vulkan: fuse adds (#15252)

* vulkan: fuse adds

Fuse adds that have the same shape, which are common in MoE models.
It will currently fuse up to 6 adds, because we assume no more than
8 descriptors per dispatch. But this could be changed.

* check runtimeDescriptorArray feature

* disable multi_add for Intel due to likely driver bug

.devops/cloud-v-pipeline

@@ -1,22 +0,0 @@
-node('x86_runner1'){            // Running on x86 runner containing latest vector qemu, latest vector gcc and all the necessary libraries
-    stage('Cleanup'){
-        cleanWs()               // Cleaning previous CI build in workspace
-    }
-    stage('checkout repo'){
-        retry(5){               // Retry if the cloning fails due to some reason
-            checkout scm        // Clone the repo on Runner
-        }
-    }
-    stage('Compiling llama.cpp'){
-        sh'''#!/bin/bash
-            make RISCV=1 RISCV_CROSS_COMPILE=1 # Compiling llama for RISC-V
-        '''
-    }
-    stage('Running llama.cpp'){
-        sh'''#!/bin/bash
-            module load gnu-bin2/0.1            # loading latest versions of vector qemu and vector gcc
-            qemu-riscv64 -L /softwares/gnu-bin2/sysroot  -cpu rv64,v=true,vlen=256,elen=64,vext_spec=v1.0 ./llama-cli -m /home/alitariq/codellama-7b.Q4_K_M.gguf -p "Anything" -n 9 > llama_log.txt            # Running llama.cpp on vector qemu-riscv64
-            cat llama_log.txt                   # Printing results
-        '''
-    }
-}

.devops/cpu.Dockerfile

@@ -4,8 +4,6 @@ FROM ubuntu:$UBUNTU_VERSION AS build
 
 ARG TARGETARCH
 
-ARG GGML_CPU_ARM_ARCH=armv8-a
-
 RUN apt-get update && \
     apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
@@ -13,10 +11,8 @@ WORKDIR /app
 
 COPY . .
 
-RUN if [ "$TARGETARCH" = "amd64" ]; then \
+RUN if [ "$TARGETARCH" = "amd64" ] || [ "$TARGETARCH" = "arm64" ]; then \
         cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_BACKEND_DL=ON -DGGML_CPU_ALL_VARIANTS=ON; \
-    elif [ "$TARGETARCH" = "arm64" ]; then \
-        cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DGGML_NATIVE=OFF -DLLAMA_BUILD_TESTS=OFF -DGGML_CPU_ARM_ARCH=${GGML_CPU_ARM_ARCH}; \
     else \
         echo "Unsupported architecture"; \
         exit 1; \

.devops/cuda.Dockerfile

@@ -60,8 +60,7 @@ RUN apt-get update \
     git \
     python3 \
     python3-pip \
-    && pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt \
+    && pip install --break-system-packages -r requirements.txt \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.github/ISSUE_TEMPLATE/010-bug-compilation.yml

@@ -40,7 +40,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true

.github/ISSUE_TEMPLATE/011-bug-results.yml

@@ -42,7 +42,7 @@ body:
     attributes:
         label: GGML backends
         description: Which GGML backends do you know to be affected?
-        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL]
+        options: [AMX, BLAS, CPU, CUDA, HIP, Metal, Musa, RPC, SYCL, Vulkan, OpenCL, zDNN]
         multiple: true
     validations:
       required: true

.github/labeler.yml

@@ -22,6 +22,11 @@ Vulkan:
         - any-glob-to-any-file:
             - ggml/include/ggml-vulkan.h
             - ggml/src/ggml-vulkan/**
+IBM zDNN:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml/include/ggml-zdnn.h
+            - ggml/src/ggml-zdnn/**
 documentation:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/build-riscv-native.yml

@@ -0,0 +1,43 @@
+name: Build on RISCV Linux Machine by Cloud-V
+on:
+  workflow_dispatch:
+  workflow_call:
+
+jobs:
+  bianbu-riscv64-native: # Bianbu 2.2
+    runs-on: self-hosted
+
+    steps:
+      - name: Install prerequisites
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y libatomic1
+      - uses: actions/checkout@v4
+      - name: Setup Riscv
+        run: |
+          sudo apt-get update || true
+          sudo apt-get install -y --no-install-recommends \
+                  build-essential \
+                  gcc-14-riscv64-linux-gnu \
+                  g++-14-riscv64-linux-gnu \
+                  cmake
+
+      - 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=riscv64 \
+                         -DCMAKE_C_COMPILER=riscv64-linux-gnu-gcc-14 \
+                         -DCMAKE_CXX_COMPILER=riscv64-linux-gnu-g++-14 \
+                         -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
+                         -DCMAKE_FIND_ROOT_PATH=/usr/lib/riscv64-linux-gnu \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_PROGRAM=NEVER \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_LIBRARY=ONLY \
+                         -DCMAKE_FIND_ROOT_PATH_MODE_INCLUDE=BOTH
+
+          cmake --build build --config Release -j $(nproc)

.github/workflows/build.yml

@@ -64,7 +64,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -104,7 +104,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -144,7 +144,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64-webgpu
           evict-old-files: 1d
@@ -199,7 +199,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -251,7 +251,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-sanitizer-${{ matrix.sanitizer }}
           evict-old-files: 1d
@@ -330,7 +330,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-latest-cmake-rpc
           evict-old-files: 1d
@@ -363,7 +363,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -400,7 +400,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-webgpu
           evict-old-files: 1d
@@ -443,7 +443,7 @@ jobs:
 
   ubuntu-22-cmake-hip:
     runs-on: ubuntu-22.04
-    container: rocm/dev-ubuntu-22.04:6.0.2
+    container: rocm/dev-ubuntu-22.04:6.1.2
 
     steps:
       - name: Clone
@@ -457,7 +457,7 @@ jobs:
           sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev libcurl4-openssl-dev
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-hip
           evict-old-files: 1d
@@ -471,16 +471,6 @@ jobs:
             -DGGML_HIP=ON
           cmake --build build --config Release -j $(nproc)
 
-      - name: Build with legacy HIP support
-        id: cmake_build_legacy_hip
-        run: |
-          cmake -B build2 -S . \
-            -DCMAKE_C_COMPILER=hipcc \
-            -DCMAKE_CXX_COMPILER=hipcc \
-            -DGGML_HIP_ROCWMMA_FATTN=ON \
-            -DGGML_HIP=ON
-          cmake --build build2 --config Release -j $(nproc)
-
   ubuntu-22-cmake-musa:
     runs-on: ubuntu-22.04
     container: mthreads/musa:rc4.2.0-devel-ubuntu22.04-amd64
@@ -497,7 +487,7 @@ jobs:
           apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-musa
           evict-old-files: 1d
@@ -542,7 +532,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl
           evict-old-files: 1d
@@ -590,7 +580,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-sycl-fp16
           evict-old-files: 1d
@@ -621,7 +611,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-ios
           evict-old-files: 1d
@@ -658,7 +648,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-tvos
           evict-old-files: 1d
@@ -730,7 +720,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-swift
           evict-old-files: 1d
@@ -776,7 +766,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-msys2
           variant: ccache
@@ -844,7 +834,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.build }}
           variant: ccache
@@ -958,7 +948,7 @@ jobs:
               apt install -y cmake build-essential ninja-build libgomp1 git libcurl4-openssl-dev
 
         - name: ccache
-          uses: hendrikmuhs/ccache-action@v1.2.16
+          uses: ggml-org/ccache-action@v1.2.16
           with:
             key: ubuntu-latest-cmake-cuda
             evict-old-files: 1d
@@ -987,7 +977,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -1043,7 +1033,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -1089,7 +1079,7 @@ jobs:
           & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ${{ github.job }}
           evict-old-files: 1d
@@ -1123,6 +1113,11 @@ jobs:
       - name: Checkout code
         uses: actions/checkout@v4
 
+      - name: Setup Xcode
+        uses: maxim-lobanov/setup-xcode@v1
+        with:
+          xcode-version: latest-stable
+
       - name: Build
         id: cmake_build
         run: |
@@ -1156,7 +1151,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: android-build
           evict-old-files: 1d

.github/workflows/copilot-setup-steps.yml

@@ -0,0 +1,53 @@
+name: "Copilot Setup Steps"
+
+# Automatically run the setup steps when they are changed to allow for easy validation, and
+# allow manual testing through the repository's "Actions" tab
+on:
+  workflow_dispatch:
+  push:
+    paths:
+      - .github/workflows/copilot-setup-steps.yml
+  pull_request:
+    paths:
+      - .github/workflows/copilot-setup-steps.yml
+
+jobs:
+  # The job MUST be called `copilot-setup-steps` or it will not be picked up by Copilot.
+  copilot-setup-steps:
+    runs-on: ubuntu-latest
+
+    # Set the permissions to the lowest permissions possible needed for your steps.
+    # Copilot will be given its own token for its operations.
+    permissions:
+      # If you want to clone the repository as part of your setup steps, for example to install dependencies, you'll need the `contents: read` permission. If you don't clone the repository in your setup steps, Copilot will do this for you automatically after the steps complete.
+      contents: read
+
+    # You can define any steps you want, and they will run before the agent starts.
+    # If you do not check out your code, Copilot will do this for you.
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: copilot-setup-steps
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential libcurl4-openssl-dev
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Install Python dependencies
+        run: |
+          python3 -m venv .venv
+          .venv/bin/activate
+          pip install -r requirements/requirements-all.txt -r tools/server/tests/requirements.txt
+          pip install flake8 pyright

.github/workflows/release.yml

@@ -32,7 +32,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
@@ -85,7 +85,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
@@ -147,7 +147,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-cpu-cmake
           evict-old-files: 1d
@@ -198,7 +198,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: ubuntu-22-cmake-vulkan
           evict-old-files: 1d
@@ -256,7 +256,7 @@ jobs:
           fetch-depth: 0
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-cpu-${{ matrix.arch }}
           variant: ccache
@@ -328,7 +328,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-${{ matrix.backend }}-${{ matrix.arch }}
           variant: ccache
@@ -398,7 +398,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: Install ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-cuda-${{ matrix.cuda }}
           variant: ccache
@@ -471,7 +471,7 @@ jobs:
         uses: actions/checkout@v4
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-sycl
           variant: ccache
@@ -545,7 +545,7 @@ jobs:
           git clone https://github.com/rocm/rocwmma --branch rocm-6.2.4 --depth 1
 
       - name: ccache
-        uses: hendrikmuhs/ccache-action@v1.2.16
+        uses: ggml-org/ccache-action@v1.2.16
         with:
           key: windows-latest-cmake-hip-${{ matrix.name }}-x64
           evict-old-files: 1d
@@ -600,7 +600,7 @@ jobs:
           name: llama-bin-win-hip-${{ matrix.name }}-x64.zip
 
   ios-xcode-build:
-    runs-on: macos-latest
+    runs-on: macos-15
 
     steps:
       - name: Checkout code
@@ -608,6 +608,10 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Setup Xcode
+        run: |
+          sudo xcode-select -s /Applications/Xcode_16.4.app
+
       - name: Build
         id: cmake_build
         run: |

CMakeLists.txt

@@ -12,6 +12,8 @@ if (NOT XCODE AND NOT MSVC AND NOT CMAKE_BUILD_TYPE)
     set_property(CACHE CMAKE_BUILD_TYPE PROPERTY STRINGS "Debug" "Release" "MinSizeRel" "RelWithDebInfo")
 endif()
 
+message("CMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}")
+
 # Add path to modules
 list(APPEND CMAKE_MODULE_PATH "${CMAKE_CURRENT_SOURCE_DIR}/cmake/")
 

CODEOWNERS

@@ -10,3 +10,4 @@
 /ggml/src/ggml-opt.cpp @JohannesGaessler
 /ggml/src/gguf.cpp @JohannesGaessler
 /ggml/src/ggml-vulkan/ @0cc4m
+/ggml/src/ggml-zdnn/ @taronaeo

README.md

@@ -17,6 +17,7 @@ LLM inference in C/C++
 
 ## Hot topics
 
+- **[[FEEDBACK] Better packaging for llama.cpp to support downstream consumers 🤗](https://github.com/ggml-org/llama.cpp/discussions/15313)**
 - Support for the `gpt-oss` model with native MXFP4 format has been added | [PR](https://github.com/ggml-org/llama.cpp/pull/15091) | [Collaboration with NVIDIA](https://blogs.nvidia.com/blog/rtx-ai-garage-openai-oss) | [Comment](https://github.com/ggml-org/llama.cpp/discussions/15095)
 - Hot PRs: [All](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+) | [Open](https://github.com/ggml-org/llama.cpp/pulls?q=is%3Apr+label%3Ahot+is%3Aopen)
 - Multimodal support arrived in `llama-server`: [#12898](https://github.com/ggml-org/llama.cpp/pull/12898) | [documentation](./docs/multimodal.md)

common/arg.cpp

@@ -749,6 +749,39 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
 // utils
 //
 
+// Helper function to parse tensor buffer override strings
+static void parse_tensor_buffer_overrides(const std::string & value, std::vector<llama_model_tensor_buft_override> & overrides) {
+    std::map<std::string, ggml_backend_buffer_type_t> buft_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto * buft = ggml_backend_dev_buffer_type(dev);
+        if (buft) {
+            buft_list[ggml_backend_buft_name(buft)] = buft;
+        }
+    }
+
+    for (const auto & override : string_split<std::string>(value, ',')) {
+        std::string::size_type pos = override.find('=');
+        if (pos == std::string::npos) {
+            throw std::invalid_argument("invalid value");
+        }
+        std::string tensor_name = override.substr(0, pos);
+        std::string buffer_type = override.substr(pos + 1);
+
+        if (buft_list.find(buffer_type) == buft_list.end()) {
+            printf("Available buffer types:\n");
+            for (const auto & it : buft_list) {
+                printf("  %s\n", ggml_backend_buft_name(it.second));
+            }
+            throw std::invalid_argument("unknown buffer type");
+        }
+        // keep strings alive and avoid leaking memory by storing them in a static vector
+        static std::list<std::string> buft_overrides;
+        buft_overrides.push_back(tensor_name);
+        overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
+    }
+}
+
 struct handle_model_result {
     bool found_mmproj = false;
     common_params_model mmproj;
@@ -993,6 +1026,10 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         params.tensor_buft_overrides.push_back({nullptr, nullptr});
     }
 
+    if (!params.speculative.tensor_buft_overrides.empty()) {
+        params.speculative.tensor_buft_overrides.push_back({nullptr, nullptr});
+    }
+
     if (!params.chat_template.empty() && !common_chat_verify_template(params.chat_template, params.use_jinja)) {
         throw std::runtime_error(string_format(
             "error: the supplied chat template is not supported: %s%s\n",
@@ -1201,6 +1238,7 @@ bool common_params_parse(int argc, char ** argv, common_params & params, llama_e
             common_params_print_completion(ctx_arg);
             exit(0);
         }
+        params.lr.init();
     } catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
         ctx_arg.params = params_org;
@@ -1469,6 +1507,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.swa_full = true;
         }
     ).set_env("LLAMA_ARG_SWA_FULL"));
+    add_opt(common_arg(
+        {"--swa-checkpoints"}, "N",
+        string_format("max number of SWA checkpoints per slot to create (default: %d)\n"
+            "[(more info)](https://github.com/ggml-org/llama.cpp/pull/15293)", params.n_swa_checkpoints),
+        [](common_params & params, int value) {
+            params.n_swa_checkpoints = value;
+        }
+    ).set_env("LLAMA_ARG_SWA_CHECKPOINTS").set_examples({LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--kv-unified", "-kvu"},
         string_format("use single unified KV buffer for the KV cache of all sequences (default: %s)\n"
@@ -2349,40 +2395,15 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--override-tensor", "-ot"}, "<tensor name pattern>=<buffer type>,...",
         "override tensor buffer type", [](common_params & params, const std::string & value) {
-            /* static */ std::map<std::string, ggml_backend_buffer_type_t> buft_list;
-            if (buft_list.empty()) {
-                // enumerate all the devices and add their buffer types to the list
-                for (size_t i = 0; i < ggml_backend_dev_count(); ++i) {
-                    auto * dev = ggml_backend_dev_get(i);
-                    auto * buft = ggml_backend_dev_buffer_type(dev);
-                    if (buft) {
-                        buft_list[ggml_backend_buft_name(buft)] = buft;
-                    }
-                }
-            }
-
-            for (const auto & override : string_split<std::string>(value, ',')) {
-                std::string::size_type pos = override.find('=');
-                if (pos == std::string::npos) {
-                    throw std::invalid_argument("invalid value");
-                }
-                std::string tensor_name = override.substr(0, pos);
-                std::string buffer_type = override.substr(pos + 1);
-
-                if (buft_list.find(buffer_type) == buft_list.end()) {
-                    printf("Available buffer types:\n");
-                    for (const auto & it : buft_list) {
-                        printf("  %s\n", ggml_backend_buft_name(it.second));
-                    }
-                    throw std::invalid_argument("unknown buffer type");
-                }
-                // keep strings alive and avoid leaking memory by storing them in a static vector
-                static std::list<std::string> buft_overrides;
-                buft_overrides.push_back(tensor_name);
-                params.tensor_buft_overrides.push_back({buft_overrides.back().c_str(), buft_list.at(buffer_type)});
-            }
+            parse_tensor_buffer_overrides(value, params.tensor_buft_overrides);
         }
     ));
+    add_opt(common_arg(
+        {"--override-tensor-draft", "-otd"}, "<tensor name pattern>=<buffer type>,...",
+        "override tensor buffer type for draft model", [](common_params & params, const std::string & value) {
+            parse_tensor_buffer_overrides(value, params.speculative.tensor_buft_overrides);
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"--cpu-moe", "-cmoe"},
         "keep all Mixture of Experts (MoE) weights in the CPU",
@@ -2405,6 +2426,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_env("LLAMA_ARG_N_CPU_MOE"));
+    add_opt(common_arg(
+        {"--cpu-moe-draft", "-cmoed"},
+        "keep all Mixture of Experts (MoE) weights in the CPU for the draft model",
+        [](common_params & params) {
+            params.speculative.tensor_buft_overrides.push_back({"\\.ffn_(up|down|gate)_exps", ggml_backend_cpu_buffer_type()});
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CPU_MOE_DRAFT"));
+    add_opt(common_arg(
+        {"--n-cpu-moe-draft", "-ncmoed"}, "N",
+        "keep the Mixture of Experts (MoE) weights of the first N layers in the CPU for the draft model",
+        [](common_params & params, int value) {
+            if (value < 0) {
+                throw std::invalid_argument("invalid value");
+            }
+            for (int i = 0; i < value; ++i) {
+                static std::list<std::string> buft_overrides_draft;
+                buft_overrides_draft.push_back(string_format("blk\\.%d\\.ffn_(up|down|gate)_exps", i));
+                params.speculative.tensor_buft_overrides.push_back({buft_overrides_draft.back().c_str(), ggml_backend_cpu_buffer_type()});
+            }
+        }
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_N_CPU_MOE_DRAFT"));
     add_opt(common_arg(
         {"-ngl", "--gpu-layers", "--n-gpu-layers"}, "N",
         "number of layers to store in VRAM",
@@ -2655,7 +2697,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, const std::string & value) {
             params.out_file = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS}));
+    ).set_examples({LLAMA_EXAMPLE_IMATRIX, LLAMA_EXAMPLE_CVECTOR_GENERATOR, LLAMA_EXAMPLE_EXPORT_LORA, LLAMA_EXAMPLE_TTS, LLAMA_EXAMPLE_FINETUNE}));
     add_opt(common_arg(
         {"-ofreq", "--output-frequency"}, "N",
         string_format("output the imatrix every N iterations (default: %d)", params.n_out_freq),
@@ -3130,7 +3172,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-tbd", "--threads-batch-draft"}, "N",
         "number of threads to use during batch and prompt processing (default: same as --threads-draft)",
@@ -3140,7 +3182,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                 params.speculative.cpuparams_batch.n_threads = std::thread::hardware_concurrency();
             }
         }
-    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE}));
+    ).set_examples({LLAMA_EXAMPLE_SPECULATIVE, LLAMA_EXAMPLE_SERVER}));
     add_opt(common_arg(
         {"-Cd", "--cpu-mask-draft"}, "M",
         "Draft model CPU affinity mask. Complements cpu-range-draft (default: same as --cpu-mask)",
@@ -3533,5 +3575,51 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({ LLAMA_EXAMPLE_DIFFUSION }));
 
 
+    add_opt(
+        common_arg({ "-lr", "--learning-rate" }, "ALPHA",
+                   string_format(
+                       "adamw or sgd optimizer alpha (default: %.2g); note: sgd alpha recommended ~10x (no momentum)",
+                       (double) params.lr.lr0),
+                   [](common_params & params, const std::string & value) { params.lr.lr0 = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-lr-min", "--learning-rate-min" }, "ALPHA",
+                   string_format(
+                       "(if >0) final learning rate after decay (if -decay-epochs is set, default=%.2g)",
+                       (double) params.lr.lr_min),
+                   [](common_params & params, const std::string & value) { params.lr.lr_min = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(
+        common_arg({ "-decay-epochs", "--learning-rate-decay-epochs" }, "ALPHA",
+                   string_format(
+                       "(if >0) decay learning rate to -lr-min after this many epochs (exponential decay, default=%.2g)",
+                       (double) params.lr.decay_epochs),
+                   [](common_params & params, const std::string & value) { params.lr.decay_epochs = std::stof(value); })
+            .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+                { "-wd", "--weight-decay" }, "WD",
+                string_format(
+                    "adamw or sgd optimizer weight decay (0 is off; recommend very small e.g. 1e-9) (default: %.2g).",
+                    (double) params.lr.wd),
+                [](common_params & params, const std::string & value) { params.lr.wd = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-val-split", "--val-split" }, "FRACTION",
+                       string_format("fraction of data to use as validation set for training (default: %.2g).",
+                                     (double) params.val_split),
+                       [](common_params & params, const std::string & value) { params.val_split = std::stof(value); })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-epochs", "--epochs" }, "N",
+                       string_format("optimizer max # of epochs (default: %d)", params.lr.epochs),
+                       [](common_params & params, int epochs) { params.lr.epochs = epochs; })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg({ "-opt", "--optimizer" }, "sgd|adamw", "adamw or sgd",
+                       [](common_params & params, const std::string & name) {
+                           params.optimizer = common_opt_get_optimizer(name.c_str());
+                           if (params.optimizer == GGML_OPT_OPTIMIZER_TYPE_COUNT) {
+                               throw std::invalid_argument("invalid --optimizer, valid options: adamw, sgd");
+                           }
+                       })
+                .set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+
     return ctx_arg;
 }

common/chat.cpp

@@ -296,6 +296,7 @@ json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msg
         }
         if (!msg.reasoning_content.empty()) {
             jmsg["reasoning_content"] = msg.reasoning_content;
+            jmsg["thinking"] = msg.reasoning_content; // gpt-oss
         }
         if (!msg.tool_name.empty()) {
             jmsg["name"] = msg.tool_name;
@@ -472,11 +473,12 @@ std::string common_chat_format_single(
     return ss.str();
 }
 
-std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja) {
+std::string common_chat_format_example(const struct common_chat_templates * tmpls, bool use_jinja, const std::map<std::string, std::string> & chat_template_kwargs) {
     common_chat_templates_inputs inputs;
     inputs.use_jinja = use_jinja;
     inputs.add_bos = tmpls->add_bos;
     inputs.add_eos = tmpls->add_eos;
+    inputs.chat_template_kwargs = chat_template_kwargs;
     auto add_simple_msg = [&](auto role, auto content) {
         common_chat_msg msg;
         msg.role = role;
@@ -1338,16 +1340,164 @@ static common_chat_params common_chat_params_init_gpt_oss(const common_chat_temp
     data.prompt = prompt;
     data.format = COMMON_CHAT_FORMAT_GPT_OSS;
 
