] = GGML_F32_VEC_FMA(result_vec[kk], state_vec, r_vec);
+                            }
+                        }
+                        GGML_F32_VEC_REDUCE(dst_data[t_h_i_offset], result_vec);
+
+                        // There shouldn't be left-overs though.
+                        for (; j < head_size; j++) {
+                            int64_t t_h_j_offset = t_h_offset + j;
+                            int64_t h_2d_i_j_offset = h_2d_i_offset + j;
+
+                            float r_val = r[t_h_j_offset];
+                            float w_val = w[t_h_j_offset];
+                            float k_val = k[t_h_j_offset];
+                            float b_val = b[t_h_j_offset];
+                            float kv_val = v[t_h_i_offset] * k_val;
+
+                            float prev_state_val = state_prev[h_2d_i_j_offset];
+                            state_cur[h_2d_i_j_offset] = prev_state_val * w_val + kv_val + sa * b_val;
+                            dst_data[t_h_i_offset] += state_cur[h_2d_i_j_offset] * r_val;
+                        }
+                    }
+                }
+            }
+        #endif
     #else
         for (int64_t t = 0; t < T; t++) {
             int64_t t_offset = t * t_stride;

ggml/src/ggml-cpu/quants.h

@@ -0,0 +1,116 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+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);
+void ggml_vec_dot_q4_1_q8_1(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);
+void ggml_vec_dot_q5_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);
+void ggml_vec_dot_q5_1_q8_1(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);
+void ggml_vec_dot_q8_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);
+
+void ggml_vec_dot_q2_K_q8_K(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);
+void ggml_vec_dot_q3_K_q8_K(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);
+void ggml_vec_dot_q4_K_q8_K(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);
+void ggml_vec_dot_q5_K_q8_K(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);
+void ggml_vec_dot_q6_K_q8_K(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);
+
+void ggml_vec_dot_tq1_0_q8_K(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);
+void ggml_vec_dot_tq2_0_q8_K(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);
+
+void ggml_vec_dot_iq2_xxs_q8_K(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);
+void ggml_vec_dot_iq2_xs_q8_K (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);
+void ggml_vec_dot_iq2_s_q8_K  (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);
+void ggml_vec_dot_iq3_xxs_q8_K(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);
+void ggml_vec_dot_iq1_s_q8_K  (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);
+void ggml_vec_dot_iq1_m_q8_K  (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);
+void ggml_vec_dot_iq4_nl_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);
+void ggml_vec_dot_iq4_xs_q8_K (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);
+void ggml_vec_dot_iq3_s_q8_K  (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);
+
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q4_0_q8_0_generic(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);
+void ggml_vec_dot_q4_1_q8_1_generic(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);
+void ggml_vec_dot_q5_0_q8_0_generic(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);
+void ggml_vec_dot_q5_1_q8_1_generic(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);
+void ggml_vec_dot_q8_0_q8_0_generic(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);
+void ggml_vec_dot_tq1_0_q8_K_generic(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);
+void ggml_vec_dot_tq2_0_q8_K_generic(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);
+void ggml_vec_dot_q2_K_q8_K_generic(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);
+void ggml_vec_dot_q3_K_q8_K_generic(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);
+void ggml_vec_dot_q4_K_q8_K_generic(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);
+void ggml_vec_dot_q5_K_q8_K_generic(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);
+void ggml_vec_dot_q6_K_q8_K_generic(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);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(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);
+void ggml_vec_dot_iq2_xs_q8_K_generic(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);
+void ggml_vec_dot_iq2_s_q8_K_generic(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);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(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);
+void ggml_vec_dot_iq3_s_q8_K_generic(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);
+void ggml_vec_dot_iq1_s_q8_K_generic(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);
+void ggml_vec_dot_iq1_m_q8_K_generic(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);
+void ggml_vec_dot_iq4_nl_q8_0_generic(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);
+void ggml_vec_dot_iq4_xs_q8_K_generic(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);
+
+#if defined(GGML_CPU_GENERIC)
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/src/ggml-cpu/repack.h

@@ -0,0 +1,119 @@
+#pragma once
+
+#define GGML_COMMON_DECL_CPP
+#include "ggml-common.h"
+
+#include "traits.h"
+#include "ggml.h"
+
+// GGML internal header
+
+ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void);
+
+template <int K> constexpr int QK_0() {
+    if constexpr (K == 4) {
+        return QK4_0;
+    }
+    if constexpr (K == 8) {
+        return QK8_0;
+    }
+    return -1;
+}
+
+template <int K, int N> struct block {
+    ggml_half d[N];                         // deltas for N qK_0 blocks
+    int8_t    qs[(QK_0<K>() * N * K) / 8];  // quants for N qK_0 blocks
+};
+
+// control size
+static_assert(sizeof(block<4, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 2, "wrong block<4,4> size/padding");
+static_assert(sizeof(block<4, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<4,8> size/padding");
+static_assert(sizeof(block<8, 4>) == 4 * sizeof(ggml_half) + QK8_0 * 4, "wrong block<8,4> size/padding");
+static_assert(sizeof(block<8, 8>) == 8 * sizeof(ggml_half) + QK8_0 * 8, "wrong block<8,8> size/padding");
+
+using block_q4_0x4 = block<4, 4>;
+using block_q4_0x8 = block<4, 8>;
+using block_q8_0x4 = block<8, 4>;
+using block_q8_0x8 = block<8, 8>;
+
+struct block_q4_Kx8 {
+    ggml_half d[8];      // super-block scale for quantized scales
+    ggml_half dmin[8];   // super-block scale for quantized mins
+    uint8_t scales[96];  // scales and mins, quantized with 6 bits
+    uint8_t qs[1024];    // 4--bit quants
+};
+
+static_assert(sizeof(block_q4_Kx8) == sizeof(ggml_half) * 16 + K_SCALE_SIZE * 8 + QK_K * 4, "wrong q4_K block size/padding");
+
+struct block_q8_Kx4 {
+    float d[4];              // delta
+    int8_t qs[QK_K * 4];     // quants
+    int16_t bsums[QK_K / 4]; // sum of quants in groups of 16
+};
+
+static_assert(sizeof(block_q8_Kx4) == sizeof(float) * 4 + QK_K * 4 + (QK_K / 4) * sizeof(int16_t), "wrong q8_K block size/padding");
+
+struct block_iq4_nlx4 {
+    ggml_half d[4];            // deltas for 4 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 2];  // nibbles / quants for 4 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+// Workaround for clang:
+// clang++ complains: ``error: call to 'ggml_gemm_q4_0_4x4_q8_0' is ambiguous''
+// repro: https://godbolt.org/z/oKdeWKonM (ICE), https://godbolt.org/z/1szq6P36v (ambiguous call)
+#if defined(GGML_CPU_CLANG_WORKAROUND) || !(defined(__GNUC__) && defined(__clang__)) || defined(__HIPCC__)
+void ggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+#endif // !defined(__clang__)
+
+// Native implementations
+void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+
+#if defined(GGML_CPU_GENERIC)
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#endif
+
+#if defined(__cplusplus)
+} // extern "C"
+#endif

