CANN: Simplify the environment variable setting(#13104 )

* Simplify the environment variable setting to specify the memory pool type. * Adjust the GGML_CANN_ASYNC_MODE setting to accept yes, enable, 1, or on (case-insensitive) as valid options. * update * fix CI * update * delete whitespace * fix according to review * update CANN.md * update CANN.md
webui: fix sidebar being covered by main content (#14082 )
2025-06-09 19:47:39 +08:00 · 2025-06-09 12:01:17 +02:00 · 2025-06-09 12:57:58 +03:00 · 2025-06-09 11:47:07 +02:00 · 2025-06-09 05:15:31 +01:00 · 2025-06-09 11:20:06 +08:00
204 changed files with 11367 additions and 7011 deletions
@@ -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
@@ -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
@@ -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)
@@ -839,12 +839,12 @@ 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 +878,7 @@ 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 ^
@@ -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 ^
@@ -180,7 +180,7 @@ jobs:


  server-windows:
-    runs-on: windows-2019
+    runs-on: windows-2022

    steps:
      - name: Clone
@@ -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
 #
@@ -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/`.

@@ -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 ${GG_BUILD_SYCL} ]; then
@@ -58,23 +58,20 @@ add_library(${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)

@@ -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"); }
        }
@@ -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;
        }
    }
@@ -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>
@@ -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");
    }
@@ -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;
@@ -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);
@@ -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 {
@@ -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;

@@ -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 {
@@ -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>
@@ -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);
@@ -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());
        }
@@ -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."):
@@ -3679,7 +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

@@ -3687,11 +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)
@@ -3742,6 +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:
@@ -3761,62 +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")
@@ -3836,7 +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

@@ -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}")
@@ -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.
@@ -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

@@ -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.
@@ -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 {
@@ -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
@@ -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);

@@ -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);
 }
@@ -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] {
@@ -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);
            }
        }
    }
@@ -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);
@@ -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
@@ -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();

@@ -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;
        }
    }

@@ -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);
@@ -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
@@ -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");
@@ -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) {
@@ -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());
@@ -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()

@@ -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()
@@ -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,
@@ -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
@@ -224,6 +226,7 @@ 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 +290,20 @@ 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("")
@@ -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) {
@@ -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()
@@ -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");
    }
@@ -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));
 }
@@ -82,13 +82,8 @@ 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)$"))
-
+    if (GGML_SYSTEM_ARCH STREQUAL "ARM")
        message(STATUS "ARM detected")
-
        if (MSVC AND NOT CMAKE_C_COMPILER_ID STREQUAL "Clang")
            message(FATAL_ERROR "MSVC is not supported for ARM, use clang")
        else()
@@ -170,12 +165,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                endforeach()
            endif()
        endif()
-    elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
-            (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
-            CMAKE_SYSTEM_PROCESSOR MATCHES "^(x86_64|i686|AMD64|amd64)$"))
-
+    elseif (GGML_SYSTEM_ARCH STREQUAL "x86")
        message(STATUS "x86 detected")
-
        if (MSVC)
            # instruction set detection for MSVC only
            if (GGML_NATIVE)
@@ -318,7 +309,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
            set_target_properties(${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 "${CMAKE_SYSTEM_PROCESSOR} " STREQUAL "powerpc ")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "PowerPC")
        message(STATUS "PowerPC detected")
        if (GGML_NATIVE)
            if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
@@ -327,7 +318,8 @@ 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)
@@ -344,9 +336,8 @@ 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 ARCH_FLAGS -march=loongarch64)
        if (GGML_LASX)
            list(APPEND ARCH_FLAGS -mlasx)
@@ -354,8 +345,8 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
        if (GGML_LSX)
            list(APPEND ARCH_FLAGS -mlsx)
        endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "riscv64")
-        message(STATUS "RISC-V detected")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "riscv64")
+        message(STATUS "riscv64 detected")
        if (GGML_RVV)
            if (GGML_XTHEADVECTOR)
                list(APPEND ARCH_FLAGS -march=rv64gc_xtheadvector -mabi=lp64d)
@@ -365,7 +356,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                list(APPEND ARCH_FLAGS -march=rv64gcv -mabi=lp64d)
            endif()
        endif()
-    elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
+    elseif (GGML_SYSTEM_ARCH STREQUAL "s390x")
        message(STATUS "s390x detected")
        file(READ "/proc/cpuinfo" CPUINFO_CONTENTS)
        string(REGEX REPLACE "machine[ \t\r\n]*=[ \t\r\n]*([0-9]+)" "\\1" S390X_M ${CPUINFO_CONTENTS})
@@ -6995,7 +6995,11 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

 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) {
    assert(n % QK_K == 0);
+#ifdef __ARM_FEATURE_MATMUL_INT8
+    assert((nrc == 2) || (nrc == 1));
+#else
    assert(nrc == 1);
+#endif
    UNUSED(nrc);
    UNUSED(bx);
    UNUSED(by);
@@ -7012,6 +7016,146 @@ void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi

    uint32_t utmp[4];

+#if defined(__ARM_FEATURE_MATMUL_INT8)
+    if (nrc == 2) {
+        const block_q4_K * GGML_RESTRICT x0 = x;
+        const block_q4_K * GGML_RESTRICT x1 = (const block_q4_K *) ((const uint8_t *)vx + bx);
+        const block_q8_K * GGML_RESTRICT y0 = y;
+        const block_q8_K * GGML_RESTRICT y1 = (const block_q8_K *) ((const uint8_t *)vy + by);
+
+        const uint8x16_t m4b = vdupq_n_u8(0x0f);
+
+        float32x4_t vfsum = vdupq_n_f32(0.0f);
+
+        for (int i = 0; i < nb; ++i, ++x0, ++x1, ++y0, ++y1) {
+            const uint8_t * GGML_RESTRICT qx0 = x0->qs;
+            const uint8_t * GGML_RESTRICT qx1 = x1->qs;
+            const  int8_t * GGML_RESTRICT qy0 = y0->qs;
+            const  int8_t * GGML_RESTRICT qy1 = y1->qs;
+
+            // decode scales and mins
+            int8_t x0_scales[8], x1_scales[8];
+            int16x8_t x0_mins, x1_mins;
+            {
+                uint32_t scales_mins[3];
+                memcpy(scales_mins, x0->scales, 12);
+                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
+                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
+                const uint32x2_t mins = {mins_0_3, mins_4_7};
+                x0_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
+                uint32_t scales[2];
+                scales[0] = scales_mins[0] & kmask1; // scales 0~3
+                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
+                memcpy(x0_scales, scales, 8);
+            }
+            {
+                uint32_t scales_mins[3];
+                memcpy(scales_mins, x1->scales, 12);
+                const uint32_t mins_0_3 = scales_mins[1] & kmask1;
+                const uint32_t mins_4_7 = ((scales_mins[2] >> 4) & kmask2) | (((scales_mins[1] >> 6) & kmask3) << 4);
+                const uint32x2_t mins = {mins_0_3, mins_4_7};
+                x1_mins = vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(mins)));
+                uint32_t scales[2];
+                scales[0] = scales_mins[0] & kmask1; // scales 0~3
+                scales[1] = (scales_mins[2] & kmask2) | (((scales_mins[0] >> 6) & kmask3) << 4); // scales 4~7
+                memcpy(x1_scales, scales, 8);
+            }
+
+            int32x4_t visum = {0};
+
+            // process 64 data points per iteration, totally 256 data points
+            for (int j = 0; j < QK_K / 64; ++j, qx0 += 32, qx1 += 32, qy0 += 64, qy1 += 64) {
+                const int8x16x4_t vy0 = vld1q_s8_x4(qy0);
+                const int8x16x4_t vy1 = vld1q_s8_x4(qy1);
+
+                int8x16_t vx0[4], vx1[4];
+                {
+                    const uint8x16x2_t vv = vld1q_u8_x2(qx0);
+                    vx0[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
+                    vx0[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
+                    vx0[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
+                    vx0[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
+                }
+                {
+                    const uint8x16x2_t vv = vld1q_u8_x2(qx1);
+                    vx1[0] = vreinterpretq_s8_u8(vandq_u8(vv.val[0], m4b));
+                    vx1[1] = vreinterpretq_s8_u8(vandq_u8(vv.val[1], m4b));
+                    vx1[2] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[0], 4));
+                    vx1[3] = vreinterpretq_s8_u8(vshrq_n_u8(vv.val[1], 4));
+                }
+
+                // process 32 data points (share same block scale) per iteration
+                for (int k = 0; k < 2; ++k) {
+                    const int blk = j * 2 + k;
+                    const int32x4_t block_scale = {
+                        x0_scales[blk],
+                        x0_scales[blk],
+                        x1_scales[blk],
+                        x1_scales[blk],
+                    };
+
+                    int32x4_t vr = {0};
+                    for (int l = 0; l < 2; ++l) {
+                        const int idx = k * 2 + l;
+                        const int64x2_t vx0_s64 = vreinterpretq_s64_s8(vx0[idx]);
+                        const int64x2_t vx1_s64 = vreinterpretq_s64_s8(vx1[idx]);
+                        const int64x2_t vy0_s64 = vreinterpretq_s64_s8(vy0.val[idx]);
+                        const int64x2_t vy1_s64 = vreinterpretq_s64_s8(vy1.val[idx]);
+                        const int8x16_t vx_l = vreinterpretq_s8_s64(vzip1q_s64(vx0_s64, vx1_s64));
+                        const int8x16_t vx_h = vreinterpretq_s8_s64(vzip2q_s64(vx0_s64, vx1_s64));
+                        const int8x16_t vy_l = vreinterpretq_s8_s64(vzip1q_s64(vy0_s64, vy1_s64));
+                        const int8x16_t vy_h = vreinterpretq_s8_s64(vzip2q_s64(vy0_s64, vy1_s64));
+                        vr = vmmlaq_s32(vr, vx_l, vy_l);
+                        vr = vmmlaq_s32(vr, vx_h, vy_h);
+                    }
+                    // apply block scale, will NOT overflow
+                    // block_scale * sum_256(int4*int8) <= 2^(8+8+4+8) = 28 bits
+                    visum = vmlaq_s32(visum, vr, block_scale);
+                }
+            }
+
+            // adjust bias, apply superblock scale
+            {
+                int32_t bias[4];
+                // no obvious uplift from sve sdot-16, just use neon mul add
+                const int16x8_t y0_sums = vpaddq_s16(vld1q_s16(y0->bsums), vld1q_s16(y0->bsums+8));
+                const int16x8_t y1_sums = vpaddq_s16(vld1q_s16(y1->bsums), vld1q_s16(y1->bsums+8));
+                bias[0] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x0_mins)),
+                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x0_mins))));
+                bias[1] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x0_mins)),
+                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x0_mins))));
+                bias[2] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y0_sums), vget_low_s16(x1_mins)),
+                                               vmull_s16(vget_high_s16(y0_sums), vget_high_s16(x1_mins))));
+                bias[3] = vaddvq_s32(vaddq_s32(vmull_s16(vget_low_s16(y1_sums), vget_low_s16(x1_mins)),
+                                               vmull_s16(vget_high_s16(y1_sums), vget_high_s16(x1_mins))));
+                const float32x4_t dmins = {
+                    GGML_FP16_TO_FP32(x0->dmin) * y0->d,
+                    GGML_FP16_TO_FP32(x0->dmin) * y1->d,
+                    GGML_FP16_TO_FP32(x1->dmin) * y0->d,
+                    GGML_FP16_TO_FP32(x1->dmin) * y1->d,
+                };
+                vfsum = vmlsq_f32(vfsum, vcvtq_f32_s32(vld1q_s32(bias)), dmins);
+
+                const float32x4_t superblock_scale = {
+                    GGML_FP16_TO_FP32(x0->d) * y0->d,
+                    GGML_FP16_TO_FP32(x0->d) * y1->d,
+                    GGML_FP16_TO_FP32(x1->d) * y0->d,
+                    GGML_FP16_TO_FP32(x1->d) * y1->d,
+                };
+                vfsum = vmlaq_f32(vfsum, vcvtq_f32_s32(visum), superblock_scale);
+            }
+        }
+
+        // vfsum = ABCD -> ACBD
+        // AC -> s, BD -> (s+bs)
+        vfsum = vzip1q_f32(vfsum, vextq_f32(vfsum, vfsum, 2));
+        vst1_f32(s,      vget_low_f32 (vfsum));
+        vst1_f32(s + bs, vget_high_f32(vfsum));
+
+        return;
+    }
+#endif
+
 #ifdef __ARM_FEATURE_SVE
    float sumf = 0;
    for (int i = 0; i < nb; ++i) {
@@ -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;
@@ -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;
@@ -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

@@ -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;
@@ -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
@@ -635,6 +635,7 @@ 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;
@@ -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");

@@ -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;

@@ -1065,6 +1065,10 @@ 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 +1144,6 @@ 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 +1426,6 @@ 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 +1747,7 @@ 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 +2642,8 @@ 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 +2662,7 @@ 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 +2997,12 @@ 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 +3269,9 @@ 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) {
@@ -32,6 +32,8 @@
 extern "C" {
 #endif

+void ggml_print_backtrace(void);
+
 #ifndef MIN
 #    define MIN(a, b) ((a) < (b) ? (a) : (b))
 #endif
@@ -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) {
@@ -3328,14 +3328,14 @@ 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 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 + 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 +3354,7 @@ 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;
            }
        }
    }
@@ -3634,9 +3634,6 @@ kernel void kernel_flash_attn_ext(

    // 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
@@ -3657,12 +3654,12 @@ kernel void kernel_flash_attn_ext(
                const float M0 = ss[j*TS +         1];
                const float M1 = ss[j*TS + sg*SH + 1];

-                M = max(M0, M1);
+                const float M = max(M0, M1);

                const float ms0 = exp(M0 - M);
                const float ms1 = exp(M1 - M);

-                S = S0*ms0 + S1*ms1;
+                const float S = S0*ms0 + S1*ms1;

                if (tiisg == 0) {
                    ss[j*TS + 0] = S;
@@ -3701,16 +3698,18 @@ kernel void kernel_flash_attn_ext(
        }
    }

-    device float4 * dst4 = (device float4 *) dst;
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    threadgroup s_t * sf = (threadgroup s_t *) (shmem_f16 + 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 +3718,22 @@ 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, \
+    float,  float4,    simdgroup_float8x8
+    //half,   half4,     simdgroup_half8x8
+
+#define FA_TYPES_BF \
+    bfloat, bfloat4,   simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    bfloat, bfloat4x4, simdgroup_bfloat8x8, \
+    float,             simdgroup_float8x8,  \
+    float,             simdgroup_float8x8,  \
+    float,  float4,    simdgroup_float8x8
+    //half,   half4,     simdgroup_half8x8

 typedef decltype(kernel_flash_attn_ext<FA_TYPES, half4x4, 1, dequantize_f16, half4x4, 1, dequantize_f16, 64, 64>) flash_attn_ext_t;

@@ -3739,15 +3748,15 @@ 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 +3810,7 @@ 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 +3857,12 @@ 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 +4167,7 @@ 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;

@@ -95,6 +95,12 @@ set(GGML_OPENCL_KERNELS
    sub
    sum_rows
    transpose
+    concat
+    tsembd
+    upscale
+    tanh
+    pad
+    repeat
 )

 foreach (K ${GGML_OPENCL_KERNELS})
@@ -315,6 +315,12 @@ struct ggml_backend_opencl_context {
    cl_program program_softmax_4_f16;
    cl_program program_argsort_f32_i32;
    cl_program program_sum_rows_f32;
+    cl_program program_repeat;
+    cl_program program_pad;
+    cl_program program_tanh;
+    cl_program program_upscale;
+    cl_program program_concat;
+    cl_program program_tsembd;

    cl_kernel kernel_add, kernel_add_row;
    cl_kernel kernel_mul, kernel_mul_row;
@@ -351,6 +357,15 @@ struct ggml_backend_opencl_context {
    cl_kernel kernel_im2col_f32, kernel_im2col_f16;
    cl_kernel kernel_argsort_f32_i32;
    cl_kernel kernel_sum_rows_f32;
+    cl_kernel kernel_repeat;
+    cl_kernel kernel_pad;
+    cl_kernel kernel_tanh_f32_nd;
+    cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_upscale;
+    cl_kernel kernel_upscale_bilinear;
+    cl_kernel kernel_concat_f32_contiguous;
+    cl_kernel kernel_concat_f32_non_contiguous;
+    cl_kernel kernel_timestep_embedding;

 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    // Transpose kernels
@@ -1097,6 +1112,150 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
        GGML_LOG_CONT(".");
    }

+        // repeat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "repeat.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("repeat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_repeat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_repeat = clCreateKernel(backend_ctx->program_repeat, "kernel_repeat", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: repeat kernel source not found or empty. Repeat operations will not be available.\n");
+            backend_ctx->program_repeat = nullptr;
+            backend_ctx->kernel_repeat = nullptr;
+        }
+    }
+
+    // pad
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "pad.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("pad.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_pad =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_pad = clCreateKernel(backend_ctx->program_pad, "kernel_pad", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: pad kernel source not found or empty. Pad operations will not be available.\n");
+            backend_ctx->program_pad = nullptr;
+            backend_ctx->kernel_pad = nullptr;
+        }
+    }
+
+    // tanh
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tanh.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("tanh.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tanh =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_tanh_f32_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_tanh_f16_nd = clCreateKernel(backend_ctx->program_tanh, "kernel_tanh_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: tanh kernel source not found or empty. Tanh operation will not be available.\n");
+            backend_ctx->program_tanh = nullptr;
+            backend_ctx->kernel_tanh_f32_nd = nullptr;
+            backend_ctx->kernel_tanh_f16_nd = nullptr;
+        }
+    }
+
+    // upscale
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "upscale.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("upscale.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_upscale =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_upscale = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale", &err), err));
+            if (backend_ctx->program_upscale) {
+                 cl_int err_bilinear;
+                 backend_ctx->kernel_upscale_bilinear = clCreateKernel(backend_ctx->program_upscale, "kernel_upscale_bilinear", &err_bilinear);
+                 if (err_bilinear != CL_SUCCESS) {
+                    GGML_LOG_WARN("ggml_opencl: kernel_upscale_bilinear not found in upscale.cl. Bilinear upscale will not be available. Error: %d\n", err_bilinear);
+                    backend_ctx->kernel_upscale_bilinear = nullptr;
+                 }
+            } else {
+                backend_ctx->kernel_upscale_bilinear = nullptr;
+            }
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: upscale kernel source not found or empty. Upscale operations will not be available.\n");
+            backend_ctx->program_upscale = nullptr;
+            backend_ctx->kernel_upscale = nullptr;
+            backend_ctx->kernel_upscale_bilinear = nullptr;
+        }
+    }
+
+    // concat
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "concat.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("concat.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_concat =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+            CL_CHECK((backend_ctx->kernel_concat_f32_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_contiguous", &err), err));
+            CL_CHECK((backend_ctx->kernel_concat_f32_non_contiguous = clCreateKernel(backend_ctx->program_concat, "kernel_concat_f32_non_contiguous", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: concat kernel source not found or empty. Concat operations will not be available.\n");
+            backend_ctx->program_concat = nullptr;
+            backend_ctx->kernel_concat_f32_contiguous = nullptr;
+            backend_ctx->kernel_concat_f32_non_contiguous = nullptr;
+        }
+    }
+
+    // timestep_embedding
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "tsembd.cl.h"
+        };
+#else
+
+        const std::string kernel_src = read_file("tsembd.cl");
+#endif
+        if (!kernel_src.empty()) {
+            backend_ctx->program_tsembd =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_timestep_embedding = clCreateKernel(backend_ctx->program_tsembd, "kernel_timestep_embedding", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: timestep_embedding kernel source not found or empty. This op will not be available.\n");
+            backend_ctx->program_tsembd = nullptr;
+            backend_ctx->kernel_timestep_embedding = nullptr;
+        }
+    }
+
    // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
    // transpose
@@ -1863,7 +2022,12 @@ static bool ggml_backend_opencl_cpy_tensor_async(ggml_backend_t backend, const g
 }

 static void ggml_backend_opencl_synchronize(ggml_backend_t backend) {
-    GGML_UNUSED(backend);
+    auto * backend_ctx = static_cast<ggml_backend_opencl_context *>(backend->context);
+
+    cl_event evt;
+    CL_CHECK(clEnqueueBarrierWithWaitList(backend_ctx->queue, 0, nullptr, &evt));
+    CL_CHECK(clWaitForEvents(1, &evt));
+    CL_CHECK(clReleaseEvent(evt));
 }