-    // TODO: support tool calls in GPT-OSS?
+    // These special tokens are required to parse properly, so we include them
+    // even if parse_tool_calls is false.
+    data.preserved_tokens = {
+        "<|channel|>",
+        "<|constrain|>",
+        "<|message|>",
+        "<|start|>",
+        "<|end|>",
+    };
+
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED;
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            // tool calls can appear in commentary or analysis channels
+            auto channel = builder.add_rule("channel", "\"<|channel|>\" ( \"commentary\" | \"analysis\" )");
+
+            std::vector<std::string> tool_rules_recipient_in_role;
+            std::vector<std::string> tool_rules_recipient_in_channel;
+            foreach_function(inputs.tools, [&](const json & tool) {
+                const auto & function = tool.at("function");
+                std::string name = function.at("name");
+                auto parameters = function.at("parameters");
+                builder.resolve_refs(parameters);
+
+                tool_rules_recipient_in_role.push_back(
+                    builder.add_rule(name + "-call",
+                        "\"" + name + "\"" + channel + " \" <|constrain|>json\"? \"<|message|>\" " +
+                        builder.add_schema(name + "-args", parameters)
+                    )
+                );
+
+                tool_rules_recipient_in_channel.push_back(
+                    builder.add_rule(name + "-call",
+                        "\"" + name + "\"" + " \" <|constrain|>json\"? \"<|message|>\" " +
+                        builder.add_schema(name + "-args", parameters)
+                    )
+                );
+            });
+
+            auto recipient_in_role = builder.add_rule("recipient_in_role",
+                "\"<|start|>assistant\"? \" to=functions.\" ( " +
+                string_join(tool_rules_recipient_in_role, " | ") + " )"
+            );
+
+            auto recipient_in_channel = builder.add_rule("recipient_in_channel",
+                channel + " \" to=functions.\" ( " +
+                string_join(tool_rules_recipient_in_channel, " | ") + " )"
+            );
+
+            builder.add_rule("root", recipient_in_role + " | " + recipient_in_channel);
+
+            // Trigger on tool calls that appear in the commentary channel
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+                "<\\|channel\\|>(commentary|analysis) to"
+            });
+
+            // Trigger tool calls that appear in the role section, either at the
+            // start or in the middle.
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                "^ to"
+            });
+
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN,
+                "<\\|start\\|>assistant to"
+            });
+        });
+    }
 
     return data;
 }
 static void common_chat_parse_gpt_oss(common_chat_msg_parser & builder) {
-    // TODO @ngxson : this won't work with --special enabled, we should fix that
-    builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|start|>assistant<|channel|>final<|message|>");
-    if (!builder.syntax().parse_tool_calls) {
-        builder.add_content(builder.consume_rest());
-        return;
+    static const std::string constraint = "(?: (<\\|constrain\\|>)?([a-zA-Z0-9_-]+))";
+    static const std::string recipient("(?: to=functions\\.([^<\\s]+))");
+
+    static const common_regex start_regex("<\\|start\\|>assistant");
+    static const common_regex analysis_regex("<\\|channel\\|>analysis");
+    static const common_regex final_regex("<\\|channel\\|>final" + constraint + "?");
+    static const common_regex preamble_regex("<\\|channel\\|>commentary");
+    static const common_regex tool_call1_regex(recipient + "<\\|channel\\|>(analysis|commentary)" + constraint + "?");
+    static const common_regex tool_call2_regex("<\\|channel\\|>(analysis|commentary)" + recipient + constraint + "?");
+
+    auto consume_end = [&](bool include_end = false) {
+        if (auto res = builder.try_find_literal("<|end|>")) {
+            return res->prelude + (include_end ? builder.str(res->groups[0]) : "");
+        }
+        return builder.consume_rest();
+    };
+
+    auto handle_tool_call = [&](const std::string & name) {
+        if (auto args = builder.try_consume_json_with_dumped_args({{}})) {
+            if (builder.syntax().parse_tool_calls) {
+                if (!builder.add_tool_call(name, "", args->value) || args->is_partial) {
+                    throw common_chat_msg_partial_exception("incomplete tool call");
+                }
+            } else if (args->is_partial) {
+                throw common_chat_msg_partial_exception("incomplete tool call");
+            }
+        }
+    };
+
+    auto regex_match = [](const common_regex & regex, const std::string & input) -> std::optional<common_regex_match> {
+        auto match = regex.search(input, 0, true);
+        if (match.type == COMMON_REGEX_MATCH_TYPE_FULL) {
+            return match;
+        }
+        return std::nullopt;
+    };
+
+    do {
+        auto header_start_pos = builder.pos();
+        auto content_start = builder.try_find_literal("<|message|>");
+        if (!content_start) {
+            throw common_chat_msg_partial_exception("incomplete header");
+        }
+
+        auto header = content_start->prelude;
+
+        if (auto match = regex_match(tool_call1_regex, header)) {
+            auto group = match->groups[1];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (auto match = regex_match(tool_call2_regex, header)) {
+            auto group = match->groups[2];
+            auto name = header.substr(group.begin, group.end - group.begin);
+            handle_tool_call(name);
+            continue;
+        }
+
+        if (regex_match(analysis_regex, header)) {
+            builder.move_to(header_start_pos);
+            if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE || builder.syntax().reasoning_in_content) {
+                builder.add_content(consume_end(true));
+            } else {
+                builder.try_parse_reasoning("<|channel|>analysis<|message|>", "<|end|>");
+            }
+            continue;
+        }
+
+        if(regex_match(final_regex, header) || regex_match(preamble_regex, header)) {
+            builder.add_content(consume_end());
+            continue;
+        }
+
+        // Possibly a malformed message, attempt to recover by rolling
+        // back to pick up the next <|start|>
+        LOG_DBG("%s: unknown header from message: %s\n", __func__, header.c_str());
+        builder.move_to(header_start_pos);
+    } while (builder.try_find_regex(start_regex, std::string::npos, false));
+
+    auto remaining = builder.consume_rest();
+    if (!remaining.empty()) {
+        LOG_DBG("%s: content after last message: %s\n", __func__, remaining.c_str());
     }
 }
 
@@ -1911,8 +2061,8 @@ static common_chat_params common_chat_templates_apply_jinja(
     params.enable_thinking = inputs.enable_thinking;
     params.grammar = inputs.grammar;
     params.now = inputs.now;
-    params.add_bos = inputs.add_bos;
-    params.add_eos = inputs.add_eos;
+    params.add_bos = tmpls->add_bos;
+    params.add_eos = tmpls->add_eos;
 
     params.extra_context = json::object();
     for (auto el : inputs.chat_template_kwargs) {

common/chat.h

@@ -187,7 +187,8 @@ std::string common_chat_format_single(
 // Returns an example of formatted chat
 std::string common_chat_format_example(
     const struct common_chat_templates * tmpls,
-    bool use_jinja);
+    bool use_jinja,
+    const std::map<std::string, std::string> & chat_template_kwargs);
 
 const char*               common_chat_format_name(common_chat_format format);
 const char*               common_reasoning_format_name(common_reasoning_format format);

common/common.cpp

@@ -41,6 +41,7 @@
 #endif
 #include <locale>
 #include <windows.h>
+#include <string.h>
 #include <fcntl.h>
 #include <io.h>
 #else
@@ -1565,3 +1566,56 @@ ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std
 
     return result;
 }
+
+ggml_opt_optimizer_params common_opt_lr_pars(void * userdata) {
+    ggml_opt_optimizer_params result = ggml_opt_get_default_optimizer_params(nullptr);
+    const lr_opt &            d      = *(lr_opt *) userdata;
+    result.adamw.alpha = result.sgd.alpha = d.get_lr(d.epoch);
+    result.sgd.wd = result.adamw.wd = d.wd;
+    return result;
+}
+
+// TODO make all command line args case-insensitive
+static inline bool eq_case_insensitive(char const* a, char const* b) {
+    return !
+#if defined(_MSC_VER)
+        _stricmp
+#else
+        strcasecmp
+#endif // defined(_MSC_VER)
+        (a, b);
+}
+
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char * n) {
+    if (eq_case_insensitive("adamw", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    }
+    if (eq_case_insensitive("sgd", n)) {
+        return GGML_OPT_OPTIMIZER_TYPE_SGD;
+    }
+    return GGML_OPT_OPTIMIZER_TYPE_COUNT;
+}
+
+// TODO simplify to use just log and exp
+static float const k_log_2 = std::log(2.f);
+
+void lr_opt::init() {
+    if (lr_min > 0 && lr_min < lr0) {
+        float nhalf = std::log(lr0 / lr_min) / k_log_2;
+        float e     = epochs;
+        if (decay_epochs > 0 && decay_epochs < e) {
+            e = decay_epochs;
+        } else {
+            decay_epochs = e;
+        }
+        scale_epoch = nhalf / e;
+    }
+}
+
+float lr_opt::get_lr(float epoch) const {
+    float r = lr_min <= 0 ? lr0 :
+        epoch >= decay_epochs ? lr_min :
+        lr0 * std::pow(0.5f, epoch * scale_epoch);
+    LOG_INF("epoch %.2g lr=%.2g\n", epoch, r);
+    return r;
+}

common/common.h

@@ -2,14 +2,17 @@
 
 #pragma once
 
-#include "llama-cpp.h"
-
 #include <set>
+#include <sstream>
 #include <string>
 #include <string_view>
 #include <vector>
 #include <map>
 #include <sstream>
+#include <cmath>
+
+#include "ggml-opt.h"
+#include "llama-cpp.h"
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -82,6 +85,7 @@ enum llama_example {
     LLAMA_EXAMPLE_PARALLEL,
     LLAMA_EXAMPLE_TTS,
     LLAMA_EXAMPLE_DIFFUSION,
+    LLAMA_EXAMPLE_FINETUNE,
 
     LLAMA_EXAMPLE_COUNT,
 };
@@ -202,6 +206,7 @@ struct common_params_speculative {
     float   p_split      =  0.1f; // speculative decoding split probability
     float   p_min        = 0.75f; // minimum speculative decoding probability (greedy)
     std::vector<std::pair<std::string, std::string>> replacements; // main to speculative model replacements
+    std::vector<llama_model_tensor_buft_override> tensor_buft_overrides;
 
     ggml_type cache_type_k = GGML_TYPE_F16; // KV cache data type for the K
     ggml_type cache_type_v = GGML_TYPE_F16; // KV cache data type for the V
@@ -242,6 +247,25 @@ enum common_reasoning_format {
     COMMON_REASONING_FORMAT_GRANITE,         // Extract thinking tag contents and return as `message.reasoning_content`, including in streaming deltas.
 };
 
+
+struct lr_opt {
+    float    lr0          = 1e-5; // learning rate at first epoch
+    float    lr_min       = -1;
+    float    decay_epochs = -1;   // if >0, the learning rate starts at lr0 and decays to lr_min after this many epochs
+    float    scale_epoch  = 0;
+    float    wd           = 0;
+    unsigned epochs       = 2;
+
+    unsigned epoch; // set by optimizer outer (epochs) loop
+    // learning rate decay - constant LR per epoch only for now
+    float get_lr(float e) const;
+    float get_lr() const { return get_lr(epoch); }
+    // must call after arg parse, before get_lr
+    void init();
+};
+
+struct ggml_opt_optimizer_params common_opt_lr_pars(void * userdata);
+
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =  4096; // context size
@@ -376,6 +400,11 @@ struct common_params {
     bool no_mmproj = false;         // explicitly disable multimodal model
     std::vector<std::string> image; // path to image file(s)
 
+    // finetune
+    struct lr_opt lr;
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+    float val_split = 0.05f; // fraction of the data used for the validation set
+
     // embedding
     bool embedding         = false; // get only sentence embedding
     int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
@@ -384,11 +413,12 @@ struct common_params {
     std::string cls_sep    = "\t";  // separator of classification sequences
 
     // server params
-    int32_t port           = 8080;         // server listens on this network port
-    int32_t timeout_read   = 600;          // http read timeout in seconds
-    int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
-    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t port              = 8080;         // server listens on this network port
+    int32_t timeout_read      = 600;          // http read timeout in seconds
+    int32_t timeout_write     = timeout_read; // http write timeout in seconds
+    int32_t n_threads_http    = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse     = 0;            // min chunk size to reuse from the cache via KV shifting
+    int32_t n_swa_checkpoints = 3;            // max number of SWA checkpoints per slot
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
@@ -703,3 +733,6 @@ const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
 //
 
 ggml_opt_dataset_t common_opt_dataset_init(struct llama_context * ctx, const std::vector<llama_token> & tokens, int64_t stride);
+
+// "adamw" or "sgd" (case insensitive)
+enum ggml_opt_optimizer_type common_opt_get_optimizer(const char *);

docs/build-s390x.md

@@ -76,6 +76,23 @@ cmake --build build --config Release -j $(nproc)
     cmake --build build --config Release -j $(nproc)
     ```
 
+## IBM zDNN Accelerator
+
+This provides acceleration using the IBM zAIU co-processor located in the Telum I and Telum II processors. Make sure to have the [IBM zDNN library](https://github.com/IBM/zDNN) installed.
+
+#### Compile from source from IBM
+
+You may find the official build instructions here: [Building and Installing zDNN](https://github.com/IBM/zDNN?tab=readme-ov-file#building-and-installing-zdnn)
+
+### Compilation
+
+```bash
+cmake -S . -B build             \
+    -DCMAKE_BUILD_TYPE=Release  \
+    -DGGML_ZDNN=ON
+cmake --build build --config Release -j$(nproc)
+```
+
 ## Getting GGUF Models
 
 All models need to be converted to Big-Endian. You can achieve this in three cases:
@@ -145,15 +162,15 @@ All models need to be converted to Big-Endian. You can achieve this in three cas
 
 ### 1. SIMD Acceleration
 
-Only available in IBM z15 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
+Only available in IBM z15/LinuxONE 3 or later system with the `-DGGML_VXE=ON` (turned on by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z14/arch12. In such systems, the APIs can still run but will use a scalar implementation.
 
 ### 2. NNPA Vector Intrinsics Acceleration
 
-Only available in IBM z16 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
+Only available in IBM z16/LinuxONE 4 or later system with the `-DGGML_NNPA=ON` (turned off by default) compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs can still run but will use a scalar implementation.
 
-### 3. zDNN Accelerator
+### 3. zDNN Accelerator (WIP)
 
-_Only available in IBM z16 / LinuxONE 4 or later system. No support currently available._
+Only available in IBM z17/LinuxONE 5 or later system with the `-DGGML_ZDNN=ON` compile flag. No hardware acceleration is possible with llama.cpp with older systems, such as IBM z15/arch13. In such systems, the APIs will default back to CPU routines.
 
 ### 4. Spyre Accelerator
 
@@ -229,11 +246,12 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 ## Appendix A: Hardware Support Matrix
 
-|         | Support | Minimum Compiler Version |
-| ------- | ------- | ------------------------ |
-| IBM z15 | ✅      |                          |
-| IBM z16 | ✅      |                          |
-| IBM z17 | ✅      | GCC 15.1.0               |
+|          | Support | Minimum Compiler Version |
+| -------- | ------- | ------------------------ |
+| IBM z15  | ✅      |                          |
+| IBM z16  | ✅      |                          |
+| IBM z17  | ✅      | GCC 15.1.0               |
+| IBM zDNN | ✅      |                          |
 
 -   ✅ - supported and verified to run as intended
 -   🚫 - unsupported, we are unlikely able to provide support
@@ -242,7 +260,7 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 
 |            | VX/VXE/VXE2 | NNPA | zDNN | Spyre |
 | ---------- | ----------- | ---- | ---- | ----- |
-| FP32       | ✅          | ✅   | ❓   | ❓    |
+| FP32       | ✅          | ✅   | ✅   | ❓    |
 | FP16       | ✅          | ✅   | ❓   | ❓    |
 | BF16       | 🚫          | 🚫   | ❓   | ❓    |
 | Q4_0       | ✅          | ✅   | ❓   | ❓    |
@@ -273,4 +291,4 @@ IBM VXE/VXE2 SIMD acceleration depends on the BLAS implementation. It is strongl
 -   🚫 - acceleration unavailable, will still run using scalar implementation
 -   ❓ - acceleration unknown, please contribute if you can test it yourself
 
-Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 25, 2025.
+Last Updated by **Aaron Teo (aaron.teo1@ibm.com)** on July 31, 2025.

docs/ops.md

@@ -12,91 +12,92 @@ Legend:
 - 🟡 Partially supported by this backend
 - ❌ Not supported by this backend
 
-| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan |
-|-----------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 |
-|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
-|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
-|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
-|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ |
-|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ |
-|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ |
-|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ |
+| Operation | BLAS | CANN | CPU | CUDA | Metal | OpenCL | SYCL | Vulkan | zDNN |
+|-----------|------|------|------|------|------|------|------|------|------|
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              DUP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                          L2_NORM | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ |
+|                              PAD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | ❌ |
+|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                 RMS_NORM_MUL_ADD | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                             ROPE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SET_ROWS | ❌ | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ |
+|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ✅ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | 🟡 | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ |
+|                              SUM | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ❌ |

examples/eval-callback/eval-callback.cpp

@@ -7,6 +7,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <numeric>
 
 /**
  * This the arbitrary data which will be passed to each callback.
@@ -77,6 +78,12 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
         LOG("                                     ]\n");
         LOG("                                     sum = %f\n", sum);
     }
+
+    // TODO: make this abort configurable/optional?
+    if (std::isnan(sum)) {
+        LOG_ERR("encountered NaN - aborting\n");
+        exit(0);
+    }
 }
 
 /**

examples/speculative-simple/speculative-simple.cpp

@@ -59,6 +59,8 @@ int main(int argc, char ** argv) {
     }
 
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+
     common_init_result llama_init_dft = common_init_from_params(params);
 
     //model_dft = llama_init_dft.model.get();

examples/speculative/speculative.cpp

@@ -85,6 +85,8 @@ int main(int argc, char ** argv) {
     }
 
     params.cpuparams_batch.n_threads = params.speculative.cpuparams_batch.n_threads;
+    params.tensor_buft_overrides     = params.speculative.tensor_buft_overrides;
+
     common_init_result llama_init_dft = common_init_from_params(params);
 
     model_dft = llama_init_dft.model.get();

examples/training/finetune.cpp

@@ -10,20 +10,20 @@
 #include <vector>
 
 #if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
+#pragma warning(disable: 4244 4267)  // possible loss of data
 #endif
 
 int main(int argc, char ** argv) {
     common_params params;
-
     params.escape = false;
 
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_FINETUNE)) {
         return 1;
     }
 
     if (params.use_mmap) {
-        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n", __func__);
+        LOG_INF("%s: force disabling memory mapping because it would result in-read-only pointers to the weights\n",
+                __func__);
         params.use_mmap = false;
     }
     if (params.cache_type_k != GGML_TYPE_F32) {
@@ -38,11 +38,10 @@ int main(int argc, char ** argv) {
     common_init();
     llama_backend_init();
     llama_numa_init(params.numa);
-
     // load the model and apply lora adapter, if any
-    common_init_result llama_init = common_init_from_params(params);
-    llama_model_ptr   & model = llama_init.model;
-    llama_context_ptr & ctx   = llama_init.context;
+    common_init_result   llama_init = common_init_from_params(params);
+    llama_model_ptr    & model      = llama_init.model;
+    llama_context_ptr  & ctx        = llama_init.context;
 
     if (model == NULL) {
         LOG_ERR("%s: unable to load model\n", __func__);
@@ -55,31 +54,32 @@ int main(int argc, char ** argv) {
         LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
-    constexpr float val_split = 0.05f;
+    std::vector<llama_token> tokens  = common_tokenize(ctx.get(), params.prompt, true);
+    ggml_opt_dataset_t       dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get()) / 2);
 
-    std::vector<llama_token> tokens = common_tokenize(ctx.get(), params.prompt, true);
-    ggml_opt_dataset_t dataset = common_opt_dataset_init(ctx.get(), tokens, llama_n_ctx(ctx.get())/2);
+    struct lr_opt & lr = params.lr;
+    LOG_INF("-optimizer %s -lr0 %.2g -wd %.2g -lr-min %.2g -min-epochs %.2g -epochs %d -period %.2g -val %.2g\n",
+            ggml_opt_optimizer_name(params.optimizer), (double) lr.lr0, (double) lr.wd, (double) lr.lr_min, (double) lr.decay_epochs,
+            (unsigned) lr.epochs, (double) params.n_batch / params.n_ubatch, (double) params.val_split);
 
-    struct ggml_opt_optimizer_params optimizer_params = ggml_opt_get_default_optimizer_params(nullptr);
-    optimizer_params.adamw.alpha = 1e-7f; // learning rate
-
-    struct llama_opt_params lopt_params {
-        /*n_ctx_train     =*/ 0,
-        /*param_filter    =*/ llama_opt_param_filter_all,
-        /*param_filter_ud =*/ nullptr,
-        /*get_opt_pars    =*/ ggml_opt_get_constant_optimizer_params,
-        /*get_opt_pars_ud =*/ &optimizer_params,
+    struct llama_opt_params lopt_params{
+        /*n_ctx_train     =*/0,
+        /*param_filter    =*/llama_opt_param_filter_all,
+        /*param_filter_ud =*/nullptr,
+        /*get_opt_pars    =*/common_opt_lr_pars,
+        /*get_opt_pars_ud =*/&params.lr,
+        /*optimizer_type  =*/params.optimizer,
     };
     llama_opt_init(ctx.get(), model.get(), lopt_params);
 
-    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - val_split);
+    const int64_t idata_split = ggml_opt_dataset_ndata(dataset) * (1.0f - params.val_split);
 
     ggml_opt_result_t result_train = ggml_opt_result_init();
     ggml_opt_result_t result_eval  = ggml_opt_result_init();
 
-    for (int epoch = 0; epoch < 2; ++epoch) {
+    for (lr.epoch = 0; lr.epoch < lr.epochs; ++lr.epoch) {
         llama_opt_epoch(ctx.get(), dataset, result_train, result_eval, idata_split,
-            ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
+                        ggml_opt_epoch_callback_progress_bar, ggml_opt_epoch_callback_progress_bar);
         fprintf(stderr, "\n");
 
         ggml_opt_result_reset(result_train);
@@ -88,7 +88,7 @@ int main(int argc, char ** argv) {
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_eval);
 
-    llama_model_save_to_file(model.get(), "finetuned-model.gguf");
+    llama_model_save_to_file(model.get(), params.out_file.c_str());
 
     llama_backend_free();
 

flake.nix

@@ -36,9 +36,6 @@
   # ```
   # nixConfig = {
   #   extra-substituters = [
-  #     # Populated by the CI in ggml-org/llama.cpp
-  #     "https://llama-cpp.cachix.org"
-  #
   #     # A development cache for nixpkgs imported with `config.cudaSupport = true`.
   #     # Populated by https://hercules-ci.com/github/SomeoneSerge/nixpkgs-cuda-ci.
   #     # This lets one skip building e.g. the CUDA-enabled openmpi.
@@ -47,10 +44,8 @@
   #   ];
   #
   #   # Verify these are the same keys as published on
-  #   # - https://app.cachix.org/cache/llama-cpp
   #   # - https://app.cachix.org/cache/cuda-maintainers
   #   extra-trusted-public-keys = [
-  #     "llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc="
   #     "cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E="
   #   ];
   # };

ggml/CMakeLists.txt

@@ -188,6 +188,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_WEBGPU                          "ggml: use WebGPU"                                OFF)
 option(GGML_WEBGPU_DEBUG                    "ggml: enable WebGPU debug output"                OFF)
+option(GGML_ZDNN                            "ggml: use zDNN"                                  OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
 option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)

ggml/include/ggml-opt.h

@@ -74,16 +74,26 @@ extern "C" {
         GGML_OPT_BUILD_TYPE_OPT     = 30,
     };
 