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

@@ -17,7 +17,123 @@
 //   number of elements to fit in a single register
 //
 
-#if defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_FMA)
+
+#define GGML_SIMD
+
+// F32 SVE
+#define GGML_F32_EPR 8
+#define DEFAULT_PG svptrue_b32()
+
+#define GGML_F32xt                        svfloat32_t
+#define GGML_F32xt_ZERO                   svdup_n_f32(0.0f)
+#define GGML_F32xt_SET1(x)                svdup_n_f32(x)
+#define GGML_F32xt_LOAD_IMPL(pg, a, ...)  svld1_f32(pg, a)
+#define GGML_F32xt_LOAD(...)              GGML_F32xt_LOAD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_STORE_IMPL(pg,a,b)     svst1_f32(pg, a, b)
+#define GGML_F32xt_STORE(...)             GGML_F32xt_STORE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c)  svmad_f32_m(pg, a, b, c)
+#define GGML_F32xt_FMA(...)               GGML_F32xt_FMA_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_ADD_IMPL(pg, a, b)     svadd_f32_m(pg, a, b)
+#define GGML_F32xt_ADD(...)               GGML_F32xt_ADD_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_MUL_IMPL(pg, a, b)     svmul_f32_m(pg, a, b)
+#define GGML_F32xt_MUL(...)               GGML_F32xt_MUL_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
+#define GGML_F32xt_REDUCE_ONE(...)        GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, __VA_ARGS__)
+#define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)  \
+{                                                      \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2);        \
+    sum3 = svadd_f32_m(DEFAULT_PG, sum3, sum4);        \
+    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum6);        \
+    sum7 = svadd_f32_m(DEFAULT_PG, sum7, sum8);        \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum3);        \
+    sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum7);        \
+    sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5);        \
+    (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1);  \
+}
+#define GGML_F32xt_REDUCE(...) GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, __VA_ARGS__)
+
+#define GGML_F32_VEC        GGML_F32xt
+#define GGML_F32_VEC_ZERO   GGML_F32xt_ZERO
+#define GGML_F32_VEC_SET1   GGML_F32xt_SET1
+#define GGML_F32_VEC_LOAD   GGML_F32xt_LOAD
+#define GGML_F32_VEC_STORE  GGML_F32xt_STORE
+#define GGML_F32_VEC_FMA    GGML_F32xt_FMA
+#define GGML_F32_VEC_ADD    GGML_F32xt_ADD
+#define GGML_F32_VEC_MUL    GGML_F32xt_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32xt_REDUCE
+
+// F16 NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+    #define GGML_F16_STEP 32
+    #define GGML_F16_EPR  8
+
+    #define GGML_F16x8              float16x8_t
+    #define GGML_F16x8_ZERO         vdupq_n_f16(0.0f)
+    #define GGML_F16x8_SET1(x)      vdupq_n_f16(x)
+    #define GGML_F16x8_LOAD(x)      vld1q_f16((const __fp16 *)(x))
+    #define GGML_F16x8_STORE        vst1q_f16
+    #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
+    #define GGML_F16x8_ADD          vaddq_f16
+    #define GGML_F16x8_MUL          vmulq_f16
+    #define GGML_F16x8_REDUCE(res, x)                               \
+    do {                                                            \
+        int offset = GGML_F16_ARR >> 1;                             \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        offset >>= 1;                                               \
+        for (int i = 0; i < offset; ++i) {                          \
+            (x)[i] = vaddq_f16((x)[i], (x)[offset+i]);              \
+        }                                                           \
+        const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+        const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+        (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1));         \
+    } while (0)
+
+    #define GGML_F16_VEC                GGML_F16x8
+    #define GGML_F16_VEC_ZERO           GGML_F16x8_ZERO
+    #define GGML_F16_VEC_SET1           GGML_F16x8_SET1
+    #define GGML_F16_VEC_LOAD(p, i)     GGML_F16x8_LOAD(p)
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
+    #define GGML_F16_VEC_FMA            GGML_F16x8_FMA
+    #define GGML_F16_VEC_ADD            GGML_F16x8_ADD
+    #define GGML_F16_VEC_MUL            GGML_F16x8_MUL
+    #define GGML_F16_VEC_REDUCE         GGML_F16x8_REDUCE
+#else
+    // if FP16 vector arithmetic is not supported, we use FP32 instead
+    // and take advantage of the vcvt_ functions to convert to/from FP16
+
+    #define GGML_F16_STEP 16
+    #define GGML_F16_EPR  4
+
+    #define GGML_F32Cx4              float32x4_t
+    #define GGML_F32Cx4_ZERO         vdupq_n_f32(0.0f)
+    #define GGML_F32Cx4_SET1(x)      vdupq_n_f32(x)
+    #define GGML_F32Cx4_LOAD(x)      vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
+    #define GGML_F32Cx4_STORE(x, y)  vst1_f16(x, vcvt_f16_f32(y))
+    #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
+    #define GGML_F32Cx4_ADD          vaddq_f32
+    #define GGML_F32Cx4_MUL          vmulq_f32
+    #define GGML_F32Cx4_REDUCE       GGML_F32x4_REDUCE
+
+    #define GGML_F16_VEC                GGML_F32Cx4
+    #define GGML_F16_VEC_ZERO           GGML_F32Cx4_ZERO
+    #define GGML_F16_VEC_SET1           GGML_F32Cx4_SET1
+    #define GGML_F16_VEC_LOAD(p, i)     GGML_F32Cx4_LOAD(p)
+    #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
+    #define GGML_F16_VEC_FMA            GGML_F32Cx4_FMA
+    #define GGML_F16_VEC_ADD            GGML_F32Cx4_ADD
+    #define GGML_F16_VEC_MUL            GGML_F32Cx4_MUL
+    #define GGML_F16_VEC_REDUCE         GGML_F32Cx4_REDUCE
+#endif
+
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA)
 
 #define GGML_SIMD
 

ggml/src/ggml-cpu/traits.cpp

@@ -1,4 +1,4 @@
-#include "ggml-cpu-traits.h"
+#include "traits.h"
 
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"

ggml/src/ggml-cpu/vec.cpp

@@ -17,29 +17,98 @@ void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * G
 
 #if defined(GGML_SIMD)
     float sumf = 0.0f;
-    const int np = (n & ~(GGML_F32_STEP - 1));
 
-    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
 
-    GGML_F32_VEC ax[GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t sum1 = svdup_n_f32(0.0f);
+        svfloat32_t sum2 = svdup_n_f32(0.0f);
+        svfloat32_t sum3 = svdup_n_f32(0.0f);
+        svfloat32_t sum4 = svdup_n_f32(0.0f);
+        svfloat32_t sum5 = svdup_n_f32(0.0f);
+        svfloat32_t sum6 = svdup_n_f32(0.0f);
+        svfloat32_t sum7 = svdup_n_f32(0.0f);
+        svfloat32_t sum8 = svdup_n_f32(0.0f);
+        svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
+        svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
+        for (int i = 0; i < np; i += ggml_f32_step) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            sum2 = GGML_F32_VEC_FMA(ax2, ay2, sum2);
 