 // Syncronizes the 'backend_ctx's device with others so that commands
@@ -1976,9 +2140,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_RELU:
                case GGML_UNARY_OP_GELU_QUICK:
-                    return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
+                   return ggml_is_contiguous(op->src[0]) && op->src[0]->type == GGML_TYPE_F32;
                case GGML_UNARY_OP_SIGMOID:
                    return ggml_is_contiguous(op->src[0]);
+                case GGML_UNARY_OP_TANH:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
                default:
                    return false;
            }
@@ -1988,6 +2155,17 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
        case GGML_OP_NORM:
        case GGML_OP_RMS_NORM:
            return true;
+        case GGML_OP_REPEAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32; // Assuming F32 for now, can be expanded
+        case GGML_OP_PAD:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32 &&
+                   op->src[0]->ne[3] == 1 && op->ne[3] == 1;
+        case GGML_OP_UPSCALE:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_CONCAT:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            return op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
        case GGML_OP_GROUP_NORM:
            return ggml_is_contiguous(op->src[0]);
        case GGML_OP_MUL_MAT:
@@ -2052,7 +2230,7 @@ static ggml_backend_i ggml_backend_opencl_i = {
    /* .set_tensor_async        = */ NULL,  /* ggml_backend_opencl_set_tensor_async */
    /* .get_tensor_async        = */ NULL,  /* ggml_backend_opencl_get_tensor_async */
    /* .cpy_tensor_async        = */ NULL,  /* ggml_backend_opencl_cpy_tensor_async */
-    /* .synchronize             = */ NULL,  /* ggml_backend_opencl_synchronize */
+    /* .synchronize             = */ ggml_backend_opencl_synchronize,
    /* .graph_plan_create       = */ NULL,
    /* .graph_plan_free         = */ NULL,
    /* .graph_plan_update       = */ NULL,
@@ -4108,6 +4286,536 @@ static void ggml_cl_group_norm(ggml_backend_t backend, const ggml_tensor * src0,
 #endif
 }

+static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_tanh_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_tanh_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_tanh");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0]; const int ne01 = src0->ne[1]; const int ne02 = src0->ne[2]; const int ne03 = src0->ne[3];
+    const cl_ulong nb00 = src0->nb[0]; const cl_ulong nb01 = src0->nb[1]; const cl_ulong nb02 = src0->nb[2]; const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0]; const int ne11 = dst->ne[1]; const int ne12 = dst->ne[2]; const int ne13 = dst->ne[3];
+    const cl_ulong nb10 = dst->nb[0]; const cl_ulong nb11 = dst->nb[1]; const cl_ulong nb12 = dst->nb[2]; const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(dst->type == src0->type);
+
+    UNUSED(src1_shape_def);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_repeat == nullptr) {
+        GGML_LOG_WARN("%s: repeat kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int src0_ne0 = src0->ne[0]; const int src0_ne1 = src0->ne[1]; const int src0_ne2 = src0->ne[2]; const int src0_ne3 = src0->ne[3];
+    const cl_ulong src0_nb0 = src0->nb[0]; const cl_ulong src0_nb1 = src0->nb[1]; const cl_ulong src0_nb2 = src0->nb[2]; const cl_ulong src0_nb3 = src0->nb[3];
+
+    const int dst_ne0 = dst->ne[0]; const int dst_ne1 = dst->ne[1]; const int dst_ne2 = dst->ne[2]; const int dst_ne3 = dst->ne[3];
+    const cl_ulong dst_nb0 = dst->nb[0]; const cl_ulong dst_nb1 = dst->nb[1]; const cl_ulong dst_nb2 = dst->nb[2]; const cl_ulong dst_nb3 = dst->nb[3];
+
+    cl_kernel kernel = backend_ctx->kernel_repeat;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &src0_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &src0_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &src0_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &src0_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong),  &src0_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong),  &src0_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong), &src0_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &src0_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &dst_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &dst_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &dst_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dst_ne3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &dst_nb0));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &dst_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &dst_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &dst_nb3));
+
+    size_t gws0 = dst_ne1 > 0 ? (size_t)dst_ne1 : 1;
+    size_t gws1 = dst_ne2 > 0 ? (size_t)dst_ne2 : 1;
+    size_t gws2 = dst_ne3 > 0 ? (size_t)dst_ne3 : 1;
+
+    size_t global_work_size[] = { gws0, gws1, gws2 };
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_pad(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(src0->ne[3] == 1 && dst->ne[3] == 1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_pad == nullptr) {
+        GGML_LOG_WARN("%s: pad kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int s_ne0 = src0->ne[0];
+    const int s_ne1 = src0->ne[1];
+    const int s_ne2 = src0->ne[2];
+
+    const int d_ne0 = dst->ne[0];
+    const int d_ne1 = dst->ne[1];
+    const int d_ne2 = dst->ne[2];
+
+    cl_kernel kernel = backend_ctx->kernel_pad;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &s_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &s_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &s_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne0));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne1));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne2));
+
+    size_t lws0 = 64;
+    size_t gws0 = (( (size_t)d_ne0 + lws0 - 1 ) / lws0) * lws0;
+
+    size_t global_work_size[] = { gws0, (size_t)d_ne1, (size_t)d_ne2 };
+    size_t local_work_size[]  = { lws0, 1, 1 };
+
+    size_t * local_work_size_ptr = local_work_size;
+     if (d_ne0 % lws0 != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, global_work_size, local_work_size_ptr ? local_work_size : (size_t[3]){0,0,0}, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_upscale(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    const ggml_scale_mode mode = (ggml_scale_mode) ggml_get_op_params_i32(dst, 0);
+    cl_kernel kernel = nullptr;
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        kernel = backend_ctx->kernel_upscale;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: nearest upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        kernel = backend_ctx->kernel_upscale_bilinear;
+        if (kernel == nullptr) {
+            GGML_LOG_WARN("%s: bilinear upscale kernel not available, skipping OpenCL execution.\n", __func__);
+            return;
+        }
+    } else {
+        GGML_LOG_WARN("%s: unsupported upscale mode %d, skipping OpenCL execution.\n", __func__, mode);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne00_src = src0->ne[0];
+    const int ne01_src = src0->ne[1];
+
+    const int ne10_dst = dst->ne[0];
+    const int ne11_dst = dst->ne[1];
+    const int ne12_dst = dst->ne[2];
+    const int ne13_dst = dst->ne[3];
+
+    const float sf0 = (float)dst->ne[0] / src0->ne[0];
+    const float sf1 = (float)dst->ne[1] / src0->ne[1];
+    const float sf2 = (float)dst->ne[2] / src0->ne[2];
+    const float sf3 = (float)dst->ne[3] / src0->ne[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_ulong),  &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_ulong),  &nb02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong),  &nb03));
+
+    if (mode == GGML_SCALE_MODE_NEAREST) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf3));
+    } else if (mode == GGML_SCALE_MODE_BILINEAR) {
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &ne00_src));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &ne01_src));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne10_dst));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &ne11_dst));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &ne12_dst));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne13_dst));
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(float),    &sf0));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(float),    &sf1));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(float),    &sf2));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(float),    &sf3));
+    }
+
+
+    size_t dst_total_elements = (size_t)ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+    if (dst_total_elements == 0) {
+        return;
+    }
+    size_t global_work_size[] = { dst_total_elements, 1, 1 };
+    size_t local_work_size_pref = 256;
+    size_t local_work_size[] = { MIN(local_work_size_pref, dst_total_elements), 1, 1};
+
+    size_t * local_work_size_ptr = local_work_size;
+    if (dst_total_elements % local_work_size[0] != 0 && !backend_ctx->non_uniform_workgroups) {
+        local_work_size_ptr = nullptr;
+    }
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, &evt));
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], 1, 1};
+    size_t profiling_lws[3] = {local_work_size_ptr ? local_work_size[0] : 0, 1, 1};
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 1, NULL, global_work_size, local_work_size_ptr, 0, NULL, NULL));
+#endif
+}
+
+static void ggml_cl_concat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(src1);
+    GGML_ASSERT(src1->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_concat_f32_contiguous == nullptr || backend_ctx->kernel_concat_f32_non_contiguous == nullptr) {
+        GGML_LOG_WARN("%s: concat kernels not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra0_cl = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra1_cl = (ggml_tensor_extra_cl *)src1->extra;
+    ggml_tensor_extra_cl * extrad_cl = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra0_cl->offset + src0->view_offs;
+    cl_ulong off_src1 = extra1_cl->offset + src1->view_offs;
+    cl_ulong off_dst  = extrad_cl->offset + dst->view_offs;
+
+    const int32_t dim = ((const int32_t *) dst->op_params)[0];
+    GGML_ASSERT(dim >= 0 && dim <= 3);
+
+    if (ggml_is_contiguous(src0) && ggml_is_contiguous(src1) && ggml_is_contiguous(dst)) {
+        if (dim == 3) {
+
+            size_t nbytes_src0 = ggml_nbytes(src0);
+            size_t nbytes_src1 = ggml_nbytes(src1);
+
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra0_cl->data_device, extrad_cl->data_device,
+                                         off_src0, off_dst, nbytes_src0, 0, NULL, NULL));
+            CL_CHECK(clEnqueueCopyBuffer(queue, extra1_cl->data_device, extrad_cl->data_device,
+                                         off_src1, off_dst + nbytes_src0, nbytes_src1, 0, NULL, NULL));
+        } else {
+
+            cl_kernel kernel = backend_ctx->kernel_concat_f32_contiguous;
+            size_t global_work_size[3];
+
+            for (int i3 = 0; i3 < dst->ne[3]; ++i3) {
+                cl_ulong current_off_src0 = off_src0 + (i3 * src0->nb[3]);
+                cl_ulong current_off_src1 = off_src1 + (i3 * src1->nb[3]);
+                cl_ulong current_off_dst  = off_dst  + (i3 * dst->nb[3]);
+
+                int d_ne00 = src0->ne[0]; int d_ne01 = src0->ne[1]; int d_ne02 = src0->ne[2];
+                int d_ne10 = src1->ne[0]; int d_ne11 = src1->ne[1]; int d_ne12 = src1->ne[2];
+                int d_ne0  = dst->ne[0];  int d_ne1  = dst->ne[1];  int d_ne2  = dst->ne[2];
+
+                CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &current_off_src0));
+                CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &current_off_src1));
+                CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+                CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &current_off_dst));
+                CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &d_ne00));
+                CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),       &d_ne01));
+                CL_CHECK(clSetKernelArg(kernel, 8, sizeof(int),       &d_ne02));
+                CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),       &d_ne10));
+                CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &d_ne11));
+                CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),      &d_ne12));
+                CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),      &d_ne0));
+                CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &d_ne1));
+                CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &d_ne2));
+                CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &dim));
+
+                global_work_size[0] = d_ne0;
+                global_work_size[1] = d_ne1;
+                global_work_size[2] = d_ne2;
+
+                CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size, NULL, 0, NULL, NULL));
+            }
+        }
+    } else {
+        cl_kernel kernel = backend_ctx->kernel_concat_f32_non_contiguous;
+
+        long ne00 = src0->ne[0], ne01 = src0->ne[1], ne02 = src0->ne[2], ne03 = src0->ne[3];
+        cl_ulong nb00 = src0->nb[0], nb01 = src0->nb[1], nb02 = src0->nb[2], nb03 = src0->nb[3];
+
+        cl_ulong nb10 = src1->nb[0], nb11 = src1->nb[1], nb12 = src1->nb[2], nb13 = src1->nb[3];
+
+        long d_ne0 = dst->ne[0], d_ne1 = dst->ne[1], d_ne2 = dst->ne[2], d_ne3 = dst->ne[3];
+        cl_ulong d_nb0 = dst->nb[0], d_nb1 = dst->nb[1], d_nb2 = dst->nb[2], d_nb3 = dst->nb[3];
+
+
+        CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra0_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+        CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra1_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_src1));
+        CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),    &extrad_cl->data_device));
+        CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong),  &off_dst));
+
+        CL_CHECK(clSetKernelArg(kernel, 6, sizeof(long),      &ne00));
+        CL_CHECK(clSetKernelArg(kernel, 7, sizeof(long),      &ne01));
+        CL_CHECK(clSetKernelArg(kernel, 8, sizeof(long),      &ne02));
+        CL_CHECK(clSetKernelArg(kernel, 9, sizeof(long),      &ne03));
+        CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),    &nb00));
+        CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),    &nb01));
+        CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong),    &nb02));
+        CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong),    &nb03));
+
+        CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong),    &nb10));
+        CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong),    &nb11));
+        CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),    &nb12));
+        CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),    &nb13));
+
+        CL_CHECK(clSetKernelArg(kernel, 18, sizeof(long),     &d_ne0));
+        CL_CHECK(clSetKernelArg(kernel, 19, sizeof(long),     &d_ne1));
+        CL_CHECK(clSetKernelArg(kernel, 20, sizeof(long),     &d_ne2));
+        CL_CHECK(clSetKernelArg(kernel, 21, sizeof(long),     &d_ne3));
+        CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong),    &d_nb0));
+        CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong),    &d_nb1));
+        CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong),    &d_nb2));
+        CL_CHECK(clSetKernelArg(kernel, 25, sizeof(cl_ulong),    &d_nb3));
+        CL_CHECK(clSetKernelArg(kernel, 26, sizeof(int),      &dim));
+
+        size_t global_work_size_nc[] = { d_ne1 > 0 ? (size_t)d_ne1 : 1,
+                                         d_ne2 > 0 ? (size_t)d_ne2 : 1,
+                                         d_ne3 > 0 ? (size_t)d_ne3 : 1 };
+
+        CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 3, NULL, global_work_size_nc, NULL, 0, NULL, NULL));
+    }
+}
+
+static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor * src0, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+    cl_command_queue queue = backend_ctx->queue;
+
+    if (backend_ctx->kernel_timestep_embedding == nullptr) {
+        GGML_LOG_WARN("%s: timestep_embedding kernel not available, skipping OpenCL execution.\n", __func__);
+        return;
+    }
+
+    ggml_tensor_extra_cl * extra_src0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extra_dst  = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong off_src0 = extra_src0->offset + src0->view_offs;
+    cl_ulong off_dst  = extra_dst->offset  + dst->view_offs;
+
+    const int logical_dim = dst->op_params[0];
+    const int max_period  = dst->op_params[1];
+    const int dst_nb1_bytes = dst->nb[1];
+
+    cl_kernel kernel = backend_ctx->kernel_timestep_embedding;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),    &extra_src0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong),  &off_src0));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),    &extra_dst->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong),  &off_dst));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),       &dst_nb1_bytes));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),       &logical_dim));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),       &max_period));
+
+    size_t gws0 = (size_t)(((logical_dim + 1) / 2) + 1);
+
+    size_t gws1 = (size_t)src0->ne[0];
+
+    size_t global_work_size[] = {gws0, gws1, 1};
+
+#ifdef GGML_OPENCL_PROFILING
+    cl_event evt;
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, &evt)); // Pass 2 for 2D problem
+
+    g_profiling_info.emplace_back();
+    size_t profiling_gws[3] = {global_work_size[0], global_work_size[1], 1};
+    size_t profiling_lws[3] = {0,0,0}; // Reflects NULL LWS
+    populateProfilingInfo(g_profiling_info.back(), evt, kernel, profiling_gws, profiling_lws, dst);
+#else
+    CL_CHECK(clEnqueueNDRangeKernel(queue, kernel, 2, NULL, global_work_size, NULL, 0, NULL, NULL)); // Pass 2 for 2D problem
+#endif
+}
+
 static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
    GGML_ASSERT(src0);
    GGML_ASSERT(src0->extra);
@@ -5667,6 +6375,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                    }
                    func = ggml_cl_sigmoid;
                    break;
+                case GGML_UNARY_OP_TANH:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_tanh;
+                    break;
                default:
                    return false;
            } break;
@@ -5694,6 +6408,36 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
            }
            func = ggml_cl_group_norm;
            break;
+                case GGML_OP_REPEAT:
+             if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_repeat;
+            break;
+        case GGML_OP_PAD:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_pad(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_UPSCALE:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_upscale(backend, tensor->src[0], tensor);
+            return true;
+        case GGML_OP_CONCAT:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_concat;
+            break;
+        case GGML_OP_TIMESTEP_EMBEDDING:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_timestep_embedding(backend, tensor->src[0], tensor);
+            return true;
        case GGML_OP_MUL_MAT:
            if (!any_on_device && !ggml_cl_can_mul_mat(tensor->src[0], tensor->src[1], tensor)) {
                return false;
@@ -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;
+    }
+}
@@ -0,0 +1,30 @@
+kernel void kernel_pad(
+        global const void * src0_ptr,
+        ulong src0_offset,
+        global void * dst_ptr,
+        ulong dst_offset,
+        int s_ne0, int s_ne1, int s_ne2,
+        int d_ne0, int d_ne1, int d_ne2
+) {
+    global const float * src0 = (global const float *)((global const char *)src0_ptr + src0_offset);
+    global float * dst = (global float *)((global char *)dst_ptr + dst_offset);
+
+    int nidx   = get_global_id(0);
+    int idx_d1 = get_group_id(1);
+    int idx_d2 = get_group_id(2);
+
+    if (nidx >= d_ne0) {
+        return;
+    }
+
+    int dst_el_offset = nidx + idx_d1 * d_ne0 + idx_d2 * d_ne0 * d_ne1;
+
+    bool in_src_bounds = (nidx < s_ne0) && (idx_d1 < s_ne1) && (idx_d2 < s_ne2);
+
+    if (in_src_bounds) {
+        int src_el_offset = nidx + idx_d1 * s_ne0 + idx_d2 * s_ne0 * s_ne1;
+        dst[dst_el_offset] = src0[src_el_offset];
+    } else {
+        dst[dst_el_offset] = 0.0f;
+    }
+}
@@ -0,0 +1,39 @@
+kernel void kernel_repeat(
+    global const char * src0_data_in,
+    global       char * dst_data_in,
+    ulong src0_offset,
+    ulong dst_offset,
+    int src0_ne0, int src0_ne1, int src0_ne2, int src0_ne3,
+    ulong src0_nb0, ulong src0_nb1, ulong src0_nb2, ulong src0_nb3,
+    int dst_ne0, int dst_ne1, int dst_ne2, int dst_ne3,
+    ulong dst_nb0, ulong dst_nb1, ulong dst_nb2, ulong dst_nb3
+) {
+    global const char * src0_data = src0_data_in + src0_offset;
+    global       char * dst_data  = dst_data_in + dst_offset;
+
+    const int d3 = get_global_id(2);
+    const int d2 = get_global_id(1);
+    const int d1 = get_global_id(0);
+
+    if (d3 >= dst_ne3 || d2 >= dst_ne2 || d1 >= dst_ne1) {
+        return;
+    }
+
+    const int s3 = d3 % src0_ne3;
+    const int s2 = d2 % src0_ne2;
+    const int s1 = d1 % src0_ne1;
+
+    const global char * p_src0_slice = src0_data + (ulong)s3*src0_nb3 + (ulong)s2*src0_nb2 + (ulong)s1*src0_nb1;
+    global char * p_dst_slice  = dst_data  + (ulong)d3*dst_nb3 + (ulong)d2*dst_nb2 + (ulong)d1*dst_nb1;
+
+    for (int d0 = 0; d0 < dst_ne0; ++d0) {
+        // Determine source index for dimension 0 based on tiling/broadcasting.
+        const int s0 = d0 % src0_ne0;
+
+        const global char * restrict current_src_el_ptr = p_src0_slice + (ulong)s0*src0_nb0;
+        global char * restrict current_dst_el_ptr  = p_dst_slice  + (ulong)d0*dst_nb0;
+        for (int k = 0; k < src0_nb0; ++k) {
+            current_dst_el_ptr[k] = current_src_el_ptr[k];
+        }
+    }
+}
@@ -0,0 +1,63 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#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
+
+kernel void kernel_tanh_f32_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    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 i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_tanh_f16_nd(
+    global void * p_src0_base, ulong off_src0_abs,
+    global void * p_dst_base,  ulong off_dst_abs,
+    int ne00, int ne01, int ne02, int ne03,
+    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 i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = tanh(*src_val_ptr);
+        }
+    }
+}
@@ -0,0 +1,48 @@
+kernel void kernel_timestep_embedding(
+    global const void * p_timesteps,
+    ulong off_timesteps,
+    global void * p_dst,
+    ulong off_dst,
+    int dst_nb1_bytes,
+    int logical_dim,
+    int max_period
+) {
+    int local_i;
+    int local_j;
+    int local_half_dim;
+    float local_timestep_val;
+    float local_freq;
+    float local_arg;
+    global float * local_embed_data_ptr;
+    global const float * local_timesteps_input_ptr;
+    global float * local_dst_output_base_ptr;
+
+    local_timesteps_input_ptr = (global const float *)((global char *)p_timesteps + off_timesteps);
+    local_dst_output_base_ptr = (global float *)((global char *)p_dst + off_dst);
+
+    local_i = get_global_id(1);
+    local_j = get_global_id(0);
+
+    local_half_dim = logical_dim / 2;
+    local_embed_data_ptr = (global float *)((global char *)local_dst_output_base_ptr + local_i * dst_nb1_bytes);
+
+    if (logical_dim % 2 != 0 && local_j == ((logical_dim + 1) / 2)) {
+        local_embed_data_ptr[logical_dim] = 0.0f;
+    }
+
+    if (local_j >= local_half_dim) {
+        return;
+    }
+
+    local_timestep_val = local_timesteps_input_ptr[local_i];
+
+    if (local_half_dim == 0) {
+        local_freq = 1.0f;
+    } else {
+        local_freq = exp(-log((float)max_period) * (float)local_j / (float)local_half_dim);
+    }
+
+    local_arg = local_timestep_val * local_freq;
+    local_embed_data_ptr[local_j] = cos(local_arg);
+    local_embed_data_ptr[local_j + local_half_dim] = sin(local_arg);
+}
@@ -0,0 +1,121 @@
+kernel void kernel_upscale(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne10,
+    int ne11,
+    int ne12,
+    int ne13,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10 * ne11 * ne12 * ne13;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10 = index % ne10;
+    int i11 = (index / ne10) % ne11;
+    int i12 = (index / (ne10 * ne11)) % ne12;
+    int i13 = index / (ne10 * ne11 * ne12);
+
+    int i00 = (int)(i10 / sf0);
+    int i01 = (int)(i11 / sf1);
+    int i02 = (int)(i12 / sf2);
+    int i03 = (int)(i13 / sf3);
+
+    ulong offset_src_element = (ulong)i03 * nb03 + (ulong)i02 * nb02 + (ulong)i01 * nb01 + (ulong)i00 * nb00;
+    global const float * src_element_ptr = (global const float *)(src_base + offset_src_element);
+
+    dst_base[index] = *src_element_ptr;
+}
+
+kernel void kernel_upscale_bilinear(
+    global const void * p_src0,
+    ulong off_src0,
+    global void * p_dst,
+    ulong off_dst,
+    ulong nb00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne00_src,
+    int ne01_src,
+    int ne10_dst,
+    int ne11_dst,
+    int ne12_dst,
+    int ne13_dst,
+    float sf0,
+    float sf1,
+    float sf2,
+    float sf3
+) {
+    global const char * src_base = (global const char *)p_src0 + off_src0;
+    global float * dst_base = (global float *)((global char *)p_dst + off_dst);
+
+    int index = get_global_id(0);
+    int dst_total_elements = ne10_dst * ne11_dst * ne12_dst * ne13_dst;
+
+    if (index >= dst_total_elements) {
+        return;
+    }
+
+    int i10_dst = index % ne10_dst;
+    int i11_dst = (index / ne10_dst) % ne11_dst;
+    int i12_dst = (index / (ne10_dst * ne11_dst)) % ne12_dst;
+    int i13_dst = index / (ne10_dst * ne11_dst * ne12_dst);
+
+    int i02_src = (int)(i12_dst / sf2);
+    int i03_src = (int)(i13_dst / sf3);
+
+    const float pixel_offset = 0.5f;
+
+    float y_src_f = ((float)i11_dst + pixel_offset) / sf1 - pixel_offset;
+    long y0_src = (long)floor(y_src_f);
+    long y1_src = y0_src + 1;
+
+    y0_src = max(0L, min(y0_src, (long)ne01_src - 1));
+    y1_src = max(0L, min(y1_src, (long)ne01_src - 1));
+
+    float dy = y_src_f - (float)y0_src;
+    dy = max(0.0f, min(dy, 1.0f));
+
+    float x_src_f = ((float)i10_dst + pixel_offset) / sf0 - pixel_offset;
+    long x0_src = (long)floor(x_src_f);
+    long x1_src = x0_src + 1;
+
+    x0_src = max(0L, min(x0_src, (long)ne00_src - 1));
+    x1_src = max(0L, min(x1_src, (long)ne00_src - 1));
+
+    float dx = x_src_f - (float)x0_src;
+    dx = max(0.0f, min(dx, 1.0f));
+
+    global const float * p_a = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_b = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y0_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_c = (global const float *)(src_base + (ulong)x0_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+    global const float * p_d = (global const float *)(src_base + (ulong)x1_src * nb00 + (ulong)y1_src * nb01 + (ulong)i02_src * nb02 + (ulong)i03_src * nb03);
+
+    const float val_a = *p_a;
+    const float val_b = *p_b;
+    const float val_c = *p_c;
+    const float val_d = *p_d;
+
+    float result = val_a * (1.0f - dx) * (1.0f - dy) +
+                   val_b * dx * (1.0f - dy) +
+                   val_c * (1.0f - dx) * dy +
+                   val_d * dx * dy;
+
+    dst_base[index] = result;
+}
@@ -13,7 +13,7 @@ elseif(SUPPORTS_SYCL)
        If you expected the oneAPI Release compiler, please install oneAPI & source it, like:
        source /opt/intel/oneapi/setvars.sh")
 else()
-    message(FATAL_ERROR, "C++ compiler lacks SYCL support.")
+    message(FATAL_ERROR "C++ compiler lacks SYCL support.")
 endif()
 message(STATUS "SYCL found")
 #todo: AOT
@@ -170,7 +170,7 @@ else()
        target_compile_definitions(ggml-sycl PRIVATE GGML_SYCL_NVIDIA)
    elseif (GGML_SYCL_TARGET STREQUAL "AMD")
        if (NOT GGML_SYCL_DEVICE_ARCH)
-            message(ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
+            message(FATAL_ERROR "Can't enable SYCL hip backend, GGML_SYCL_DEVICE_ARCH has not been set.")
        endif()
        target_link_libraries(ggml-sycl PRIVATE ONEMATH::onemath_blas_rocblas)
        target_compile_options(ggml-sycl PRIVATE "-fsycl-targets=amdgcn-amd-amdhsa")
@@ -265,6 +265,17 @@ static void dequantize_row_q6_K_sycl(const void *vx, dst_t *y, const int64_t k,
 #endif
 }