+    enum ggml_opt_optimizer_type {
+        GGML_OPT_OPTIMIZER_TYPE_ADAMW,
+        GGML_OPT_OPTIMIZER_TYPE_SGD,
+
+        GGML_OPT_OPTIMIZER_TYPE_COUNT
+    };
+
     // parameters that control which optimizer is used and how said optimizer tries to find the minimal loss
     struct ggml_opt_optimizer_params {
-        // AdamW optimizer parameters
         struct {
             float alpha; // learning rate
-            float beta1;
-            float beta2;
+            float beta1; // first AdamW momentum
+            float beta2; // second AdamW momentum
             float eps;   // epsilon for numerical stability
-            float wd;    // weight decay for AdamW, use 0.0f to disable
+            float wd;    // weight decay - 0.0f to disable
         } adamw;
+        struct {
+            float alpha; // learning rate
+            float wd;    // weight decay
+        } sgd;
     };
 
     // callback to calculate optimizer parameters prior to a backward pass
@@ -112,8 +122,11 @@ extern "C" {
 
         int32_t opt_period; // after how many gradient accumulation steps an optimizer step should be done
 
-        ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
-        void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+        ggml_opt_get_optimizer_params get_opt_pars;    // callback for calculating optimizer parameters
+        void *                        get_opt_pars_ud; // userdata for calculating optimizer parameters
+
+        // only GGML_OPT_OPTIMIZER_TYPE_ADAMW needs m, v momenta per parameter tensor
+        enum ggml_opt_optimizer_type optimizer;
     };
 
     // get parameters for an optimization context with defaults set where possible
@@ -142,6 +155,10 @@ extern "C" {
     // get the gradient accumulator for a node from the forward graph
     GGML_API struct ggml_tensor * ggml_opt_grad_acc(ggml_opt_context_t opt_ctx, struct ggml_tensor * node);
 
+    GGML_API enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t); //TODO consistent naming scheme
+
+    GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type);
+
     // ====== Optimization Result ======
 
     GGML_API ggml_opt_result_t ggml_opt_result_init(void);
@@ -226,12 +243,14 @@ extern "C" {
             struct ggml_tensor            * outputs,        // output tensor, must have shape [ne_label, ndata_batch] if labels are used
             ggml_opt_dataset_t              dataset,        // dataset with data and optionally also labels
             enum ggml_opt_loss_type         loss_type,      // loss to minimize
+            enum ggml_opt_optimizer_type    optimizer,      // sgd or adamw
             ggml_opt_get_optimizer_params   get_opt_pars,   // callback to get optimizer params, userdata is pointer to epoch (of type int64_t)
             int64_t                         nepoch,         // how many times the dataset should be iterated over
             int64_t                         nbatch_logical, // datapoints optimizer step, must be a multiple of ndata_batch in inputs/outputs
             float                           val_split,      // fraction of the dataset to use for validation, must be in [0.0f, 1.0f)
             bool                            silent);        // whether or not info prints to stderr should be suppressed
 
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-zdnn.h

@@ -0,0 +1,16 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+GGML_BACKEND_API ggml_backend_t ggml_backend_zdnn_init(void);
+
+GGML_BACKEND_API ggml_backend_reg_t ggml_backend_zdnn_reg(void);
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml.h

@@ -241,6 +241,8 @@
 #define GGML_ROPE_TYPE_MROPE  8
 #define GGML_ROPE_TYPE_VISION 24
 
+#define GGML_MROPE_SECTIONS   4
+
 #define GGML_UNUSED(x) (void)(x)
 
 #define GGML_PAD(x, n) (((x) + (n) - 1) & ~((n) - 1))
@@ -540,6 +542,7 @@ extern "C" {
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
         GGML_OP_OPT_STEP_ADAMW,
+        GGML_OP_OPT_STEP_SGD,
 
         GGML_OP_GLU,
 
@@ -1660,7 +1663,7 @@ extern "C" {
             struct ggml_tensor  * b,
             struct ggml_tensor  * c,
             int                   n_dims,
-            int                   sections[4],
+            int                   sections[GGML_MROPE_SECTIONS],
             int                   mode,
             int                   n_ctx_orig,
             float                 freq_base,
@@ -1686,6 +1689,22 @@ extern "C" {
             float                 beta_fast,
             float                 beta_slow);
 
+    GGML_API struct ggml_tensor * ggml_rope_multi_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c,
+            int                   n_dims,
+            int                   sections[GGML_MROPE_SECTIONS],
+            int                   mode,
+            int                   n_ctx_orig,
+            float                 freq_base,
+            float                 freq_scale,
+            float                 ext_factor,
+            float                 attn_factor,
+            float                 beta_fast,
+            float                 beta_slow);
+
     GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_rope_custom(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -2293,7 +2312,14 @@ extern "C" {
             struct ggml_tensor  * grad,
             struct ggml_tensor  * m,
             struct ggml_tensor  * v,
-            struct ggml_tensor  * adamw_params); // parameters such a the learning rate
+            struct ggml_tensor  * adamw_params); // parameters such as the learning rate
+
+    // stochastic gradient descent step (with weight decay)
+    GGML_API struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor *  a,
+        struct ggml_tensor *  grad,
+        struct ggml_tensor *  sgd_params); // alpha, weight decay
 
     //
     // automatic differentiation

ggml/src/CMakeLists.txt

@@ -382,6 +382,7 @@ ggml_add_backend(RPC)
 ggml_add_backend(SYCL)
 ggml_add_backend(Vulkan)
 ggml_add_backend(WebGPU)
+ggml_add_backend(zDNN)
 ggml_add_backend(OpenCL)
 
 foreach (target ggml-base ggml)

ggml/src/ggml-backend-reg.cpp

@@ -49,6 +49,10 @@
 #include "ggml-webgpu.h"
 #endif
 
+#ifdef GGML_USE_ZDNN
+#include "ggml-zdnn.h"
+#endif
+
 #ifdef GGML_USE_OPENCL
 #include "ggml-opencl.h"
 #endif
@@ -180,6 +184,9 @@ struct ggml_backend_registry {
 #ifdef GGML_USE_WEBGPU
         register_backend(ggml_backend_webgpu_reg());
 #endif
+#ifdef GGML_USE_ZDNN
+        register_backend(ggml_backend_zdnn_reg());
+#endif
 #ifdef GGML_USE_OPENCL
         register_backend(ggml_backend_opencl_reg());
 #endif

ggml/src/ggml-cpu/CMakeLists.txt

@@ -460,7 +460,7 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
             # NOTE: Only available from GCC 15.1.0 onwards. Any z17 machine with compile issues must first verify their GCC version.
             #       binutils must also be updated to the latest for the -march=z17 flag to work. Otherwise, use -march=arch15.
             message(STATUS "z17 target")
-            list(APPEND ARCH_FLAGS -march=z17)
+            list(APPEND ARCH_FLAGS -march=arch15)
         else()
             message(STATUS "Unknown target")
             message(WARNING "Unknown target. If you are compiling for z14 and earlier, you might have to add -DGGML_VXE=OFF.")

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

@@ -40,18 +40,22 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64)
 // repack.cpp
 #define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
 // repack.cpp
@@ -80,12 +84,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__loongarch64)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -103,12 +109,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__riscv)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -133,11 +141,13 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__s390x__)
 // quants.c
 #define quantize_row_q8_K_generic quantize_row_q8_K
@@ -164,12 +174,14 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #elif defined(__wasm__)
 // quants.c
 #define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
@@ -195,10 +207,12 @@
 #define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
 #define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
 #define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
 #define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
 #define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
 #define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
 #define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
 #define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
 #define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
 #endif

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

@@ -2022,6 +2022,11 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_opt_step_adamw(params, tensor);
             }
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            {
+                ggml_compute_forward_opt_step_sgd(params, tensor);
+            }
+            break;
         case GGML_OP_NONE:
             {
                 // nop
@@ -2325,6 +2330,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 n_tasks = n_threads;
             } break;

ggml/src/ggml-cpu/ops.cpp

@@ -10330,6 +10330,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const int ir1 = MIN(ir0 + dr, nr);
 
     const float * adamw_params_ptr = ggml_get_data_f32(adamw_params);
+
     const float alpha  = adamw_params_ptr[0];
     const float beta1  = adamw_params_ptr[1];
     const float beta2  = adamw_params_ptr[2];
@@ -10337,7 +10338,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
     const float wd     = adamw_params_ptr[4];
     const float beta1h = adamw_params_ptr[5];
     const float beta2h = adamw_params_ptr[6];
-
+    const float keep   = 1.f - alpha * wd;
     for (int ir = ir0; ir < ir1; ++ir) {
         const int64_t i03 = ir/(ne02*ne01);
         const int64_t i02 = (ir - i03*ne02*ne01)/ne01;
@@ -10360,7 +10361,7 @@ static void ggml_compute_forward_opt_step_adamw_f32(
             // The weight decay is applied independently of the Adam momenta m and v.
             // This is NOT equivalent to l2 regularization that adds w[i00]*w[i00] to the loss.
             // See: https://arxiv.org/pdf/1711.05101v3.pdf
-            w[i00] = w[i00]*(1.0f - alpha*wd) - alpha*mh/vh;
+            w[i00] = w[i00] * keep - alpha * mh / vh;
         }
     }
 }
@@ -10382,3 +10383,63 @@ void ggml_compute_forward_opt_step_adamw(
             }
     }
 }
+
+static void ggml_compute_forward_opt_step_sgd_f32(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0       = dst->src[0];
+    const ggml_tensor * src0_grad  = dst->src[1];
+    const ggml_tensor * sgd_params = dst->src[2];
+
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(sgd_params) == 2);
+
+    const int ith = params->ith;
+    const int nth = params->nth;
+
+    const int nr = ggml_nrows(src0);
+
+    GGML_TENSOR_UNARY_OP_LOCALS
+    GGML_ASSERT(nb00 == sizeof(float));
+
+    // rows per thread
+    const int dr = (nr + nth - 1) / nth;
+
+    // row range for this thread
+    const int ir0 = dr * ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+
+    // using adamw param subset we care about - alpha, wd - could have a separate struct
+    const float * sgd_params_ptr   = ggml_get_data_f32(sgd_params);
+    const float   alpha            = sgd_params_ptr[0];
+    const float   keep             = 1.f - alpha * sgd_params_ptr[1];
+
+    for (int ir = ir0; ir < ir1; ++ir) {
+        const int64_t i03 = ir / (ne02 * ne01);
+        const int64_t i02 = (ir - i03 * ne02 * ne01) / ne01;
+        const int64_t i01 = (ir - i03 * ne02 * ne01 - i02 * ne01);
+
+        const size_t offset = i03 * nb03 + i02 * nb02 + i01 * nb01;
+
+        float *       w = (float *) ((char *) src0->data + offset);                   // weight
+        const float * g = (const float *) ((const char *) src0_grad->data + offset);  // grad
+
+        for (int i00 = 0; i00 < ne00; ++i00) {
+            w[i00] = w[i00] * keep - alpha * g[i00];
+        }
+    }
+}
+
+void ggml_compute_forward_opt_step_sgd(const ggml_compute_params * params, ggml_tensor * dst) {
+    const ggml_tensor * src0 = dst->src[0];
+
+    switch (src0->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_opt_step_sgd_f32(params, dst);
+            }
+            break;
+        default:
+            {
+                GGML_ABORT("fatal error - sgd is F32 only");
+            }
+    }
+}

ggml/src/ggml-cpu/ops.h

@@ -107,7 +107,7 @@ void ggml_compute_forward_cross_entropy_loss(const struct ggml_compute_params *
 void ggml_compute_forward_cross_entropy_loss_back(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_opt_step_adamw(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 void ggml_compute_forward_mul_mat(const struct ggml_compute_params * params, struct ggml_tensor * dst);
-
+void ggml_compute_forward_opt_step_sgd(const struct ggml_compute_params * params, struct ggml_tensor * dst);
 #ifdef __cplusplus
 }
 #endif

ggml/src/ggml-cpu/repack.cpp

@@ -206,8 +206,9 @@ void ggml_gemv_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
     UNUSED(s);
     UNUSED(bs);
@@ -307,30 +308,28 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     UNUSED(ncols_interleaved);
     UNUSED(blocklen);
 
-    {
-        float sumf[8];
-        int sumi;
+    float sumf[8];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
-                            const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
+                        const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -494,43 +493,73 @@ void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     const int ncols_interleaved = 4;
     const int blocklen = 4;
 
-    assert (n % qk == 0);
-    assert (nc % ncols_interleaved == 0);
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
 
-    UNUSED(s);
     UNUSED(bs);
-    UNUSED(vx);
-    UNUSED(vy);
     UNUSED(nr);
-    UNUSED(nc);
-    UNUSED(nb);
-    UNUSED(ncols_interleaved);
-    UNUSED(blocklen);
 
-    {
-        float sumf[4];
-        int sumi;
+    float sumf[4];
+    int sumi;
 
-        const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
-        for (int x = 0; x < nc / ncols_interleaved; x++) {
-            const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb);
 
-            for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
-            for (int l = 0; l < nb; l++) {
-                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
-                    for (int j = 0; j < ncols_interleaved; j++) {
-                        sumi = 0;
-                        for (int i = 0; i < blocklen; ++i) {
-                            const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
-                            const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
-                            sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
-                        }
-                        sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
                     }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
                 }
             }
-            for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
         }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
+    }
+}
+
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(nr == 1);
+    assert(n % qk == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    UNUSED(bs);
+    UNUSED(nr);
+
+    float sumf[8];
+    int sumi;
+
+    const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
+    for (int x = 0; x < nc / ncols_interleaved; x++) {
+        const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+
+        for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
+        for (int l = 0; l < nb; l++) {
+            for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                for (int j = 0; j < ncols_interleaved; j++) {
+                    sumi = 0;
+                    for (int i = 0; i < blocklen; ++i) {
+                        const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                        const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                        sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2]));
+                    }
+                    sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d);
+                }
+            }
+        }
+        for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
     }
 }
 
@@ -934,6 +963,50 @@ void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs
     }
 }
 
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc) {
+    const int qk = QK8_0;
+    const int nb = n / qk;
+    const int ncols_interleaved = 8;
+    const int blocklen = 8;
+
+    assert(n % qk == 0);
+    assert(nr % 4 == 0);
+    assert(nc % ncols_interleaved == 0);
+
+    float sumf[4][8];
+    int sumi;
+
+    for (int y = 0; y < nr / 4; y++) {
+        const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
+        for (int x = 0; x < nc / ncols_interleaved; x++) {
+            const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb);
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
+            }
+            for (int l = 0; l < nb; l++) {
+                for (int k = 0; k < (qk / (2 * blocklen)); k++) {
+                    for (int m = 0; m < 4; m++) {
+                        for (int j = 0; j < ncols_interleaved; j++) {
+                            sumi = 0;
+                            for (int i = 0; i < blocklen; ++i) {
+                                const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F];
+                                const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4];
+                                sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
+                                         (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4]));
+                            }
+                            sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m]);
+                        }
+                    }
+                }
+            }
+            for (int m = 0; m < 4; m++) {
+                for (int j = 0; j < ncols_interleaved; j++)
+                    s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
+            }
+        }
+    }
+}
+
 } // extern "C"
 
 static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
@@ -1285,15 +1358,16 @@ static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_s
 
 static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
     GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
-    //GGML_ASSERT(interleave_block == 4 || interleave_block == 8);
     GGML_ASSERT(interleave_block == 4);
 
-    block_iq4_nlx4 * dst = (block_iq4_nlx4 *)t->data;
-    const block_iq4_nl * src = (const block_iq4_nl *)data;
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx4 * dst = (      block_iq4_nlx4 *)t->data;
+
     block_iq4_nl dst_tmp[4];
+
     int nrow = ggml_nrows(t);
     int nrows_interleaved = 4;
-    int nblocks = t->ne[0] / QK4_0;
+    int nblocks = t->ne[0] / QK4_NL;
 
     GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
 
@@ -1315,6 +1389,63 @@ static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_b
     GGML_UNUSED(data_size);
 }
 
+static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) {
+    block_iq4_nlx8 out;
+
+    for (int i = 0; i < 8; i++) {
+        out.d[i] = in[i].d;
+    }
+
+    const int end = QK4_NL * 4 / blck_size_interleave;
+
+    if (blck_size_interleave == 8) {
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 8;
+            int src_offset = (i / 8) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t));
+        }
+    } else {
+        GGML_ASSERT(false);
+    }
+
+    return out;
+}
+
+static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT data, size_t data_size) {
+    GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL);
+    GGML_ASSERT(interleave_block == 8);
+
+    const block_iq4_nl   * src = (const block_iq4_nl   *)data;
+          block_iq4_nlx8 * dst = (      block_iq4_nlx8 *)t->data;
+
+    block_iq4_nl dst_tmp[8];
+
+    int nrow = ggml_nrows(t);
+    int nrows_interleaved = 8;
+    int nblocks = t->ne[0] / QK4_NL;
+
+    GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl));
+
+    if (t->ne[1] % nrows_interleaved != 0) {
+        return -1;
+    }
+
+    for (int b = 0; b < nrow; b += nrows_interleaved) {
+        for (int64_t x = 0; x < nblocks; x++) {
+            for (int i = 0; i < nrows_interleaved; i++) {
+                dst_tmp[i] = src[x + i * nblocks];
+            }
+            *dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block);
+        }
+        src += nrows_interleaved * nblocks;
+    }
+    return 0;
+
+    GGML_UNUSED(data_size);
+}
+
 namespace ggml::cpu::repack {
 // repack
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS>
@@ -1350,6 +1481,10 @@ template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void *
 //    return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size);
 //}
 
+template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) {
+    return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size);
+}
+
 // gemv
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemv(int, float *, size_t, const void *, const void *, int, int);
@@ -1378,6 +1513,10 @@ template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 // gemm
 template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE>
 void gemm(int, float *, size_t, const void *, const void *, int, int);
@@ -1406,6 +1545,10 @@ template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size
     ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
 }
 
+template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
+    ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc);
+}
+
 class tensor_traits_base : public ggml::cpu::tensor_traits {
   public:
     virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0;
@@ -1680,6 +1823,7 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
 
     // instance for IQ4
     static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
+    static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0;
 
     if (cur->type == GGML_TYPE_Q4_0) {
         if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_0)) {
@@ -1710,6 +1854,11 @@ static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(cons
             }
         }
     } else if (cur->type == GGML_TYPE_IQ4_NL) {
+        if (ggml_cpu_has_avx2()) {
+            if (cur->ne[1] % 8 == 0) {
+                return &iq4_nl_8x8_q8_0;
+            }
+        }
         if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) {
             if (cur->ne[1] % 4 == 0) {
                 return &iq4_nl_4x4_q8_0;

ggml/src/ggml-cpu/repack.h

@@ -67,6 +67,13 @@ struct block_iq4_nlx4 {
 
 static_assert(sizeof(block_iq4_nlx4) == 4 * sizeof(ggml_half) + QK4_NL * 2, "wrong iq4_nlx4 block size/padding");
 
+struct block_iq4_nlx8 {
+    ggml_half d[8];            // deltas for 8 iq4_nl blocks
+    uint8_t   qs[QK4_NL * 4];  // nibbles / quants for 8 iq4_nl blocks
+};
+
+static_assert(sizeof(block_iq4_nlx8) == 8 * sizeof(ggml_half) + QK4_NL * 4, "wrong iq4_nlx8 block size/padding");
+
 #if defined(__cplusplus)
 extern "C" {
 #endif
@@ -80,12 +87,14 @@ void ggml_gemv_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const vo
 void ggml_gemv_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 // Native implementations
 void ggml_quantize_mat_q8_0_4x4_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
@@ -97,12 +106,14 @@ void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
 void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemv_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_4x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 void ggml_gemm_iq4_nl_4x4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
+void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, const void * GGML_RESTRICT vy, int nr, int nc);
 
 #if defined(__cplusplus)
 } // extern "C"

ggml/src/ggml-cuda/common.cuh

@@ -87,6 +87,10 @@
 #define GGML_CUDA_CC_IS_QY2(cc)      (cc >= GGML_CUDA_CC_QY2 && cc < GGML_CUDA_CC_NG)
 #define GGML_CUDA_CC_IS_NG(cc)       (cc >= GGML_CUDA_CC_NG)
 
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#    define GGML_CUDA_USE_CUB
+#endif  // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+
 #ifdef __CUDA_ARCH_LIST__
 constexpr bool ggml_cuda_has_arch_impl(int) {
     return false;
@@ -420,26 +424,6 @@ static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
 #endif // FP16_AVAILABLE
 }
 
-// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
-template<bool norm>
-static __global__ void reduce_rows_f32(const float * x, float * dst, const int ncols) {
-    const int row = blockIdx.x;
-    const int col = threadIdx.x;
-
-    float sum = 0.0f;
-    for (int i = col; i < ncols; i += blockDim.x) {
-        sum += x[row * ncols + i];
-    }
-
-    sum = warp_reduce_sum(sum);
-
-    if (col != 0) {
-        return;
-    }
-
-    dst[row] = norm ? sum / ncols : sum;
-}
-
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ int warp_reduce_all(int x) {
 #ifdef GGML_USE_HIP
@@ -480,25 +464,21 @@ static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b
 }
 
 static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if defined(GGML_USE_HIP) && HIP_VERSION >= 50700000
+#if defined(GGML_USE_HIP)
     return half2(__hmax(a.x, b.x), __hmax(a.y, b.y));
-#elif !defined(GGML_USE_HIP) && CUDART_VERSION >= CUDART_HMAX
+#elif CUDART_VERSION >= CUDART_HMAX
     return __hmax2(a, b);
-#elif !defined(GGML_USE_HIP)
+#else
     half2 ret;
     reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
     reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
     return ret;
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
 #endif
 }
 
 template<int width = WARP_SIZE>
 static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#if !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 #pragma unroll
    for (int offset = width/2; offset > 0; offset >>= 1) {
        x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, offset, width));
@@ -507,7 +487,7 @@ static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
 #else
    GGML_UNUSED(x);
    NO_DEVICE_CODE;
-#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || (defined(GGML_USE_HIP) && HIP_VERSION >= 50700000)
+#endif // !defined(GGML_USE_HIP) && __CUDA_ARCH__ >= GGML_CUDA_CC_PASCAL || defined(GGML_USE_HIP)
 }
 