-            sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+            sum3 = GGML_F32_VEC_FMA(ax3, ay3, sum3);
+
+            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+            sum4 = GGML_F32_VEC_FMA(ax4, ay4, sum4);
+
+            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+            sum5 = GGML_F32_VEC_FMA(ax5, ay5, sum5);
+
+            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+            sum6 = GGML_F32_VEC_FMA(ax6, ay6, sum6);
+
+            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+            sum7 = GGML_F32_VEC_FMA(ax7, ay7, sum7);
+
+            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+            sum8 = GGML_F32_VEC_FMA(ax8, ay8, sum8);
         }
-    }
+        // leftovers
+        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+        const int np2 = (n & ~(ggml_f32_epr - 1));
+        for (int i = np; i < np2; i += ggml_f32_epr) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            sum1 = GGML_F32_VEC_FMA(ax1, ay1, sum1);
+        }
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np2 < n) {
+            svbool_t pg = svwhilelt_b32(np2, n);
+            ax1 = svld1_f32(pg, x + np2);
+            ay1 = svld1_f32(pg, y + np2);
+            sum1 = svmad_f32_m(pg, ax1, ay1, sum1);
+        }
+        // reduce sum1,sum2 to sum1
+        GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    // reduce sum0..sum3 to sum0
-    GGML_F32_VEC_REDUCE(sumf, sum);
+        GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        sumf += x[i]*y[i];
-    }
+        GGML_F32_VEC ax[GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+                sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+            }
+        }
+
+        // reduce sum0..sum3 to sum0
+        GGML_F32_VEC_REDUCE(sumf, sum);
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            sumf += x[i]*y[i];
+        }
+    #endif
 #else
     // scalar
     ggml_float sumf = 0.0;

ggml/src/ggml-cpu/vec.h

@@ -5,6 +5,7 @@
 #include "ggml-impl.h"
 #include "simd-mappings.h"
 #include "ggml.h"
+#include "ggml-cpu.h"
 
 #if defined(GGML_USE_ACCELERATE)
 #include <Accelerate/Accelerate.h>
@@ -148,27 +149,108 @@ inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GG
 
 inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
 #if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
 
-    GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
 
-    GGML_F32_VEC ax[GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+        svfloat32_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+        for (int i = 0; i < np; i += ggml_f32_step) {
 
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
 
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            GGML_F32_VEC_STORE(y + i, ay1);
+
+            ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_FMA(ax2, vx, ay2);
+
+            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
+
+            ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+            ay3 = GGML_F32_VEC_FMA(ax3, vx, ay3);
+
+            GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
+
+            ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+            ay4 = GGML_F32_VEC_FMA(ax4, vx, ay4);
+
+            GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
+
+            ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+            ay5 = GGML_F32_VEC_FMA(ax5, vx, ay5);
+
+            GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
+
+            ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+            ay6 = GGML_F32_VEC_FMA(ax6, vx, ay6);
+
+            GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
+
+            ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+            ay7 = GGML_F32_VEC_FMA(ax7, vx, ay7);
+
+            GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
+
+            ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+            ay8 = GGML_F32_VEC_FMA(ax8, vx, ay8);
+
+            GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
         }
-    }
+        // leftovers
+        // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+        const int np2 = (n & ~(ggml_f32_epr - 1));
+        for (int i = np; i < np2; i += ggml_f32_epr) {
+            ax1 = GGML_F32_VEC_LOAD(x + i);
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_FMA(ax1, vx, ay1);
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] += x[i]*v;
-    }
+            GGML_F32_VEC_STORE(y + i, ay1);
+        }
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np2 < n) {
+            svbool_t pg =svwhilelt_b32(np2, n);
+            ax1 = svld1_f32(pg, x + np2);
+            ay1 = svld1_f32(pg, y + np2);
+            ay1 = svmad_f32_m(pg, ax1, vx, ay1);
+
+            svst1_f32(pg, y + np2, ay1);
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
+
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+        GGML_F32_VEC ax[GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] += x[i]*v;
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -220,36 +302,45 @@ inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int
     }
 
 #if defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
-
-    GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
-
-    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-        vx[k] = GGML_F32_VEC_SET1(v[k][0]);
-    }
-
-    GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
-    GGML_F32_VEC ay[GGML_F32_ARR];
-
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
-
-            for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-                ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
-                ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
+    #if defined(__ARM_FEATURE_SVE)
+        // scalar Route to scalar implementation       //TODO: Write SVE code
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            for (int i = 0; i < n; ++i) {
+                y[i] += x[k][i]*v[k][0];
             }
-
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
         }
-    }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    // leftovers
-    for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
-        for (int i = np; i < n; ++i) {
-            y[i] += x[k][i]*v[k][0];
+        GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
+
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            vx[k] = GGML_F32_VEC_SET1(v[k][0]);
         }
-    }
+
+        GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+                for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+                    ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
+                    ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
+                }
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+            for (int i = np; i < n; ++i) {
+                y[i] += x[k][i]*v[k][0];
+            }
+        }
+    #endif
 #else
     // scalar
     for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
@@ -265,25 +356,53 @@ inline static void ggml_vec_scale_f32(const int n, float * y, const float   v) {
 #if defined(GGML_USE_ACCELERATE)
     vDSP_vsmul(y, 1, &v, y, 1, n);
 #elif defined(GGML_SIMD)
-    const int np = (n & ~(GGML_F32_STEP - 1));
+    #if defined(__ARM_FEATURE_SVE)
+        const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+        const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+        const int ggml_f32_step = 2 * ggml_f32_epr;
 
-    GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+        const int np = (n & ~(ggml_f32_step - 1));
+        svfloat32_t ay1;
+        svfloat32_t ay2;
+        for (int i = 0; i < np; i += ggml_f32_step) {
+            ay1 = GGML_F32_VEC_LOAD(y + i);
+            ay1 = GGML_F32_VEC_MUL(ay1, vx);
+            GGML_F32_VEC_STORE(y + i, ay1);
 
-    GGML_F32_VEC ay[GGML_F32_ARR];
-
-    for (int i = 0; i < np; i += GGML_F32_STEP) {
-        for (int j = 0; j < GGML_F32_ARR; j++) {
-            ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
-            ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
-
-            GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+            ay2 = GGML_F32_VEC_MUL(ay2, vx);
+            GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
         }
-    }
+        // leftovers
+        // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+        if (np < n) {
+            svbool_t pg = svwhilelt_b32(np, n);
+            ay1 = svld1_f32(pg, y + np);
+            ay1 = svmul_f32_m(pg, ay1, vx);
+            svst1_f32(pg, y + np, ay1);
+        }
+    #else
+        const int np = (n & ~(GGML_F32_STEP - 1));
 
-    // leftovers
-    for (int i = np; i < n; ++i) {
-        y[i] *= v;
-    }
+        GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+        GGML_F32_VEC ay[GGML_F32_ARR];
+
+        for (int i = 0; i < np; i += GGML_F32_STEP) {
+            for (int j = 0; j < GGML_F32_ARR; j++) {
+                ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+                ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
+
+                GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+            }
+        }
+
+        // leftovers
+        for (int i = np; i < n; ++i) {
+            y[i] *= v;
+        }
+    #endif
 #else
     // scalar
     for (int i = 0; i < n; ++i) {
@@ -528,6 +647,42 @@ inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
 #error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
 #endif
 