+template <typename dst_t>
+static void dequantize_row_q6_K_sycl_reorder(const void * vx, dst_t * y, const int64_t k, dpct::queue_ptr stream) {
+    const int64_t nb = k / QK_K;
+
+    dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, nb) * sycl::range<3>(1, 1, 64), sycl::range<3>(1, 1, 64)),
+        [=](sycl::nd_item<3> item_ct1) { dequantize_block_q6_K_reorder(vx, y, item_ct1, nb); });
+}
+
 template <typename dst_t>
 static void dequantize_row_iq1_s_sycl(const void *vx, dst_t *y, const int64_t k,
                                        dpct::queue_ptr stream) {
@@ -530,7 +541,11 @@ to_fp16_sycl_t ggml_get_to_fp16_sycl(ggml_type type, ggml_tensor * dst) {
        case GGML_TYPE_Q5_K:
            return dequantize_row_q5_K_sycl;
        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
        case GGML_TYPE_IQ1_S:
            return dequantize_row_iq1_s_sycl;
        case GGML_TYPE_IQ1_M:
@@ -587,7 +602,11 @@ to_fp32_sycl_t ggml_get_to_fp32_sycl(ggml_type type, ggml_tensor *dst) {
        case GGML_TYPE_Q5_K:
            return dequantize_row_q5_K_sycl;
        case GGML_TYPE_Q6_K:
-            return dequantize_row_q6_K_sycl;
+            if (dst->src[0]->extra && ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                return dequantize_row_q6_K_sycl_reorder;
+            } else {
+                return dequantize_row_q6_K_sycl;
+            }
        case GGML_TYPE_IQ1_S:
            return dequantize_row_iq1_s_sycl;
        case GGML_TYPE_IQ1_M:
@@ -1,8 +1,12 @@
 #include "cpy.hpp"

 #include <float.h>
+#include <string>

 #include "dequantize.hpp"
+#include "ggml-sycl/common.hpp"
+#include "ggml-sycl/presets.hpp"
+#include "ggml.h"

 static __dpct_inline__ int best_index_int8(int n, const int8_t * val, float x) {
    if (x <= val[0]) {
@@ -116,6 +120,15 @@ static void cpy_blck_f32_q8_0(const char * cxi, char * cdsti) {
    }
 }

+/* quantized type same copy */
+template<typename T>
+static void cpy_blck_q_q(const char * cxi, char * cdsti) {
+    const T * xi = (const T *) cxi;
+    T * dsti = (T *) cdsti;
+    *dsti = *xi;
+}
+
+
 static void cpy_blck_q8_0_f32(const char * cxi, char * cdsti) {
    float * cdstf = (float *) (cdsti);

@@ -311,6 +324,34 @@ template <dequantize_kernel_t dequant, int qk> static void cpy_blck_q_f32(const
    }
 }

+
+template <typename T, int qk>
+static void cpy_q_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02,
+                      const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11,
+                      const int ne12, const int nb10, const int nb11, const int nb12, const int nb13,
+                      const sycl::nd_item<3> & item_ct1) {
+    const int i = (item_ct1.get_local_range(2) * item_ct1.get_group(2) + item_ct1.get_local_id(2)) * qk;
+
+    if (i >= ne) {
+        return;
+    }
+
+    const int i03      = i / (ne00 * ne01 * ne02);
+    const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
+    const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
+    const int i00      = i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00 - i01 * ne00;
+    const int x_offset = (i00 / qk) * nb00 + i01 * nb01 + i02 * nb02 + i03 * nb03;
+
+
+    const int i13        = i / (ne10 * ne11 * ne12);
+    const int i12        = (i - i13 * ne10 * ne11 * ne12) / (ne10 * ne11);
+    const int i11        = (i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11) / ne10;
+    const int i10        = i - i13 * ne10 * ne11 * ne12 - i12 * ne10 * ne11 - i11 * ne10;
+    const int dst_offset = (i10 / qk) * nb10 + i11 * nb11 + i12 * nb12 + i13 * nb13;
+
+    cpy_blck_q_q<T>(cx + x_offset, cdst + dst_offset);
+}
+
 template <cpy_kernel_t cpy_blck, int qk>
 static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00, const int ne01, const int ne02,
                      const int nb00, const int nb01, const int nb02, const int nb03, const int ne10, const int ne11,
@@ -322,6 +363,7 @@ static void cpy_f32_q(const char * cx, char * cdst, const int ne, const int ne00
        return;
    }

+
    const int i03      = i / (ne00 * ne01 * ne02);
    const int i02      = (i - i03 * ne00 * ne01 * ne02) / (ne00 * ne01);
    const int i01      = (i - i03 * ne00 * ne01 * ne02 - i02 * ne01 * ne00) / ne00;
@@ -615,6 +657,70 @@ static void ggml_cpy_i32_i32_sycl(const char * cx, char * cdst, const int ne, co
    }
 }

+static void ggml_cpy_q8_0_q8_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q8_0, QK8_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_0_q5_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_0, QK5_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q5_1_q5_1(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE),
+                              sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q5_1, QK5_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_0_q4_0(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+    const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+    stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_0, QK4_0>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
+
+static void ggml_cpy_q4_1_q4_1(const char * cx, char * cdst, const int ne, const int ne00, const int ne01,
+                                   const int ne02, const int nb00, const int nb01, const int nb02, const int nb03,
+                                   const int ne10, const int ne11, const int ne12, const int nb10, const int nb11,
+                                   const int nb12, const int nb13, queue_ptr stream) {
+
+   const int num_blocks = ceil_div(ne, SYCL_CPY_BLOCK_SIZE);
+   stream->parallel_for(
+        sycl::nd_range<3>(sycl::range<3>(1, 1, num_blocks) * sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE), sycl::range<3>(1, 1, SYCL_CPY_BLOCK_SIZE)), [=](sycl::nd_item<3> item_ct1) {
+            cpy_q_q<block_q4_1, QK4_1>(cx, cdst, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, item_ct1);
+        });
+}
+
 void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, const ggml_tensor * src1) try {
    // Unlike other operators ggml_sycl_cpy takes 2 distinct tensors instead of a dst ggml_tensor and rely on its src field
    scope_op_debug_print scope_dbg_print(__func__, src1, /*num_src=*/0,
@@ -632,8 +738,10 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co

    char * src0_ddc = (char *) src0->data;
    char * src1_ddc = (char *) src1->data;
-
-    if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
+    if ((src0->type == src1->type) && (ggml_is_contiguous(src0) && ggml_is_contiguous(src1))) {
+        GGML_SYCL_DEBUG("%s: memcpy path\n", __func__);
+        main_stream->memcpy(src1_ddc, src0_ddc, ggml_nbytes(src0));
+    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32) {
        ggml_cpy_f32_f32_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10,
                              nb11, nb12, nb13, main_stream);
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F16) {
@@ -684,6 +792,16 @@ void ggml_sycl_cpy(ggml_backend_sycl_context & ctx, const ggml_tensor * src0, co
    } else if (src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_IQ4_NL) {
        ggml_cpy_f32_iq4_nl_sycl(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12,
                                 nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q8_0 && src1->type == GGML_TYPE_Q8_0) {
+        ggml_cpy_q8_0_q8_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_0 && src1->type == GGML_TYPE_Q5_0) {
+        ggml_cpy_q5_0_q5_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q5_1 && src1->type == GGML_TYPE_Q5_1) {
+        ggml_cpy_q5_1_q5_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_0 && src1->type == GGML_TYPE_Q4_0) {
+        ggml_cpy_q4_0_q4_0(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
+    } else if (src0->type == GGML_TYPE_Q4_1 && src1->type == GGML_TYPE_Q4_1) {
+        ggml_cpy_q4_1_q4_1(src0_ddc, src1_ddc, ne, ne00, ne01, ne02, nb00, nb01, nb02, nb03, ne10, ne11, ne12, nb10, nb11, nb12, nb13, main_stream);
    } else {
        GGML_LOG_ERROR("%s: unsupported type combination (%s to %s)\n", __func__, ggml_type_name(src0->type),
                       ggml_type_name(src1->type));
@@ -538,6 +538,38 @@ static void dequantize_block_q6_K(const void * __restrict__ vx, dst_t * __restri
 #endif
 }

+template <typename dst_t>
+static void dequantize_block_q6_K_reorder(const void * __restrict__ vx, dst_t * __restrict__ yy,
+                                          const sycl::nd_item<3> & item_ct1, int64_t n_blocks) {
+    const int64_t ib = item_ct1.get_group(2);
+
+    const int64_t tid = item_ct1.get_local_id(2);
+    const int64_t ip  = tid / 32;       // ip is 0 or 1
+    const int64_t il  = tid - 32 * ip;  // 0...32
+    const int64_t is  = 8 * ip + il / 16;
+
+    const uint8_t *   base_ptr           = static_cast<const uint8_t *>(vx);
+    const auto        ql_offset          = ib * (QK_K / 2);
+    const auto        qh_offset          = (QK_K / 2) * n_blocks + (QK_K / 4) * ib;
+    const auto        base_scales_offset = (QK_K / 2) * n_blocks + (QK_K / 4) * n_blocks + (QK_K / 16) * ib;
+    const auto        base_d_offset      = ((QK_K / 2) + (QK_K / 4) + (QK_K / 16)) * n_blocks;
+    const uint8_t *   ql_ptr             = base_ptr + ql_offset;
+    const uint8_t *   qh_ptr             = base_ptr + qh_offset;
+    const uint8_t *   scales_ptr         = base_ptr + base_scales_offset;
+    const ggml_half * d                  = (const ggml_half *) (base_ptr + base_d_offset) + ib;
+
+    dst_t * y = yy + ib * QK_K + 128 * ip + il;
+
+    const uint8_t * ql = ql_ptr + 64 * ip + il;
+    const uint8_t   qh = *(qh_ptr + 32 * ip + il);
+    const int8_t *  sc = reinterpret_cast<const int8_t *>(scales_ptr + is);
+
+    y[0]  = *d * sc[0] * ((int8_t) ((ql[0] & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
+    y[32] = *d * sc[2] * ((int8_t) ((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
+    y[64] = *d * sc[4] * ((int8_t) ((ql[0] >> 4) | (((qh >> 4) & 3) << 4)) - 32);
+    y[96] = *d * sc[6] * ((int8_t) ((ql[32] >> 4) | (((qh >> 6) & 3) << 4)) - 32);
+}
+
 template<typename dst_t>
 static void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy,
                                     const sycl::nd_item<3> &item_ct1,
@@ -354,7 +354,8 @@ ggml_backend_sycl_buffer_init_tensor(ggml_backend_buffer_t buffer,
        assert(tensor->view_src->buffer->buft == buffer->buft);
        return GGML_STATUS_SUCCESS;
    }
-    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K) && !g_ggml_sycl_disable_optimize) {
+    if ((tensor->type == GGML_TYPE_Q4_0 || tensor->type == GGML_TYPE_Q4_K || tensor->type == GGML_TYPE_Q6_K) &&
+        !g_ggml_sycl_disable_optimize) {
        ggml_tensor_extra_gpu * extra = new ggml_tensor_extra_gpu{};
        tensor->extra                 = extra;
        ctx->tensor_extras.push_back(extra);  //used to release it when destroy ctx.
@@ -1434,6 +1435,59 @@ static void quantize_q8_1(const float * __restrict__ x, void * __restrict__ vy,
    reinterpret_cast<sycl::half &>(y[ib].ds.y()) = sum;
 }