 #if CUDART_VERSION < CUDART_HMASK

ggml/src/ggml-cuda/convert.cu

@@ -31,8 +31,8 @@ static __global__ void dequantize_block(const void * __restrict__ vx, dst_t * __
     dequantize_kernel(vx, ib, iqs, v);
 
     const int64_t iy0 = ((i03*ne02 + i02)*ne01 + i01)*ne00 + iybs + iqs;
-    y[iy0 + 0]        = float(v.x);
-    y[iy0 + y_offset] = float(v.y);
+    y[iy0 + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    y[iy0 + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template <bool need_check>
@@ -630,7 +630,7 @@ static __global__ void convert_unary(
 
     const int64_t ix = i03*s03 + i02*s02 + i01*s01 + i00;
     const int64_t iy = ((i03*ne02 + i02)*ne01 + i01)*ne00 + i00;
-    y[iy] = float(x[ix]);
+    y[iy] = ggml_cuda_cast<dst_t>(x[ix]);
 }
 
 template <typename src_t, typename dst_t>

ggml/src/ggml-cuda/convert.cuh

@@ -29,3 +29,16 @@ typedef to_t_nc_cuda_t<nv_bfloat16> to_bf16_nc_cuda_t;
 to_fp32_nc_cuda_t ggml_get_to_fp32_nc_cuda(ggml_type type);
 to_fp16_nc_cuda_t ggml_get_to_fp16_nc_cuda(ggml_type type);
 to_bf16_nc_cuda_t ggml_get_to_bf16_nc_cuda(ggml_type type);
+
+template<typename dst_t, typename src_t>
+ __host__ __device__ inline dst_t ggml_cuda_cast(src_t x) {
+    if constexpr (std::is_same_v<dst_t, src_t>) {
+        return x;
+    } else if constexpr(std::is_same_v<dst_t, nv_bfloat16>) {
+        return __float2bfloat16(float(x));
+    } else if constexpr(std::is_same_v<src_t, nv_bfloat16>) {
+        return __bfloat162float(x);
+    } else {
+        return float(x);
+    }
+}

ggml/src/ggml-cuda/cpy-utils.cuh

@@ -1,15 +1,7 @@
 #pragma once
 
 #include "ggml-common.h"
-
-template<typename src_t, typename dst_t>
-static __device__ __forceinline__ void convert_flt(const src_t * src, dst_t * dst) {
-    if constexpr (std::is_same_v<src_t, dst_t>) {
-        *dst = *src;
-    } else {
-        *dst = float(*src);
-    }
-}
+#include "convert.cuh"
 
 static __device__ __forceinline__ int best_index_int8(int n, const int8_t * val, float x) {
     if (x <= val[0]) return 0;
@@ -221,5 +213,5 @@ static __device__ void cpy_blck_f32_iq4_nl(const char * cxi, char * cdsti) {
 
 template<typename src_t, typename dst_t>
 static __device__ void cpy_1_flt(const char * cxi, char * cdsti) {
-    convert_flt((const src_t *)cxi, (dst_t *)cdsti);
+    *(dst_t *) cdsti = ggml_cuda_cast<dst_t>(*(const src_t *) cxi);
 }

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

@@ -539,11 +539,15 @@ static __global__ void flash_attn_mask_to_KV_max(
         all_inf = warp_reduce_all(all_inf);
 
         if (!all_inf) {
-            KV_max_sj += FATTN_KQ_STRIDE;
             break;
         }
     }
 
+    // If the break in the loop was not triggered, KV_max_sj is now -FATTN_KQ_STRIDE.
+    // If the break was triggered it's the lower edge of the tile with the first non-masked values.
+    // In either case, walk back the decrementation by FATTN_KQ_STRIDE.
+    KV_max_sj += FATTN_KQ_STRIDE;
+
     if (threadIdx.x != 0) {
         return;
     }

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

@@ -15,7 +15,6 @@ namespace wmma = mtmusa::wmma;
 namespace wmma = nvcuda::wmma;
 #endif // GGML_USE_MUSA
 #elif defined(GGML_HIP_ROCWMMA_FATTN) && defined(FP16_MMA_AVAILABLE)
-#undef HIP_ENABLE_WARP_SYNC_BUILTINS // conflicts with rocWMMA headers
 #include <rocwmma/rocwmma.hpp>
 namespace wmma = rocwmma;
 #endif // !defined(GGML_USE_HIP)

ggml/src/ggml-cuda/getrows.cu

@@ -1,5 +1,6 @@
 #include "getrows.cuh"
 #include "dequantize.cuh"
+#include "convert.cuh"
 
 template<int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static __global__ void k_get_rows(
@@ -34,8 +35,8 @@ static __global__ void k_get_rows(
     dfloat2 v;
     dequantize_kernel(src0_row, ib, iqs, v);
 
-    dst_row[iybs + iqs + 0]        = float(v.x);
-    dst_row[iybs + iqs + y_offset] = float(v.y);
+    dst_row[iybs + iqs + 0]        = ggml_cuda_cast<dst_t>(v.x);
+    dst_row[iybs + iqs + y_offset] = ggml_cuda_cast<dst_t>(v.y);
 }
 
 template<typename src0_t, typename dst_t>
@@ -62,7 +63,7 @@ static __global__ void k_get_rows_float(
     dst_t * dst_row = dst + i10*s1 + i11*s2 + i12*s3;
     const src0_t * src0_row = (const src0_t *)((const char *) src0 + i01*nb01 + i11*nb02 + i12*nb03);
 
-    dst_row[i00] = float(src0_row[i00]);
+    dst_row[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 }
 
 template<typename grad_t, typename dst_t>

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

@@ -28,6 +28,7 @@
 #include "ggml-cuda/mmvq.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
+#include "ggml-cuda/opt-step-sgd.cuh"
 #include "ggml-cuda/out-prod.cuh"
 #include "ggml-cuda/pad.cuh"
 #include "ggml-cuda/pool2d.cuh"
@@ -180,30 +181,6 @@ static int ggml_cuda_parse_id(char devName[]) {
 #endif // defined(GGML_USE_HIP)
 
 static ggml_cuda_device_info ggml_cuda_init() {
-#if defined(GGML_USE_HIP)
-    // Workaround for a rocBLAS bug when using multiple graphics cards:
-    // https://github.com/ROCmSoftwarePlatform/rocBLAS/issues/1346
-    {
-        int major_version = 0;
-        size_t version_length = 0;
-        if (rocblas_get_version_string_size(&version_length) == rocblas_status_success) {
-            std::vector<char> version(version_length+1, '\0');
-            if (rocblas_get_version_string(version.data(), version.size()) == rocblas_status_success) {
-                version.resize(::strlen(version.data()));
-                int parsed_value = 0;
-                if (std::from_chars(version.data(), version.data() + version.size(), parsed_value).ec == std::errc()) {
-                    major_version = parsed_value;
-                }
-            }
-        }
-        if (major_version < 4) {
-            GGML_LOG_DEBUG(GGML_CUDA_NAME " calling rocblas_initialize as a workaround for a rocBLAS bug\n");
-            rocblas_initialize();
-            CUDA_CHECK(cudaDeviceSynchronize());
-        }
-    }
-#endif
-
     ggml_cuda_device_info info = {};
 
     cudaError_t err = cudaGetDeviceCount(&info.device_count);
@@ -2503,6 +2480,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
         case GGML_OP_OPT_STEP_ADAMW:
             ggml_cuda_opt_step_adamw(ctx, dst);
             break;
+        case GGML_OP_OPT_STEP_SGD:
+            ggml_cuda_opt_step_sgd(ctx, dst);
+            break;
         default:
             return false;
     }
@@ -3560,6 +3540,7 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
         case GGML_OP_CROSS_ENTROPY_LOSS:
         case GGML_OP_CROSS_ENTROPY_LOSS_BACK:
         case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return true;
         default:
             return false;

ggml/src/ggml-cuda/mean.cu

@@ -1,4 +1,14 @@
 #include "mean.cuh"
+#include "reduce_rows.cuh"
+
+#ifdef GGML_CUDA_USE_CUB
+#include <cub/cub.cuh>
+using namespace cub;
+#endif  // GGML_CUDA_USE_CUB
+
+template <typename T> __global__ void divide_by_count(T * result, size_t count) {
+    *result /= static_cast<T>(count);
+}
 
 void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const ggml_tensor * src0   = dst->src[0];
@@ -13,7 +23,51 @@ void ggml_cuda_op_mean(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+// Special case for reducing vectors
+#ifdef GGML_CUDA_USE_CUB
+#ifdef USE_CUDA_GRAPH
+    cudaStreamCaptureStatus iscapturing;
+    CUDA_CHECK(cudaStreamIsCapturing(stream, &iscapturing));
+#endif // USE_CUDA_GRAPH
+    if ((nrows == 1) &&
+#ifdef USE_CUDA_GRAPH
+            // CUDA_GRAPHS_DISABLED
+            ((ncols > 65536) &&
+             ((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+              ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+              ctx.cuda_graph->disable_due_to_failed_graph_capture)) ||
+        // CUDA_GRAPHS ENABLED
+        ((ncols > 32768) &&
+         !((ctx.cuda_graph->instance == nullptr) && (iscapturing == cudaStreamCaptureStatusNone) ||
+           ctx.cuda_graph->disable_due_to_gpu_arch || ctx.cuda_graph->disable_due_to_too_many_updates ||
+           ctx.cuda_graph->disable_due_to_failed_graph_capture))) {
+#else
+        (ncols > 65536)) {
+#endif // USE_CUDA_GRAPH
+        // Single row - use device-wide reduction
+        size_t           tmp_size = 0;
+        ggml_cuda_pool & pool     = ctx.pool();
+
+        DeviceReduce::Sum(nullptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
+        DeviceReduce::Sum(tmp_alloc.ptr, tmp_size, src0_d, dst_d, ncols, stream);
+
+        // Divide by ncols
+        divide_by_count<float><<<1, 1, 0, stream>>>(dst_d, ncols);
+        return;
+    }
+#endif // GGML_CUDA_USE_CUB
+
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/ true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/true><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }

ggml/src/ggml-cuda/mmvf.cu

@@ -1,5 +1,6 @@
 #include "ggml.h"
 #include "common.cuh"
+#include "convert.cuh"
 #include "mmvf.cuh"
 
 template <typename T, typename type_acc, int ncols_dst, int block_size>
@@ -93,8 +94,8 @@ static __global__ void mul_mat_vec_f(
 #pragma unroll
             for (int j = 0; j < ncols_dst; ++j) {
                 const float2 tmpy = y2[j*stride_col_y2 + col2];
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
-                sumf[j] += float(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[0]) * tmpy.x;
+                sumf[j] += ggml_cuda_cast<float>(reinterpret_cast<const nv_bfloat16 *>(&tmpx)[1]) * tmpy.y;
             }
         }
     } else {

ggml/src/ggml-cuda/opt-step-sgd.cu

@@ -0,0 +1,49 @@
+#include "ggml-impl.h"
+#include "opt-step-sgd.cuh"
+
+#include <cstdint>
+
+static __global__ void opt_step_sgd_f32(
+    float * __restrict__ x, const float * __restrict__ g,
+    const float * __restrict__ pars, const int64_t k) {
+
+    const int64_t i = (int64_t) blockIdx.x*blockDim.x + threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+    x[i] = x[i] * (1.0f - pars[0] * pars[1]) - pars[0] * g[i];
+}
+
+static void opt_step_sgd_f32_cuda(
+    float * x, const float * g, const float * __restrict__ pars, const int64_t k, cudaStream_t stream) {
+
+    const dim3 block_dims(CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    const dim3 block_nums((k + CUDA_OPT_STEP_SGD_BLOCK_SIZE - 1) / CUDA_OPT_STEP_SGD_BLOCK_SIZE, 1, 1);
+    opt_step_sgd_f32<<<block_nums, block_dims, 0, stream>>>(x, g, pars, k);
+}
+
+void ggml_cuda_opt_step_sgd(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * src0      = dst->src[0];
+    const ggml_tensor * src0_grad = dst->src[1];
+    const ggml_tensor * params    = dst->src[2];
+
+    GGML_ASSERT(src0->type      == GGML_TYPE_F32);
+    GGML_ASSERT(src0_grad->type == GGML_TYPE_F32);
+    GGML_ASSERT(params->type    == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(src0));
+    GGML_ASSERT(ggml_is_contiguous(src0_grad));
+    GGML_ASSERT(ggml_is_contiguous(params));
+    GGML_ASSERT(ggml_are_same_shape(src0, src0_grad));
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    float       * src0_d      = (float       *) src0->data;
+    const float * src0_grad_d = (const float *) src0_grad->data;
+    const float * params_d    = (const float *) params->data;
+
+    cudaStream_t stream = ctx.stream();
+
+    const int64_t ne = ggml_nelements(src0);
+
+    opt_step_sgd_f32_cuda(src0_d, src0_grad_d, params_d, ne, stream);
+}

ggml/src/ggml-cuda/opt-step-sgd.cuh

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

ggml/src/ggml-cuda/reduce_rows.cuh

@@ -0,0 +1,53 @@
+#include "common.cuh"
+
+// Row reduction kernel template - compute sum (norm=false) or mean (norm=true)
+template <bool norm>
+static __global__ void reduce_rows_f32(const float * __restrict__ x, float * __restrict__ dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int col = threadIdx.x;
+
+    float     sum        = 0.0f;
+    const int num_unroll = 8;
+    float     temp[num_unroll];
+    float     sum_temp[num_unroll] = { 0.0f };
+    for (int i = col; i < ncols;) {
+        for (int j = 0; j < num_unroll; ++j) {
+            if (i < ncols) {
+                temp[j] = x[row * ncols + i];
+            } else {
+                temp[j] = 0;
+            }
+            i += blockDim.x;
+        }
+        for (int j = 0; j < num_unroll; ++j) {
+            sum_temp[j] += temp[j];
+        }
+    }
+    for (int j = 0; j < num_unroll; ++j) {
+        sum += sum_temp[j];
+    }
+
+    // sum up partial sums
+    sum = warp_reduce_sum(sum);
+    if (blockDim.x > WARP_SIZE) {
+        assert((blockDim.x <= 1024) && (blockDim.x % WARP_SIZE) == 0);
+        __shared__ float s_sum[32];
+        const int        warp_id = threadIdx.x / WARP_SIZE;
+        const int        lane_id = threadIdx.x % WARP_SIZE;
+        if (lane_id == 0) {
+            s_sum[warp_id] = sum;
+        }
+        __syncthreads();
+        sum = 0.0f;
+        if (lane_id < (blockDim.x / WARP_SIZE)) {
+            sum = s_sum[lane_id];
+        }
+        sum = warp_reduce_sum(sum);
+    }
+
+    if (col != 0) {
+        return;
+    }
+
+    dst[row] = norm ? sum / ncols : sum;
+}

ggml/src/ggml-cuda/set-rows.cu

@@ -3,11 +3,6 @@
 
 typedef void (*set_rows_kernel_t)(const char * src, char * dst);
 
-template<typename src_t, typename dst_t>
-__device__ __forceinline__ void set_rows_1(const src_t * src_f, dst_t * dst_f) {
-    convert_flt(src_f, dst_f);
-}
-
 // Generic quantized set_rows kernel template
 template<typename block_type, int qk, void (*quantize_func)(const float*, block_type*)>
 static __global__ void k_set_rows_quant(
@@ -117,9 +112,7 @@ static __global__ void k_set_rows(
     const src_t * src0_row = src0 + i01*s01 + i02*s02 + i03*s03;
     dst_t * dst_row_ptr    = dst + dst_row*s1 + i02*s2 + i03*s3;
 
-    const src_t* src_elem = src0_row + i00;
-    dst_t* dst_elem = dst_row_ptr + i00;
-    set_rows_1(src_elem, dst_elem);
+    dst_row_ptr[i00] = ggml_cuda_cast<dst_t>(src0_row[i00]);
 
     GGML_UNUSED(ne10);
     GGML_UNUSED(ne13);

ggml/src/ggml-cuda/sum.cu

@@ -1,19 +1,15 @@
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
-#define USE_CUB
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA) && CUDART_VERSION >= 11070
+#include "sum.cuh"
+#include "sumrows.cuh"
 
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
 #include <cub/cub.cuh>
 using namespace cub;
-#endif // USE_CUB
-
-#include "sumrows.cuh"
-#include "sum.cuh"
+#endif  // GGML_CUDA_USE_CUB
 
 #include <cstdint>
 
 void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int64_t ne, cudaStream_t stream) {
-#ifdef USE_CUB
+#ifdef GGML_CUDA_USE_CUB
     size_t tmp_size = 0;
     DeviceReduce::Sum(nullptr,       tmp_size, x, dst, ne, stream);
     ggml_cuda_pool_alloc<uint8_t> tmp_alloc(pool, tmp_size);
@@ -23,7 +19,7 @@ void sum_f32_cuda(ggml_cuda_pool & pool, const float * x, float * dst, const int
     // For AMD there is rocPRIM which could be used as a drop-in replacement via hipcub but this would require C++11 -> C++14.
     sum_rows_f32_cuda(x, dst, ne, 1, stream);
     GGML_UNUSED(pool);
-#endif // USE_CUB
+#endif // GGML_CUDA_USE_CUB
 }
 
 void ggml_cuda_op_sum(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {

ggml/src/ggml-cuda/sumrows.cu

@@ -1,9 +1,17 @@
+#include "reduce_rows.cuh"
 #include "sumrows.cuh"
 
 void sum_rows_f32_cuda(const float * x, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
-    const dim3 block_dims(WARP_SIZE, 1, 1);
+    const int  id  = ggml_cuda_get_device();
+    const int  nsm = ggml_cuda_info().devices[id].nsm;
     const dim3 block_nums(nrows, 1, 1);
-    reduce_rows_f32</*norm*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    if ((nrows / nsm) < 2) {
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    } else {
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(x, dst, ncols);
+    }
 }
 
 void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
@@ -19,8 +27,17 @@ void ggml_cuda_op_sum_rows(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
     const int64_t ncols = src0->ne[0];
     const int64_t nrows = ggml_nrows(src0);
 
-    const dim3 block_dims(WARP_SIZE, 1, 1);
     const dim3 block_nums(nrows, 1, 1);
 
-    reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    const int id  = ggml_cuda_get_device();
+    const int nsm = ggml_cuda_info().devices[id].nsm;
+    if ((nrows / nsm) < 2) {
+        // Increase num threads to 512 for small nrows to better hide the latency
+        const dim3 block_dims(512, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    } else {
+        // Enough active SMs to hide latency, use smaller blocks to allow better scheduling
+        const dim3 block_dims(ncols < 1024 ? 32 : 128, 1, 1);
+        reduce_rows_f32</*norm=*/false><<<block_nums, block_dims, 0, stream>>>(src0_d, dst_d, ncols);
+    }
 }

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

@@ -1,12 +1,10 @@
 #pragma once
 
-#define HIP_ENABLE_WARP_SYNC_BUILTINS 1
+#define HIP_DISABLE_WARP_SYNC_BUILTINS 1
 #include <hip/hip_runtime.h>
 #include <hipblas/hipblas.h>
 #include <hip/hip_fp16.h>
-#include <hip/hip_bfloat16.h>
-// for rocblas_initialize()
-#include "rocblas/rocblas.h"
+#include <hip/hip_bf16.h>
 
 #define CUBLAS_GEMM_DEFAULT HIPBLAS_GEMM_DEFAULT
 #define CUBLAS_GEMM_DEFAULT_TENSOR_OP HIPBLAS_GEMM_DEFAULT
@@ -137,7 +135,7 @@
 #define CUBLAS_STATUS_INTERNAL_ERROR HIPBLAS_STATUS_INTERNAL_ERROR
 #define CUBLAS_STATUS_NOT_SUPPORTED HIPBLAS_STATUS_NOT_SUPPORTED
 
-#if HIP_VERSION >= 70000000
+#if HIP_VERSION >= 60500000
 #define CUBLAS_COMPUTE_16F HIPBLAS_COMPUTE_16F
 #define CUBLAS_COMPUTE_32F HIPBLAS_COMPUTE_32F
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_COMPUTE_32F_FAST_16F
@@ -149,7 +147,7 @@
 #define CUBLAS_COMPUTE_32F_FAST_16F HIPBLAS_R_32F
 #define cublasComputeType_t hipblasDatatype_t
 #define cudaDataType_t hipblasDatatype_t
-#endif // HIP_VERSION >= 7000000
+#endif // HIP_VERSION >= 6050000
 
 #if !defined(__HIP_PLATFORM_AMD__)
 #error "The HIP backend supports only AMD targets"
@@ -181,8 +179,7 @@
 #define RDNA4
 #endif
 
-#if defined(__gfx1100__) || defined(__gfx1101__) || defined(__gfx1102__) || defined(__gfx1103__) || \
-    defined(__gfx1150__) || defined(__gfx1151__)
+#if defined(__GFX11__)
 #define RDNA3
 #endif
 
@@ -199,8 +196,8 @@
     #define __has_builtin(x) 0
 #endif
 
-typedef hip_bfloat16 nv_bfloat16;
-typedef short2 nv_bfloat162; // FIXME there is no 2x BF16 type being defined in bfloat16.h, ad-hoc compilation fix
+typedef __hip_bfloat16 nv_bfloat16;
+typedef __hip_bfloat162 nv_bfloat162;
 
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
@@ -251,17 +248,3 @@ static __device__ __forceinline__ unsigned int __vcmpne4(unsigned int a, unsigne
     }
     return c;
 }
-
-#if HIP_VERSION < 50600000
-// __shfl_xor() for half2 was added in ROCm 5.6
-static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
-    typedef union half2_b32 {
-        half2 val;
-        int   b32;
-    } half2_b32_t;
-    half2_b32_t tmp;
-    tmp.val = var;
-    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
-    return tmp.val;
-}
-#endif // HIP_VERSION < 50600000

ggml/src/ggml-hip/CMakeLists.txt

@@ -46,8 +46,8 @@ if (GGML_HIP_ROCWMMA_FATTN)
     endif()
 endif()
 
-if (${hip_VERSION} VERSION_LESS 5.5)
-    message(FATAL_ERROR "At least ROCM/HIP V5.5 is required")
+if (${hip_VERSION} VERSION_LESS 6.1)
+    message(FATAL_ERROR "At least ROCM/HIP V6.1 is required")
 endif()
 
 message(STATUS "HIP and hipBLAS found")

ggml/src/ggml-opencl/CMakeLists.txt

@@ -82,7 +82,9 @@ set(GGML_OPENCL_KERNELS
     mul_mv_q4_0_f32_1d_8x_flat
     mul_mv_q4_0_f32_1d_16x_flat
     mul_mv_q6_k
+    mul_mv_mxfp4_f32
     mul_mv_id_q4_0_f32_8x_flat
+    mul_mv_id_mxfp4_f32
     mul_mm_f32_f32_l4_lm
     mul_mm_f16_f32_l4_lm
     mul
@@ -110,6 +112,9 @@ set(GGML_OPENCL_KERNELS
     mul_mat_f16_f32
     conv2d
     conv2d_f16_f32
+    flash_attn_f32_f16
+    flash_attn_f16
+    flash_attn_f32
 )
 
 foreach (K ${GGML_OPENCL_KERNELS})

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

@@ -25,6 +25,7 @@
 #include <vector>
 #include <string>
 #include <cmath>
+#include <map>
 #include <memory>
 #include <charconv>
 #include <mutex>
@@ -365,6 +366,7 @@ struct ggml_backend_opencl_context {
     cl_program program_mul_mv_q4_0_f32_1d_8x_flat;
     cl_program program_mul_mv_q4_0_f32_1d_16x_flat;
     cl_program program_mul_mv_q6_K;
+    cl_program program_mul_mv_mxfp4_f32;
     cl_program program_mul_mv_f16_f16;
     cl_program program_mul_mv_f16_f32_1row;
     cl_program program_mul_mv_f16_f32_l4;
@@ -398,6 +400,7 @@ struct ggml_backend_opencl_context {
     cl_program program_conv_2d_f16_f32;
     cl_program program_tsembd;
     cl_program program_mul_mv_id_q4_0_f32_8x_flat;
+    cl_program program_mul_mv_id_mxfp4_f32;
     cl_program program_mul_mm_f32_f32_l4_lm;
     cl_program program_mul_mm_f16_f32_l4_lm;
 