+/* Below function was borrowed from the GitHub repository:
+https://github.com/openvinotoolkit/openvino/blob/master/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/common.hpp */
+#if defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+    inline static svfloat32_t exp_ps_sve(svbool_t pg, svfloat32_t src) {
+        // Constants
+        const svfloat32_t log2_e = svdup_n_f32(1.4426950409f);
+        const svfloat32_t ln2 = svdup_n_f32(0.6931473921f);
+        const svfloat32_t half_ln2_sq = svdup_n_f32(0.2413862043f);
+        const svuint32_t not_mask17 = svdup_n_u32(~((1u << 17) - 1));
+        const svfloat32_t one = svdup_n_f32(1.0f);
+        const svfloat32_t inactive1 = svdup_n_f32(0.0f);
+        const svint32_t inactive2 = svdup_n_s32(0);
+
+        // Algorithm starts here
+        svfloat32_t t0 = svmul_f32_m(pg, src, log2_e);  // y = x * log2(e)
+        svfloat32_t t1 = svrintm_f32_m(inactive1, pg, t0);         // rount to int (float)
+        svint32_t t2 = svcvt_s32_f32_m(inactive2, pg, t1);         // n
+
+        t1 = svsub_f32_m(pg, t0, t1);   // a = y - floor(y)
+        t1 = svadd_f32_m(pg, t1, one);  // b = a + 1
+
+        svuint32_t t3 = svlsr_n_u32_m(pg, svreinterpret_u32_f32(t1), 17);  // v = b >> 17 (u32)
+        svfloat32_t t4 = svexpa_f32(t3);                                   // c = fexpa(v)
+        t4 = svscale_f32_m(pg, t4, t2);                                    // fexpa(v) * 2^(n)
+
+        // and_(t2.d, t1.d, not_mask17.d)
+        svfloat32_t t5 = svreinterpret_f32_u32(svand_u32_m(pg, svreinterpret_u32_f32(t1), not_mask17));
+        t5 = svsub_f32_m(pg, t1, t5);                // z
+        t0 = svmla_f32_m(pg, ln2, t5, half_ln2_sq);  // ln2 + half_ln2_sq * z
+        t0 = svmla_f32_m(pg, one, t5, t0);           // 1 + (ln2 * z) + (half_ln2_sq * z * z)
+        t0 = svmul_f32_m(pg, t0, t4);                // Final result
+
+        return t0;
+    }
+#endif
+
 #if defined(__ARM_NEON) && defined(__aarch64__)
 
 // adapted from arm limited optimized routine

ggml/src/ggml-cuda/common.cuh

@@ -466,9 +466,6 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
 }
 
-// TODO: move to ggml-common.h
-static constexpr __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
-
 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);
 
 static __device__ __forceinline__ float get_alibi_slope(
@@ -635,6 +632,7 @@ struct ggml_cuda_device_info {
         int     nsm;                // number of streaming multiprocessors
         size_t  smpb;               // max. shared memory per block
         size_t  smpbo;              // max. shared memory per block (with opt-in)
+        bool    integrated;         // Device is integrated as opposed to discrete
         bool    vmm;                // virtual memory support
         size_t  vmm_granularity;    // granularity of virtual memory
         size_t  total_vram;

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

@@ -623,8 +623,8 @@ static __global__ void flash_attn_combine_results(
     __builtin_assume(tid < D);
 
     extern __shared__ float2 meta[];
-    if (tid < 2*parallel_blocks) {
-        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + tid];
+    for (int i = tid; i < 2*parallel_blocks; i += D) {
+        ((float *) meta)[i] = ((const float *)VKQ_meta) [blockIdx.z*(2*parallel_blocks) + i];
     }
 
     __syncthreads();

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

@@ -652,9 +652,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         float KQ_max_scale[cols_per_thread];
 #pragma unroll
         for (int col = 0; col < cols_per_thread; ++col) {
-            KQ_max_scale[col] = expf(KQ_max[col] - KQ_max_new[col]);
+            const float KQ_max_diff = KQ_max[col] - KQ_max_new[col];
+            KQ_max_scale[col] = expf(KQ_max_diff);
             KQ_max[col] = KQ_max_new[col];
 
+            *((uint32_t *) &KQ_max_scale[col]) *= KQ_max_diff >= SOFTMAX_FTZ_THRESHOLD;
+
             // Scale previous KQ_rowsum to account for a potential increase in KQ_max:
             KQ_rowsum[col] = KQ_max_scale[col]*KQ_rowsum[col] + KQ_rowsum_add[col];
         }
@@ -1246,7 +1249,7 @@ static __global__ void flash_attn_ext_f16(
         NO_DEVICE_CODE;
         return;
     }
-#endif __CUDA_ARCH__ == GGML_CUDA_CC_TURING
+#endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
     static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 

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

@@ -243,10 +243,10 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += prop.totalGlobalMem;
-
-        info.devices[id].nsm       = prop.multiProcessorCount;
-        info.devices[id].smpb      = prop.sharedMemPerBlock;
-        info.devices[id].warp_size = prop.warpSize;
+        info.devices[id].integrated = prop.integrated;
+        info.devices[id].nsm        = prop.multiProcessorCount;
+        info.devices[id].smpb       = prop.sharedMemPerBlock;
+        info.devices[id].warp_size  = prop.warpSize;
 #if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__)
         info.devices[id].smpbo = prop.sharedMemPerBlock;
 
@@ -615,9 +615,8 @@ static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
     ggml_cuda_set_device(ctx->device);
-    CUDA_CHECK(cudaDeviceSynchronize());
-    CUDA_CHECK(cudaMemset(ctx->dev_ptr, value, buffer->size));
-    CUDA_CHECK(cudaDeviceSynchronize());
+    CUDA_CHECK(cudaMemsetAsync(ctx->dev_ptr, value, buffer->size, cudaStreamPerThread));
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
 static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
@@ -1065,6 +1064,10 @@ static const char * ggml_backend_cuda_host_buffer_type_name(ggml_backend_buffer_
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_buft_is_cuda_host(ggml_backend_buffer_type_t buft) {
+    return buft->iface.get_name == ggml_backend_cuda_host_buffer_type_name;
+}
+
 static void ggml_backend_cuda_host_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     CUDA_CHECK(cudaFreeHost(buffer->context));
 }
@@ -1140,7 +1143,6 @@ typedef void (*ggml_cuda_op_mul_mat_t)(
 static cudaError_t ggml_cuda_cpy_tensor_2d(
     void * dst, const struct ggml_tensor * src, int64_t i3, int64_t i2, int64_t i1_low, int64_t i1_high, cudaStream_t stream) {
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src->buffer));
     const char * src_ptr = (const char *) src->data;
     char       * dst_ptr = (char       *) dst;
 
@@ -1423,8 +1425,6 @@ static void ggml_cuda_op_mul_mat(
     const int64_t nb2 = dst->nb[2];
     const int64_t nb3 = dst->nb[3];
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(dst->buffer));
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src1->buffer));
     ggml_backend_cuda_buffer_context * src1_ctx = (ggml_backend_cuda_buffer_context *) src1->buffer->context;
     ggml_backend_cuda_buffer_context * dst_ctx  = (ggml_backend_cuda_buffer_context *) dst->buffer->context;
 
@@ -1746,7 +1746,7 @@ static void ggml_cuda_mul_mat_batched_cublas(ggml_backend_cuda_context & ctx, co
     GGML_ASSERT(!ggml_is_transposed(src0));
     GGML_ASSERT(!ggml_is_transposed(src1));
 