+template <int ElementsPerWI>
+static __dpct_inline__ void quantize_and_reorder_q8_1(const float * __restrict__ x, void * reordered_q8_tensor,
+                                                      const int kx, const int kx_padded, const sycl::nd_item<1> & it) {
+    /*
+        Quantizes and reorders the resultant q8 tensor in a per row fashion
+        Each sub-group calculates one quant block. i.e. QK8_1 quant values and the d and sum values
+    */
+
+    auto subgroup_id = it.get_group(0);
+    auto wi_id       = it.get_local_id(0);
+
+    const int num_blocks_per_row = kx / QK8_1;
+    auto      row                = subgroup_id / num_blocks_per_row;
+    auto      col                = subgroup_id % num_blocks_per_row;
+
+    auto row_offset = row * (kx_padded / QK8_1) * sizeof(block_q8_1);
+    auto col_offset = QK8_1 * col + wi_id * ElementsPerWI;
+
+    auto quant_ptr = (int8_t *) ((char *) reordered_q8_tensor + row_offset + col_offset);
+    auto ds_ptr    = (sycl::half2 *) ((char *) reordered_q8_tensor + row_offset + kx + col * sizeof(sycl::half2));
+
+    sycl::vec<float, ElementsPerWI>  wi_f32_vals;
+    sycl::vec<int8_t, ElementsPerWI> quantized_values;
+
+    auto float_ptr_offset = subgroup_id * QK8_1 + ElementsPerWI * wi_id;
+    wi_f32_vals           = *reinterpret_cast<const sycl::vec<float, ElementsPerWI> *>(x + float_ptr_offset);
+
+    float sum  = 0.0f;
+    float amax = 0.0f;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        sum += wi_f32_vals[i];
+        amax                = sycl::fmax(amax, sycl::fabs(wi_f32_vals[i]));
+        quantized_values[i] = 0;
+    }
+    sum     = sycl::reduce_over_group(it.get_group(), sum, sycl::plus<float>());
+    amax    = sycl::reduce_over_group(it.get_group(), amax, sycl::maximum<float>());
+    float d = amax == 0 ? 1 : amax / 127;
+
+#pragma unroll(ElementsPerWI)
+    for (int i = 0; i < ElementsPerWI; i++) {
+        quantized_values[i] = sycl::round(wi_f32_vals[i] / d);
+    }
+
+    d = amax == 0 ? 0 : d;
+
+    *reinterpret_cast<sycl::vec<int8_t, ElementsPerWI> *>(quant_ptr) = quantized_values;
+    if (wi_id == 0) {
+        *ds_ptr = sycl::half2(sycl::half(d), sycl::half(sum));
+    }
+}
+
 static void mul_mat_p021_f16_f32(
    const void * __restrict__ vx, const float * __restrict__ y, float * __restrict__ dst,
    const int ncols_x, const int nrows_x, const int nchannels_x, const int nchannels_y,
@@ -1718,23 +1772,30 @@ static  void pool2d_nchw_kernel(
        o_ptr[cur_oh * ow + cur_ow] = res;
 }

-static void quantize_row_q8_1_sycl(const float *x, void *vy, const int kx,
-                                   const int ky, const int kx_padded,
-                                   queue_ptr stream) {
-    const int block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
-    const sycl::range<3> num_blocks(1, ky, block_num_x);
-    int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
-    static_assert(QK8_1 % WARP_SIZE == 0);
-    const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
-    {
-        dpct::has_capability_or_fail(stream->get_device(),
-                                     {sycl::aspect::fp16});
+static void quantize_row_q8_1_sycl(const float * x, void * vy, const int kx, const int ky, const int kx_padded,
+                                   bool reorder_q8_tensor, queue_ptr stream) {
+    if (reorder_q8_tensor) {
+        auto local_range      = std::size_t(WARP_SIZE);
+        auto num_quant_blocks = ky * (kx / QK8_1);
+        auto global_range     = num_quant_blocks * local_range;
+        stream->parallel_for(sycl::nd_range<1>({ global_range }, { local_range }),
+                             [=](sycl::nd_item<1> it) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                 quantize_and_reorder_q8_1<QK8_1 / WARP_SIZE>(x, vy, kx, kx_padded, it);
+                             });
+    } else {
+        const int            block_num_x = (kx_padded + SYCL_QUANTIZE_BLOCK_SIZE - 1) / SYCL_QUANTIZE_BLOCK_SIZE;
+        const sycl::range<3> num_blocks(1, ky, block_num_x);
+        int constexpr QUANT_BLOCK_TILE = QK8_1 / WARP_SIZE;
+        static_assert(QK8_1 % WARP_SIZE == 0);
+        const sycl::range<3> block_size(1, 1, SYCL_QUANTIZE_BLOCK_SIZE / QUANT_BLOCK_TILE);
+        {
+            dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });

-        stream->parallel_for(
-            sycl::nd_range<3>(num_blocks * block_size, block_size),
-            [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
-                quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
-            });
+            stream->parallel_for(sycl::nd_range<3>(num_blocks * block_size, block_size),
+                                 [=](sycl::nd_item<3> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                                     quantize_q8_1<QUANT_BLOCK_TILE>(x, vy, kx, kx_padded, item_ct1);
+                                 });
+        }
    }
 }

@@ -2446,9 +2507,10 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
            dev[i].src1_ddq = dev[i].src1_ddq_alloc.alloc(ctx.pool(i), nrows1*src1_padded_col_size*q8_1_ts/q8_1_bs);

            if (src1_on_device && src1_is_contiguous) {
+                bool reorder_q8_tensor = src0->extra && ((ggml_tensor_extra_gpu *)src0->extra)->optimized_feature.reorder;
                scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                     /*num_src=*/2, " : converting src1 to Q8_1");
-                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, stream);
+                quantize_row_q8_1_sycl(dev[i].src1_ddf, dev[i].src1_ddq, ne10, nrows1, src1_padded_col_size, reorder_q8_tensor, stream);
                /*
                DPCT1010:90: SYCL uses exceptions to report errors and does not
                use the error codes. The call was replaced with 0. You need to
@@ -2554,7 +2616,7 @@ static void ggml_sycl_op_mul_mat(ggml_backend_sycl_context & ctx, const ggml_ten
                if (convert_src1_to_q8_1 && !src1_is_contiguous) {
                    scope_op_debug_print scope_dbg_print(__func__, "/quantize_row_q8_1_sycl", dst,
                                                         /*num_src=*/2, " : converting src1 to Q8_1");
-                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, stream);
+                    quantize_row_q8_1_sycl(src1_ddf_i, src1_ddq_i, ne10, src1_ncols, src1_padded_col_size, false, stream);
                    /*
                    DPCT1010:92: SYCL uses exceptions to report errors and does
                    not use the error codes. The call was replaced with 0. You
@@ -2928,6 +2990,7 @@ inline bool ggml_sycl_supports_reorder_mul_mat_sycl(enum ggml_type type) {
        case GGML_TYPE_Q4_0:
            return true;
        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
            return !g_ggml_sycl_prioritize_dmmv;
        default:
            return false;
@@ -2947,6 +3010,7 @@ inline bool ggml_sycl_supports_reorder_mmvq(enum ggml_type type) {
    switch (type) {
        case GGML_TYPE_Q4_0:
        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q6_K:
            return true;
        default:
            return false;
@@ -3031,6 +3095,50 @@ static void reorder_qw_q4_k(uint8_t * data_device, size_t size, size_t offset, d
    sycl::free(tmp_buf, *stream);
 }

+static void reorder_qw_q6_k(uint8_t * data_device, size_t size, size_t offset, dpct::queue_ptr stream) {
+    GGML_ASSERT(size % sizeof(block_q6_K) == 0);
+    GGML_ASSERT(offset % sizeof(block_q6_K) == 0);
+
+    const int nblocks = size / sizeof(block_q6_K);
+
+    auto * tmp_buf = sycl::malloc_shared<uint8_t>(size, *stream);
+    SYCL_CHECK(CHECK_TRY_ERROR((*stream).memcpy(tmp_buf, data_device, size).wait()));
+
+    auto *       ql_ptr     = data_device;
+    auto *       qh_ptr     = ql_ptr + (QK_K / 2) * nblocks;
+    auto *       scales_ptr = qh_ptr + (QK_K / 4) * nblocks;
+    sycl::half * dm_ptr     = (sycl::half *) (scales_ptr + (QK_K / 16) * nblocks);
+
+    stream
+        ->parallel_for(nblocks,
+                       [=](auto i) {
+                           const block_q6_K * x  = (const block_q6_K *) tmp_buf;
+                           const int          ib = i;
+
+                           const uint8_t * ql              = x[ib].ql;
+                           const uint8_t * qh              = x[ib].qh;
+                           uint8_t *       base_ql_ptr     = ql_ptr + (QK_K / 2) * ib;
+                           uint8_t *       base_qh_ptr     = qh_ptr + (QK_K / 4) * ib;
+                           uint8_t *       base_scales_ptr = scales_ptr + (QK_K / 16) * ib;
+
+                           for (int j = 0; j < QK_K / 2; ++j) {
+                               base_ql_ptr[j] = ql[j];
+                           }
+                           for (int j = 0; j < QK_K / 4; ++j) {
+                               base_qh_ptr[j] = qh[j];
+                           }
+
+                           for (int j = 0; j < QK_K / 16; ++j) {
+                               base_scales_ptr[j] = x[ib].scales[j];
+                           }
+
+                           dm_ptr[ib] = x[ib].d;
+                       })
+        .wait_and_throw();
+
+    sycl::free(tmp_buf, *stream);
+}
+
 static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
    uint8_t * data_device = (uint8_t *) src0->data;
    size_t ncols = src0->ne[0];
@@ -3044,6 +3152,9 @@ static void reorder_qw(const ggml_tensor * src0, dpct::queue_ptr stream) {
        case GGML_TYPE_Q4_K:
            reorder_qw_q4_k(data_device, size, 0, stream);
            break;
+        case GGML_TYPE_Q6_K:
+            reorder_qw_q6_k(data_device, size, 0, stream);
+            break;
        default:
            GGML_ABORT("reorder_qw() called with unsupported type");
            break;
@@ -4165,6 +4276,9 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
            {
                ggml_type src0_type = op->src[0]->type;
                ggml_type src1_type = op->src[1]->type;
+                if (src0_type == src1_type && (ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1])) && src0_type != GGML_TYPE_BF16) {
+                    return true;
+                }
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_F32) {
                    return true;
                }
@@ -4210,6 +4324,21 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                if (src0_type == GGML_TYPE_F32 && src1_type == GGML_TYPE_IQ4_NL) {
                    return true;
                }
+                if(src0_type == GGML_TYPE_Q8_0 && src1_type == GGML_TYPE_Q8_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_0 && src1_type == GGML_TYPE_Q5_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q5_1 && src1_type == GGML_TYPE_Q5_1) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_0 && src1_type == GGML_TYPE_Q4_0) {
+                    return true;
+                }
+                if(src0_type == GGML_TYPE_Q4_1 && src1_type == GGML_TYPE_Q4_1) {
+                    return true;
+                }
                return false;
            }
        case GGML_OP_CONCAT:
@@ -4257,14 +4386,6 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
        case GGML_OP_SOFT_MAX:
            return true;
        case GGML_OP_ROPE:
-            {
-                const int mode = ((const int32_t *) op->op_params)[2];
-                // mode is not used as a bitmask in practice, the various rope type modes are independent implementations
-                if (mode == GGML_ROPE_TYPE_MROPE) {
-                    return false;
-                }
-                return true;
-            }
        case GGML_OP_IM2COL:
            return true;
        case GGML_OP_UPSCALE:
@@ -29,24 +29,23 @@ static void mul_mat_vec_q_reorder(const void * __restrict__ vx, const void * __r
    static_assert(blocks_per_subgroup > 0);
    static_assert(block_elements_per_subgroup > 0);

-    const block_q8_1 * y = (const block_q8_1 *) vy;
-
    float partial_sum = 0.0f;
    for (int i = sg.get_local_linear_id() / block_elements_per_subgroup; i < blocks_per_row; i += blocks_per_subgroup) {
-        const int ibx       = row * blocks_per_row + i;  // x block index
-        // TODO: Generalize offsets, right now only works for quantizations that don't split high and low bits
-        const int bx_offset = block_type::get_block_offset(ibx);
-        const int d_offset  = block_type::get_d_offset(nrows, ncols, ibx);
+        const int ibx = row * blocks_per_row + i;  // x block index

+        const auto         bx_offset      = block_type::get_block_offset(ibx, nblocks);
+        const auto         d_offset       = block_type::get_d_offset(nrows, ncols, ibx);
        // Y block index that aligns with ibx
        const int iby = i * block_type::block_to_q8_1_ratio();
+        const int8_t* q8_1_quant_ptr = (const int8_t*)vy + iby * QK8_1;
+        const sycl::half2* q8_1_ds_ptr = (const sycl::half2*)((const char*)vy + ncols + iby * sizeof(sycl::half2));

 #pragma unroll
        for (int elem = 0; elem < block_elements_per_subgroup; elem += WARP_SIZE) {
            // x block quant index when casting the quants to int
            const int iqs = elem + block_traits::vdr_mmvq * (sg.get_local_linear_id() % block_elements_per_subgroup);

-            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, &y[iby], iqs, nblocks);
+            partial_sum += reorder_vec_dot_q_sycl()(vx, bx_offset, d_offset, q8_1_quant_ptr, q8_1_ds_ptr, iqs);
        }
    }

@@ -785,6 +784,24 @@ static void mul_mat_vec_q5_K_q8_1_sycl(const void *vx, const void *vy,
    }
 }

+static void reorder_mul_mat_vec_q6_k_q8_1_sycl(const void * vx, const void * vy, float * dst, const int ncols,
+                                               const int nrows, dpct::queue_ptr stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int        block_num_y   = ceil_div(nrows, GGML_SYCL_MMV_Y);
+    constexpr size_t num_subgroups = 16;
+    GGML_ASSERT(block_num_y % num_subgroups == 0);
+
+    const sycl::range<3> global_size(1, GGML_SYCL_MMV_Y, block_num_y * WARP_SIZE);
+    const sycl::range<3> workgroup_size(1, GGML_SYCL_MMV_Y, num_subgroups * WARP_SIZE);
+
+    stream->submit([&](sycl::handler & cgh) {
+        cgh.parallel_for(sycl::nd_range<3>(global_size, workgroup_size),
+                         [=](sycl::nd_item<3> nd_item) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {
+                             mul_mat_vec_q_reorder<reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K>>(vx, vy, dst, ncols, nrows,
+                                                                                           nd_item);
+                         });
+    });
+}
 static void mul_mat_vec_q6_K_q8_1_sycl(const void *vx, const void *vy,
                                       float *dst, const int ncols,
                                       const int nrows,
@@ -1070,7 +1087,14 @@ void ggml_sycl_op_mul_mat_vec_q(ggml_backend_sycl_context & ctx, const ggml_tens
                mul_mat_vec_q5_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
                break;
            case GGML_TYPE_Q6_K:
-                mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                if ((ggml_tensor_extra_gpu *) dst->src[0]->extra &&
+                    ((ggml_tensor_extra_gpu *) dst->src[0]->extra)->optimized_feature.reorder) {
+                    GGML_SYCL_DEBUG("Calling reorder_mul_mat_vec_q6_k_q8_1_sycl\n");
+                    reorder_mul_mat_vec_q6_k_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                } else {
+                    GGML_SYCL_DEBUG("Calling mul_mat_vec_q6_k_q8_1_sycl\n");
+                    mul_mat_vec_q6_K_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
+                }
                break;
            case GGML_TYPE_IQ1_S:
                mul_mat_vec_iq1_s_q8_1_sycl(src0_dd_i, src1_ddq_i_bs, dst_dd_i_bs, ne00, row_diff, stream);
@@ -14,12 +14,13 @@
 #ifndef GGML_SYCL_QUANTS_HPP
 #define GGML_SYCL_QUANTS_HPP

+#include <utility>
+
 #include "ggml-common.h"
 #include "ggml.h"

 namespace ggml_sycl_reordered {

-
 // The reordered block moves quants (qs) and  scales(d) to two
 // uniform regions of memory that is contiguous in the same tensor.
 // What this means is that instead of having:
@@ -32,7 +33,6 @@ namespace ggml_sycl_reordered {

 template <ggml_type type> struct block_q_t;

-
 // qk number of weights / quants in a block
 // qr number of weights in a byte (described as 'before dequantization')
 //    for quantization types that has low and high bits split, qr is calculated with
@@ -47,10 +47,12 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
        static constexpr uint32_t vdr_mmvq = 2;
    };

-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }

-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
-        return (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half);
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        return { (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half), 0 };
    }

    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
@@ -64,20 +66,46 @@ template <> struct block_q_t<GGML_TYPE_Q4_K> {
        static constexpr uint32_t vdr_mmvq = 2;
    };

-    static constexpr int get_block_offset(const int block_index) { return block_index * (traits::qk / traits::qr); }
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
+        return { block_index * (traits::qk / traits::qr), 0 };
+    }

-    static constexpr int get_d_offset(int nrows, int ncols, const int block_index) {
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
        auto nblocks = (nrows * (ncols / traits::qk));
-        return (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2));
+        return { nblocks * (QK_K / 2),
+                 (nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2)) };
    }

    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }

    constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
-
-    constexpr size_t get_dm_offset(int nblocks) { return get_total_qs_bytes(nblocks) + nblocks * K_SCALE_SIZE; }
 };

+template <> struct block_q_t<GGML_TYPE_Q6_K> {
+    struct traits {
+        static constexpr uint32_t qk       = QK_K;
+        static constexpr uint32_t qi       = QI6_K;
+        static constexpr uint32_t qr       = QR6_K;
+        static constexpr uint32_t vdr_mmvq = 1;
+    };
+
+    static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
+        auto low_bits_index  = block_index * (traits::qk / traits::qr);
+        // the index of high bits it's after all low bits
+        auto high_bits_index = n_blocks * (QK_K / 2) + (block_index * (QK_K / 4));
+        return { low_bits_index, high_bits_index };
+    }
+
+    static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
+        auto nblocks        = (nrows * (ncols / traits::qk));
+        auto total_qs_bytes = nblocks * (QK_K / 2) + nblocks * (QK_K / 4);
+        auto block_scales   = total_qs_bytes + block_index * (QK_K / 16);
+        auto sb_scale       = total_qs_bytes + nblocks * (QK_K / 16);
+        return { block_scales, sb_scale };
+    }
+
+    static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
+};
 }  // namespace ggml_sycl_reordered

 #endif  // GGML_SYCL_QUANTS_HPP
@@ -49,10 +49,7 @@ static void rope_norm(const T * x, T * dst, const int ne0, const int ne1, const

    if (i0 >= n_dims) {
        const int i = row * ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
        return;
    }