@@ -422,6 +425,14 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_diag_mask_inf, kernel_diag_mask_inf_8;
     cl_kernel kernel_soft_max, kernel_soft_max_4;
     cl_kernel kernel_soft_max_f16, kernel_soft_max_4_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f16_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_q1;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16;
+    std::map<std::pair<int, int>, cl_kernel> kernels_flash_attn_f32_f16_q1;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bm;
+    std::map<std::pair<int, int>, int>       kernels_flash_attn_bn;
     cl_kernel kernel_get_rows_f32, kernel_get_rows_f16, kernel_get_rows_q4_0;
     cl_kernel kernel_set_rows_f32, kernel_set_rows_f16;
     cl_kernel kernel_rope_norm_f32, kernel_rope_norm_f16, kernel_rope_neox_f32, kernel_rope_neox_f16;
@@ -439,6 +450,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_convert_block_q4_0_noshuffle;
     cl_kernel kernel_mul_mat_q4_0_f32_1d_8x_flat, kernel_mul_mat_q4_0_f32_1d_16x_flat;
     cl_kernel kernel_mul_mv_q6_K_f32;
+    cl_kernel kernel_mul_mv_mxfp4_f32;
     cl_kernel kernel_im2col_f32, kernel_im2col_f16;
     cl_kernel kernel_argsort_f32_i32;
     cl_kernel kernel_sum_rows_f32;
@@ -455,6 +467,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_conv_2d_f16_f32;
     cl_kernel kernel_timestep_embedding;
     cl_kernel kernel_mul_mv_id_q4_0_f32_8x_flat;
+    cl_kernel kernel_mul_mv_id_mxfp4_f32;
     cl_kernel kernel_mul_mm_f32_f32_l4_lm;
     cl_kernel kernel_mul_mm_f16_f32_l4_lm;
 
@@ -577,6 +590,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_transpose_32;
     cl_kernel kernel_transpose_32_16;
     cl_kernel kernel_transpose_16;
+    cl_kernel kernel_transpose_16_4x1;
 
     cl_mem A_s_d_max;            // max scale buffer size for transpose
     cl_mem A_q_d_max;            // max weight buffer size for transpose
@@ -971,6 +985,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_mxfp4_f32, "kernel_mul_mv_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // mul_mv_f16_f16
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1287,6 +1317,73 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // flash_attn
+    {
+        #ifdef GGML_OPENCL_EMBED_KERNELS
+                const std::string kernel_src_f16 {
+                    #include "flash_attn_f16.cl.h"
+                };
+                const std::string kernel_src_f32 {
+                    #include "flash_attn_f32.cl.h"
+                };
+                const std::string kernel_src_f32_f16 {
+                    #include "flash_attn_f32_f16.cl.h"
+                };
+        #else
+                const std::string kernel_src_f16 = read_file("flash_attn_f16.cl");
+                const std::string kernel_src_f32 = read_file("flash_attn_f32.cl");
+                const std::string kernel_src_f32_f16 = read_file("flash_attn_f32_f16.cl");
+        #endif
+
+        if (!kernel_src_f16.empty() && !kernel_src_f32.empty() && !kernel_src_f32_f16.empty()) {
+            const struct { int dk; int dv; int bm; int bn; } fa_dims[] = {
+                { 64,  64, 64, 64}, { 80,  80, 64, 32}, { 96,  96, 64, 32},
+                {112, 112, 32, 32}, {128, 128, 32, 32}, {192, 128, 16, 16},
+                {192, 192, 16, 16}, {256, 256, 16, 16},
+            };
+
+            for (size_t i = 0; i < sizeof(fa_dims)/sizeof(fa_dims[0]); ++i) {
+                const int dk = fa_dims[i].dk;
+                const int dv = fa_dims[i].dv;
+                const int bm = fa_dims[i].bm;
+                const int bn = fa_dims[i].bn;
+                std::string OPTS = compile_opts +
+                    " -D DK=" + std::to_string(dk) +
+                    " -D DV=" + std::to_string(dv) +
+                    " -D BLOCK_M=" + std::to_string(bm) +
+                    " -D BLOCK_N=" + std::to_string(bn);
+
+                cl_program prog_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f16.c_str(), OPTS);
+                cl_kernel k_f16, k_f16_q1;
+                CL_CHECK((k_f16 = clCreateKernel(prog_f16, "flash_attn_f16", &err), err));
+                CL_CHECK((k_f16_q1 = clCreateKernel(prog_f16, "flash_attn_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f16[{dk, dv}] = k_f16;
+                backend_ctx->kernels_flash_attn_f16_q1[{dk, dv}] = k_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f16));
+
+                cl_program prog_f32 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32.c_str(), OPTS);
+                cl_kernel k_f32, k_f32_q1;
+                CL_CHECK((k_f32 = clCreateKernel(prog_f32, "flash_attn_f32", &err), err));
+                CL_CHECK((k_f32_q1 = clCreateKernel(prog_f32, "flash_attn_f32_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32[{dk, dv}] = k_f32;
+                backend_ctx->kernels_flash_attn_f32_q1[{dk, dv}] = k_f32_q1;
+                CL_CHECK(clReleaseProgram(prog_f32));
+
+                cl_program prog_f32_f16 = build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src_f32_f16.c_str(), OPTS);
+                cl_kernel k_f32_f16, k_f32_f16_q1;
+                CL_CHECK((k_f32_f16 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16", &err), err));
+                CL_CHECK((k_f32_f16_q1 = clCreateKernel(prog_f32_f16, "flash_attn_f32_f16_q1", &err), err));
+                backend_ctx->kernels_flash_attn_f32_f16[{dk, dv}] = k_f32_f16;
+                backend_ctx->kernels_flash_attn_f32_f16_q1[{dk, dv}] = k_f32_f16_q1;
+                CL_CHECK(clReleaseProgram(prog_f32_f16));
+
+                backend_ctx->kernels_flash_attn_bm[{dk, dv}] = bm;
+                backend_ctx->kernels_flash_attn_bn[{dk, dv}] = bn;
+            }
+            GGML_LOG_CONT(".");
+        }
+    }
+
     // argsort
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1611,6 +1708,22 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // mul_mv_id_mxfp4_f32
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "mul_mv_id_mxfp4_f32.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("mul_mv_id_mxfp4_f32.cl");
+#endif
+        backend_ctx->program_mul_mv_id_mxfp4_f32 =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_mul_mv_id_mxfp4_f32 = clCreateKernel(backend_ctx->program_mul_mv_id_mxfp4_f32, "kernel_mul_mv_id_mxfp4_f32", &err), err));
+        GGML_LOG_CONT(".");
+    }
+
     // Adreno kernels
 #ifdef GGML_OPENCL_USE_ADRENO_KERNELS
     // transpose
@@ -1628,6 +1741,7 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         CL_CHECK((backend_ctx->kernel_transpose_32_16 = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32_16", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_32    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_32", &err), err));
         CL_CHECK((backend_ctx->kernel_transpose_16    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16", &err), err));
+        CL_CHECK((backend_ctx->kernel_transpose_16_4x1    = clCreateKernel(backend_ctx->program_transpose, "kernel_transpose_16_4x1", &err), err));
         GGML_LOG_CONT(".");
     }
 
@@ -2552,13 +2666,14 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 return true;
             } else if (op->src[0]->type == GGML_TYPE_F32) {
                 return op->src[1]->type == GGML_TYPE_F32;
-            } else if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+            } else if (op->src[0]->type == GGML_TYPE_Q4_0 || op->src[0]->type == GGML_TYPE_MXFP4 ||
                        op->src[0]->type == GGML_TYPE_Q6_K) {
                 return op->src[1]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
             }
             return false;
         case GGML_OP_MUL_MAT_ID:
-            if (op->src[0]->type == GGML_TYPE_Q4_0) {
+            if (op->src[0]->type == GGML_TYPE_Q4_0 ||
+                op->src[0]->type == GGML_TYPE_MXFP4) {
                 if (op->src[1]->type == GGML_TYPE_F32) {
                     return ggml_is_contiguous(op->src[0]) && ggml_is_contiguous(op->src[1]);
                 }
@@ -2597,6 +2712,45 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SUM_ROWS:
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous(op->src[0]);
+        case GGML_OP_FLASH_ATTN_EXT:
+            {
+                if (op->src[4]) {
+                    return false;
+                }
+
+                const ggml_tensor * q = op->src[0];
+                const ggml_tensor * k = op->src[1];
+                const ggml_tensor * v = op->src[2];
+
+                const int dk = q->ne[0];
+                const int dv = v->ne[0];
+
+                const struct { int dk; int dv; } supported_dims[] = {
+                    { 64,  64}, { 80,  80}, { 96,  96},
+                    {112, 112}, {128, 128}, {192, 128},
+                    {192, 192}, {256, 256},
+                };
+
+                bool dims_supported = false;
+                for (size_t i = 0; i < sizeof(supported_dims)/sizeof(supported_dims[0]); ++i) {
+                    if (supported_dims[i].dk == dk && supported_dims[i].dv == dv) {
+                        dims_supported = true;
+                        break;
+                    }
+                }
+                if (!dims_supported) {
+                    return false;
+                }
+
+                const bool is_f32_f32 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F32 &&
+                                        v->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32;
+                const bool is_f16_f16 = q->type == GGML_TYPE_F16 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16;
+                const bool is_f32_f16 = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16 &&
+                                        v->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F32;
+
+                return is_f32_f32 || is_f16_f16 || is_f32_f16;
+            }
         default:
             return false;
     }
@@ -2944,7 +3098,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         // cl_mem qT_d = clCreateBuffer(context, CL_MEM_READ_WRITE, q_size_bytes, NULL, &err);
         CL_CHECK(err);
 
-        // size_t d_size_bytes = M * (K / 32) / 2 * sizeof(float);
+        bool K_tile_trans = true;
+        if ((K / 32) % 4 != 0){
+            K_tile_trans =false;
+        }
         size_t d_size_bytes = M * (K / 32) * 2;
         region.origin = 0;
         region.size = d_size_bytes;
@@ -2985,10 +3142,15 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         qT_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
 
-        img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
         memset(&img_desc_1d, 0, sizeof(img_desc_1d));
+        if (K_tile_trans) {
+            img_fmt_1d = { CL_RGBA, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32 / 4;
+        } else {
+            img_fmt_1d = { CL_R, CL_HALF_FLOAT };
+            img_desc_1d.image_width = M * K / 32;
+        }
         img_desc_1d.image_type = CL_MEM_OBJECT_IMAGE1D_BUFFER;
-        img_desc_1d.image_width = M * K / 32 / 4;
         img_desc_1d.buffer = extra->d;
         d_d_image1D = clCreateImage(context, 0, &img_fmt_1d, &img_desc_1d, NULL, &err);
         CL_CHECK(err);
@@ -3024,6 +3186,10 @@ static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer,
         int width_s = K / 32 / 4;
 
         kernel = backend_ctx->kernel_transpose_16;
+        if (!K_tile_trans) {
+            kernel = backend_ctx->kernel_transpose_16_4x1;
+            width_s = K / 32;
+        }
         CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem), &d_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_mem), &dT_d_image1D));
         CL_CHECK(clSetKernelArg(kernel, 2, sizeof(int), &height_s));
@@ -5400,6 +5566,133 @@ static void ggml_cl_timestep_embedding(ggml_backend_t backend, const ggml_tensor
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, NULL, dst);
 }
 
+static void ggml_cl_flash_attn(ggml_backend_t backend, const ggml_tensor * q, const ggml_tensor * k, ggml_tensor * dst) {
+    const ggml_tensor * v = dst->src[2];
+    const ggml_tensor * mask = dst->src[3];
+    GGML_ASSERT(q->extra);
+    GGML_ASSERT(k->extra);
+    GGML_ASSERT(v->extra);
+    GGML_ASSERT(dst->extra);
+    if (mask) {
+        GGML_ASSERT(mask->extra);
+    }
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    const int n_q = q->ne[1];
+    const int n_kv = k->ne[1];
+    const int d_head_q = q->ne[0];
+    const int d_head_v = v->ne[0];
+    const int n_head = q->ne[2];
+    const int n_head_kv = k->ne[2];
+    const int n_batch = q->ne[3];
+
+    cl_kernel kernel = NULL;
+
+    const bool is_f16 = q->type == GGML_TYPE_F16;
+    const bool is_mixed = q->type == GGML_TYPE_F32 && k->type == GGML_TYPE_F16;
+    const std::pair<int, int> dk_dv = {d_head_q, d_head_v};
+
+    if (n_q == 1) {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16_q1.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16_q1.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32_q1.at(dk_dv);
+        }
+    } else {
+        if (is_mixed) {
+            kernel = backend_ctx->kernels_flash_attn_f32_f16.at(dk_dv);
+        } else if (is_f16) {
+            kernel = backend_ctx->kernels_flash_attn_f16.at(dk_dv);
+        } else {
+            kernel = backend_ctx->kernels_flash_attn_f32.at(dk_dv);
+        }
+    }
+    GGML_ASSERT(kernel != NULL);
+
+    ggml_tensor_extra_cl * extra_q = (ggml_tensor_extra_cl *)q->extra;
+    ggml_tensor_extra_cl * extra_k = (ggml_tensor_extra_cl *)k->extra;
+    ggml_tensor_extra_cl * extra_v = (ggml_tensor_extra_cl *)v->extra;
+    ggml_tensor_extra_cl * extra_o = (ggml_tensor_extra_cl *)dst->extra;
+    ggml_tensor_extra_cl * extra_mask = mask ? (ggml_tensor_extra_cl *)mask->extra : NULL;
+
+    cl_ulong offset_q = extra_q->offset + q->view_offs;
+    cl_ulong offset_k = extra_k->offset + k->view_offs;
+    cl_ulong offset_v = extra_v->offset + v->view_offs;
+    cl_ulong offset_o = extra_o->offset + dst->view_offs;
+    cl_mem   mask_buffer = extra_mask ? extra_mask->data_device : NULL;
+    cl_ulong offset_mask = extra_mask ? extra_mask->offset + mask->view_offs : 0;
+
+    const cl_ulong q_nb1 = q->nb[1], q_nb2 = q->nb[2], q_nb3 = q->nb[3];
+    const cl_ulong k_nb1 = k->nb[1], k_nb2 = k->nb[2], k_nb3 = k->nb[3];
+    const cl_ulong v_nb1 = v->nb[1], v_nb2 = v->nb[2], v_nb3 = v->nb[3];
+    const cl_ulong o_nb1 = dst->nb[1], o_nb2 = dst->nb[2], o_nb3 = dst->nb[3];
+    const cl_ulong mask_nb1 = mask ? mask->nb[1] : 0;
+    const cl_ulong mask_nb2 = mask ? mask->nb[2] : 0;
+    const cl_ulong mask_nb3 = mask ? mask->nb[3] : 0;
+    const int mask_ne2 = mask ? mask->ne[2] : 0;
+    const int mask_ne3 = mask ? mask->ne[3] : 0;
+
+    float scale, max_bias, logit_softcap;
+    const float * params = (const float *)dst->op_params;
+    scale         = params[0];
+    max_bias      = params[1];
+    logit_softcap = params[2];
+
+    const int is_causal = (mask == NULL && n_q > 1 && n_q == n_kv);
+
+    const int n_head_log2_val = n_head > 0 ? 1u << (int)floorf(log2f((float)n_head)) : 0;
+    const float n_head_log2_f = n_head_log2_val > 0 ? (float)n_head_log2_val : 1.0f;
+    const float m0 = powf(2.0f, -(max_bias) / n_head_log2_f);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2_f);
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra_q->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset_q));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extra_k->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offset_k));
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(cl_mem),   &extra_v->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(cl_ulong), &offset_v));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(cl_mem),   &extra_o->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(cl_ulong), &offset_o));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(float),    &scale));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(int),      &n_q));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),     &n_kv));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(int),     &is_causal));
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &n_head));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &q_nb1)); CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &q_nb2)); CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &q_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &k_nb1)); CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong), &k_nb2)); CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong), &k_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &v_nb1)); CL_CHECK(clSetKernelArg(kernel, 20, sizeof(cl_ulong), &v_nb2)); CL_CHECK(clSetKernelArg(kernel, 21, sizeof(cl_ulong), &v_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 22, sizeof(cl_ulong), &o_nb1)); CL_CHECK(clSetKernelArg(kernel, 23, sizeof(cl_ulong), &o_nb2)); CL_CHECK(clSetKernelArg(kernel, 24, sizeof(cl_ulong), &o_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 25, sizeof(float),    &max_bias));
+    CL_CHECK(clSetKernelArg(kernel, 26, sizeof(float),    &m0));
+    CL_CHECK(clSetKernelArg(kernel, 27, sizeof(float),    &m1));
+    CL_CHECK(clSetKernelArg(kernel, 28, sizeof(int),      &n_head_log2_val));
+    CL_CHECK(clSetKernelArg(kernel, 29, sizeof(float),    &logit_softcap));
+    CL_CHECK(clSetKernelArg(kernel, 30, sizeof(int),      &n_head_kv));
+    CL_CHECK(clSetKernelArg(kernel, 31, sizeof(cl_mem),   &mask_buffer));
+    CL_CHECK(clSetKernelArg(kernel, 32, sizeof(cl_ulong), &offset_mask));
+    CL_CHECK(clSetKernelArg(kernel, 33, sizeof(cl_ulong), &mask_nb1));
+    CL_CHECK(clSetKernelArg(kernel, 34, sizeof(cl_ulong), &mask_nb2));
+    CL_CHECK(clSetKernelArg(kernel, 35, sizeof(cl_ulong), &mask_nb3));
+    CL_CHECK(clSetKernelArg(kernel, 36, sizeof(int),      &mask_ne2));
+    CL_CHECK(clSetKernelArg(kernel, 37, sizeof(int),      &mask_ne3));
+
+    if (n_q == 1) {
+        const size_t wg_size = 64;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    } else {
+        const int block_m = backend_ctx->kernels_flash_attn_bm.at(dk_dv);
+        const size_t wg_size = block_m;
+        size_t local_work_size[] = { wg_size, 1 };
+        size_t global_work_size[] = { (size_t)((n_q + block_m - 1) / block_m) * wg_size, (size_t)(n_head * n_batch) };
+        backend_ctx->enqueue_ndrange_kernel(kernel, 2, global_work_size, local_work_size, dst);
+    }
+}
+
 static void ggml_cl_mul_mat_f16_f32_tiled(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
@@ -6254,11 +6547,47 @@ static void ggml_cl_mul_mat(ggml_backend_t backend, const ggml_tensor * src0, co
             CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &r2));
             CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &r3));
             break;
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                nth0 = 16;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                nth0 = 64;
+                nth1 = 2;
+                ndst = nth1*2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(int),      &ne00));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &nb01));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(cl_ulong), &nb02));
+            CL_CHECK(clSetKernelArg(kernel,  9, sizeof(cl_ulong), &nb03));
+            CL_CHECK(clSetKernelArg(kernel, 10, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 12, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(float)*nth0,nullptr));
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");
     }
 
-    if (src0t == GGML_TYPE_Q4_0 ||
+    if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_MXFP4 ||
         src0t == GGML_TYPE_Q4_1 ||
         src0t == GGML_TYPE_Q8_0 ||
         src0t == GGML_TYPE_Q2_K) {
@@ -6307,10 +6636,12 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
 
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
     ggml_tensor_extra_cl * extra1 = (ggml_tensor_extra_cl *)src1->extra;
     ggml_tensor_extra_cl * extra2 = (ggml_tensor_extra_cl *)src2->extra;
     ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
 
+    cl_ulong offset0 = extra0->offset + src0->view_offs;
     cl_ulong offset1 = extra1->offset + src1->view_offs;
     cl_ulong offset2 = extra2->offset + src2->view_offs;
     cl_ulong offsetd = extrad->offset + dst->view_offs;
@@ -6325,7 +6656,9 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
     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 = src1->ne[0];
     const int ne11 = src1->ne[1];
@@ -6334,6 +6667,7 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
     const cl_ulong nb11 = src1->nb[1];
     const cl_ulong nb12 = src1->nb[2];
+    const cl_ulong nb13 = src1->nb[3];
 
     const int ne20 = src2->ne[0];
     const int ne21 = src2->ne[1];
@@ -6401,6 +6735,49 @@ static void ggml_cl_mul_mat_id(ggml_backend_t backend, const ggml_tensor * src0,
 
             break;
         }
+        case GGML_TYPE_MXFP4: {
+            kernel = backend_ctx->kernel_mul_mv_id_mxfp4_f32;
+
+            if (backend_ctx->gpu_family == INTEL) {
+                sgs  = 16;
+                nsg  = 2;
+                ndst = 2;
+            } else if (backend_ctx->gpu_family == ADRENO) {
+                sgs  = 64;
+                nsg  = 2;
+                ndst = 2;
+            } else {
+                GGML_ASSERT(false && "TODO: Unknown GPU");
+            }
+
+            CL_CHECK(clSetKernelArg(kernel,  0, sizeof(cl_mem),   &extra0->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  1, sizeof(cl_ulong), &offset0));
+            CL_CHECK(clSetKernelArg(kernel,  2, sizeof(cl_mem),   &extra1->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  3, sizeof(cl_ulong), &offset1));
+            CL_CHECK(clSetKernelArg(kernel,  4, sizeof(cl_mem),   &extra2->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  5, sizeof(cl_ulong), &offset2));
+            CL_CHECK(clSetKernelArg(kernel,  6, sizeof(cl_mem),   &extrad->data_device));
+            CL_CHECK(clSetKernelArg(kernel,  7, sizeof(cl_ulong), &offsetd));
+            CL_CHECK(clSetKernelArg(kernel,  8, sizeof(int),      &ne00));
+            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),      &ne11));
+            CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),      &ne12));
+            CL_CHECK(clSetKernelArg(kernel, 14, sizeof(cl_ulong), &nb11));
+            CL_CHECK(clSetKernelArg(kernel, 15, sizeof(cl_ulong), &nb12));
+            CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong), &nb13));
+            CL_CHECK(clSetKernelArg(kernel, 17, sizeof(int),      &ne20));
+            CL_CHECK(clSetKernelArg(kernel, 18, sizeof(int),      &ne21));
+            CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong), &nb21));
+            CL_CHECK(clSetKernelArg(kernel, 20, sizeof(int),      &ne0));
+            CL_CHECK(clSetKernelArg(kernel, 21, sizeof(int),      &ne1));
+            CL_CHECK(clSetKernelArg(kernel, 22, sizeof(int),      &r2));
+            CL_CHECK(clSetKernelArg(kernel, 23, sizeof(int),      &r3));
+            CL_CHECK(clSetKernelArg(kernel, 24, sizeof(float)*sgs,nullptr));
+
+            break;
+        }
         default:
             GGML_ASSERT(false && "not implemented");;
     }
@@ -7472,6 +7849,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_sum_rows;
             break;
+        case GGML_OP_FLASH_ATTN_EXT:
+            if (!any_on_device) {
+                return false;
+            }
+            ggml_cl_flash_attn(backend, tensor->src[0], tensor->src[1], tensor);
+            return true;
         default:
             return false;
     }

ggml/src/ggml-opencl/kernels/flash_attn_f16.cl

@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE half
+#define DATA_TYPE4 half4
+#define CONVERT_ACC4(x) convert_float4(x)
+#define CONVERT_DATA4(x) convert_half4(x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}