-    GGML_ASSERT(ggml_backend_buffer_is_cuda(src0->buffer));
+    GGML_ASSERT(!ggml_backend_buft_is_cuda_split(src0->buffer->buft));
     GGML_ASSERT(src0->type == GGML_TYPE_F16);
 
     // Byte offsets and tensor dimensions are currently used in an inconsistent way for dst.
@@ -2641,6 +2641,8 @@ static void update_cuda_graph_executable(ggml_backend_cuda_context * cuda_ctx) {
 
 static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph * cgraph,
     bool & graph_evaluated_or_captured, bool & use_cuda_graph, bool & cuda_graph_update_required) {
+    // flag used to determine whether it is an integrated_gpu
+    const bool integrated = ggml_cuda_info().devices[cuda_ctx->device].integrated;
 
     while (!graph_evaluated_or_captured) {
         // Only perform the graph execution if CUDA graphs are not enabled, or we are capturing the graph.
@@ -2659,7 +2661,7 @@ static void evaluate_and_capture_cuda_graph(ggml_backend_cuda_context * cuda_ctx
                     if (node->src[j] != nullptr) {
                         assert(node->src[j]->buffer);
                         assert(node->src[j]->buffer->buft == ggml_backend_cuda_buffer_type(cuda_ctx->device) ||
-                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft));
+                               ggml_backend_buft_is_cuda_split(node->src[j]->buffer->buft) || (integrated && ggml_backend_buft_is_cuda_host(node->src[j]->buffer->buft)));
                     }
                 }
 #endif
@@ -2994,9 +2996,12 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
             {
                 struct ggml_tensor * a = op->src[0];
                 struct ggml_tensor * b = op->src[1];
-                // for small weight matrices the active device can end up without any rows, don't use row split in those cases
-                // this avoids some edge cases (and the performance would not be good anyways)
                 if (a->buffer && ggml_backend_buft_is_cuda_split(a->buffer->buft)) {
+                    if (a->ne[2] > 1 || a->ne[3] > 1) {
+                        return false;
+                    }
+                    // for small weight matrices the active device can end up without any rows, don't use row split in those cases
+                    // this avoids some edge cases (and the performance would not be good anyways)
                     ggml_backend_cuda_split_buffer_type_context * buft_ctx = (ggml_backend_cuda_split_buffer_type_context *) a->buffer->buft->context;
                     int64_t row_low;
                     int64_t row_high;
@@ -3263,7 +3268,9 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 }
 
 static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
-    return (ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
+    const bool integrated = ggml_cuda_info().devices[dev_ctx->device].integrated;
+    return (((ggml_backend_buft_is_cuda(buft) || ggml_backend_buft_is_cuda_split(buft)) && buft->device == dev) || (integrated && ggml_backend_buft_is_cuda_host(buft)));
 }
 
 static int64_t get_op_batch_size(const ggml_tensor * op) {

ggml/src/ggml-impl.h

@@ -32,6 +32,8 @@
 extern "C" {
 #endif
 
+void ggml_print_backtrace(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif

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

@@ -4766,6 +4766,8 @@ static bool ggml_metal_encode_node(
                     GGML_ASSERT(nqptg  % 8  == 0);
                     GGML_ASSERT(ncpsg  % 32 == 0);
 
+                    const int is_q = ggml_is_quantized(src1->type) ? 1 : 0;
+
                     // 2*(2*ncpsg + nqptg)*(nsg)
                     // ncpsg soft_max values + ncpsg mask values + a diagonal scaling matrix (in float)
                     //
@@ -4773,7 +4775,7 @@ static bool ggml_metal_encode_node(
                     // the shared memory needed for the simdgroups to load the KV cache
                     // each thread loads (dequantizes) 16 head elements, there are 32 threads in th SG
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + 16*32*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(2*ne00 + 2*(2*ncpsg + nqptg)*(nsg)) + is_q*(16*32*(nsg)))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
 
@@ -4810,9 +4812,9 @@ static bool ggml_metal_encode_node(
                     // and store the soft_max values and the mask
                     //
                     // ne00*(nsg)
-                    // each simdgroup has a full f16 head vector in shared mem to accumulate results
+                    // each simdgroup has a full f32 head vector in shared mem to accumulate results
                     //
-#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + ne20*(nsg))*(sizeof(float)/2), 16))
+#define FATTN_SMEM(nsg) (GGML_PAD((nqptg*(GGML_PAD(ne00, 128) + 4*ncpsg*(nsg)) + 2*ne20*(nsg))*(sizeof(float)/2), 16))
 
                     int64_t nsgmax = 2;
                     while (true) {

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

@@ -3328,14 +3328,12 @@ kernel void kernel_flash_attn_ext(
     constexpr short NW  = N_SIMDWIDTH;
     constexpr short SH  = (2*C + Q); // shared memory per simdgroup (s_t == float)
 
-    const short TS = nsg*SH;   // shared memory size per query in (s_t == float)
-    const short T  = DK + 2*TS; // shared memory size per query in (half)
+    const short TS = nsg*SH;      // shared memory size per query in (s_t == float)
+    const short T  = 2*DK + 2*TS; // shared memory size per query in (half)
 
-    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +              0*DK); // holds the query data
-    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +              0*DK); // same as above but in q4_t
-    threadgroup o_t  * so  = (threadgroup o_t  *) (shmem_f16 +              0*DK); // reuse query data for accumulation
-    threadgroup o4_t * so4 = (threadgroup o4_t *) (shmem_f16 +              0*DK); // same as above but in o4_t
-    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + Q*DK); // scratch buffer for attention, mask and diagonal matrix
+    threadgroup q_t  * sq  = (threadgroup q_t  *) (shmem_f16 +                0*DK); // holds the query data
+    threadgroup q4_t * sq4 = (threadgroup q4_t *) (shmem_f16 +                0*DK); // same as above but in q4_t
+    threadgroup s_t  * ss  = (threadgroup s_t  *) (shmem_f16 + 2*sgitg*SH + 2*Q*DK); // scratch buffer for attention, mask and diagonal matrix
 
     threadgroup k_t    * sk    = (threadgroup k_t    *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // scratch buffer to load K in shared memory
     threadgroup k4x4_t * sk4x4 = (threadgroup k4x4_t *) (shmem_f16 + sgitg*(4*16*KV) + Q*T); // same as above but in k4x4_t
@@ -3354,7 +3352,7 @@ kernel void kernel_flash_attn_ext(
             if (iq1 + j < args.ne01) {
                 sq4[j*DK4 + i] = (q4_t) q4[i];
             } else {
-                sq4[j*DK4 + i] = (q4_t) 0.0f;
+                sq4[j*DK4 + i] = 0;
             }
         }
     }
@@ -3548,20 +3546,20 @@ kernel void kernel_flash_attn_ext(
 
             // O = diag(ms)*O
             {
-                s8x8_t mm;
-                simdgroup_load(mm, ss + 2*C, TS, 0, false);
+                s8x8_t ms;
+                simdgroup_load(ms, ss + 2*C, TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    simdgroup_multiply(lo[i], mm, lo[i]);
+                    simdgroup_multiply(lo[i], ms, lo[i]);
                 }
             }
 