@@ -93,10 +90,7 @@ static void rope_neox(const T * x, T * dst, const int ne0, const int ne1, const

    if (i0 >= n_dims) {
        const int i = row * ne0 + i0;
-
-        dst[i + 0] = x[i + 0];
-        dst[i + 1] = x[i + 1];
-
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
        return;
    }

@@ -122,6 +116,63 @@ static void rope_neox(const T * x, T * dst, const int ne0, const int ne1, const
    dst[i + n_dims / 2] = x0 * sin_theta + x1 * cos_theta;
 }

+template <typename T, bool has_ff>
+static void rope_multi(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                        const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
+                        const float ext_factor, const float attn_factor, const rope_corr_dims corr_dims,
+                        const float theta_scale, const float * freq_factors, const mrope_sections sections,
+                        const sycl::nd_item<3> & item_ct1) {
+    // get index pos
+    const int i0 = 2 * (item_ct1.get_group(1) * item_ct1.get_local_range(1) + item_ct1.get_local_id(1));
+    if (i0 >= ne0) {
+        return;
+    }
+    const int    row_dst   = (item_ct1.get_group(2) * item_ct1.get_local_range(2)) + item_ct1.get_local_id(2);
+
+    if (i0 >= n_dims) {
+        const int i = row_dst*ne0 + i0;
+        *reinterpret_cast<sycl::vec<T, 2> *>(dst + i) = *reinterpret_cast<const sycl::vec<T, 2> *>(x + i);
+        return;
+    }
+
+    const int    row_x     = row_dst % ne1;
+    const int    channel_x = row_dst / ne1;
+    const int    idst      = (row_dst * ne0) + (i0 / 2);
+    const size_t ix        = ((size_t) channel_x * s2) + ((size_t) row_x * s1) + (i0 / 2);
+
+    const int sect_dims = sections.v[0] + sections.v[1] + sections.v[2] + sections.v[3];
+    const int sec_w = sections.v[1] + sections.v[0];
+    const int sector = (i0 / 2) % sect_dims;
+
+
+    float theta_base = 0.0;
+    if (sector < sections.v[0]) {
+        theta_base = pos[channel_x]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sections.v[0] && sector < sec_w) {
+        theta_base = pos[channel_x + ne2 * 1]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w && sector < sec_w + sections.v[2]) {
+        theta_base = pos[channel_x + ne2 * 2]*sycl::pow(theta_scale, i0/2.0f);
+    }
+    else if (sector >= sec_w + sections.v[2]) {
+        theta_base = pos[channel_x + ne2 * 3]*sycl::pow(theta_scale, i0/2.0f);
+    }
+
+    const float freq_factor = has_ff ? freq_factors[i0 / 2] : 1.0f;
+    float       cos_theta;
+    float       sin_theta;
+    rope_yarn(theta_base / freq_factor, freq_scale, corr_dims, i0, ext_factor, attn_factor, &cos_theta, &sin_theta);
+    const float x0 = x[ix + 0];
+    const float x1 = x[ix + n_dims/2];
+
+    // store results in dst
+    dst[idst + 0]      = x0 * cos_theta - x1 * sin_theta;
+    dst[idst + n_dims/2] = x0 * sin_theta + x1 * cos_theta;
+}
+
+
+
 template <typename T, bool has_ff>
 static void rope_vision(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
                        const size_t s2, const int n_dims, const int32_t * pos, const float freq_scale,
@@ -171,7 +222,7 @@ static void rope_norm_sycl(const T * x, T * dst, const int ne0, const int ne1, c
                           const float * freq_factors, queue_ptr stream) {
    GGML_ASSERT(ne0 % 2 == 0);
    const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
    const sycl::range<3> block_nums(1, num_blocks_x, nr);

    const float theta_scale = powf(freq_base, -2.0f / n_dims);
@@ -208,7 +259,7 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
                           const rope_corr_dims corr_dims, const float * freq_factors, queue_ptr stream) {
    GGML_ASSERT(ne0 % 2 == 0);
    const sycl::range<3> block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int            num_blocks_x = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int            num_blocks_x = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
    const sycl::range<3> block_nums(1, num_blocks_x, nr);

    const float theta_scale = powf(freq_base, -2.0f / n_dims);
@@ -228,6 +279,40 @@ static void rope_neox_sycl(const T * x, T * dst, const int ne0, const int ne1, c
    }
 }

+template <typename T>
+static void rope_multi_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
+                             const size_t s2, const int n_dims, const int nr, const int32_t * pos,
+                             const float freq_scale, const float freq_base, const float ext_factor,
+                             const float attn_factor, const rope_corr_dims corr_dims, const float * freq_factors,
+                             const mrope_sections sections, queue_ptr stream) {
+    GGML_ASSERT(ne0 % 2 == 0);
+    const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
+    const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
+    const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
+    const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);
+
+    const float theta_scale = std::pow(freq_base, -2.0f / n_dims);
+    // Add FP16 capability check if T could be sycl::half
+    if constexpr (std::is_same_v<T, sycl::half>) {
+        dpct::has_capability_or_fail(stream->get_device(), { sycl::aspect::fp16 });
+    }
+    // launch kernel
+    if (freq_factors == nullptr) {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_multi<T, false>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                  corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    } else {
+        stream->parallel_for(nd_range, [=](sycl::nd_item<3> item_ct1) {
+            rope_multi<T, true>(x, dst, ne0, ne1, ne2, s1, s2, n_dims, pos, freq_scale, ext_factor, attn_factor,
+                                 corr_dims, theta_scale, freq_factors, sections, item_ct1);
+        });
+    }
+}
+
+
+
+
 // rope vision
 template <typename T>
 static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1, const int ne2, const size_t s1,
@@ -237,7 +322,7 @@ static void rope_vision_sycl(const T * x, T * dst, const int ne0, const int ne1,
                             const mrope_sections sections, queue_ptr stream) {
    GGML_ASSERT(ne0 % 2 == 0);
    const sycl::range<3>    block_dims(1, SYCL_ROPE_BLOCK_SIZE, 1);
-    const int               n_blocks_y = (ne0 + 2 * SYCL_ROPE_BLOCK_SIZE - 1) / (2 * SYCL_ROPE_BLOCK_SIZE);
+    const int               n_blocks_y = ceil_div(ne0, (2 * SYCL_ROPE_BLOCK_SIZE));
    const sycl::range<3>    grid_dims(1, n_blocks_y, nr);
    const sycl::nd_range<3> nd_range(grid_dims * block_dims, block_dims);

@@ -298,8 +383,17 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
    memcpy(&sections.v,  (int32_t *) dst->op_params + 11, sizeof(int)*4);

    const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;

+    if (is_mrope) {
+        GGML_ASSERT(sections.v[0] > 0 || sections.v[1] > 0 || sections.v[2] > 0);
+    }
+
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne00/2);
+    }
+
    const int32_t * pos = (const int32_t *) dst->src[1]->data;

    const float * freq_factors = nullptr;
@@ -326,6 +420,19 @@ inline void ggml_sycl_op_rope(ggml_backend_sycl_context & ctx, ggml_tensor *dst)
        } else {
            GGML_ABORT("fatal error");
        }
+    } else if (is_mrope && !is_vision) {
+        GGML_SYCL_DEBUG("%s: mrope path\n", __func__);
+        if (dst->src[0]->type == GGML_TYPE_F16) {
+            rope_multi_sycl((const sycl::half *)dst->src[0]->data, (sycl::half *)dst->data, ne00, ne01, ne02, s01,
+                s02, n_dims, nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims,
+                freq_factors, sections, main_stream);
+        } else if (dst->src[0]->type == GGML_TYPE_F32) {
+            rope_multi_sycl((const float *) dst->src[0]->data, (float *) dst->data, ne00, ne01, ne02, s01, s02, n_dims,
+                             nr, pos, freq_scale, freq_base, ext_factor, attn_factor, corr_dims, freq_factors, sections,
+                             main_stream);
+        } else {
+            GGML_ABORT("Fatal error: Tensor type unsupported!");
+        }
    } else if (is_vision) {
        GGML_SYCL_DEBUG("%s: vision path\n", __func__);
        if (dst->src[0]->type == GGML_TYPE_F16) {
@@ -284,22 +284,23 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_0> {
        return d4 * (sumi * ds8f.x() - (8 * q4_0_traits::vdr_mmvq / q4_0_traits::qi) * ds8f.y());
    }

-    __dpct_inline__ float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int /* nblocks */) {
-        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset;
-        const ggml_half d     = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset));
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const uint8_t * bq4_0 = static_cast<const uint8_t *>(vbq) + ibx_offset.first;
+        const ggml_half d = *(reinterpret_cast<const ggml_half *>(static_cast<const uint8_t *>(vbq) + d_offset.first));
        int             v[q4_0_traits::vdr_mmvq];
        int             u[2 * q4_0_traits::vdr_mmvq];

-#pragma unroll

+#pragma unroll
        for (size_t i = 0; i < q4_0_traits::vdr_mmvq; ++i) {
            v[i]         = get_int_from_uint8(bq4_0, iqs + i);
-            u[2 * i + 0] = get_int_from_int8_aligned(bq8_1->qs, iqs + i);
-            u[2 * i + 1] = get_int_from_int8_aligned(bq8_1->qs, iqs + i + q4_0_traits::qi);
+            u[2 * i + 0] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i);
+            u[2 * i + 1] = get_int_from_int8_aligned(q8_1_quant_ptr, iqs + i + q4_0_traits::qi);
        }

-        return vec_dot_q4_0_q8_1_impl(v, u, d, bq8_1->ds);
+        return vec_dot_q4_0_q8_1_impl(v, u, d, *q8_1_ds);
    };
 };

@@ -346,24 +347,115 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
    using q4_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q4_K>;
    using q4_k_traits = typename q4_k_block::traits;

-    float operator()(const void * __restrict__ vbq, const int ibx_offset, const int d_offset,
-                     const block_q8_1 * __restrict__ bq8_1, const int & iqs, int nblocks) {
-        const int ib = ibx_offset / (QK_K / 2);
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
+                                     const sycl::half2 * q8_1_ds, const int & iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);

        const uint8_t *    base           = static_cast<const uint8_t *>(vbq);
-        const uint8_t *    qs             = base + ibx_offset;
-        const int          total_qs_bytes = nblocks * (QK_K / 2);
-        const uint8_t *    scs            = base + total_qs_bytes + ib * K_SCALE_SIZE;
-        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset);
+        const uint8_t *    qs             = base + ibx_offset.first;
+        const uint8_t *    scs            = base + d_offset.first + ib * K_SCALE_SIZE;
+        const ggml_half2 * dms            = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);

        const int        bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
        const int *      q4         = (const int *) (qs + 16 * bq8_offset + 4 * ((iqs / 2) % 4));
        const uint16_t * scales     = (const uint16_t *) scs;

-        return vec_dot_q4_K_q8_1_common(q4, scales, *dms, bq8_1, iqs);
+        int   v[2];
+        int   u[2 * QR4_K];
+        float d8[QR4_K];
+
+        v[0] = q4[0];
+        v[1] = q4[4];
+
+        uint16_t  aux[2];
+        const int j = (QR4_K * ((iqs / 2) / (QI8_1 / 2))) / 2;
+        if (j < 2) {
+            aux[0] = scales[j + 0] & 0x3f3f;
+            aux[1] = scales[j + 2] & 0x3f3f;
+        } else {
+            aux[0] = ((scales[j + 2] >> 0) & 0x0f0f) | ((scales[j - 2] & 0xc0c0) >> 2);
+            aux[1] = ((scales[j + 2] >> 4) & 0x0f0f) | ((scales[j - 0] & 0xc0c0) >> 2);
+        }
+
+        const uint8_t * sc = (const uint8_t *) aux;
+        const uint8_t * m  = sc + 2;
+
+        for (int i = 0; i < QR4_K; ++i) {
+            const int8_t* quant_base_ptr = q8_1_quant_ptr + (bq8_offset + i) * QK8_1;
+            sycl::half2 ds_values = *(q8_1_ds + bq8_offset + i);
+
+            d8[i]                   = ds_values[0];
+
+            const int * q8 = (const int *) quant_base_ptr + ((iqs / 2) % 4);
+            u[2 * i + 0]   = q8[0];
+            u[2 * i + 1]   = q8[4];
+        }
+
+        return vec_dot_q4_K_q8_1_impl_vmmq(v, u, sc, m, *dms, d8);
    }
 };

+template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
+    static constexpr ggml_type gtype = GGML_TYPE_Q6_K;
+
+    using q6_k_block  = ggml_sycl_reordered::block_q_t<GGML_TYPE_Q6_K>;
+    using q6_k_traits = typename q6_k_block::traits;
+
+    __dpct_inline__ float vec_dot_q6_K_q8_1_impl_mmvq(const int vl, const int vh, const int * __restrict__ u,
+                                                      const int8_t * __restrict__ scales, const float d,
+                                                      const float * __restrict__ d8) {
+        float sumf = 0.0f;
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            const int sc = scales[4 * i];
+
+            const int vil = (vl >> (4 * i)) & 0x0F0F0F0F;
+
+            const int vih = ((vh >> (4 * i)) << 4) & 0x30303030;
+
+            const int vi = dpct::vectorized_binary<sycl::char4>((vil | vih), 0x20202020,
+                                                                dpct::sub_sat());  // vi = (vil | vih) - 32
+
+            sumf += d8[i] * (dpct::dp4a(vi, u[i], 0) * sc);                        // SIMD dot product
+        }
+
+        return d * sumf;
+    }
+
+    __dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
+                     const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr, const sycl::half2 * q8_1_ds,
+                     const int iqs) {
+        const int ib = ibx_offset.first / (QK_K / 2);
+
+        const uint8_t *   base   = static_cast<const uint8_t *>(vbq);
+        const uint8_t *   ql     = base + ibx_offset.first;
+        const uint8_t *   qh     = base + ibx_offset.second;
+        const int8_t *    scales = reinterpret_cast<const int8_t *>(base + d_offset.first);
+        const ggml_half * d      = (const ggml_half *) (base + d_offset.second) + ib;
+
+        const int bq8_offset   = 2 * QR6_K * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 4);
+        const int scale_offset = (QI6_K / 4) * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 8);
+        const int vh_shift     = 2 * ((iqs % (QI6_K / 2)) / (QI6_K / 4));
+
+        const int vl = get_int_from_uint8(ql, iqs);
+        const int vh = get_int_from_uint8(qh, (QI6_K / 4) * (iqs / (QI6_K / 2)) + iqs % (QI6_K / 4)) >> vh_shift;
+
+        const int8_t * scs = scales + scale_offset;
+
+        int   u[QR6_K];
+        float d8[QR6_K];
+
+#pragma unroll
+        for (int i = 0; i < QR6_K; ++i) {
+            u[i] = get_int_from_int8_aligned(q8_1_quant_ptr + (bq8_offset + 2 * i) * QK8_1, iqs % QI8_1);
+            const sycl::half2 ds_values = *(q8_1_ds + bq8_offset + 2 * i);
+            d8[i]                       = ds_values[0];
+        }
+        return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scs, *d, d8);
+    }
+};
 #define VDR_Q4_0_Q8_1_MMVQ 2
 #define VDR_Q4_0_Q8_1_MMQ  4

@@ -196,6 +196,7 @@ enum vk_device_architecture {
    AMD_RDNA1,
    AMD_RDNA2,
    AMD_RDNA3,
+    INTEL_XE2,
 };

 static vk_device_architecture get_device_architecture(const vk::PhysicalDevice& device) {
@@ -246,6 +247,34 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
            }
            return vk_device_architecture::AMD_RDNA2;
        }
+    } else if (props.vendorID == VK_VENDOR_ID_INTEL) {
+        const std::vector<vk::ExtensionProperties> ext_props = device.enumerateDeviceExtensionProperties();
+
+        bool subgroup_size_control = false;
+
+        for (const auto& properties : ext_props) {
+            if (strcmp("VK_EXT_subgroup_size_control", properties.extensionName) == 0) {
+                subgroup_size_control = true;
+            }
+        }
+
+        if (!subgroup_size_control) {
+            return vk_device_architecture::OTHER;
+        }
+
+        vk::PhysicalDeviceProperties2 props2;
+        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
+        props2.pNext = &subgroup_size_control_props;
+        device.getProperties2(&props2);
+
+        if (subgroup_size_control_props.minSubgroupSize == 16) {
+            // Xe2 architecture uses SIMD16 while previous Xe and Gen architecture uses SIMD8.
+            // Minimum subgroup size matches the SIMD width so we distinguish architecture by checking this value.
+            // https://www.intel.com/content/www/us/en/content-details/824434/2024-intel-tech-tour-xe2-and-lunar-lake-s-gpu.html
+            // https://www.intel.com/content/www/us/en/docs/oneapi/optimization-guide-gpu/2025-0/intel-xe-gpu-architecture.html
+            return vk_device_architecture::INTEL_XE2;
+        }
    }
    return vk_device_architecture::OTHER;
 }
@@ -396,6 +425,7 @@ struct vk_device_struct {
    vk_pipeline pipeline_count_equal_i32;
    vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
    vk_pipeline pipeline_timestep_embedding_f32;
+    vk_pipeline pipeline_conv_transpose_1d_f32;
    vk_pipeline pipeline_pool2d_f32;
    vk_pipeline pipeline_rwkv_wkv6_f32;
    vk_pipeline pipeline_rwkv_wkv7_f32;
@@ -444,7 +474,7 @@ struct vk_device_struct {
    // for GGML_VK_PERF_LOGGER
    std::unique_ptr<vk_perf_logger> perf_logger;
    vk::QueryPool query_pool;
-    uint32_t num_queries;
+    int32_t num_queries;

    ~vk_device_struct() {
        VK_LOG_DEBUG("destroy device " << name);
@@ -706,6 +736,21 @@ struct vk_op_timestep_embedding_push_constants {
    uint32_t max_period;
 };

+struct vk_op_conv_transpose_1d_push_constants {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+};
+
 struct vk_op_pool2d_push_constants {
    uint32_t IW; uint32_t IH;
    uint32_t OW; uint32_t OH;
@@ -1652,7 +1697,7 @@ static std::array<uint32_t, 2> fa_rows_cols(FaCodePath path, uint32_t D, uint32_
        return {64, 32};
    }
    return {64, 64};
-};
+}

 static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vector<uint32_t>& warptile, bool mul_mat_id, ggml_type src0_type) {

@@ -2726,6 +2771,8 @@ static void ggml_vk_load_shaders(vk_device& device) {

    ggml_vk_create_pipeline(device, device->pipeline_timestep_embedding_f32, "timestep_embedding_f32", timestep_embedding_f32_len, timestep_embedding_f32_data, "main", 2, sizeof(vk_op_timestep_embedding_push_constants), {256, 1, 1}, {}, 1);

+    ggml_vk_create_pipeline(device, device->pipeline_conv_transpose_1d_f32, "conv_transpose_1d_f32", conv_transpose_1d_f32_len, conv_transpose_1d_f32_data, "main", 3, sizeof(vk_op_conv_transpose_1d_push_constants), {1, 1, 1}, {}, 1);
+
    ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);

    ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
@@ -4061,7 +4108,33 @@ static vk_submission ggml_vk_begin_submission(vk_device& device, vk_queue& q, bo
    return s;
 }