ggml/src/ggml-opencl/kernels/flash_attn_f32.cl

@@ -0,0 +1,343 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define DATA_TYPE float
+#define DATA_TYPE4 float4
+#define CONVERT_ACC4(x) (x)
+#define CONVERT_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_ACC4(l_v[j][i]) + p1 * CONVERT_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global DATA_TYPE4* q_ptr = (const global DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global DATA_TYPE4* k_ptr = (const global DATA_TYPE4*)(k_base + k_row_offset);
+        const global DATA_TYPE4* v_ptr = (const global DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global DATA_TYPE* mask_ptr = (const global DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global DATA_TYPE4 *o_row = (global DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (DATA_TYPE4)(0.0f);
+    }
+}

ggml/src/ggml-opencl/kernels/flash_attn_f32_f16.cl

@@ -0,0 +1,346 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define ACC_TYPE float
+#define ACC_TYPE4 float4
+#define Q_DATA_TYPE4 float4
+#define KV_DATA_TYPE4 half4
+#define O_DATA_TYPE4 float4
+#define MASK_DATA_TYPE half
+#define CONVERT_Q_ACC4(x) (x)
+#define CONVERT_KV_ACC4(x) convert_float4(x)
+#define CONVERT_O_DATA4(x) (x)
+
+#define DK_VEC (DK/4)
+#define DV_VEC (DV/4)
+#define WG_SIZE (BLOCK_M)
+#define Q1_WG_SIZE 64
+
+inline float get_alibi_slope(
+    const float max_bias, const uint h, const uint n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return pow(base, exph);
+}
+__kernel void flash_attn_f32_f16(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int block_q_idx = get_group_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int my_query_row = block_q_idx * BLOCK_M + tid;
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    if (my_query_row < n_q) {
+        const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2 + my_query_row * q_nb1;
+        const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+        #pragma unroll
+        for (int i = 0; i < DK_VEC; ++i) {
+            q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+        }
+    }
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) {
+        o_acc[i] = (ACC_TYPE4)(0.0f);
+    }
+    ACC_TYPE m_i = -INFINITY;
+    ACC_TYPE l_i = 0.0f;
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    __local KV_DATA_TYPE4 l_k[BLOCK_N][DK_VEC];
+    __local KV_DATA_TYPE4 l_v[BLOCK_N][DV_VEC];
+
+    for (int k_start = 0; k_start < n_kv; k_start += BLOCK_N) {
+        for (int i = tid; i < BLOCK_N * DK_VEC; i += WG_SIZE) {
+            const int row = i / DK_VEC;
+            const int col = i % DK_VEC;
+            const int k_row_idx = k_start + row;
+            if (k_row_idx < n_kv) {
+                const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_row_idx * k_nb1;
+                l_k[row][col] = ((__global KV_DATA_TYPE4*)(k_base + k_row_offset))[col];
+            }
+        }
+        for (int i = tid; i < BLOCK_N * DV_VEC; i += WG_SIZE) {
+            const int row = i / DV_VEC;
+            const int col = i % DV_VEC;
+            const int v_row_idx = k_start + row;
+            if (v_row_idx < n_kv) {
+                const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + v_row_idx * v_nb1;
+                l_v[row][col] = ((__global KV_DATA_TYPE4*)(v_base + v_row_offset))[col];
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (my_query_row >= n_q) {
+            continue;
+        }
+
+        for (int j = 0; j < BLOCK_N; j += 2) {
+            const int k_row0 = k_start + j;
+            const int k_row1 = k_start + j + 1;
+
+            ACC_TYPE4 dot_acc0 = (ACC_TYPE4)(0.0f);
+            ACC_TYPE4 dot_acc1 = (ACC_TYPE4)(0.0f);
+            #pragma unroll
+            for (int k = 0; k < DK_VEC; k++) {
+                dot_acc0 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j][k]), dot_acc0);
+                dot_acc1 = mad(q_priv[k], CONVERT_KV_ACC4(l_k[j+1][k]), dot_acc1);
+            }
+            ACC_TYPE score0 = (dot_acc0.s0 + dot_acc0.s1 + dot_acc0.s2 + dot_acc0.s3) * scale;
+            ACC_TYPE score1 = (dot_acc1.s0 + dot_acc1.s1 + dot_acc1.s2 + dot_acc1.s3) * scale;
+
+            if (is_causal) {
+                if (k_row0 > (n_kv - n_q + my_query_row)) score0 = -INFINITY;
+                if (k_row1 > (n_kv - n_q + my_query_row)) score1 = -INFINITY;
+            }
+
+            if (k_row0 >= n_kv) score0 = -INFINITY;
+            if (k_row1 >= n_kv) score1 = -INFINITY;
+
+            if (mask_base != NULL) {
+                const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base + my_query_row * mask_nb1);
+                if (k_row0 < n_kv) score0 += slope * (ACC_TYPE)mask_ptr[k_row0];
+                if (k_row1 < n_kv) score1 += slope * (ACC_TYPE)mask_ptr[k_row1];
+            }
+
+            if (logit_softcap > 0.0f) {
+                score0 = logit_softcap * tanh(score0 / logit_softcap);
+                score1 = logit_softcap * tanh(score1 / logit_softcap);
+            }
+
+            const ACC_TYPE m_new = max(m_i, max(score0, score1));
+            const ACC_TYPE p0 = exp(score0 - m_new);
+            const ACC_TYPE p1 = exp(score1 - m_new);
+            const ACC_TYPE scale_prev = exp(m_i - m_new);
+
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_acc[i] = o_acc[i] * scale_prev + p0 * CONVERT_KV_ACC4(l_v[j][i]) + p1 * CONVERT_KV_ACC4(l_v[j+1][i]);
+            }
+            l_i = l_i * scale_prev + p0 + p1;
+            m_i = m_new;
+        }
+    }
+
+    if (my_query_row < n_q) {
+        const ulong o_row_offset = batch_idx * o_nb3 + my_query_row * o_nb2 + head_idx * o_nb1;
+        global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+        if (l_i > 0.0f) {
+            const ACC_TYPE l_inv = 1.0f / l_i;
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = CONVERT_O_DATA4(o_acc[i] * l_inv);
+            }
+        } else {
+            #pragma unroll
+            for (int i = 0; i < DV_VEC; ++i) {
+                o_row[i] = (O_DATA_TYPE4)(0.0f);
+            }
+        }
+    }
+}
+
+__kernel void flash_attn_f32_f16_q1(
+    const global void * q_void, ulong q_offset,
+    const global void * k_void, ulong k_offset,
+    const global void * v_void, ulong v_offset,
+    global void * o_void, ulong o_offset,
+    const float scale,
+    const int n_q,
+    const int n_kv,
+    const int is_causal,
+    const int n_head,
+    const ulong q_nb1, const ulong q_nb2, const ulong q_nb3,
+    const ulong k_nb1, const ulong k_nb2, const ulong k_nb3,
+    const ulong v_nb1, const ulong v_nb2, const ulong v_nb3,
+    const ulong o_nb1, const ulong o_nb2, const ulong o_nb3,
+    const float max_bias,
+    const float m0,
+    const float m1,
+    const int n_head_log2,
+    const float logit_softcap,
+    const int n_head_kv,
+    const global void* mask_void,
+    const ulong mask_offset,
+    const ulong mask_nb1,
+    const ulong mask_nb2,
+    const ulong mask_nb3,
+    const int mask_ne2,
+    const int mask_ne3
+) {
+    const int tid = get_local_id(0);
+    const int head_batch_idx = get_global_id(1);
+
+    const int batch_idx = head_batch_idx / n_head;
+    const int head_idx = head_batch_idx % n_head;
+
+    const int gqa_ratio = n_head / n_head_kv;
+    const int head_kv_idx = head_idx / gqa_ratio;
+
+    const global char* q_base = (const global char*)q_void + q_offset;
+    const global char* k_base = (const global char*)k_void + k_offset;
+    const global char* v_base = (const global char*)v_void + v_offset;
+    global char* o_base = (global char*)o_void + o_offset;
+
+    const global char* mask_base = NULL;
+    if (mask_void != NULL) {
+        const int mask_head_idx = head_idx % mask_ne2;
+        const int mask_batch_idx = batch_idx % mask_ne3;
+        mask_base = (const global char*)mask_void + mask_offset + mask_batch_idx * mask_nb3 + mask_head_idx * mask_nb2;
+    }
+
+    ACC_TYPE4 q_priv[DK_VEC];
+    const ulong q_row_offset = batch_idx * q_nb3 + head_idx * q_nb2;
+    const global Q_DATA_TYPE4* q_ptr = (const global Q_DATA_TYPE4*)(q_base + q_row_offset);
+    #pragma unroll
+    for (int i = 0; i < DK_VEC; ++i) {
+        q_priv[i] = CONVERT_Q_ACC4(q_ptr[i]);
+    }
+
+    float slope = get_alibi_slope(max_bias, head_idx, n_head_log2, m0, m1);
+
+    ACC_TYPE m_i = -INFINITY;
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        m_i = max(m_i, score);
+    }
+
+    __local ACC_TYPE local_m[Q1_WG_SIZE];
+    local_m[tid] = m_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_m[tid] = max(local_m[tid], local_m[tid + s]);
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    const ACC_TYPE m_final = local_m[0];
+
+    ACC_TYPE4 o_acc[DV_VEC];
+    #pragma unroll
+    for (int i = 0; i < DV_VEC; ++i) o_acc[i] = (ACC_TYPE4)(0.0f);
+    ACC_TYPE l_i = 0.0f;
+
+    for (int k_idx = tid; k_idx < n_kv; k_idx += Q1_WG_SIZE) {
+        const ulong k_row_offset = batch_idx * k_nb3 + head_kv_idx * k_nb2 + k_idx * k_nb1;
+        const ulong v_row_offset = batch_idx * v_nb3 + head_kv_idx * v_nb2 + k_idx * v_nb1;
+        const global KV_DATA_TYPE4* k_ptr = (const global KV_DATA_TYPE4*)(k_base + k_row_offset);
+        const global KV_DATA_TYPE4* v_ptr = (const global KV_DATA_TYPE4*)(v_base + v_row_offset);
+        ACC_TYPE4 dot_acc = (ACC_TYPE4)(0.0f);
+        #pragma unroll
+        for (int k = 0; k < DK_VEC; k++) {
+            dot_acc = mad(q_priv[k], CONVERT_KV_ACC4(k_ptr[k]), dot_acc);
+        }
+        ACC_TYPE score = (dot_acc.s0 + dot_acc.s1 + dot_acc.s2 + dot_acc.s3) * scale;
+        if (mask_base != NULL) {
+            const global MASK_DATA_TYPE* mask_ptr = (const global MASK_DATA_TYPE*)(mask_base);
+            score += slope * (ACC_TYPE)mask_ptr[k_idx];
+        }
+        if (logit_softcap > 0.0f) {
+            score = logit_softcap * tanh(score / logit_softcap);
+        }
+        const ACC_TYPE p = exp(score - m_final);
+        l_i += p;
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; i++) {
+            o_acc[i] = mad(p, CONVERT_KV_ACC4(v_ptr[i]), o_acc[i]);
+        }
+    }
+
+    __local ACC_TYPE local_l[Q1_WG_SIZE];
+    __local ACC_TYPE4 local_o_comp[Q1_WG_SIZE];
+    local_l[tid] = l_i;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    #pragma unroll
+    for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) local_l[tid] += local_l[tid + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const ulong o_row_offset = batch_idx * o_nb3 + head_idx * o_nb1;
+    global O_DATA_TYPE4 *o_row = (global O_DATA_TYPE4 *)(o_base + o_row_offset);
+    const ACC_TYPE l_final = local_l[0];
+
+    if (l_final > 0.0f) {
+        const ACC_TYPE l_inv = 1.0f / l_final;
+        for (int i = 0; i < DV_VEC; i++) {
+            local_o_comp[tid] = o_acc[i];
+            barrier(CLK_LOCAL_MEM_FENCE);
+            #pragma unroll
+            for (int s = Q1_WG_SIZE / 2; s > 0; s >>= 1) {
+                if (tid < s) local_o_comp[tid] += local_o_comp[tid + s];
+                barrier(CLK_LOCAL_MEM_FENCE);
+            }
+            if (tid == 0) {
+                o_row[i] = CONVERT_O_DATA4(local_o_comp[0] * l_inv);
+            }
+        }
+    } else if (tid == 0) {
+        #pragma unroll
+        for (int i = 0; i < DV_VEC; ++i) o_row[i] = (O_DATA_TYPE4)(0.0f);
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_id_mxfp4_f32.cl

@@ -0,0 +1,189 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+inline void mul_mv_mxfp4_f32(
+    global char * src0,
+    global char * src1,
+    global char * dst,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = 0;
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_id_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * src2,
+    ulong         offset2,
+    global char * dst,
+    ulong         offsetd,
+    int           ne00,
+    ulong         nb01,
+    ulong         nb02,
+    ulong         nb03,
+    int           ne11,
+    int           ne12,
+    ulong         nb11,
+    ulong         nb12,
+    ulong         nb13,
+    int           ne20,
+    int           ne21,
+    ulong         nb21,
+    int           ne0,
+    int           ne1,
+    int           r2,
+    int           r3,
+    local  char * shmem
+) {
+    src0 = (global char *)((global char *)src0 + offset0);
+    src1 = (global char *)((global char *)src1 + offset1);
+    src2 = (global char *)((global char *)src2 + offset2);
+    dst  = (global char *)((global char *)dst  + offsetd);
+
+    const int iid1 = get_group_id(2)/ne20;
+    const int idx  = get_group_id(2)%ne20;
+
+    int i02 = ((global int *) (src2 + iid1*nb21))[idx];
+
+    int i11 = idx % ne11;
+    int i12 = iid1;
+
+    int i1 = idx;
+    int i2 = i12;
+
+    global char * src0_cur = src0 + i02*nb02;
+    global char * src1_cur = src1 + i11*nb11 + i12*nb12;
+
+    global char * dst_cur = dst + (i1*ne0 + i2*ne1*ne0)*sizeof(float);
+
+    mul_mv_mxfp4_f32(src0_cur, src1_cur, dst_cur,
+        ne00, nb01, nb02, nb03, ne12, nb11, nb12, nb13, ne0, ne1, r2, r3, shmem);
+}

ggml/src/ggml-opencl/kernels/mul_mv_mxfp4_f32.cl

@@ -0,0 +1,144 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define QK_MXFP4 32
+typedef struct {
+    uchar e; // E8M0
+    uchar qs[QK_MXFP4/2];
+} block_mxfp4;
+
+constant static float kvalues_mxfp4_f[16] = {
+    0, .5f, 1.f, 1.5f, 2.f, 3.f, 4.f, 6.f, -0, -.5f, -1.f, -1.5f, -2.f, -3.f, -4.f, -6.f
+};
+
+static inline float e8m0_to_fp32(uchar x) {
+    int bits;
+
+    if (x == 0) {
+        bits = 0x00400000;
+    } else {
+        bits = (uint) x << 23;
+    }
+
+    return as_float(bits);
+}
+
+#ifdef INTEL_GPU
+#define N_R0_MXFP4 2 // number of rows each subgroup works on
+#define N_SG_MXFP4 2 // number of subgroups in a work group
+#define N_SIMDWIDTH 16 // subgroup size
+#elif defined (ADRENO_GPU)
+#define N_R0_MXFP4 2
+#define N_SG_MXFP4 2
+#define N_SIMDWIDTH 64
+#endif
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_mxfp4_f32(
+    global char * src0,
+    ulong         offset0,
+    global char * src1,
+    ulong         offset1,
+    global char * dst,
+    ulong         offsetd,
+    int ne00,
+    ulong nb01,
+    ulong nb02,
+    ulong nb03,
+    int ne12,
+    ulong nb11,
+    ulong nb12,
+    ulong nb13,
+    int ne0,
+    int ne1,
+    int r2,
+    int r3,
+    local  char * shmem
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global char*)((global char*)dst  + offsetd);
+
+    local float * shmem_f32 = (local float *) shmem;
+    int nb = ne00/QK_MXFP4;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+
+    int first_row = (r0 * N_SG_MXFP4 + get_sub_group_id()) * N_R0_MXFP4;
+
+    uint i12 = im%ne12;
+    uint i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_mxfp4 * x = (global block_mxfp4 *) (src0 + offset_src0);
+    global float       * y = (global float       *) (src1 + offset_src1);
+
+    const short ix = get_sub_group_local_id()/2;  // 0...15
+    const short it = get_sub_group_local_id()%2;  // 0 or 1
+
+    shmem_f32[get_sub_group_local_id()] = kvalues_mxfp4_f[get_sub_group_local_id()%16];
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    float4 yl[4];
+    float sumf[N_R0_MXFP4] = {0.f};
+
+    global float * yb = y + ix * QK_MXFP4 + it * 8;
+
+    for (int ib = ix; ib < nb; ib += N_SIMDWIDTH/2) {
+        global float4 * y4 = (global float4 *)yb;
+        yl[0] = y4[0];
+        yl[1] = y4[4];
+        yl[2] = y4[1];
+        yl[3] = y4[5];
+
+        for (short row = 0; row < N_R0_MXFP4; row++) {
+            global block_mxfp4 * xb = x + row*nb + ib;
+            global uchar       * q2 = (global uchar *)(xb->qs + 8*it);
+
+            float4 acc1 = yl[0]*(float4)(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
+            float4 acc2 = yl[1]*(float4)(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
+            float4 acc3 = yl[2]*(float4)(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
+            float4 acc4 = yl[3]*(float4)(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+
+            acc1 = (acc1 + acc3) + (acc2 + acc4);
+
+            sumf[row] += e8m0_to_fp32(xb->e) * ((acc1.s0 + acc1.s1) + (acc1.s2 + acc1.s3));
+        }
+
+        yb += (N_SIMDWIDTH/2) * QK_MXFP4;
+    }
+
+    global float * dst_f32 = (global float *) dst + (ulong)im*ne0*ne1 + (ulong)r1*ne0;
+
+    for (int row = 0; row < N_R0_MXFP4 && first_row + row < ne0; ++row) {
+        float sum_all = sub_group_reduce_add(sumf[row]);
+        if (get_sub_group_local_id() == 0) {
+            dst_f32[first_row + row] = sum_all;
+        }
+    }
+}

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

@@ -24,6 +24,26 @@ kernel void kernel_transpose_16(
     write_imageh(output, (i_2+3)*rows+j, (half4)(temp0.s3, temp1.s3, temp2.s3, temp3.s3));
 }
 
+// Padded kernel for irregular shape
+kernel void kernel_transpose_16_4x1(
+    __read_only image1d_buffer_t input,
+    __write_only image1d_buffer_t output,
+    const uint rows,
+    const uint cols
+) {
+
+    const int i = get_global_id(0);
+    const int j = get_global_id(1);
+    const int j_2 = j << 2;
+
+    half temp0 = read_imageh(input, (j_2 + 0) * cols + i).x;
+    half temp1 = read_imageh(input, (j_2 + 1) * cols + i).x;
+    half temp2 = read_imageh(input, (j_2 + 2) * cols + i).x;
+    half temp3 = read_imageh(input, (j_2 + 3) * cols + i).x;
+
+    write_imageh(output, i * rows + j, (half4)(temp0, temp1, temp2, temp3));
+}
+
 // 32-bit transpose, loading/storing a 4x4 tile of elements
 kernel void kernel_transpose_32(
     __read_only image1d_buffer_t input,

ggml/src/ggml-opt.cpp

@@ -64,9 +64,11 @@ struct ggml_opt_context {
     int32_t opt_i              = 0;
     bool    loss_per_datapoint = false;
 
-    ggml_opt_get_optimizer_params get_opt_pars = nullptr;
-    void * get_opt_pars_ud                     = nullptr;
-    struct ggml_tensor * adamw_params          = nullptr;
+    ggml_opt_get_optimizer_params get_opt_pars    = nullptr;
+    void *                        get_opt_pars_ud = nullptr;
+    struct ggml_tensor *          opt_step_params = nullptr; // Stores output of get_opt_pars.
+
+    enum ggml_opt_optimizer_type optimizer = GGML_OPT_OPTIMIZER_TYPE_ADAMW;
 };
 
 struct ggml_opt_result {
@@ -229,9 +231,13 @@ struct ggml_opt_optimizer_params ggml_opt_get_default_optimizer_params(void * us
     result.adamw.eps   = 1e-8f;
     result.adamw.wd    = 0.0f;
 
+    result.sgd.alpha   = 1e-3f;
+    result.sgd.wd      = 0.0f;
+
     return result;
 }
 
+
 struct ggml_opt_optimizer_params ggml_opt_get_constant_optimizer_params(void * userdata) {
     return *((struct ggml_opt_optimizer_params *) userdata);
 }
@@ -249,6 +255,7 @@ struct ggml_opt_params ggml_opt_default_params(
         /*opt_period      =*/ 1,
         /*get_opt_pars    =*/ ggml_opt_get_default_optimizer_params,
         /*get_opt_pars_ud =*/ nullptr,
+        /*optimizer       =*/ GGML_OPT_OPTIMIZER_TYPE_ADAMW,
     };
 }
 
@@ -316,9 +323,14 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     GGML_ASSERT(opt_ctx->ctx_compute && "no compute context set, either use static graphs or set one with ggml_opt_prepare_alloc");
     GGML_ASSERT((!opt_ctx->static_graphs || opt_ctx->inputs->data) && "when using static graphs the inputs must be allocated statically");
 
+    const enum ggml_opt_optimizer_type optimizer = opt_ctx->optimizer;
+
     const bool accumulate = opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_GRAD &&
         !(opt_ctx->static_graphs && opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT && opt_ctx->opt_period == 1);
 
+    const bool need_momenta = opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT &&
+        opt_ctx->optimizer == GGML_OPT_OPTIMIZER_TYPE_ADAMW;
+
     ggml_set_input(opt_ctx->inputs);
     ggml_set_output(opt_ctx->outputs);
 
@@ -340,8 +352,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
         //   - pred (if using static graphs)
         //   - ncorrect (if using static graphs, 2 tensors).
         constexpr size_t n_loss = 1;
-        const size_t tensors_per_param = (accumulate ? 1 : 0) +
-            (opt_ctx->build_type_alloc == GGML_OPT_BUILD_TYPE_OPT ? 2 : 0);
+        const size_t tensors_per_param = (accumulate ? 1 : 0) + (need_momenta ? 2 : 0);
         const size_t tensors_const = opt_ctx->static_graphs ? 9 : 0;
         const size_t size_meta = (n_loss + tensors_per_param*n_param + tensors_const) * ggml_tensor_overhead();
         struct ggml_init_params params = {
@@ -458,7 +469,7 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
             }
         }
 
-        if (opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
+        if (need_momenta && opt_ctx->build_type_alloc >= GGML_OPT_BUILD_TYPE_OPT) {
             opt_ctx->grad_m.resize(n_nodes);
             opt_ctx->grad_v.resize(n_nodes);
             for (int i = 0; i < n_nodes; ++i) {
@@ -492,23 +503,36 @@ static void ggml_opt_build(ggml_opt_context_t opt_ctx) {
     // gb_opt == graph backward optimize, forward pass, then backward pass to calculate gradients, then optimizer step.
     opt_ctx->gb_opt = ggml_graph_dup(opt_ctx->ctx_compute, opt_ctx->gb_grad, /*force_grads =*/ true);
 