             // O = O + (Q*K^T)*V
             {
                 for (short cc = 0; cc < C/8; ++cc) {
-                    s8x8_t ms;
-                    simdgroup_load(ms, ss + 8*cc, TS, 0, false);
+                    s8x8_t vs;
+                    simdgroup_load(vs, ss + 8*cc, TS, 0, false);
 
                     if (is_same<vd4x4_t, v4x4_t>::value) {
                         // we can read directly from global memory
@@ -3572,7 +3570,7 @@ kernel void kernel_flash_attn_ext(
                             v8x8_t mv;
                             simdgroup_load(mv, pv + i*8, args.nb21/sizeof(v_t), 0, false); // TODO: use ne20
 
-                            simdgroup_multiply_accumulate(lo[i], ms, mv, lo[i]);
+                            simdgroup_multiply_accumulate(lo[i], vs, mv, lo[i]);
                         }
                     } else {
                         for (short ii = 0; ii < DV16; ii += 4) {
@@ -3593,10 +3591,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             } else {
                                 if (ii + tx < DV16) {
@@ -3611,10 +3609,10 @@ kernel void kernel_flash_attn_ext(
                                     v8x8_t mv;
 
                                     simdgroup_load(mv, sv + 16*k + 0*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], ms, mv, lo[2*(ii + k) + 0]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 0], vs, mv, lo[2*(ii + k) + 0]);
 
                                     simdgroup_load(mv, sv + 16*k + 1*8, 4*16, 0, false);
-                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], ms, mv, lo[2*(ii + k) + 1]);
+                                    simdgroup_multiply_accumulate(lo[2*(ii + k) + 1], vs, mv, lo[2*(ii + k) + 1]);
                                 }
                             }
                         }
@@ -3624,93 +3622,89 @@ kernel void kernel_flash_attn_ext(
         }
 
         // these are needed for reducing the results from the simdgroups (reuse the ss buffer)
-        for (short j = 0; j < Q; ++j) {
-            if (tiisg == 0) {
-                ss[j*TS + 0] = S[j];
-                ss[j*TS + 1] = M[j];
-            }
+        for (short j = tiisg; j < Q; j += NW) {
+            ss[j*TS + 0] = S[j];
+            ss[j*TS + 1] = M[j];
         }
     }
 
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup float  * so  = (threadgroup float  *) (shmem_f16 + 0*DK); // reuse query data for accumulation
+    threadgroup float4 * so4 = (threadgroup float4 *) (shmem_f16 + 0*DK);
+
+    // store result to shared memory in F32
+    if (sgitg == 0) {
+        for (short i = 0; i < DV8; ++i) {
+            //simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            simdgroup_float8x8 t(1.0f);
+            simdgroup_multiply(t, lo[i], t);
+            simdgroup_store(t, so + i*8, DV, 0, false);
+        }
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
     // reduce the warps sequentially
     for (ushort sg = 1; sg < nsg; ++sg) {
-        float S = { 0.0f };
-        float M = { -__FLT_MAX__/2 };
-
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // each simdgroup stores its output to shared memory, reusing sq
         if (sgitg == sg) {
-            for (short i = 0; i < DV8; ++i) {
-                simdgroup_store(lo[i], so + i*8, DV, 0, false);
+            for (short j = tiisg; j < Q; j += NW) {
+                const float S0 = ss[j*TS - 1*SH + 0];
+                const float S1 = ss[j*TS        + 0];
+
+                const float M0 = ss[j*TS - 1*SH + 1];
+                const float M1 = ss[j*TS        + 1];
+
+                const float M = max(M0, M1);
+
+                float ms0 = exp(M0 - M);
+                float ms1 = exp(M1 - M);
+
+                const float S = S0*ms0 + S1*ms1;
+
+                ss[j*TS + 0] = S;
+                ss[j*TS + 1] = M;
+
+                ss[j*TS + 2*C + j - 1*SH] = ms0;
+                ss[j*TS + 2*C + j       ] = ms1;
             }
-        }
 
-        threadgroup_barrier(mem_flags::mem_threadgroup);
-
-        // the first simdgroup accumulates the results from the other simdgroups
-        if (sgitg == 0) {
-            for (short j = 0; j < Q; ++j) {
-                const float S0 = ss[j*TS +         0];
-                const float S1 = ss[j*TS + sg*SH + 0];
-
-                const float M0 = ss[j*TS +         1];
-                const float M1 = ss[j*TS + sg*SH + 1];
-
-                M = max(M0, M1);
-
-                const float ms0 = exp(M0 - M);
-                const float ms1 = exp(M1 - M);
-
-                S = S0*ms0 + S1*ms1;
-
-                if (tiisg == 0) {
-                    ss[j*TS + 0] = S;
-                    ss[j*TS + 1] = M;
-
-                    ss[j*TS + 2*C + j        ] = ms0;
-                    ss[j*TS + 2*C + j + sg*SH] = ms1;
-                }
-            }
+            //simdgroup_barrier(mem_flags::mem_threadgroup);
 
             // O_0 = diag(ms0)*O_0 + diag(ms1)*O_1
             {
                 s8x8_t ms0;
                 s8x8_t ms1;
 
-                simdgroup_load(ms0, ss + 2*C,         TS, 0, false);
-                simdgroup_load(ms1, ss + 2*C + sg*SH, TS, 0, false);
+                simdgroup_load(ms0, ss + 2*C - 1*SH, TS, 0, false);
+                simdgroup_load(ms1, ss + 2*C,        TS, 0, false);
 
                 #pragma unroll(DV8)
                 for (short i = 0; i < DV8; ++i) {
-                    o8x8_t t;
+                    simdgroup_float8x8 t;
 
                     simdgroup_load    (t, so + i*8, DV, 0, false);
-                    simdgroup_multiply(t, ms1, t);
+                    simdgroup_multiply(t, ms0, t);
 
-                    simdgroup_multiply_accumulate(lo[i], ms0, lo[i], t);
+                    simdgroup_multiply_accumulate(t, ms1, lo[i], t);
+                    simdgroup_store(t, so + i*8, DV, 0, false);
                 }
             }
         }
+
+        threadgroup_barrier(mem_flags::mem_threadgroup);
     }
 
-    // store result to shared memory (reuse sq)
-    if (sgitg == 0) {
-        for (short i = 0; i < DV8; ++i) {
-            simdgroup_store(lo[i], so + i*8, DV, 0, false);
-        }
-    }
-
-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 2*(nsg-1)*SH + 2*Q*DK);
 
     // final rescale with 1/S and store to global memory
-    if (sgitg == 0) {
-        for (short j = 0; j < Q && iq1 + j < args.ne01; ++j) {
-            const float S = ss[j*TS + 0];
+    for (short j = sgitg; j < Q && iq1 + j < args.ne01; j += nsg) {
+        const float S = 1.0f/sf[j*TS + 0];
 
-            for (short i = tiisg; i < DV4; i += NW) {
-                dst4[((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4 + i] = (float4) so4[j*DV4 + i]/S;
-            }
+        device float4 * dst4 = (device float4 *) dst + ((uint64_t)iq3*args.ne2*args.ne1 + iq2 + (uint64_t)(iq1 + j)*args.ne1)*DV4;
+
+        for (short i = tiisg; i < DV4; i += NW) {
+            dst4[i] = (float4) so4[j*DV4 + i]*S;
         }
     }
 }
@@ -3719,12 +3713,22 @@ kernel void kernel_flash_attn_ext(
 //       template to be able to explore different combinations
 //
 #define FA_TYPES \
-    half,  half4,   simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    half,  half4x4, simdgroup_half8x8,  \
-    float,          simdgroup_float8x8, \
-    float,          simdgroup_float8x8, \
-    half,  half4,   simdgroup_half8x8
+    float,  float4,    simdgroup_float8x8, \
+    half,   half4x4,   simdgroup_half8x8,  \
+    half,   half4x4,   simdgroup_half8x8,  \
+    float,             simdgroup_float8x8, \
+    float,             simdgroup_float8x8, \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    half,   half4,     simdgroup_half8x8
+    //float,  float4,    simdgroup_float8x8
 