-static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, size_t push_constant_size, const void* push_constants, std::array<uint32_t, 3> elements) {
+template <typename T> size_t push_constant_size(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    GGML_UNUSED(t);
+    return sizeof(T);
+}
+template <typename T> size_t push_constant_size(const std::vector<T> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * t.size();
+}
+template <typename T, uint32_t N> size_t push_constant_size(const std::array<T, N> &t) {
+    GGML_UNUSED(t);
+    return sizeof(T) * N;
+}
+
+template <typename T> const T *push_constant_data(const T &t) {
+    static_assert(std::is_class<T>::value, "T must be a struct/class");
+    return &t;
+}
+template <typename T> const T *push_constant_data(const std::vector<T> &t) {
+    return t.data();
+}
+template <typename T, uint32_t N> const T *push_constant_data(const std::array<T, N> &t) {
+    return t.data();
+}
+
+template <typename T>
+static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context& subctx, vk_pipeline& pipeline, std::initializer_list<vk::DescriptorBufferInfo> const& descriptor_buffer_infos, const T &push_constants, std::array<uint32_t, 3> elements) {
    const uint32_t wg0 = CEIL_DIV(elements[0], pipeline->wg_denoms[0]);
    const uint32_t wg1 = CEIL_DIV(elements[1], pipeline->wg_denoms[1]);
    const uint32_t wg2 = CEIL_DIV(elements[2], pipeline->wg_denoms[2]);
@@ -4077,7 +4150,7 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
    vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
    ctx->device->device.updateDescriptorSets({ write_descriptor_set }, {});

-    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size, push_constants);
+    subctx->s->buffer.pushConstants(pipeline->layout, vk::ShaderStageFlagBits::eCompute, 0, push_constant_size(push_constants), push_constant_data(push_constants));
    subctx->s->buffer.bindPipeline(vk::PipelineBindPoint::eCompute, pipeline->pipeline);
    subctx->s->buffer.bindDescriptorSets(vk::PipelineBindPoint::eCompute,
                                pipeline->layout,
@@ -4540,7 +4613,7 @@ static void ggml_vk_matmul(
    ggml_vk_sync_buffers(subctx);
    if (split_k == 1) {
        const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, k, ne02, ne12, broadcast2, broadcast3, padded_n };
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, sizeof(vk_mat_mat_push_constants), &pc, { m, n, batch });
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, batch });
        return;
    }

@@ -4548,10 +4621,10 @@ static void ggml_vk_matmul(

    const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, CEIL_DIV(k, split_k), ne02, ne12, broadcast2, broadcast3, padded_n };
    // Make sure enough workgroups get assigned for split k to work
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, sizeof(vk_mat_mat_push_constants), &pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, batch });
    ggml_vk_sync_buffers(subctx);
    const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2.size() * sizeof(uint32_t), pc2.data(), { m * n * batch, 1, 1 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_matmul_split_k_reduce, { split_k_buffer, d }, pc2, { m * n * batch, 1, 1 });
 }

 static vk_pipeline ggml_vk_guess_matmul_id_pipeline(ggml_backend_vk_context * ctx, vk_matmul_pipeline& mmp, uint32_t m, uint32_t n, bool aligned, ggml_type src0_type) {
@@ -4599,7 +4672,7 @@ static void ggml_vk_matmul_id(
    ggml_vk_sync_buffers(subctx);
    const vk_mat_mat_id_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d,
                                              nei0, nei1, nbi1, ne11, padded_n };
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, sizeof(vk_mat_mat_id_push_constants), &pc, { m, nei1, n_as });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d, ids }, pc, { m, nei1, n_as });
 }

 static bool ggml_vk_dim01_contiguous(const ggml_tensor * tensor) {
@@ -4720,7 +4793,7 @@ static void ggml_vk_cpy_to_contiguous(ggml_backend_vk_context * ctx, vk_context&
    };
    init_pushconst_fastdiv(pc);
    ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(vk_op_unary_push_constants), &pc, elements);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, elements);
 }

 static vk_pipeline ggml_vk_get_quantize_pipeline(ggml_backend_vk_context * ctx, ggml_type type) {
@@ -4739,7 +4812,7 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
    vk_pipeline pipeline = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);

    ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, sizeof(uint32_t), &ne, { ne, 1, 1 });
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 1>{ne}, { ne, 1, 1 });
 }

 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -4939,7 +5012,7 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
    } else if (qx_needs_dequant) {
        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0, { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
    }
    if (y_non_contig) {
        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -5155,7 +5228,7 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
    ggml_vk_sync_buffers(subctx);
    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
                              { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 }, vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23} },
-                              sizeof(vk_mat_vec_push_constants), &pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
+                              pc, { groups_x, (uint32_t)(ne12 * ne13), groups_z });
 }

 static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5243,7 +5316,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
    }

    ggml_vk_sync_buffers(subctx);
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 6 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, workgroups_z });
+    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, workgroups_z });
 }

 static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5326,7 +5399,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
    const std::array<uint32_t, 9> pc = { (uint32_t)ne00, (uint32_t)ne01, row_stride_x, channel_stride_x, channel_stride_y, (uint32_t)(ne12 / ne02), (uint32_t)ne12, (uint32_t)(qy_shader_offset / ggml_type_size(src1->type)), (uint32_t)(d_shader_offset / ggml_type_size(dst->type)) };
    ggml_vk_sync_buffers(subctx);
    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
-        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, 7 * sizeof(uint32_t), &pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
+        { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz }, vk_subbuffer{ d_Qy, qy_buffer_offset, qy_sz + qy_shader_offset }, vk_subbuffer{ d_D, d_buffer_offset, d_sz + d_shader_offset } }, pc, { 1, (uint32_t)ne01, (uint32_t)ne12 });
 }

 static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -5542,7 +5615,7 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
        const std::vector<uint32_t> pc = { (uint32_t)ne01, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)ne10, (uint32_t)(ggml_nelements(src0)) };
        ggml_vk_sync_buffers(subctx);
        ggml_vk_dispatch_pipeline(ctx, subctx, to_fp16_vk_0,
-            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc.size() * sizeof(uint32_t), pc.data(), { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
+            { vk_subbuffer{ d_Qx, qx_buf_offset, qx_sz * ne02 * ne03 }, vk_subbuffer{ d_X, 0, x_sz * ne02 * ne03 } }, pc, { (uint32_t)(x_ne * ne02 * ne03), 1, 1});
    }
    if (y_non_contig) {
        ggml_vk_cpy_to_contiguous(ctx, subctx, to_fp16_vk_1, src1, { d_Qy, qy_buf_offset, VK_WHOLE_SIZE }, { d_Y, 0, VK_WHOLE_SIZE });
@@ -5762,7 +5835,7 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
    ggml_vk_dispatch_pipeline(ctx, subctx, dmmv,
        { vk_subbuffer{ d_X, x_buf_offset, x_sz * ne02 * ne03 },
        vk_subbuffer{ d_Y, y_buf_offset, y_sz * ne12 * ne13 }, vk_subbuffer{ d_D, d_buf_offset, d_sz * ne22 * ne23}, vk_subbuffer{ d_ids, ids_buf_offset, ids_sz } },
-        sizeof(vk_mat_vec_id_push_constants), &pc, { groups_x, (uint32_t)nei0, groups_z });
+        pc, { groups_x, (uint32_t)nei0, groups_z });
 }

 static void ggml_vk_mul_mat_id(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, const ggml_tensor * src2, ggml_tensor * dst, bool dryrun = false) {
@@ -6112,7 +6185,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                    // there's no more than one tile of rows (i.e. workgroups_x would have been
                                    // one). We reuse workgroups_x to mean the number of splits, so we need to
                                    // cancel out the divide by wg_denoms[0].
-                                    sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });
+                                    pc, { workgroups_x * pipeline->wg_denoms[0], workgroups_y, workgroups_z });

        ggml_vk_sync_buffers(subctx);
        const std::array<uint32_t, 3> pc2 = { D, (uint32_t)ne1, split_k };
@@ -6121,7 +6194,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                        vk_subbuffer{ctx->prealloc_split_k, 0, VK_WHOLE_SIZE},
                                        vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                    },
-                                    pc2.size() * uint32_t{sizeof(uint32_t)}, pc2.data(), { (uint32_t)ne1, 1, 1 });
+                                    pc2, { (uint32_t)ne1, 1, 1 });
    } else {
        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
                                    {
@@ -6131,7 +6204,7 @@ static void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx
                                        vk_subbuffer{d_M, m_buf_offset, VK_WHOLE_SIZE},
                                        vk_subbuffer{d_D, d_buf_offset, VK_WHOLE_SIZE},
                                    },
-                                    sizeof(vk_flash_attn_push_constants), &pc, { workgroups_x, workgroups_y, workgroups_z });
+                                    pc, { workgroups_x, workgroups_y, workgroups_z });
    }
 }

@@ -6392,6 +6465,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
            return ctx->device->pipeline_timestep_embedding_f32;
        }
        return nullptr;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_conv_transpose_1d_f32;
+        }
+        return nullptr;
    case GGML_OP_POOL_2D:
        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
            return ctx->device->pipeline_pool2d_f32;
@@ -6726,6 +6804,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
            uint32_t half_ceil = (dim + 1) / 2;
            elements = { half_ceil, (uint32_t)src0->ne[0], 1 };
        } break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        {
+            elements = {uint32_t(src0->ne[1]), 1, 1}; // parallelize in {Cout, 1, 1}
+        } break;
    case GGML_OP_POOL_2D:
        {
            const uint32_t N = dst->ne[3];
@@ -6800,7 +6882,7 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
        }

        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, subbuf_y, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else if (op == GGML_OP_ROPE || op == GGML_OP_ROPE_BACK) {
        // Empty src2 is possible in rope, but the shader needs a buffer
        vk_subbuffer subbuf_z;
@@ -6811,26 +6893,26 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
        }

        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, subbuf_z, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else if (op == GGML_OP_IM2COL) {
        // im2col uses only src1 and dst buffers
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else if (op == GGML_OP_COUNT_EQUAL) {
        ggml_vk_sync_buffers(subctx);
        // count_equal assumes that destination buffer is initialized with zeroes
        ggml_vk_buffer_memset_async(subctx, d_D, d_buf_offset, 0, d_sz);
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else if (use_src2) {
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_Z, z_buf_offset, z_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else if (use_src1) {
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_Y, y_buf_offset, y_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    } else {
        ggml_vk_sync_buffers(subctx);
-        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, sizeof(PC), &pc, elements);
+        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { vk_subbuffer{ d_X, x_buf_offset, x_sz }, vk_subbuffer{ d_D, d_buf_offset, d_sz } }, pc, elements);
    }
 }

@@ -6999,7 +7081,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
            vk_subbuffer{ d_srcs[4], src_offsets[4], src_sizes[4] },
            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
            vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
+        }, pc, elements);
    } else if (version == 7) {
        ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
            vk_subbuffer{ d_srcs[0], src_offsets[0], src_sizes[0] },
@@ -7010,7 +7092,7 @@ static void ggml_vk_op_f32_wkv(ggml_backend_vk_context * ctx, vk_context& subctx
            vk_subbuffer{ d_srcs[5], src_offsets[5], src_sizes[5] },
            vk_subbuffer{ d_srcs[6], src_offsets[6], src_sizes[6] },
            vk_subbuffer{ d_D, dst_offset, dst_size }
-        }, sizeof(vk_op_rwkv_wkv7_push_constants), &pc, elements);
+        }, pc, elements);
    } else {
        // shouldn't happen
        GGML_ASSERT(false);
@@ -7147,7 +7229,7 @@ static void ggml_vk_op_f32_opt_step_adamw(ggml_backend_vk_context * ctx, vk_cont
        vk_subbuffer{ d_GM, gm_offset, gm_size },
        vk_subbuffer{ d_GV, gv_offset, gv_size },
        vk_subbuffer{ d_P, p_offset, p_size },
-    }, sizeof(vk_op_push_constants), &pc, elements);
+    }, pc, elements);
 }

 static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
@@ -7529,6 +7611,37 @@ static void ggml_vk_timestep_embedding(ggml_backend_vk_context * ctx, vk_context
    }, dryrun);
 }

+static void ggml_vk_conv_transpose_1d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
+    // src0: (K, Cout, Cin, 1) -- kernel
+    // src1: (L, Cin, 1, 1) -- input
+    // dst: (*, Cout, 1, 1)
+
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+
+    GGML_TENSOR_BINARY_OP_LOCALS
+
+    GGML_ASSERT(nb00 == sizeof(float));
+    GGML_ASSERT(nb10 == sizeof(float));
+
+    const int32_t s0 = dst->op_params[0];
+
+    vk_op_conv_transpose_1d_push_constants p{};
+    p.Cout = static_cast<uint32_t>(ne01);
+    p.Cin = static_cast<uint32_t>(ne02);
+    p.K = static_cast<uint32_t>(ne00);
+    p.L = static_cast<uint32_t>(ne10);
+    p.KL = static_cast<uint32_t>(ne0);
+    p.nb01 = static_cast<uint32_t>(nb01 / nb00);
+    p.nb02 = static_cast<uint32_t>(nb02 / nb00);
+    p.nb11 = static_cast<uint32_t>(nb11 / nb10);
+    p.nb1 = static_cast<uint32_t>(nb1 / nb0);
+    p.s0 = static_cast<uint32_t>(s0);
+
+    ggml_vk_op_f32(ctx, subctx, src0, src1, nullptr, dst, GGML_OP_CONV_TRANSPOSE_1D, std::move(p), dryrun);
+}
+
 static void ggml_vk_pool_2d(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
    uint32_t op = static_cast<uint32_t>(dst->op_params[0]);
    const int32_t k1 = dst->op_params[1];
@@ -8005,7 +8118,7 @@ static void ggml_vk_test_dequant(ggml_backend_vk_context * ctx, size_t ne, ggml_
    vk_context subctx = ggml_vk_create_context(ctx, ctx->device->compute_queue);
    ggml_vk_ctx_begin(ctx->device, subctx);
    const std::vector<uint32_t> pc = { 1, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne, (uint32_t)ne };
-    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc.size() * sizeof(int), pc.data(), { (uint32_t)ne, 1, 1});
+    ggml_vk_dispatch_pipeline(ctx, subctx, p, { vk_subbuffer{ qx_buf, 0, qx_sz }, vk_subbuffer{ x_buf, 0, x_sz_f16 } }, pc, { (uint32_t)ne, 1, 1});
    ggml_vk_ctx_end(subctx);

    auto begin = std::chrono::high_resolution_clock::now();
@@ -8600,6 +8713,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
    case GGML_OP_COUNT_EQUAL:
    case GGML_OP_IM2COL:
    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
    case GGML_OP_POOL_2D:
    case GGML_OP_CONV_2D_DW:
    case GGML_OP_RWKV_WKV6:
@@ -8664,6 +8778,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
        case GGML_OP_COUNT_EQUAL:
        case GGML_OP_IM2COL:
        case GGML_OP_TIMESTEP_EMBEDDING:
+        case GGML_OP_CONV_TRANSPOSE_1D:
        case GGML_OP_POOL_2D:
        case GGML_OP_CONV_2D_DW:
        case GGML_OP_LEAKY_RELU:
@@ -8835,6 +8950,10 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
    case GGML_OP_TIMESTEP_EMBEDDING:
        ggml_vk_timestep_embedding(ctx, compute_ctx, src0, node, dryrun);

+        break;
+    case GGML_OP_CONV_TRANSPOSE_1D:
+        ggml_vk_conv_transpose_1d(ctx, compute_ctx, src0, src1, node, dryrun);
+
        break;
    case GGML_OP_POOL_2D:
        ggml_vk_pool_2d(ctx, compute_ctx, src0, node, dryrun);
@@ -8963,6 +9082,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
    case GGML_OP_COUNT_EQUAL:
    case GGML_OP_IM2COL:
    case GGML_OP_TIMESTEP_EMBEDDING:
+    case GGML_OP_CONV_TRANSPOSE_1D:
    case GGML_OP_POOL_2D:
    case GGML_OP_CONV_2D_DW:
    case GGML_OP_RWKV_WKV6:
@@ -9513,8 +9633,8 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
            if (ctx->device->query_pool) {
                ctx->device->device.destroyQueryPool(ctx->device->query_pool);
            }
-            VkQueryPoolCreateInfo query_create_info = { VK_STRUCTURE_TYPE_QUERY_POOL_CREATE_INFO };
-            query_create_info.queryType = VK_QUERY_TYPE_TIMESTAMP;
+            vk::QueryPoolCreateInfo query_create_info;
+            query_create_info.queryType = vk::QueryType::eTimestamp;
            query_create_info.queryCount = cgraph->n_nodes + 100;
            ctx->device->query_pool = ctx->device->device.createQueryPool(query_create_info);
            ctx->device->num_queries = query_create_info.queryCount;
@@ -9600,7 +9720,7 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg

        // Get the results and pass them to the logger
        std::vector<uint64_t> timestamps(cgraph->n_nodes + 1);
-        ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait);
+        VK_CHECK(ctx->device->device.getQueryPoolResults(ctx->device->query_pool, 0, cgraph->n_nodes + 1, (cgraph->n_nodes + 1)*sizeof(uint64_t), timestamps.data(), sizeof(uint64_t), vk::QueryResultFlagBits::e64 | vk::QueryResultFlagBits::eWait), "get timestamp results");
        for (int i = 0; i < cgraph->n_nodes; i++) {
            if (!ggml_vk_is_empty(cgraph->nodes[i])) {
                ctx->device->perf_logger->log_timing(cgraph->nodes[i], uint64_t((timestamps[i+1] - timestamps[i]) * ctx->device->properties.limits.timestampPeriod));
@@ -10024,6 +10144,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
        case GGML_OP_LEAKY_RELU:
        case GGML_OP_OPT_STEP_ADAMW:
            return true;
+        case GGML_OP_CONV_TRANSPOSE_1D:
+            return op->src[0]->type == GGML_TYPE_F32 && op->src[1]->type == GGML_TYPE_F32;
        default:
            return false;
    }
@@ -10170,8 +10292,9 @@ static bool ggml_vk_instance_portability_enumeration_ext_available(const std::ve
 static bool ggml_vk_khr_cooperative_matrix_support(const vk::PhysicalDeviceProperties& props, const vk::PhysicalDeviceDriverProperties& driver_props, vk_device_architecture arch) {
    switch (props.vendorID) {
    case VK_VENDOR_ID_INTEL:
-        // Intel drivers don't support coopmat properly yet
-        return false;
+        // Only allowing Xe2 GPU at the moment since Xe2 GPU can gain significant performance boost,
+        // while some older hardware (ex. Arc A770) has performance regressions
+        return arch == vk_device_architecture::INTEL_XE2;
    case VK_VENDOR_ID_AMD:
        if (driver_props.driverID == vk::DriverId::eAmdProprietary || driver_props.driverID == vk::DriverId::eAmdOpenSource) {
            // Workaround for AMD proprietary driver reporting support on all GPUs
@@ -10515,6 +10638,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
        const int32_t dim = tensor->op_params[0];
        const int32_t max_period = tensor->op_params[1];
        tensor_clone = ggml_timestep_embedding(ggml_ctx, src_clone[0], dim, max_period);
+    } else if (tensor->op == GGML_OP_CONV_TRANSPOSE_1D){
+        const int32_t s0 = tensor->op_params[0];
+        const int32_t p0 = tensor->op_params[1];
+        const int32_t d0 = tensor->op_params[2];
+        tensor_clone = ggml_conv_transpose_1d(ggml_ctx, src_clone[0], src_clone[1], s0, p0, d0);
    } else if (tensor->op == GGML_OP_POOL_2D) {
        enum ggml_op_pool op = static_cast<ggml_op_pool>(tensor->op_params[0]);
        const int32_t k0 = tensor->op_params[1];
@@ -0,0 +1,98 @@
+#version 450
+
+#include "types.comp"
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];};   // src0 - kernel:    [K, Cout, Cin]
+layout (binding = 1) readonly buffer B {B_TYPE data_b[];};   // src1 - input:     [L, Cin]
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};     // dst - result      [KL, Cout]
+
+layout(local_size_x = 128 , local_size_y = 1, local_size_z = 1) in;
+
+layout (push_constant) uniform parameter {
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t K;
+    uint32_t L;
+    uint32_t KL;
+
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb11;
+    uint32_t nb1;
+
+    int32_t s0;
+} p;
+
+
+uint32_t Cout_idx = gl_WorkGroupID.x;
+const uint32_t bs = gl_WorkGroupSize.x;
+uint32_t tid = gl_LocalInvocationID.x;
+// Code is more straightforward if we assume it is bs*s0+K instead of (bs-1)*s0+K.
+uint32_t tmp_len = bs*p.s0+p.K;
+shared D_TYPE tmp[4096];
+
+uint splitWork(uint workSize){
+    return (bs + workSize -1) / bs;
+}
+
+void main(){
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        if(idx < tmp_len){
+            tmp[idx] = 0.0;
+        }
+    }
+
+    uint32_t L_blocks = splitWork(p.L);
+    for(uint32_t L_block_id = 0; L_block_id < L_blocks; L_block_id++){
+        if(L_block_id > 0){
+            barrier();
+            // Shift values in tmp to the current processing window
+            for(int i = 0; i < splitWork(tmp_len); i++){
+                uint32_t idx = i*bs+tid;
+                if(idx >= bs*p.s0 && idx < tmp_len){
+                    tmp[idx-bs*p.s0] = tmp[idx];
+                    tmp[idx] = 0.0;
+                }else if(idx >= p.K && idx < bs*p.s0){
+                    tmp[idx] = 0.0;
+                }
+            }
+        }
+        barrier();
+
+        // Save contributions of the block to tmp
+        uint32_t L_idx = L_block_id*bs + tid;
+        for(uint32_t K_idx = 0; K_idx < p.K; K_idx++){
+            D_TYPE dp = 0.0;
+            for(uint32_t Cin_idx = 0; Cin_idx < p.Cin; Cin_idx++){
+                A_TYPE elemKrn = data_a[K_idx + Cout_idx * p.nb01 + Cin_idx * p.nb02];
+                if(L_idx < p.L){
+                    B_TYPE elemInp = data_b[L_idx + Cin_idx*p.nb11];
+                    dp = fma(elemKrn, elemInp, dp);
+                }
+            }
+            tmp[tid*p.s0 + K_idx] += dp;
+            barrier();
+        }
+
+        // Save the computed values except the last block that can have different size
+        uint32_t KLb_idx = L_block_id*bs*p.s0;
+        if(L_block_id < L_blocks-1){
+            for(uint32_t s0_idx = 0; s0_idx < p.s0; s0_idx++){
+                uint32_t sh_idx = p.s0*tid+s0_idx;
+                uint32_t KL_idx = KLb_idx+sh_idx;
+                if(KL_idx < p.KL){
+                    data_d[KL_idx + Cout_idx*p.nb1] = tmp[sh_idx];
+                }
+            }
+        }
+    }
+
+    for(uint32_t i = 0; i < splitWork(tmp_len); i++){
+        uint32_t idx = i*bs+tid;
+        uint32_t KL_idx = (L_blocks-1)*bs*p.s0+idx;
+        if(KL_idx < p.KL){
+            data_d[KL_idx + Cout_idx*p.nb1] = tmp[idx];
+        }
+    }
+}
@@ -622,6 +622,8 @@ void process_shaders() {

    string_to_spv("timestep_embedding_f32", "timestep_embedding.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));

+    string_to_spv("conv_transpose_1d_f32", "conv_transpose_1d.comp", {{"A_TYPE", "float"},  {"B_TYPE", "float"}, {"D_TYPE", "float"}});
+
    string_to_spv("pool2d_f32", "pool2d.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"D_TYPE", "float"}}));

    string_to_spv("rwkv_wkv6_f32", "wkv6.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
@@ -133,7 +133,7 @@ static void ggml_print_backtrace_symbols(void) {
 }
 #endif

-static void ggml_print_backtrace(void) {
+void ggml_print_backtrace(void) {
    const char * GGML_NO_BACKTRACE = getenv("GGML_NO_BACKTRACE");
    if (GGML_NO_BACKTRACE) {
        return;
@@ -160,6 +160,10 @@ static void ggml_print_backtrace(void) {
    const int parent_pid = getpid();
    const int child_pid = fork();
    if (child_pid < 0) { // error
+#if defined(__linux__)
+        close(lock[1]);
+        close(lock[0]);
+#endif
        return;
    } else if (child_pid == 0) { // child
        char attach[32];
@@ -167,6 +171,7 @@ static void ggml_print_backtrace(void) {
 #if defined(__linux__)
        close(lock[1]);
        (void) !read(lock[0], lock, 1);
+        close(lock[0]);
 #endif
        // try gdb
        execlp("gdb", "gdb", "--batch",
@@ -195,7 +200,7 @@ static void ggml_print_backtrace(void) {
    }
 }
 #else
-static void ggml_print_backtrace(void) {
+void ggml_print_backtrace(void) {
    // platform not supported
 }
 #endif
@@ -216,6 +221,8 @@ void ggml_abort(const char * file, int line, const char * fmt, ...) {
    abort();
 }

+// ggml_print_backtrace is registered with std::set_terminate by ggml.cpp
+
 //
 // logging
 //
@@ -0,0 +1,26 @@
+#include "ggml-impl.h"
+
+#include <cstdlib>
+#include <exception>
+
+static std::terminate_handler previous_terminate_handler;
+
+GGML_NORETURN static void ggml_uncaught_exception() {
+    ggml_print_backtrace();
+    if (previous_terminate_handler) {
+        previous_terminate_handler();
+    }
+    abort(); // unreachable unless previous_terminate_handler was nullptr
+}
+
+static bool ggml_uncaught_exception_init = []{
+    const char * GGML_NO_BACKTRACE = getenv("GGML_NO_BACKTRACE");
+    if (GGML_NO_BACKTRACE) {
+        return false;
+    }
+    const auto prev{std::get_terminate()};
+    GGML_ASSERT(prev != ggml_uncaught_exception);
+    previous_terminate_handler = prev;
+    std::set_terminate(ggml_uncaught_exception);
+    return true;
+}();
@@ -347,11 +347,28 @@ struct gguf_context * gguf_init_from_file_impl(FILE * file, struct gguf_init_par
    int64_t n_tensors = 0;

    if (ok && gr.read(ctx->version)) {
-        if (ctx->version == 1) {
+        if (ok && ctx->version == 0) {
+            GGML_LOG_ERROR("%s: bad GGUF version: %" PRIu32 "\n", __func__, ctx->version);
+            ok = false;
+        }
+
+        /*
+         * bit layout is different when reading non-native endian models.
+         * assuming that the GGUF version is 3, the non-native endian model
+         * would read it as 0x30000000. we can use the AND operation against
+         * the last 4 hexadecimal digits to check if the model is the same
+         * endianness as the host system.
+        */
+        if (ok && (ctx->version & 0x0000FFFF) == 0x00000000) {
+            GGML_LOG_ERROR("%s: failed to load model: this GGUF file version %" PRIu32 " is extremely large, is there a mismatch between the host and model endianness?\n", __func__, ctx->version);
+            ok = false;
+        }
+
+        if (ok && ctx->version == 1) {
            GGML_LOG_ERROR("%s: GGUFv1 is no longer supported, please use a more up-to-date version\n", __func__);
            ok = false;
        }
-        if (ctx->version > GGUF_VERSION) {
+        if (ok && ctx->version > GGUF_VERSION) {
            GGML_LOG_ERROR("%s: this GGUF file is version %" PRIu32 " but this software only supports up to version %d\n",
                __func__, ctx->version, GGUF_VERSION);
            ok = false;
@@ -177,6 +177,9 @@ class Keys:
        EMBEDDING_LENGTH = "{arch}.convnext.embedding_length"
        BLOCK_COUNT      = "{arch}.convnext.block_count"

+    class Classifier:
+        OUTPUT_LABELS = "{arch}.classifier.output_labels"
+
    class Tokenizer:
        MODEL                = "tokenizer.ggml.model"
        PRE                  = "tokenizer.ggml.pre"
@@ -1033,6 +1036,7 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = {
        MODEL_TENSOR.POS_EMBD,
        MODEL_TENSOR.OUTPUT_NORM,
        MODEL_TENSOR.ATTN_OUT_NORM,
+        MODEL_TENSOR.ATTN_QKV,
        MODEL_TENSOR.ATTN_Q,
        MODEL_TENSOR.ATTN_K,
        MODEL_TENSOR.ATTN_V,
@@ -49,6 +49,7 @@ class TensorInfo:
 class GGUFValue:
    value: Any
    type: GGUFValueType
+    sub_type: GGUFValueType | None = None


 class WriterState(Enum):
@@ -238,7 +239,7 @@ class GGUFWriter:

            for key, val in kv_data.items():
                kv_bytes += self._pack_val(key, GGUFValueType.STRING, add_vtype=False)
-                kv_bytes += self._pack_val(val.value, val.type, add_vtype=True)
+                kv_bytes += self._pack_val(val.value, val.type, add_vtype=True, sub_type=val.sub_type)

            fout.write(kv_bytes)

@@ -268,11 +269,11 @@ class GGUFWriter:
            fout.flush()
        self.state = WriterState.TI_DATA

-    def add_key_value(self, key: str, val: Any, vtype: GGUFValueType) -> None:
+    def add_key_value(self, key: str, val: Any, vtype: GGUFValueType, sub_type: GGUFValueType | None = None) -> None:
        if any(key in kv_data for kv_data in self.kv_data):
            raise ValueError(f'Duplicated key name {key!r}')

-        self.kv_data[0][key] = GGUFValue(value=val, type=vtype)
+        self.kv_data[0][key] = GGUFValue(value=val, type=vtype, sub_type=sub_type)

    def add_uint8(self, key: str, val: int) -> None:
        self.add_key_value(key,val, GGUFValueType.UINT8)
@@ -934,6 +935,9 @@ class GGUFWriter:
    def add_eom_token_id(self, id: int) -> None:
        self.add_uint32(Keys.Tokenizer.EOM_ID, id)

+    def add_classifier_output_labels(self, labels: Sequence[str]) -> None:
+        self.add_array(Keys.Classifier.OUTPUT_LABELS.format(arch=self.arch), labels)
+
    # for vision models

    def add_clip_has_vision_encoder(self, value: bool) -> None:
@@ -1022,7 +1026,7 @@ class GGUFWriter:
            pack_prefix = '<' if self.endianess == GGUFEndian.LITTLE else '>'
        return struct.pack(f'{pack_prefix}{fmt}', value)

-    def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool) -> bytes:
+    def _pack_val(self, val: Any, vtype: GGUFValueType, add_vtype: bool, sub_type: GGUFValueType | None = None) -> bytes:
        kv_data = bytearray()

        if add_vtype:
@@ -1043,7 +1047,9 @@ class GGUFWriter:
            if len(val) == 0:
                raise ValueError("Invalid GGUF metadata array. Empty array")

-            if isinstance(val, bytes):
+            if sub_type is not None:
+                ltype = sub_type
+            elif isinstance(val, bytes):
                ltype = GGUFValueType.UINT8
            else:
                ltype = GGUFValueType.get_type(val[0])
@@ -1521,19 +1521,21 @@ class GGUFEditorWindow(QMainWindow):
                    continue

                # Apply changes if any
+                sub_type = None
                if field.name in self.metadata_changes:
                    value_type, value = self.metadata_changes[field.name]
                    if value_type == GGUFValueType.ARRAY:
                        # Handle array values
-                        element_type, array_values = value
-                        writer.add_array(field.name, array_values)
-                    else:
-                        writer.add_key_value(field.name, value, value_type)
+                        sub_type, value = value
                else:
                    # Copy original value
                    value = field.contents()
-                    if value is not None and field.types:
-                        writer.add_key_value(field.name, value, field.types[0])
+                    value_type = field.types[0]
+                    if value_type == GGUFValueType.ARRAY:
+                        sub_type = field.types[-1]
+
+                if value is not None:
+                    writer.add_key_value(field.name, value, value_type, sub_type=sub_type)

            # Add new metadata
            for key, (value_type, value) in self.metadata_changes.items():
@@ -1541,7 +1543,12 @@ class GGUFEditorWindow(QMainWindow):
                if self.reader.get_field(key) is not None:
                    continue

-                writer.add_key_value(key, value, value_type)
+                sub_type = None
+                if value_type == GGUFValueType.ARRAY:
+                    # Handle array values
+                    sub_type, value = value
+
+                writer.add_key_value(key, value, value_type, sub_type=sub_type)

            # Add tensors (including data)
            for tensor in self.reader.tensors:
@@ -24,6 +24,7 @@ class MetadataDetails(NamedTuple):
    type: gguf.GGUFValueType
    value: Any
    description: str = ''
+    sub_type: gguf.GGUFValueType | None = None


 def get_field_data(reader: gguf.GGUFReader, key: str) -> Any:
@@ -57,7 +58,9 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
            logger.debug(f'Removing {field.name}')
            continue

-        old_val = MetadataDetails(field.types[0], field.contents())
+        val_type = field.types[0]
+        sub_type = field.types[-1] if val_type == gguf.GGUFValueType.ARRAY else None
+        old_val = MetadataDetails(val_type, field.contents(), sub_type=sub_type)
        val = new_metadata.get(field.name, old_val)

        if field.name in new_metadata:
@@ -67,7 +70,7 @@ def copy_with_new_metadata(reader: gguf.GGUFReader, writer: gguf.GGUFWriter, new
            logger.debug(f'Copying {field.name}')

        if val.value is not None:
-            writer.add_key_value(field.name, val.value, val.type)
+            writer.add_key_value(field.name, val.value, val.type, sub_type=sub_type if val.sub_type is None else val.sub_type)

    if gguf.Keys.Tokenizer.CHAT_TEMPLATE in new_metadata:
        logger.debug('Adding chat template(s)')
@@ -157,6 +157,7 @@ class TensorNameMap:
            "h.{bid}.attn.c_attn",                                                 # gpt2
            "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
            "encoder.layers.{bid}.attn.Wqkv",                                      # nomic-bert
+            "encoder.layers.{bid}.mixer.Wqkv",                                     # jina
            "model.layers.{bid}.self_attn.qkv_proj",                               # phi3
            "encoder.layers.{bid}.self_attention.query_key_value",                 # chatglm
            "transformer.layers.{bid}.attn.qkv_proj",                              # openelm
@@ -168,6 +169,7 @@ class TensorNameMap:
            "model.layers.{bid}.self_attn.q_proj_no_perm",               # llama-custom
            "layers.{bid}.attention.wq",                                 # llama-pth
            "encoder.layer.{bid}.attention.self.query",                  # bert
+            "transformer.layer.{bid}.attention.q_lin",                   # distillbert
            "transformer.h.{bid}.attn.q_proj",                           # gpt-j
            "model.layers.layers.{bid}.self_attn.q_proj",                # plamo
            "model.layers.{bid}.attention.wq",                           # internlm2
@@ -182,6 +184,7 @@ class TensorNameMap:
            "model.layers.{bid}.self_attn.k_proj_no_perm",             # llama-custom
            "layers.{bid}.attention.wk",                               # llama-pth
            "encoder.layer.{bid}.attention.self.key",                  # bert
+            "transformer.layer.{bid}.attention.k_lin",                 # distillbert
            "transformer.h.{bid}.attn.k_proj",                         # gpt-j
            "transformer.h.{bid}.attn.k",                              # refact
            "model.layers.layers.{bid}.self_attn.k_proj",              # plamo
@@ -196,6 +199,7 @@ class TensorNameMap:
            "model.layers.{bid}.self_attn.v_proj",                       # llama-hf nemotron olmoe olmo2 phimoe
            "layers.{bid}.attention.wv",                                 # llama-pth
            "encoder.layer.{bid}.attention.self.value",                  # bert
+            "transformer.layer.{bid}.attention.v_lin",                   # distillbert
            "transformer.h.{bid}.attn.v_proj",                           # gpt-j
            "transformer.h.{bid}.attn.v",                                # refact
            "model.layers.layers.{bid}.self_attn.v_proj",                # plamo
@@ -216,6 +220,7 @@ class TensorNameMap:
            "model.layers.{bid}.self_attn.linear_attn",                     # deci
            "layers.{bid}.attention.wo",                                    # llama-pth
            "encoder.layer.{bid}.attention.output.dense",                   # bert
+            "transformer.layer.{bid}.attention.out_lin",                    # distillbert
            "transformer.h.{bid}.attn.out_proj",                            # gpt-j
            "language_model.encoder.layers.{bid}.self_attention.dense",     # persimmon
            "model.layers.{bid}.self_attn.dense",                           # persimmon
@@ -224,6 +229,7 @@ class TensorNameMap:
            "model.layers.layers.{bid}.self_attn.o_proj",                   # plamo
            "model.layers.{bid}.attention.wo",                              # internlm2
            "encoder.layers.{bid}.attn.out_proj",                           # nomic-bert
+            "encoder.layers.{bid}.mixer.out_proj",                          # jina
            "transformer.decoder_layer.{bid}.multi_head_attention.linear",  # Grok
            "transformer.blocks.{bid}.norm_attn_norm.attn.out_proj",        # dbrx
            "encoder.layers.{bid}.self_attention.dense",                    # chatglm
@@ -235,6 +241,7 @@ class TensorNameMap:
        # Attention output norm
        MODEL_TENSOR.ATTN_OUT_NORM: (
            "encoder.layer.{bid}.attention.output.LayerNorm",  # bert
+            "transformer.layer.{bid}.sa_layer_norm",           # distillbert
            "encoder.layers.{bid}.norm1",                      # nomic-bert
            "transformer.decoder_layer.{bid}.rms_norm_1",      # Grok
            "transformer.blocks.{bid}.norm_attn_norm.norm_2",  # dbrx
@@ -311,6 +318,7 @@ class TensorNameMap:
            "model.layers.{bid}.mlp.up_proj",                         # llama-hf refact nemotron olmo2
            "layers.{bid}.feed_forward.w3",                           # llama-pth
            "encoder.layer.{bid}.intermediate.dense",                 # bert
+            "transformer.layer.{bid}.ffn.lin1",                       # distillbert
            "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
            "transformer.h.{bid}.mlp.linear_3",                       # refact
            "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
@@ -394,6 +402,7 @@ class TensorNameMap:
            "model.layers.{bid}.mlp.down_proj",                       # llama-hf nemotron olmo2
            "layers.{bid}.feed_forward.w2",                           # llama-pth
            "encoder.layer.{bid}.output.dense",                       # bert
+            "transformer.layer.{bid}.ffn.lin2",                       # distillbert
            "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
            "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
            "model.layers.{bid}.mlp.dense_4h_to_h",                   # persimmon
@@ -455,6 +464,7 @@ class TensorNameMap:

        MODEL_TENSOR.LAYER_OUT_NORM: (
            "encoder.layer.{bid}.output.LayerNorm",         # bert
+            "transformer.layer.{bid}.output_layer_norm",    # distillbert
            "encoder.layers.{bid}.norm2",                   # nomic-bert
            "transformer.decoder_layer.{bid}.rms_norm_3",   # Grok
            "encoder.layer.{bid}.mlp.layernorm",            # jina-bert-v2
@@ -825,6 +835,7 @@ class TensorNameMap:
        MODEL_TENSOR.CLS: (
            "classifier",       # jina
            "classifier.dense", # roberta
+            "pre_classifier",   # distillbert
        ),

        MODEL_TENSOR.CLS_OUT: (
@@ -231,7 +231,7 @@ class SafetensorRemote:
        response.raise_for_status()

        # Get raw byte data
-        return response.content[:size]
+        return response.content[slice(size if size > -1 else None)]

    @classmethod
    def check_file_exist(cls, url: str) -> bool:
@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "gguf"
-version = "0.16.3"
+version = "0.17.0"
 description = "Read and write ML models in GGUF for GGML"
 authors = ["GGML <ggml@ggml.ai>"]
 packages = [
@@ -61,7 +61,10 @@ extern "C" {
    struct llama_model;
    struct llama_context;
    struct llama_sampler;
-    struct llama_kv_cache;
+
+    typedef struct llama_memory_i * llama_memory_t;
+
+    struct llama_kv_cache; // DEPRECATED (use llama_memory instead)

    typedef int32_t llama_pos;
    typedef int32_t llama_token;
@@ -259,9 +262,9 @@ extern "C" {
        llama_token  *  token;
        float        *  embd;
        llama_pos    *  pos;
-        int32_t      *  n_seq_id;
-        llama_seq_id ** seq_id;
-        int8_t       *  logits; // TODO: rename this to "output"
+        int32_t      *  n_seq_id; // TODO: remove, should belong to only 1 sequence
+        llama_seq_id ** seq_id;   // TODO: become llama_seq_id * seq_id;
+        int8_t       *  logits;   // TODO: rename this to "output"
    } llama_batch;

    enum llama_model_kv_override_type {
@@ -366,6 +369,8 @@ extern "C" {
        bool no_perf;     // measure performance timings
        bool op_offload;  // offload host tensor operations to device
        bool swa_full;    // use full-size SWA cache (https://github.com/ggml-org/llama.cpp/pull/13194#issuecomment-2868343055)
+                          // NOTE: setting to false when n_seq_max > 1 can cause bad performance in some cases
+                          //       ref: https://github.com/ggml-org/llama.cpp/pull/13845#issuecomment-2924800573
    };

    // model quantization parameters
@@ -491,9 +496,11 @@ extern "C" {
    DEPRECATED(LLAMA_API int32_t llama_n_vocab    (const struct llama_vocab * vocab), "use llama_vocab_n_tokens instead");

    LLAMA_API const struct llama_model * llama_get_model   (const struct llama_context * ctx);
-    LLAMA_API    struct llama_kv_cache * llama_get_kv_self (      struct llama_context * ctx);
+    LLAMA_API           llama_memory_t   llama_get_memory  (const struct llama_context * ctx);
    LLAMA_API  enum llama_pooling_type   llama_pooling_type(const struct llama_context * ctx); // TODO: rename to llama_get_pooling_type

+    DEPRECATED(LLAMA_API struct llama_kv_cache * llama_get_kv_self(struct llama_context * ctx), "use llama_get_memory instead");
+
    LLAMA_API const struct llama_vocab * llama_model_get_vocab(const struct llama_model * model);
    LLAMA_API enum llama_rope_type       llama_model_rope_type(const struct llama_model * model);

@@ -502,10 +509,18 @@ extern "C" {
    LLAMA_API int32_t llama_model_n_layer    (const struct llama_model * model);
    LLAMA_API int32_t llama_model_n_head     (const struct llama_model * model);
    LLAMA_API int32_t llama_model_n_head_kv  (const struct llama_model * model);
+    LLAMA_API int32_t llama_model_n_swa      (const struct llama_model * model);

    // Get the model's RoPE frequency scaling factor
    LLAMA_API float llama_model_rope_freq_scale_train(const struct llama_model * model);

+    // Returns the number of classifier outputs (only valid for classifier models)
+    // Undefined behavior for non-classifier models
+    LLAMA_API uint32_t llama_model_n_cls_out(const struct llama_model * model);
+
+    // Returns label of classifier output by index (<n_cls_out). Returns nullptr if no label provided
+    LLAMA_API const char * llama_model_cls_label(const struct llama_model * model, uint32_t i);
+
    LLAMA_API enum llama_vocab_type llama_vocab_type(const struct llama_vocab * vocab);

    LLAMA_API int32_t llama_vocab_n_tokens(const struct llama_vocab * vocab);
@@ -606,7 +621,81 @@ extern "C" {
                         int32_t   il_end);

    //
-    // KV cache
+    // Memory
+    //
+
+    // Clear the memory contents
+    // If data == true, the data buffers will also be cleared together with the metadata
+    LLAMA_API void llama_memory_clear(
+            llama_memory_t mem,
+                      bool data);
+
+    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
+    // seq_id < 0 : match any sequence
+    // p0 < 0     : [0,  p1]
+    // p1 < 0     : [p0, inf)
+    LLAMA_API bool llama_memory_seq_rm(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Copy all tokens that belong to the specified sequence to another sequence
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_cp(
+            llama_memory_t mem,
+              llama_seq_id seq_id_src,
+              llama_seq_id seq_id_dst,
+                 llama_pos p0,
+                 llama_pos p1);
+
+    // Removes all tokens that do not belong to the specified sequence
+    LLAMA_API void llama_memory_seq_keep(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_add(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                 llama_pos delta);
+
+    // Integer division of the positions by factor of `d > 1`
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_memory_seq_div(
+            llama_memory_t mem,
+              llama_seq_id seq_id,
+                 llama_pos p0,
+                 llama_pos p1,
+                       int d);
+
+    // Returns the smallest position present in the memory for the specified sequence
+    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_min(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Returns the largest position present in the memory for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the memory
+    // Return -1 if the sequence is empty
+    LLAMA_API llama_pos llama_memory_seq_pos_max(
+            llama_memory_t mem,
+              llama_seq_id seq_id);
+
+    // Check if the memory supports shifting
+    LLAMA_API bool llama_memory_can_shift(llama_memory_t mem);
+
+    //
+    // KV cache for self-attention (TODO: deprecate in favor of llama_memory)
    //

    // Returns the number of tokens in the KV cache (slow, use only for debug)
@@ -619,93 +708,103 @@ extern "C" {
               "Use llama_kv_self_seq_pos_max() and llama_kv_self_seq_pos_min() instead (https://github.com/ggml-org/llama.cpp/issues/13793)");

    // Clear the KV cache - both cell info is erased and KV data is zeroed
-    LLAMA_API void llama_kv_self_clear(
-            struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_clear(
+                struct llama_context * ctx),
+            "Use llama_memory_clear() instead");

    // Removes all tokens that belong to the specified sequence and have positions in [p0, p1)
    // Returns false if a partial sequence cannot be removed. Removing a whole sequence never fails
    // seq_id < 0 : match any sequence
    // p0 < 0     : [0,  p1]
    // p1 < 0     : [p0, inf)
-    LLAMA_API bool llama_kv_self_seq_rm(
+    DEPRECATED(LLAMA_API bool llama_kv_self_seq_rm(
            struct llama_context * ctx,
                    llama_seq_id   seq_id,
                       llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_rm() instead");

    // Copy all tokens that belong to the specified sequence to another sequence
    // Note that this does not allocate extra KV cache memory - it simply assigns the tokens to the new sequence
    // p0 < 0 : [0,  p1]
    // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_cp(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_cp(
            struct llama_context * ctx,
                    llama_seq_id   seq_id_src,
                    llama_seq_id   seq_id_dst,
                       llama_pos   p0,
-                       llama_pos   p1);
+                       llama_pos   p1),
+            "Use llama_memory_seq_cp() instead");

    // Removes all tokens that do not belong to the specified sequence
-    LLAMA_API void llama_kv_self_seq_keep(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_keep(
            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_keep() instead");

    // Adds relative position "delta" to all tokens that belong to the specified sequence and have positions in [p0, p1)
    // If the KV cache is RoPEd, the KV data is updated accordingly:
    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
    // p0 < 0 : [0,  p1]
    // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_add(
+    DEPRECATED(LLAMA_API void llama_kv_self_seq_add(
            struct llama_context * ctx,
                    llama_seq_id   seq_id,
                       llama_pos   p0,
                       llama_pos   p1,
-                       llama_pos   delta);
+                       llama_pos   delta),
+            "Use llama_memory_seq_add() instead");

    // Integer division of the positions by factor of `d > 1`
    // If the KV cache is RoPEd, the KV data is updated accordingly:
    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
    // p0 < 0 : [0,  p1]
    // p1 < 0 : [p0, inf)
-    LLAMA_API void llama_kv_self_seq_div(
+    DEPRECATED(void llama_kv_self_seq_div(
            struct llama_context * ctx,
                    llama_seq_id   seq_id,
                       llama_pos   p0,
                       llama_pos   p1,
-                             int   d);
+                             int   d),
+            "Use llama_memory_seq_div() instead");

    // Returns the smallest position present in the KV cache for the specified sequence
    // This is typically non-zero only for SWA caches
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
    // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_min(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_min(
            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_min() instead");

    // Returns the largest position present in the KV cache for the specified sequence
+    // Note that all positions in the range [pos_min, pos_max] are guaranteed to be present in the KV cache
    // Return -1 if the sequence is empty
-    LLAMA_API llama_pos llama_kv_self_seq_pos_max(
+    DEPRECATED(LLAMA_API llama_pos llama_kv_self_seq_pos_max(
            struct llama_context * ctx,
-                    llama_seq_id   seq_id);
+                    llama_seq_id   seq_id),
+            "Use llama_memory_seq_pos_max() instead");

    // Defragment the KV cache
    // This will be applied:
    //   - lazily on next llama_decode()
-    //   - explicitly with llama_kv_self_update()
-    LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_defrag(struct llama_context * ctx),
+            "simply remove this call, the context will automatically decide when to do a defragmentation based on 'defrag_thold'");

    // Check if the context supports KV cache shifting
-    LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx);
+    DEPRECATED(LLAMA_API bool llama_kv_self_can_shift(const struct llama_context * ctx),
+            "use llama_memory_can_shift() instead");

    // Apply the KV cache updates (such as K-shifts, defragmentation, etc.)
-    LLAMA_API void llama_kv_self_update(struct llama_context * ctx);
+    DEPRECATED(LLAMA_API void llama_kv_self_update(struct llama_context * ctx),
+            "simply remove this call, updates are applied lazily on the next llama_decode()");

    //
    // State / sessions
    //

    // Returns the *actual* size in bytes of the state
-    // (logits, embedding and kv_cache)
+    // (logits, embedding and memory)
    // Only use when saving the state, not when restoring it, otherwise the size may be too small.
    LLAMA_API size_t llama_state_get_size(struct llama_context * ctx);
    LLAMA_API DEPRECATED(size_t llama_get_state_size(struct llama_context * ctx),
@@ -761,12 +860,12 @@ extern "C" {
                          size_t   n_token_count),
        "use llama_state_save_file instead");

-    // Get the exact size needed to copy the KV cache of a single sequence
+    // Get the exact size needed to copy the state of a single sequence
    LLAMA_API size_t llama_state_seq_get_size(
            struct llama_context * ctx,
                    llama_seq_id   seq_id);

-    // Copy the KV cache of a single sequence into the specified buffer
+    // Copy the state of a single sequence into the specified buffer
    LLAMA_API size_t llama_state_seq_get_data(
            struct llama_context * ctx,
                         uint8_t * dst,
@@ -832,16 +931,16 @@ extern "C" {
    // For encode-decoder contexts, processes the batch using the encoder.
    // Can store the encoder output internally for later use by the decoder's cross-attention layers.
    //   0 - success
-    // < 0 - error. the KV cache state is restored to the state before this call
+    // < 0 - error. the memory state is restored to the state before this call
    LLAMA_API int32_t llama_encode(
            struct llama_context * ctx,
              struct llama_batch   batch);

    // Process a batch of tokens.
-    // Requires KV cache.
+    // Requires the context to have a memory.
    // For encode-decoder contexts, processes the batch using the decoder.
    // Positive return values does not mean a fatal error, but rather a warning.
-    // Upon non-zero return values, the KV cache state is restored to the state before this call
+    // Upon non-zero return values, the memory state is restored to the state before this call
    //    0 - success
    //    1 - could not find a KV slot for the batch (try reducing the size of the batch or increase the context)
    //    2 - aborted
@@ -912,7 +1011,7 @@ extern "C" {

    // Get the embeddings for a sequence id
    // Returns NULL if pooling_type is LLAMA_POOLING_TYPE_NONE
-    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[1] with the rank of the sequence
+    // when pooling_type == LLAMA_POOLING_TYPE_RANK, returns float[n_cls_out] with the rank(s) of the sequence
    // otherwise: float[n_embd] (1-dimensional)
    LLAMA_API float * llama_get_embeddings_seq(struct llama_context * ctx, llama_seq_id seq_id);

@@ -1,6 +1,6 @@
 ied 4 ½ months
 __ggml_vocab_test__
-Führer
+Äpfel
 __ggml_vocab_test__

 __ggml_vocab_test__
@@ -1,5 +1,5 @@
 29464 2094 1018 1092 2706
- 11865 17875
+ 9706 7959 2140



@@ -1,112 +0,0 @@
-ied 4 ½ months
-__ggml_vocab_test__
-Führer
-__ggml_vocab_test__
-
-__ggml_vocab_test__
- 
-__ggml_vocab_test__
-  
-__ggml_vocab_test__
-   
-__ggml_vocab_test__
-	
-__ggml_vocab_test__
-
-
-__ggml_vocab_test__
-
-
-
-__ggml_vocab_test__
-
-
-
-
-__ggml_vocab_test__
-	
-
-__ggml_vocab_test__
-Hello world
-__ggml_vocab_test__
- Hello world
-__ggml_vocab_test__
-Hello World
-__ggml_vocab_test__
- Hello World
-__ggml_vocab_test__
- Hello World!
-__ggml_vocab_test__
-Hello, world!
-__ggml_vocab_test__
- Hello, world!
-__ggml_vocab_test__
- this is 🦙.cpp
-__ggml_vocab_test__
-w048 7tuijk dsdfhu
-__ggml_vocab_test__
-нещо на Български
-__ggml_vocab_test__
-កាន់តែពិសេសអាចខលចេញ
-__ggml_vocab_test__
-🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)
-__ggml_vocab_test__
-Hello
-__ggml_vocab_test__
- Hello
-__ggml_vocab_test__
-  Hello
-__ggml_vocab_test__
-   Hello
-__ggml_vocab_test__
-    Hello
-__ggml_vocab_test__
-    Hello
-    Hello
-__ggml_vocab_test__
- (
-__ggml_vocab_test__
-
- =
-__ggml_vocab_test__
-' era
-__ggml_vocab_test__
-Hello, y'all! How are you 😁 ?我想在apple工作1314151天～
-__ggml_vocab_test__
-!!!!!!
-__ggml_vocab_test__
-3
-__ggml_vocab_test__
-33
-__ggml_vocab_test__
-333
-__ggml_vocab_test__
-3333
-__ggml_vocab_test__
-33333
-__ggml_vocab_test__
-333333
-__ggml_vocab_test__
-3333333
-__ggml_vocab_test__
-33333333
-__ggml_vocab_test__
-333333333
-__ggml_vocab_test__
-Cửa Việt
-__ggml_vocab_test__
- discards
-__ggml_vocab_test__
-
- 
-
- 
-
-
- 	 		 	
-  
-   
-    
-     
-🚀 (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
-__ggml_vocab_test__
@@ -1,46 +0,0 @@
- 17245 16604 16403 16604 33583 18355
- 16421 51153
-
- 16604
- 16650
- 16650 16604
- 16581
- 16582
- 16582 16582
- 16582 16582 16582
- 16581 16582
- 31596 17394
- 34926 17394
- 31596 18671
- 34926 18671
- 34926 18671 16384
- 31596 16395 17394 16384
- 34926 16395 17394 16384
- 16811 16704 20410 16483 16631 16397 52854
- 16470 16399 16403 16407 16604 16406 35764 38185 51595 22592 26639
- 29479 23955 17012 20103 25527 27670 17408 19005 21473 24774
- 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 21954 16607 21954 16633 21954 16611 29409 16607 21954 16615
- 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 16604 16391 24664 17153 57169 16721 16872 17073 17304 28729 16392
- 31596
- 34926
- 16650 31596
- 16650 34926
- 16696 31596
- 16696 31596 16582 16696 31596
- 16604 16391
- 16582 16604 16412
- 16390 22623
- 31596 16395 16712 16390 16828 16384 17674 16769 16732 23686 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636
- 16384 16384 16384 16384 16384 16384
- 16402
- 16402 16402
- 16402 16402 16402
- 16402 16402 16402 16402
- 16402 16402 16402 16402 16402
- 16402 16402 16402 16402 16402 16402
- 16402 16402 16402 16402 16402 16402 16402
- 16402 16402 16402 16402 16402 16402 16402 16402
- 16402 16402 16402 16402 16402 16402 16402 16402 16402
- 16418 19038 16639 16448 24315 33727 16467
- 18765 17981
- 16582 16604 16582 16582 16604 16582 16582 16582 16604 16581 16604 16581 16581 16604 16581 16582 16650 16582 16650 16604 16582 16696 16582 16696 16604 16582 52351 16604 16391 25825 16392 23686 16498 39161 18885 16618 16488 30853 16604 16391 54124 17153 25134 16656 18476 26169 16895 16392 62193 16611 20410 16483 16631 18885 16483 16631 16604 16402 16604 16402 16402 16604 16402 16402 16402 16604 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16604 16402 16402 16402 16402 16402 16402 16402 16402 16604 16402 16397 16402 16604 16402 16397 16397 16402 16604 16402 16397 16397 16397 16402 16604 54254 42231 48084 29409 16617 61889 29409 16608 21954 16628 21954 16499 58445 29409 16607 58445 21954 16479 42231 21954 16611 27683 16607 16604 16414 24427 16623 41809 16495 28999 36469 45292 30197 16400 16402 16400 16403 16400 16404 16400 43969 65211 16636 16604 16396 16396 16396 16396 16396 16396 16412 16412 16412 16412 16412 16412 16412 27268 23955 17012 20103 25527 27670 17408 19005 21473 24774 16604 16390 16390 16390 16390 16390 16390 16447 16447 16447 16447 16447 16447 16447 16385 16385 16385 16385 16397 16397 16397 16397 16397 16397 16384 16384 16384 16384 16384 16384 16414 16414 16414 16414 16414 16414 16687 16390 16690 16992 16604 16390 61797 16733 16390 16466 16986 16395 16604 16390 17879 16732 17811 16414 16604 16390 16428 16804 17811 16687 16390 16683 17190 16728 16395 16604 16390 16419 16732 16945 16991 25251 16414 17119 16390 38127 16641 16390 16459 16427
@@ -1,6 +1,6 @@
 ied 4 ½ months
 __ggml_vocab_test__
-Führer
+Äpfel
 __ggml_vocab_test__

 __ggml_vocab_test__
@@ -1,5 +1,5 @@
 2536 228 27 228 22957 6983
- 45 193433
+ 90711 87 20910

 228
 1667
--- a/Show More
+++ b/Show More