-    opt_ctx->adamw_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, 7);
-    ggml_set_input(opt_ctx->adamw_params);
-    ggml_set_name(opt_ctx->adamw_params, "adamw_params");
-
+    opt_ctx->opt_step_params = ggml_new_tensor_1d(opt_ctx->ctx_cpu, GGML_TYPE_F32, need_momenta ? 7 : 2);
+    ggml_tensor * adamw_params = opt_ctx->opt_step_params;
+    ggml_set_input(adamw_params);
+    const char * optimizer_name = ggml_opt_optimizer_name(opt_ctx->optimizer);
+    ggml_format_name(adamw_params, "%s_params", optimizer_name);
     for (int i = opt_ctx->gf->n_nodes-1; i >= 0; --i) {
         struct ggml_tensor * node = opt_ctx->gb_opt->nodes[i];
         struct ggml_tensor * grad = ggml_graph_get_grad(opt_ctx->gb_opt, node);
 
         if (grad && (node->flags & GGML_TENSOR_FLAG_PARAM)) {
-            struct ggml_tensor * m        = opt_ctx->grad_m[i];
-            struct ggml_tensor * v        = opt_ctx->grad_v[i];
-            struct ggml_tensor * opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, opt_ctx->adamw_params);
-
-            ggml_set_name(m,        (std::string("AdamW m for ")    + std::string(node->name)).c_str());
-            ggml_set_name(v,        (std::string("AdamW v for ")    + std::string(node->name)).c_str());
-            ggml_set_name(opt_step, (std::string("AdamW step for ") + std::string(node->name)).c_str());
-
+            struct ggml_tensor * m = nullptr;
+            struct ggml_tensor * v = nullptr;
+            if (need_momenta) {
+                m = opt_ctx->grad_m[i];
+                v = opt_ctx->grad_v[i];
+                ggml_format_name(m, "AdamW m for %s", node->name);
+                ggml_format_name(v, "AdamW v for %s", node->name);
+            }
+            struct ggml_tensor * opt_step;
+            switch (optimizer) {
+                case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+                    opt_step = ggml_opt_step_adamw(opt_ctx->ctx_compute, node, grad, m, v, adamw_params);
+                    break;
+                case GGML_OPT_OPTIMIZER_TYPE_SGD:
+                    opt_step = ggml_opt_step_sgd(opt_ctx->ctx_compute, node, grad, adamw_params);
+                    break;
+                default:
+                    GGML_ABORT("fatal error");
+            }
+            ggml_format_name(opt_step, "%s step for %s", optimizer_name, node->name);
             ggml_build_forward_expand(opt_ctx->gb_opt, opt_step);
         }
     }
@@ -534,6 +558,7 @@ ggml_opt_context_t ggml_opt_init(struct ggml_opt_params params) {
     result->opt_period       = params.opt_period;
     result->get_opt_pars     = params.get_opt_pars;
     result->get_opt_pars_ud  = params.get_opt_pars_ud;
+    result->optimizer        = params.optimizer;
 
     GGML_ASSERT(result->opt_period >= 1);
 
@@ -756,29 +781,43 @@ void ggml_opt_alloc(ggml_opt_context_t opt_ctx, bool backward) {
 void ggml_opt_eval(ggml_opt_context_t opt_ctx, ggml_opt_result_t result) {
     GGML_ASSERT(opt_ctx->eval_ready);
     if (opt_ctx->allocated_graph == opt_ctx->gb_opt) {
-        struct ggml_opt_optimizer_params opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
+        const ggml_opt_optimizer_params & opt_pars = opt_ctx->get_opt_pars(opt_ctx->get_opt_pars_ud);
 
-        GGML_ASSERT(opt_pars.adamw.alpha >  0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
-        GGML_ASSERT(opt_pars.adamw.eps   >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    >= 0.0f);
-        GGML_ASSERT(opt_pars.adamw.wd    <= 1.0f);
+        switch (opt_ctx->optimizer) {
+            case GGML_OPT_OPTIMIZER_TYPE_ADAMW: {
+                GGML_ASSERT(opt_pars.adamw.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta1 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.beta2 <= 1.0f);
+                GGML_ASSERT(opt_pars.adamw.eps >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.adamw.wd <= 1.0f);
 
-        // beta1, beta2 after applying warmup
-        const float beta1h = 1.0f/(1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
-        const float beta2h = 1.0f/(1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
+                // beta1, beta2 after applying warmup
+                const float beta1h = 1.0f / (1.0f - powf(opt_pars.adamw.beta1, opt_ctx->iter));
+                const float beta2h = 1.0f / (1.0f - powf(opt_pars.adamw.beta2, opt_ctx->iter));
 
-        float * adamw_par_data = ggml_get_data_f32(opt_ctx->adamw_params);
-        adamw_par_data[0] = opt_pars.adamw.alpha;
-        adamw_par_data[1] = opt_pars.adamw.beta1;
-        adamw_par_data[2] = opt_pars.adamw.beta2;
-        adamw_par_data[3] = opt_pars.adamw.eps;
-        adamw_par_data[4] = opt_pars.adamw.wd;
-        adamw_par_data[5] = beta1h;
-        adamw_par_data[6] = beta2h;
+                float * adamw_par_data = ggml_get_data_f32(opt_ctx->opt_step_params);
+                adamw_par_data[0] = opt_pars.adamw.alpha;
+                adamw_par_data[1] = opt_pars.adamw.beta1;
+                adamw_par_data[2] = opt_pars.adamw.beta2;
+                adamw_par_data[3] = opt_pars.adamw.eps;
+                adamw_par_data[4] = opt_pars.adamw.wd;
+                adamw_par_data[5] = beta1h;
+                adamw_par_data[6] = beta2h;
+            } break;
+            case GGML_OPT_OPTIMIZER_TYPE_SGD: {
+                GGML_ASSERT(opt_pars.sgd.alpha > 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd >= 0.0f);
+                GGML_ASSERT(opt_pars.sgd.wd <= 1.0f);
+                float * sgd = ggml_get_data_f32(opt_ctx->opt_step_params);
+                sgd[0] = opt_pars.sgd.alpha;
+                sgd[1] = opt_pars.sgd.wd;
+            } break;
+            default:
+                GGML_ABORT("fatal error");
+        }
     }
 
     ggml_backend_sched_graph_compute(opt_ctx->backend_sched, opt_ctx->allocated_graph_copy);
@@ -963,6 +1002,7 @@ void ggml_opt_fit(
         ggml_tensor                   * outputs,
         ggml_opt_dataset_t              dataset,
         enum ggml_opt_loss_type         loss_type,
+        enum ggml_opt_optimizer_type    optimizer,
         ggml_opt_get_optimizer_params   get_opt_pars,
         int64_t                         nepoch,
         int64_t                         nbatch_logical,
@@ -993,6 +1033,7 @@ void ggml_opt_fit(
     params.opt_period      = opt_period;
     params.get_opt_pars    = get_opt_pars;
     params.get_opt_pars_ud = &epoch;
+    params.optimizer       = optimizer;
     ggml_opt_context_t opt_ctx = ggml_opt_init(params);
 
     // Shuffling the data is generally useful but there is only a point if not all data is used in a single batch.
@@ -1035,3 +1076,18 @@ void ggml_opt_fit(
     ggml_opt_result_free(result_train);
     ggml_opt_result_free(result_val);
 }
+
+enum ggml_opt_optimizer_type ggml_opt_context_optimizer_type(ggml_opt_context_t c) {
+    return c->optimizer;
+}
+
+GGML_API const char * ggml_opt_optimizer_name(enum ggml_opt_optimizer_type o) {
+    switch (o) {
+        case GGML_OPT_OPTIMIZER_TYPE_ADAMW:
+            return "adamw";
+        case GGML_OPT_OPTIMIZER_TYPE_SGD:
+            return "sgd";
+        default:
+            return "undefined";
+    };
+}

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

@@ -29,9 +29,12 @@
 #include <cstring>
 #include <fstream>
 #include <filesystem>
+#include <algorithm>
 
 namespace fs = std::filesystem;
 
+static constexpr size_t MAX_CHUNK_SIZE = 1024ull * 1024ull * 1024ull; // 1 GiB
+
 #ifdef _WIN32
 typedef SOCKET sockfd_t;
 using ssize_t = __int64;
@@ -323,11 +326,14 @@ static std::shared_ptr<socket_t> create_server_socket(const char * host, int por
 static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
     size_t bytes_sent = 0;
     while (bytes_sent < size) {
-        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size - bytes_sent, 0);
+        size_t size_to_send = std::min(size - bytes_sent, MAX_CHUNK_SIZE);
+        ssize_t n = send(sockfd, (const char *)data + bytes_sent, size_to_send, 0);
         if (n < 0) {
+            GGML_LOG_ERROR("send failed (bytes_sent=%zu, size_to_send=%zu)\n",
+                           bytes_sent, size_to_send);
             return false;
         }
-        bytes_sent += n;
+        bytes_sent += (size_t)n;
     }
     return true;
 }
@@ -335,11 +341,18 @@ static bool send_data(sockfd_t sockfd, const void * data, size_t size) {
 static bool recv_data(sockfd_t sockfd, void * data, size_t size) {
     size_t bytes_recv = 0;
     while (bytes_recv < size) {
-        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size - bytes_recv, 0);
-        if (n <= 0) {
+        size_t size_to_recv = std::min(size - bytes_recv, MAX_CHUNK_SIZE);
+        ssize_t n = recv(sockfd, (char *)data + bytes_recv, size_to_recv, 0);
+        if (n < 0) {
+            GGML_LOG_ERROR("recv failed (bytes_recv=%zu, size_to_recv=%zu)\n",
+                           bytes_recv, size_to_recv);
             return false;
         }
-        bytes_recv += n;
+        if (n == 0) {
+            GGML_LOG_ERROR("recv returned 0 (peer closed?)\n");
+            return false;
+        }
+        bytes_recv += (size_t)n;
     }
     return true;
 }

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

@@ -2705,9 +2705,9 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                                                 " : converting src1 to fp16");
 
         // iterate tensor dims and find the slowest moving dim and stride
-        int64_t last_dim=0;
-        int64_t last_str=0;
-        int64_t largest_str=0;
+        int last_dim=0;
+        int last_str=0;
+        size_t largest_str=0;
         for(int i = 0; i< 4; i++){
             // last stride is always the largest
             if(src1->nb[i] == largest_str){
@@ -2783,7 +2783,7 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
             auto launch_gemm_for_batches = [&ctx, queue](const sycl::half *src0,
                                                 const sycl::half *src1, float *dst,
                                                 int64_t a0, int64_t a1, int64_t batcha,
-                                                int64_t b0, int64_t b1, int64_t batchb,
+                                                int64_t /*b0*/, int64_t b1, int64_t batchb,
                                                 int64_t sa0, int64_t sa1, int64_t sa2,
                                                 int64_t sb0, int64_t sb1, int64_t sb2,
                                                 int64_t sd2) {
@@ -2832,14 +2832,26 @@ static void ggml_sycl_mul_mat_batched_sycl(ggml_backend_sycl_context & ctx, cons
                 }
             };
 
-            bool cont_batches_a = nb02 * ne02 == nb03;
-            bool cont_batches_b = nb12 * ne12 == nb13;
-            if (cont_batches_a && cont_batches_b) {
+            const bool cont_batches_dim2_a = nb02 * ne02 == nb03;
+            const bool cont_batches_dim2_b = nb12 * ne12 == nb13;
+            const bool cont_batches_dim3_a = ne02 == 1 && nb02 * ne01 == nb03;
+            const bool cont_batches_dim3_b = ne12 == 1 && nb12 * ne11 == nb13;
+            if (cont_batches_dim2_a && cont_batches_dim2_b) {
+                // A batch is considered contiguous if the dimension 2 is not strided
                 int64_t batches0 = ne02 * ne03;
                 int64_t batches1 = ne12 * ne13;
                 launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
                         ne10, ne11, batches1, str_a0, str_a1, str_a2, str_b0, str_b1,
                         str_b2, nb2 / sizeof(float));
+            } else if (cont_batches_dim3_a && cont_batches_dim3_b) {
+                // This case is similar to the one above with the difference that only the batch in dimension 3 is used and the dimension 2 is of size 1.
+                int64_t batches0 = ne02 * ne03;
+                int64_t batches1 = ne12 * ne13;
+                int64_t str_a3 = nb03 / type_size_src0;
+                int64_t str_b3 = nb13 / type_size_src1;
+                launch_gemm_for_batches(src0_f16, src1_f16, dst_ddf, ne00, ne01, batches0,
+                        ne10, ne11, batches1, str_a0, str_a1, str_a3, str_b0, str_b1,
+                        str_b3, nb2 / sizeof(float));
             } else {
                 for (int64_t b_a = 0; b_a < ne03; b_a++) {
                     const sycl::half *src0_f16_shifted
@@ -4215,6 +4227,15 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
                     // FIXME: keep a list of supported types to avoid breaking the backend when a new type is added
                     return false;
                 }
+                // TODO: The configuration below needs more work to be supported with oneDNN
+                if (ggml_is_permuted(a) && !ggml_is_contiguous(a) && a->ne[2] > 1 && a->ne[3] > 1) {
+                    return false;
+                }
+                // TODO: This specific configuration can fail with oneDNN and needs more debugging
+                if (!ggml_is_permuted(a) && ggml_is_permuted(b) && b->ne[2] > 1 && b->ne[3] > 1 &&
+                    a->ne[0] > 128 && a->ne[2] == 1 && src0_type == GGML_TYPE_F16) {
+                    return false;
+                }
                 return true;
             }
         case GGML_OP_OUT_PROD:

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

@@ -103,6 +103,8 @@ static bool is_pow2(uint32_t x) { return x > 1 && (x & (x-1)) == 0; }
 struct ggml_backend_vk_context;
 
 #define MAX_PARAMETER_COUNT 8
+// Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
+#define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 2)
 
 struct vk_pipeline_struct {
     std::string name;
@@ -368,6 +370,7 @@ struct vk_device_struct {
     bool float_controls_rte_fp16;
     bool subgroup_add;
     bool subgroup_shuffle;
+    bool multi_add;
 
     bool integer_dot_product;
 
@@ -449,6 +452,9 @@ struct vk_device_struct {
     vk_pipeline pipeline_div[2][2][2];
     vk_pipeline pipeline_div_norepeat[2][2][2];
 
+    // indexed by num_additional_fused_ops == num_adds - 1
+    vk_pipeline pipeline_multi_add[MAX_FUSED_ADDS];
+
     vk_pipeline pipeline_add_id_f32;
 
     vk_pipeline pipeline_concat_f32, pipeline_concat_f16, pipeline_concat_i32;
@@ -510,6 +516,7 @@ struct vk_device_struct {
     vk_pipeline pipeline_rwkv_wkv6_f32;
     vk_pipeline pipeline_rwkv_wkv7_f32;
     vk_pipeline pipeline_opt_step_adamw_f32;
+    vk_pipeline pipeline_opt_step_sgd_f32;
     vk_pipeline pipeline_conv2d_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_f16_f32[CONV_SHAPE_COUNT];
     vk_pipeline pipeline_conv2d_dw_whcn_f32;
@@ -800,6 +807,14 @@ struct vk_op_binary_push_constants {
     float param1; float param2; int32_t param3;
 };
 
+struct vk_op_multi_add_push_constants {
+    // shape for dst
+    uint32_t ne20; uint32_t ne21; uint32_t ne22; uint32_t ne23;
+
+    // strides for srcs+dst
+    uint32_t nb[8][4];
+};
+
 struct vk_op_add_id_push_constants {
     uint32_t ne0;
     uint32_t ne1;
@@ -1110,17 +1125,23 @@ class vk_perf_logger {
             return;
         }
         if (node->op == GGML_OP_MUL_MAT || node->op == GGML_OP_MUL_MAT_ID) {
-            const uint64_t m    = node->src[0]->ne[1];
-            const uint64_t n    = node->src[1]->ne[1];
-            const uint64_t k    = node->src[1]->ne[0];
-            std::string    name = ggml_op_name(node->op);
-            if (n == 1) {
-                name += "_VEC m=" + std::to_string(m) + " k=" + std::to_string(k);
-            } else {
-                name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            const uint64_t m     = node->src[0]->ne[1];
+            const uint64_t n     = node->ne[1];
+            const uint64_t k     = node->src[1]->ne[0];
+            const uint64_t batch = node->src[1]->ne[2] * node->src[1]->ne[3];
+            std::string    name  = ggml_op_name(node->op);
+            if ((node->op == GGML_OP_MUL_MAT && n <= mul_mat_vec_max_cols) ||
+                (node->op == GGML_OP_MUL_MAT_ID && node->src[2]->ne[1] == 1)) {
+                name += "_VEC";
+            }
+            name += " ";
+            name += ggml_type_name(node->src[0]->type);
+            name += " m=" + std::to_string(m) + " n=" + std::to_string(n) + " k=" + std::to_string(k);
+            if (batch > 1) {
+                name += " batch=" + std::to_string(batch);
             }
             timings[name].push_back(time);
-            flops[name].push_back(m * n * (k + (k - 1)));
+            flops[name].push_back(m * n * (k + (k - 1)) * batch);
             return;
         }
         if (node->op == GGML_OP_CONV_2D) {
@@ -2380,26 +2401,26 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4)
         }
 #endif
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f16, , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f16,  , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
-        CREATE_MM(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f16,   , mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f16, mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f16,  mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
+        CREATE_MM2(pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f16,   mmqid_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, 4)
 #undef CREATE_MM
 #undef CREATE_MM2
     } else
@@ -2495,51 +2516,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
         }
 #endif
 
-        if (device->coopmat_acc_f16_support) {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        } else {
-            CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-
-            CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-            CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   , mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        }
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MM
     } else
@@ -2624,27 +2621,27 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, 4, _id);
 
-        CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f16acc, matmul_id_q4_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1].f16acc, matmul_id_q4_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0].f16acc, matmul_id_q5_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1].f16acc, matmul_id_q5_1_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0].f16acc, matmul_id_q8_0_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_q4_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_1], matmul_id_q4_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_0], matmul_id_q5_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_1, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_1], matmul_id_q5_1_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q8_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q8_0], matmul_id_q8_0_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 
-        CREATE_MM(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K].f16acc, matmul_id_q2_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K].f16acc, matmul_id_q3_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K].f16acc, matmul_id_q4_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K].f16acc, matmul_id_q5_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K].f16acc, matmul_id_q6_k_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S].f16acc,   matmul_id_iq1_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M].f16acc,   matmul_id_iq1_m_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS].f16acc, matmul_id_iq2_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS].f16acc,  matmul_id_iq2_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S].f16acc,   matmul_id_iq2_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS].f16acc, matmul_id_iq3_xxs_f32, _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S].f16acc,   matmul_id_iq3_s_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS].f16acc,  matmul_id_iq4_xs_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f16acc,  matmul_id_iq4_nl_f32,  _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
-        CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f16acc,   matmul_id_mxfp4_f32,   _f16acc, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q2_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q2_K], matmul_id_q2_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q3_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q3_K], matmul_id_q3_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q4_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_K], matmul_id_q4_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q5_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q5_K], matmul_id_q5_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_Q6_K, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q6_K], matmul_id_q6_k_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_S],   matmul_id_iq1_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ1_M,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ1_M],   matmul_id_iq1_m_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XXS], matmul_id_iq2_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_XS],  matmul_id_iq2_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ2_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ2_S],   matmul_id_iq2_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_XXS, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_XXS], matmul_id_iq3_xxs_f32, mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ3_S,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ3_S],   matmul_id_iq3_s_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_XS,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_XS],  matmul_id_iq4_xs_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL],  matmul_id_iq4_nl_f32,  mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
+        CREATE_MM2(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4],   matmul_id_mxfp4_f32,   mmq_wg_denoms, warptile_mmq, vk_mat_mat_id_push_constants, 4, _id);
 #undef CREATE_MM2
 #undef CREATE_MMQ
 #undef CREATE_MM
@@ -3011,6 +3008,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
     CREATE_BINARY(div, _norepeat, {1})
 #undef CREATE_BINARY
 
+    if (device->multi_add) {
+        for (uint32_t i = 0; i < MAX_FUSED_ADDS; ++i) {
+            ggml_vk_create_pipeline(device, device->pipeline_multi_add[i], "multi_add_f32_" + std::to_string(i+1), multi_add_f32_len, multi_add_f32_data, "main", MAX_PARAMETER_COUNT, sizeof(vk_op_multi_add_push_constants), {512, 1, 1}, {i+2}, 1);
+        }
+    }
+
     ggml_vk_create_pipeline(device, device->pipeline_add_id_f32, "add_id_f32", add_id_f32_len, add_id_f32_data, "main", 4, sizeof(vk_op_add_id_push_constants), {1, 1, 1}, {}, 1);
 
     ggml_vk_create_pipeline(device, device->pipeline_acc_f32, "acc_f32", acc_f32_len, acc_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
@@ -3123,6 +3126,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
     ggml_vk_create_pipeline(device, device->pipeline_opt_step_adamw_f32, "opt_step_adamw_f32", opt_step_adamw_f32_len, opt_step_adamw_f32_data, "main", 5, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
+    ggml_vk_create_pipeline(device, device->pipeline_opt_step_sgd_f32, "opt_step_sgd_f32", opt_step_sgd_f32_len, opt_step_sgd_f32_data, "main", 3, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
+
     // conv2d
     for (uint32_t s = 0; s < CONV_SHAPE_COUNT; ++s) {
         uint32_t conv2d_WG_SIZE  = 256;
@@ -3548,6 +3553,12 @@ static vk_device ggml_vk_get_device(size_t idx) {
 
         device->pipeline_robustness = pl_robustness_features.pipelineRobustness;
 
+        device->multi_add = vk12_props.shaderRoundingModeRTEFloat16 &&
+                            device->properties.limits.maxPushConstantsSize >= sizeof(vk_op_multi_add_push_constants) &&
+                            vk12_features.runtimeDescriptorArray &&
+                            device->vendor_id != VK_VENDOR_ID_INTEL &&
+                            getenv("GGML_VK_DISABLE_MULTI_ADD") == nullptr;
+
         if (device->subgroup_size_control) {
             device->subgroup_min_size = subgroup_size_control_props.minSubgroupSize;
             device->subgroup_max_size = subgroup_size_control_props.maxSubgroupSize;
@@ -4461,7 +4472,17 @@ static vk_matmul_pipeline ggml_vk_get_mul_mat_mat_id_pipeline(ggml_backend_vk_co
             return nullptr;
     }
 
-    return ctx->device->fp16 ? ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc : ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    // XXX TODO 'prec' is not actually allowed in mul_mat_id.
+    bool prefer_fp16acc = ctx->device->fp16 /*&& prec == GGML_PREC_DEFAULT*/;
+    bool support_fp16acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc != nullptr;
+    bool support_fp32acc = ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc != nullptr;
+
+    if (support_fp16acc && (prefer_fp16acc || !support_fp32acc)) {
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f16acc;
+    } else {
+        GGML_ASSERT(support_fp32acc);
+        return ctx->device->pipeline_dequant_mul_mat_mat_id[src0_type].f32acc;
+    }
 }
 
 static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec_id(ggml_backend_vk_context * ctx, ggml_type a_type, ggml_type b_type) {
@@ -6892,6 +6913,9 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
         switch (op) {
         case GGML_OP_ADD:
         {
+            if (ctx->num_additional_fused_ops > 0) {
+                return ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+            }
             auto pipelines = ggml_are_same_shape(src0, src1) ? ctx->device->pipeline_add_norepeat : ctx->device->pipeline_add;
             return pipelines[src0->type == GGML_TYPE_F16][src1->type == GGML_TYPE_F16][dst->type == GGML_TYPE_F16];
         }
@@ -7193,6 +7217,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
             return ctx->device->pipeline_opt_step_adamw_f32;
         }
         return nullptr;
+    case GGML_OP_OPT_STEP_SGD:
+        if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
+            return ctx->device->pipeline_opt_step_sgd_f32;
+        }
+        return nullptr;
     case GGML_OP_LEAKY_RELU:
         if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
             return ctx->device->pipeline_leaky_relu_f32;
@@ -7692,6 +7721,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
         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 } }, pc, elements);
+    } else if (op == GGML_OP_OPT_STEP_SGD) {
+        // OPT_STEP_SGD works on src0, it does not need dst
+        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 } }, 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 } }, pc, elements);
@@ -7739,6 +7772,107 @@ static void ggml_vk_acc(ggml_backend_vk_context * ctx, vk_context& subctx, const
     }, dryrun);
 }
 