 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;
 
@@ -3739,15 +3743,15 @@ template [[host_name("kernel_flash_attn_ext_f16_h256")]]         kernel flash_at
 template [[host_name("kernel_flash_attn_ext_f16_hk576_hv512")]]  kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, half4x4,    1, dequantize_f16,  half4x4,    1, dequantize_f16,  576, 512>;
 
 #if defined(GGML_METAL_USE_BF16)
-template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
-template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
-template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
-template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
-template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
-template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
-template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
+template [[host_name("kernel_flash_attn_ext_bf16_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 64,  64>;
+template [[host_name("kernel_flash_attn_ext_bf16_h80" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 80,  80>;
+template [[host_name("kernel_flash_attn_ext_bf16_h96" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 96,  96>;
+template [[host_name("kernel_flash_attn_ext_bf16_h112")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 112, 112>;
+template [[host_name("kernel_flash_attn_ext_bf16_h128")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 128, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h192")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 192>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk192_hv128")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 192, 128>;
+template [[host_name("kernel_flash_attn_ext_bf16_h256")]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 256, 256>;
+template [[host_name("kernel_flash_attn_ext_bf16_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES_BF, bfloat4x4,  1, dequantize_bf16, bfloat4x4,  1, dequantize_bf16, 576, 512>;
 #endif
 
 template [[host_name("kernel_flash_attn_ext_q4_0_h64" )]]        kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q4_0, 2, dequantize_q4_0, block_q4_0, 2, dequantize_q4_0, 64,  64>;
@@ -3801,6 +3805,7 @@ template [[host_name("kernel_flash_attn_ext_q8_0_h256")]]        kernel flash_at
 template [[host_name("kernel_flash_attn_ext_q8_0_hk576_hv512")]] kernel flash_attn_ext_t kernel_flash_attn_ext<FA_TYPES, block_q8_0, 2, dequantize_q8_0, block_q8_0, 2, dequantize_q8_0, 576, 512>;
 
 #undef FA_TYPES
+#undef FA_TYPES_BF
 
 template<
     typename q4_t,  // query types in shared memory
@@ -3847,12 +3852,12 @@ kernel void kernel_flash_attn_ext_vec(
 
     const short T = DK + nsg*SH; // shared memory size per query in (half)
 
-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                  0*DK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                  0*DK); // same as above but in q4_t
-    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 + sgitg*SH       + Q*DK); // scratch buffer for attention
-    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 + sgitg*SH       + Q*DK); // same as above but in s4_t
-    threadgroup float * sm  = (threadgroup float *) (shmem_f16 + sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
-    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + sgitg*DV       + Q*T);  // scratch buffer for the results
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                    0*DK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                    0*DK); // same as above but in q4_t
+    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH       + Q*DK); // scratch buffer for attention
+    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH       + Q*DK); // same as above but in s4_t
+    threadgroup float * sm  = (threadgroup float *) (shmem_f16 +   sgitg*SH + 2*C + Q*DK); // scratch buffer for mask
+    threadgroup o4_t  * sr4 = (threadgroup o4_t  *) (shmem_f16 + 2*sgitg*DV       + Q*T);  // scratch buffer for the results
 
     // store the result for all queries in local memory (the O matrix from the paper)
     o4_t lo[DV4/NL];
@@ -4157,7 +4162,7 @@ kernel void kernel_flash_attn_ext_vec(
            half4,  \
     float,         \
     float, float4, \
-           half4
+           float4
 
 typedef decltype(kernel_flash_attn_ext_vec<FA_TYPES, half4, 1, dequantize_f16_t4, half4, 1, dequantize_f16_t4, 128, 128, 4>) flash_attn_ext_vec_t;
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -55,14 +55,17 @@ endfunction()
 
 set(GGML_OPENCL_KERNELS
     add
+    argsort
     clamp
     cpy
     cvt
     diag_mask_inf
+    div
     gelu
     gemv_noshuffle_general
     gemv_noshuffle
     get_rows
+    group_norm
     im2col_f32
     im2col_f16
     mul_mat_Ab_Bi_8x4
@@ -83,12 +86,21 @@ set(GGML_OPENCL_KERNELS
     rms_norm
     rope
     scale
+    sigmoid
     silu
     softmax_4_f32
     softmax_4_f16
     softmax_f32
     softmax_f16
+    sub
+    sum_rows
     transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    pad
+    repeat
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -0,0 +1,86 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(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")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define SWAP(x, y, T) { T tmp = (x); (x) = (y); (y) = tmp; }
+
+enum ggml_sort_order {
+    GGML_SORT_ORDER_ASC,
+    GGML_SORT_ORDER_DESC,
+};
+
+kernel void kernel_argsort_f32_i32(
+    global float * src0,
+    ulong          offset0,
+    global int   * dst,
+    ulong          offsetd,
+    const int      ne00,
+    const int      ne00_pad,
+    const int      order,
+    local int    * dst_row
+) {
+    // bitonic sort
+    int col = get_local_id(0);
+    int row = get_group_id(1);
+
+    if (col >= ne00_pad) {
+        return;
+    }
+
+    src0 = (global char  *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    global float * x_row = src0 + row * ne00;
+
+    // initialize indices
+    dst_row[col] = col;
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    for (int k = 2; k <= ne00_pad; k *= 2) {
+        for (int j = k / 2; j > 0; j /= 2) {
+            int ixj = col ^ j;
+            if (ixj > col) {
+                if ((col & k) == 0) {
+                    if (dst_row[col] >= ne00 ||
+                        (dst_row[ixj] < ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj], int);
+                    }
+                } else {
+                    if (dst_row[ixj] >= ne00 ||
+                        (dst_row[col] < ne00 && (order == GGML_SORT_ORDER_ASC ?
+                            x_row[dst_row[col]] < x_row[dst_row[ixj]] :
+                            x_row[dst_row[col]] > x_row[dst_row[ixj]]))
+                    ) {
+                        SWAP(dst_row[col], dst_row[ixj], int);
+                    }
+                }
+            }
+            barrier(CLK_LOCAL_MEM_FENCE);
+        }
+    }
+
+    // copy the result to dst without the padding
+    if (col < ne00) {
+        dst[row * ne00 + col] = dst_row[col];
+    }
+}

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

@@ -0,0 +1,109 @@
+kernel void kernel_concat_f32_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+    int d_ne00, int d_ne01, int d_ne02, // src0->ne[0..2] for the slice
+    int d_ne10, int d_ne11, int d_ne12, // src1->ne[0..2] for the slice (d_ne1X must match d_ne0X on non-concat axes)
+    int d_ne0,  int d_ne1,  int d_ne2,  // dst->ne[0..2] for the slice
+    int dim
+) {
+    global const float * src0 = (global const float*)((global char*)p_src0 + off_src0);
+    global const float * src1 = (global const float*)((global char*)p_src1 + off_src1);
+    global float * dst        = (global float*)((global char*)p_dst + off_dst);
+
+    int i0 = get_global_id(0); // Index along dst's 0th dimension
+    int i1 = get_global_id(1); // Index along dst's 1st dimension
+    int i2 = get_global_id(2); // Index along dst's 2nd dimension
+
+    if (i0 >= d_ne0 || i1 >= d_ne1 || i2 >= d_ne2) {
+        return;
+    }
+
+    ulong dst_idx = (ulong)i2 * d_ne0 * d_ne1 + (ulong)i1 * d_ne0 + i0;
+    ulong src_idx;
+
+    if (dim == 0) {
+        if (i0 < d_ne00) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + (i0 - d_ne00);
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 1) {
+        if (i1 < d_ne01) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+            src_idx = (ulong)i2 * d_ne10 * d_ne11 + (ulong)(i1 - d_ne01) * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    } else if (dim == 2) {
+        if (i2 < d_ne02) { // Data from src0
+            src_idx = (ulong)i2 * d_ne00 * d_ne01 + (ulong)i1 * d_ne00 + i0;
+            dst[dst_idx] = src0[src_idx];
+        } else { // Data from src1
+
+            src_idx = (ulong)(i2 - d_ne02) * d_ne10 * d_ne11 + (ulong)i1 * d_ne10 + i0;
+            dst[dst_idx] = src1[src_idx];
+        }
+    }
+}
+
+kernel void kernel_concat_f32_non_contiguous(
+    global const char * p_src0, ulong off_src0,
+    global const char * p_src1, ulong off_src1,
+    global char * p_dst, ulong off_dst,
+
+    long ne00, long ne01, long ne02, long ne03,
+    ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+
+    ulong nb10, ulong nb11, ulong nb12, ulong nb13, // Strides for src1
+
+    long d_ne0, long d_ne1, long d_ne2, long d_ne3,
+    ulong d_nb0, ulong d_nb1, ulong d_nb2, ulong d_nb3,
+    int dim
+) {
+    global const char * src0_base = p_src0 + off_src0;
+    global const char * src1_base = p_src1 + off_src1;
+    global char * dst_base        = p_dst + off_dst;
+
+    long current_i1 = get_global_id(0); // Index for dst_dim_1
+    long current_i2 = get_global_id(1); // Index for dst_dim_2
+    long current_i3 = get_global_id(2); // Index for dst_dim_3
+
+    if (current_i1 >= d_ne1 || current_i2 >= d_ne2 || current_i3 >= d_ne3) {
+        return;
+    }
+
+    global const float * x_val_ptr;
+    global float * y_val_ptr;
+
+    for (long current_i0 = 0; current_i0 < d_ne0; ++current_i0) {
+        bool use_src0;
+        long s_i0 = current_i0, s_i1 = current_i1, s_i2 = current_i2, s_i3 = current_i3;
+
+        if (dim == 0) {
+            use_src0 = (current_i0 < ne00);
+            if (!use_src0) { s_i0 = current_i0 - ne00; }
+        } else if (dim == 1) {
+            use_src0 = (current_i1 < ne01);
+            if (!use_src0) { s_i1 = current_i1 - ne01; }
+        } else if (dim == 2) {
+            use_src0 = (current_i2 < ne02);
+            if (!use_src0) { s_i2 = current_i2 - ne02; }
+        } else { // dim == 3
+            use_src0 = (current_i3 < ne03);
+            if (!use_src0) { s_i3 = current_i3 - ne03; }
+        }
+
+        if (use_src0) {
+            x_val_ptr = (global const float *)(src0_base + (ulong)s_i3*nb03 + (ulong)s_i2*nb02 + (ulong)s_i1*nb01 + (ulong)s_i0*nb00);
+        } else {
+            x_val_ptr = (global const float *)(src1_base + (ulong)s_i3*nb13 + (ulong)s_i2*nb12 + (ulong)s_i1*nb11 + (ulong)s_i0*nb10);
+        }
+
+        y_val_ptr = (global float *)(dst_base + (ulong)current_i3*d_nb3 + (ulong)current_i2*d_nb2 + (ulong)current_i1*d_nb1 + (ulong)current_i0*d_nb0);
+        *y_val_ptr = *x_val_ptr;
+    }
+}

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

@@ -0,0 +1,72 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// div
+//------------------------------------------------------------------------------
+kernel void kernel_div(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        ulong nb0,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst + offsetd;
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    int i13 = i03 % ne13;
+    int i12 = i02 % ne12;
+    int i11 = i01 % ne11;
+
+    global char * src0_ptr = src0 + i03*nb03 + i02*nb02 + i01*nb01;
+    global char * src1_ptr = src1 + i13*nb13 + i12*nb12 + i11*nb11;
+    global char * dst_ptr  = dst  + i03*nb3  + i02*nb2  + i01*nb1;
+
+    for (int i0 = get_local_id(0); i0 < ne0; i0 += get_local_size(0)) {
+        const int i10 = i0 % ne10;
+        *((global float *)(dst_ptr + i0*nb0)) = *((global float *)(src0_ptr + i0*nb00)) / *((global float *)(src1_ptr + i10*nb10));
+    }
+}
+
+// assumption: src1 is a row
+// broadcast src1 into src0
+kernel void kernel_div_row(
+        global float4 * src0,
+        ulong offset0,
+        global float4 * src1,
+        ulong offset1,
+        global float4 * dst,
+        ulong offsetd,
+        int ne
+) {
+    src0 = (global float4*)((global char*)src0 + offset0);
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst = (global float4*)((global char*)dst + offsetd);
+
+    // This performs better than using %.
+    uint gid = get_global_id(0);
+    uint idx1 = gid - (gid/ne)*ne; // get_global_id(0) % ne
+    dst[gid] = src0[gid] / src1[idx1];
+}

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

@@ -0,0 +1,72 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(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")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Workgroup must be a subgroup
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_32
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_group_norm(
+        global float * src0,
+        ulong offset0,
+        global float * dst,
+        ulong offsetd,
+        int ne,
+        int group_size,
+        float eps
+) {
+    src0 = (global float  *)((global char *)src0 + offset0);
+    dst  = (global float *)((global char *)dst  + offsetd);
+
+    int start = get_group_id(0) * group_size;
+    int end   = start + group_size;
+
+    start += get_local_id(0);
+
+    if (end >= ne) {
+        end = ne;
+    }
+
+    float tmp = 0.0f;
+
+    for (int j = start; j < end; j += get_local_size(0)) {
+        tmp += src0[j];
+    }
+
+    tmp = sub_group_reduce_add(tmp);
+
+    const float mean = tmp / group_size;
+    tmp = 0.0f;
+
+    for (int j = start; j < end; j += get_local_size(0)) {
+        float xi = src0[j] - mean;
+        dst[j] = xi;
+        tmp += xi * xi;
+    }
+
+    tmp = sub_group_reduce_add(tmp);
+
+    const float variance = tmp / group_size;
+    const float scale = 1.0f/sqrt(variance + eps);
+    for (int j = start; j < end; j += get_local_size(0)) {
+        dst[j] *= scale;
+    }
+}