+static void ggml_vk_multi_add(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_cgraph * cgraph, int node_idx, bool dryrun = false) {
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    const ggml_tensor *dst = cgraph->nodes[node_idx + ctx->num_additional_fused_ops];
+
+    // Make a list of all the tensors used by the op.
+    // Last element of the list is the dest tensor.
+    const ggml_tensor *tensors[MAX_PARAMETER_COUNT];
+    uint32_t num_srcs = ctx->num_additional_fused_ops + 2;
+    uint32_t num_tensors = num_srcs + 1;
+    GGML_ASSERT(num_tensors <= MAX_PARAMETER_COUNT);
+
+    tensors[0] = first_node->src[0];
+    tensors[1] = first_node->src[1];
+    for (int32_t i = 0; i < ctx->num_additional_fused_ops; ++i) {
+        // check whether the previous result is src[0] or src[1]
+        if (cgraph->nodes[node_idx + i] == cgraph->nodes[node_idx + i + 1]->src[0]) {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[1];
+        } else {
+            tensors[i+2] = cgraph->nodes[node_idx + i + 1]->src[0];
+        }
+    }
+    tensors[num_srcs] = dst;
+
+    vk_op_multi_add_push_constants pc;
+    pc.ne20 = (uint32_t)dst->ne[0];
+    pc.ne21 = (uint32_t)dst->ne[1];
+    pc.ne22 = (uint32_t)dst->ne[2];
+    pc.ne23 = (uint32_t)dst->ne[3];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        const ggml_tensor *t = tensors[i];
+        pc.nb[i][0] = (uint32_t)t->nb[0] / sizeof(float);
+        pc.nb[i][1] = (uint32_t)t->nb[1] / sizeof(float);
+        pc.nb[i][2] = (uint32_t)t->nb[2] / sizeof(float);
+        pc.nb[i][3] = (uint32_t)t->nb[3] / sizeof(float);
+    }
+
+    vk_pipeline pipeline = ctx->device->pipeline_multi_add[ctx->num_additional_fused_ops];
+
+    if (pipeline == nullptr) {
+        std::cerr << "ggml_vulkan: Error: Missing multi_add";
+        GGML_ABORT("fatal error");
+    }
+
+    if (dryrun) {
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
+        return;
+    }
+
+    ggml_backend_vk_buffer_context * buf_ctx[MAX_PARAMETER_COUNT];
+    vk_buffer buf[MAX_PARAMETER_COUNT];
+    size_t offset[MAX_PARAMETER_COUNT];
+    bool uma[MAX_PARAMETER_COUNT];
+
+    for (uint32_t i = 0; i < num_tensors; ++i) {
+        buf_ctx[i] = (ggml_backend_vk_buffer_context *)tensors[i]->buffer->context;
+        buf[i] = nullptr;
+        offset[i] = 0;
+        uma[i] = false;
+
+        if (ctx->device->uma) {
+            ggml_vk_host_get(ctx->device, tensors[i]->data, buf[i], offset[i]);
+            uma[i] = buf[i] != nullptr;
+        }
+        if (!uma[i]) {
+            buf[i] = buf_ctx[i]->dev_buffer;
+            offset[i] = vk_tensor_offset(tensors[i]) + tensors[i]->view_offs;
+        }
+        GGML_ASSERT(buf[i] != nullptr);
+    }
+    // If any remaining descriptors are unused, just point them at src[0]
+    for (uint32_t i = num_tensors; i < MAX_PARAMETER_COUNT; ++i) {
+        buf[i] = buf[0];
+        offset[i] = 0;
+    }
+
+    std::array<uint32_t, 3> elements;
+
+    uint32_t ne = ggml_nelements(dst);
+    if (ne > 262144) {
+        elements = { 512, 512, CEIL_DIV(ne, 262144) };
+    } else if (ne > 512) {
+        elements = { 512, CEIL_DIV(ne, 512), 1 };
+    } else {
+        elements = { ne, 1, 1 };
+    }
+
+    ggml_vk_sync_buffers(subctx);
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
+        {
+            vk_subbuffer{ buf[0], offset[0], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[1], offset[1], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[2], offset[2], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[3], offset[3], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[4], offset[4], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[5], offset[5], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[6], offset[6], VK_WHOLE_SIZE },
+            vk_subbuffer{ buf[7], offset[7], VK_WHOLE_SIZE },
+        }, pc, elements);
+}
+
 static void ggml_vk_add(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     const uint32_t src0_type_size = ggml_type_size(src0->type);
     const uint32_t src1_type_size = ggml_type_size(src1->type);
@@ -8045,6 +8179,12 @@ static void ggml_vk_opt_step_adamw(ggml_backend_vk_context * ctx, vk_context& su
     );
 }
 
+static void ggml_vk_opt_step_sgd(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) {
+    const size_t n = ggml_nelements(dst->src[0]);
+
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, src1, src2, dst, GGML_OP_OPT_STEP_SGD, { (uint32_t)n, 0, 0.0f, 0.0f }, dryrun);
+}
+
 static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
     int * op_params = (int *)dst->op_params;
 
@@ -8336,7 +8476,7 @@ static void ggml_vk_rope(ggml_backend_vk_context * ctx, vk_context& subctx, cons
         (uint32_t)src0->ne[0], (uint32_t)n_dims, freq_scale, (uint32_t)src0->ne[1],
         freq_base, ext_factor, attn_factor, {corr_dims[0], corr_dims[1]}, theta_scale,
         src2 != nullptr, (uint32_t)src0->ne[2], s1, s2,
-        sections[0], sections[1], sections[2], sections[3], backprop
+        { sections[0], sections[1], sections[2], sections[3] }, backprop
     }, dryrun);
 }
 
@@ -8368,7 +8508,7 @@ static void ggml_vk_sum_rows(ggml_backend_vk_context * ctx, vk_context& subctx,
 }
 
 static void ggml_vk_argmax(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
-    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], 0, 0.0f, 0.0f }, dryrun);
+    ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_ARGMAX, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], 0.0f, 0.0f }, dryrun);
 }
 
 static void ggml_vk_count_equal(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@@ -9598,11 +9738,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_LEAKY_RELU:
     case GGML_OP_FLASH_ATTN_EXT:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         break;
     default:
         std::cerr << "ggml_vulkan: Error: Missing op: " << ggml_op_name(node->op) << std::endl;
         GGML_ABORT("fatal error");
-        return false;
     }
 
     vk_context compute_ctx;
@@ -9662,6 +9802,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
         case GGML_OP_CONV_2D:
         case GGML_OP_CONV_2D_DW:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_SGD:
             {
                 // These operations all go through ggml_vk_op_f32, so short-circuit and
                 // do the only thing needed for the dryrun.
@@ -9692,8 +9833,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
 
         break;
     case GGML_OP_ADD:
-        ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
-
+        if (ctx->num_additional_fused_ops) {
+            ggml_vk_multi_add(ctx, compute_ctx, cgraph, node_idx, dryrun);
+        } else {
+            ggml_vk_add(ctx, compute_ctx, src0, src1, node, dryrun);
+        }
         break;
     case GGML_OP_SUB:
         ggml_vk_sub(ctx, compute_ctx, src0, src1, node, dryrun);
@@ -9911,6 +10055,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_cgraph * cgr
     case GGML_OP_OPT_STEP_ADAMW:
         ggml_vk_opt_step_adamw(ctx, compute_ctx, node, dryrun);
 
+        break;
+
+    case GGML_OP_OPT_STEP_SGD:
+        ggml_vk_opt_step_sgd(ctx, compute_ctx, src0, src1, src2, node, dryrun);
+
         break;
     default:
         return false;
@@ -10014,8 +10163,8 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_cgraph *
     case GGML_OP_REPEAT:
     case GGML_OP_REPEAT_BACK:
     case GGML_OP_OPT_STEP_ADAMW:
+    case GGML_OP_OPT_STEP_SGD:
         buf = tensor->buffer;
-
         break;
     case GGML_OP_UNARY:
         switch (ggml_get_unary_op(tensor)) {
@@ -10570,6 +10719,58 @@ static bool ggml_vk_can_fuse(const struct ggml_cgraph * cgraph, int node_idx, st
     return true;
 }
 
+static uint32_t ggml_vk_fuse_multi_add(ggml_backend_vk_context * ctx, const struct ggml_cgraph * cgraph, int node_idx) {
+
+    const ggml_tensor *first_node = cgraph->nodes[node_idx];
+    if (first_node->op != GGML_OP_ADD) {
+        return 0;
+    }
+
+    if (!ctx->device->multi_add) {
+        return 0;
+    }
+
+    int32_t num_adds = 1;
+    while (node_idx + num_adds < cgraph->n_nodes &&
+           cgraph->nodes[node_idx + num_adds]->op == GGML_OP_ADD &&
+           num_adds < MAX_FUSED_ADDS) {
+        num_adds++;
+    }
+
+    // The shader currently requires same shapes (but different strides are allowed),
+    // everything f32, and no misalignment
+    for (int32_t i = 0; i < num_adds; ++i) {
+        const ggml_tensor *next_node = cgraph->nodes[node_idx + i];
+        if (!ggml_are_same_shape(first_node, next_node->src[0]) ||
+            !ggml_are_same_shape(first_node, next_node->src[1]) ||
+            next_node->type != GGML_TYPE_F32 ||
+            next_node->src[0]->type != GGML_TYPE_F32 ||
+            next_node->src[1]->type != GGML_TYPE_F32 ||
+            get_misalign_bytes(ctx, next_node) ||
+            get_misalign_bytes(ctx, next_node->src[0]) ||
+            get_misalign_bytes(ctx, next_node->src[1])) {
+            num_adds = i;
+        }
+    }
+
+    // Verify we can fuse these
+    ggml_op adds[MAX_FUSED_ADDS];
+    for (int32_t i = 0; i < num_adds; ++i) {
+        adds[i] = GGML_OP_ADD;
+    }
+
+    // decrease num_adds if they can't all be fused
+    while (num_adds > 1 && !ggml_can_fuse(cgraph, node_idx, adds, num_adds)) {
+        num_adds--;
+    }
+
+    // a single add is not "fused", so just return zero
+    if (num_adds == 1) {
+        return 0;
+    }
+    return num_adds;
+}
+
 static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
     VK_LOG_DEBUG("ggml_backend_vk_graph_compute(" << cgraph->n_nodes << " nodes)");
     ggml_backend_vk_context * ctx = (ggml_backend_vk_context *)backend->context;
@@ -10583,8 +10784,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
 
     uint64_t total_mat_mul_bytes = 0;
     for (int i = 0; i < cgraph->n_nodes; i++) {
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
         ggml_vk_build_graph(ctx, cgraph, i, nullptr, 0, true, false, false, false);
         if (cgraph->nodes[i]->op == GGML_OP_MUL_MAT || cgraph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
@@ -10659,8 +10865,13 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg
             mul_mat_bytes += ggml_nbytes(cgraph->nodes[i]->src[0]);
         }
 
-        if (!ctx->device->disable_fusion && ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
-            ctx->num_additional_fused_ops = 1;
+        if (!ctx->device->disable_fusion) {
+            uint32_t num_adds = ggml_vk_fuse_multi_add(ctx, cgraph, i);
+            if (num_adds) {
+                ctx->num_additional_fused_ops = num_adds - 1;
+            } else if (ggml_vk_can_fuse(cgraph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
+                ctx->num_additional_fused_ops = 1;
+            }
         }
 
         // Signal the almost_ready fence when the graph is mostly complete (< 20% remaining)
@@ -10881,7 +11092,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_GLU:
             switch (ggml_get_glu_op(op)) {
                 case GGML_GLU_OP_GEGLU:
@@ -10897,7 +11107,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 default:
                     return false;
             }
-            break;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
             {
@@ -10961,7 +11170,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 }
 
                 return true;
-            } break;
+            }
         case GGML_OP_FLASH_ATTN_EXT:
             {
                 ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
@@ -11051,7 +11260,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_SET_ROWS:
             {
                 switch (op->type) {
@@ -11068,7 +11277,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     default:
                         return false;
                 }
-            } break;
+            }
         case GGML_OP_CONT:
         case GGML_OP_CPY:
         case GGML_OP_DUP:
@@ -11120,7 +11329,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return true;
                 }
                 return false;
-            } break;
+            }
         case GGML_OP_REPEAT:
             return ggml_type_size(op->type) == sizeof(float) && ggml_type_size(op->src[0]->type) == sizeof(float);
         case GGML_OP_REPEAT_BACK:
@@ -11154,6 +11363,9 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_SIN:
         case GGML_OP_COS:
         case GGML_OP_CLAMP:
+        case GGML_OP_LEAKY_RELU:
+        case GGML_OP_OPT_STEP_ADAMW:
+        case GGML_OP_OPT_STEP_SGD:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_UPSCALE:
         case GGML_OP_ACC:
@@ -11175,8 +11387,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_POOL_2D:
         case GGML_OP_RWKV_WKV6:
         case GGML_OP_RWKV_WKV7:
-        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;
@@ -11694,6 +11904,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         } else {
             tensor_clone = ggml_glu_split(ggml_ctx, src_clone[0], src_clone[1], (ggml_glu_op) tensor->op_params[0]);
         }
+        ggml_set_op_params_i32(tensor_clone, 2, ggml_get_op_params_i32(tensor, 2));
+        ggml_set_op_params_i32(tensor_clone, 3, ggml_get_op_params_i32(tensor, 3));
     } else if (tensor->op == GGML_OP_CPY || tensor->op == GGML_OP_DUP) {
         if (src1 == nullptr) {
             tensor_clone = ggml_dup(ggml_ctx, src_clone[0]);
@@ -11774,6 +11986,12 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_cgraph *
         src_clone[0]->flags = src0->flags;
         tensor_clone = ggml_opt_step_adamw(ggml_ctx, src_clone[0], src_clone[1],
         src_clone[2], src_clone[3], src_clone[4]);
+    } else if (tensor->op == GGML_OP_OPT_STEP_SGD) {
+        src_clone[0]->flags = src0->flags;
+        tensor_clone = ggml_opt_step_sgd(ggml_ctx, src_clone[0], src_clone[1],
+        src_clone[2]);
+    } else if (tensor->op == GGML_OP_ADD_ID) {
+        tensor_clone = ggml_add_id(ggml_ctx, src_clone[0], src_clone[1], src_clone[2]);
     }
     else {
         std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;

ggml/src/ggml-vulkan/vulkan-shaders/argmax.comp

@@ -5,6 +5,8 @@
 
 #extension GL_EXT_control_flow_attributes : enable
 
+#define FLT_MAX 3.402823466e+38F
+
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 layout (binding = 0) readonly buffer A {A_TYPE data_a[];};
@@ -19,19 +21,26 @@ void main() {
     const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
     const uint col = gl_LocalInvocationID.x;
 
-    if (col >= p.KX) {
+    if (row >= p.KY) {
         return;
     }
-    A_TYPE amax = data_a[row*p.KX + col];
-    tmp[col] = col;
+
+    A_TYPE amax = -FLT_MAX;
+    uint acol = col;
+
+    if (col < p.KX) {
+        amax = data_a[row*p.KX + col];
+    }
 
     for (uint i = col + BLOCK_SIZE; i < p.KX; i += BLOCK_SIZE) {
         A_TYPE val = data_a[row*p.KX + i];
         if (val > amax) {
             amax = val;
-            tmp[col] = i;
+            acol = i;
         }
     }
+
+    tmp[col] = acol;
     tmpmax[col] = amax;
 
     barrier();

ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_cm1.comp

@@ -210,7 +210,7 @@ void main() {
 
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+                Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             }
         }
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
@@ -233,7 +233,7 @@ void main() {
                 vec4 Vf = vec4(data_vv4[v_offset / 4 + (j * Bc + c * cols_per_iter + col_tid) * v_stride / 4 + d * D_split + d_tid]);
 #endif
                 [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-                    Of[r][d] += float16_t(Pf[r]) * ACC_TYPEV4(Vf);
+                    Of[r][d] += ACC_TYPE(Pf[r]) * ACC_TYPEV4(Vf);
                 }
             }
         }
@@ -288,7 +288,7 @@ void main() {
     [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
         [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
-            Of[r][d] = float16_t(eMf[r]) * Of[r][d];
+            Of[r][d] = ACC_TYPE(eMf[r]) * Of[r][d];
             tmpshv4[tid] = Of[r][d];
 
             barrier();
@@ -357,7 +357,7 @@ void main() {
 
     [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < rows_per_thread; ++r) {
-            Of[r][d] *= float16_t(Lfrcp[r]);
+            Of[r][d] *= ACC_TYPE(Lfrcp[r]);
         }
     }
 

ggml/src/ggml-vulkan/vulkan-shaders/generic_binary_head.comp

@@ -2,6 +2,7 @@
 #extension GL_EXT_control_flow_attributes : require
 
 #include "rte.comp"
+#include "utils.comp"
 
 layout (push_constant) uniform parameter
 {
@@ -28,25 +29,9 @@ uint get_aoffset() { return p.misalign_offsets >> 16; }
 uint get_boffset() { return (p.misalign_offsets >> 8) & 0xFF; }
 uint get_doffset() { return p.misalign_offsets & 0xFF; }
 
-// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
-uint fastmod(uint a, uint b) {
-    if ((b & (b-1)) == 0) {
-        return a & (b-1);
-    }
-    return a % b;
-}
-
-uint fastdiv(uint a, uint b) {
-    return (a < b) ? 0 : (a / b);
-}
 
 void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03) {
-    i03 = fastdiv(idx, (p.ne02*p.ne01*p.ne00));
-    const uint i03_offset = i03 * p.ne02*p.ne01*p.ne00;
-    i02 = fastdiv((idx - i03_offset), (p.ne01*p.ne00));
-    const uint i02_offset = i02*p.ne01*p.ne00;
-    i01 = (idx - i03_offset - i02_offset) / p.ne00;
-    i00 = idx - i03_offset - i02_offset - i01*p.ne00;
+    get_indices(idx, i00, i01, i02, i03, p.ne00, p.ne01, p.ne02, p.ne03);
 }
 
 uint src0_idx(uint i00, uint i01, uint i02, uint i03) {

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp

@@ -801,7 +801,7 @@ void main() {
             }
 #else
             const uint row_i = ic * BN + loadc_b + l;
-            if (row_i < _ne1) {
+            if (row_i < _ne1 && block + loadr_b < end_k) {
                 const u16vec2 row_idx = row_ids[row_i];
                 buf_b[(loadc_b + l) * SHMEM_STRIDE + loadr_b] = TO_FLOAT_TYPE(data_b[pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + loadr_b]);
             } else {
@@ -875,7 +875,9 @@ void main() {
 
                 const u16vec2 row_idx = row_ids[row_i];
 
-                data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                if (dr + cm_row * TM + store_r < p.M) {
+                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr + cm_row * TM + store_r] = D_TYPE(coopmat_stage[warp_i * TM * TN + (col + store_c) * TM + store_r]);
+                }
             }
         }
     }
@@ -925,7 +927,9 @@ void main() {
 #endif // MUL_MAT_ID
                 [[unroll]] for (uint cr = 0; cr < TM; cr++) {
 #ifdef MUL_MAT_ID
-                    data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    if (dr_warp + cr < p.M) {
+                        data_d[row_idx.y * p.batch_stride_d + row_idx.x * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);
+                    }
 #else
                     if (dr_warp + cr < p.M && dc_warp + cc < p.N) {
                         data_d[offsets + (dc_warp + cc) * p.stride_d + dr_warp + cr] = D_TYPE(sums[(wsic * TN + cc) * (WMITER * TM) + wsir * TM + cr]);

ggml/src/ggml-vulkan/vulkan-shaders/multi_add.comp

@@ -0,0 +1,68 @@
+#version 450
+
+#extension GL_EXT_shader_16bit_storage : require
+#extension GL_EXT_nonuniform_qualifier : enable
+#extension GL_EXT_control_flow_attributes : require
+
+#include "rte.comp"
+#include "types.comp"
+#include "utils.comp"
+
+layout (push_constant) uniform parameter2
+{
+    // shape for dst
+    uint ne20; uint ne21; uint ne22; uint ne23;
+
+    // strides for srcs+dst
+    uint nb[8][4];
+} p;
+
+layout (binding = 0) readonly buffer A {A_TYPE data_a[];} a[];
+layout (binding = 0) writeonly buffer D {D_TYPE data_d[];} d[];
+
+layout(constant_id = 0) const uint num_srcs = 2;
+
+uint src_idx(uint s, uint i00, uint i01, uint i02, uint i03) {
+    return i03*p.nb[s][3] + i02*p.nb[s][2] + i01*p.nb[s][1] + i00*p.nb[s][0];
+}
+
+uint dst_idx(uint i00, uint i01, uint i02, uint i03) {
+    uint nb20 = p.nb[num_srcs][0];
+    uint nb21 = p.nb[num_srcs][1];
+    uint nb22 = p.nb[num_srcs][2];
+    uint nb23 = p.nb[num_srcs][3];
+    return i03*nb23 + i02*nb22 + i01*nb21 + i00*nb20;
+}
+
+uint get_idx() {
+    return gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+}
+
+const uint num_threads = 256;
+
+layout(local_size_x = num_threads, local_size_y = 1, local_size_z = 1) in;
+
+void main() {
+    uint idx = get_idx();
+
+    uint ne = p.ne20 * p.ne21 * p.ne22 * p.ne23;
+
+    // num_threads * num_iter must equal 512, to match the wg_denoms and get_idx calculation
+    const uint num_iter = 2;
+
+    [[unroll]] for (uint i = 0; i < num_iter; ++i) {
+        if (idx >= ne) {
+            continue;
+        }
+        uint i00, i01, i02, i03;
+        get_indices(idx, i00, i01, i02, i03, p.ne20, p.ne21, p.ne22, p.ne23);
+
+        FLOAT_TYPE sum = FLOAT_TYPE(0);
+        [[unroll]] for (uint s = 0; s < num_srcs; ++s) {
+            sum += FLOAT_TYPE(a[s].data_a[src_idx(s, i00, i01, i02, i03)]);
+        }
+        d[num_srcs].data_d[dst_idx(i00, i01, i02, i03)] = D_TYPE(sum);
+
+        idx += num_threads;
+    }
+}

ggml/src/ggml-vulkan/vulkan-shaders/opt_step_sgd.comp

@@ -0,0 +1,22 @@
+#version 450
+
+#include "generic_head.comp"
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) buffer X {A_TYPE data_x[];};
+layout (binding = 1) readonly buffer G {A_TYPE data_grad[];};
+layout (binding = 2) readonly buffer P {float data_params[2];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    const float alpha = data_params[0];
+    const float keep = 1.f - alpha * data_params[1];
+
+    data_x[i] = data_x[i] * keep - alpha * data_grad[i];
+}

ggml/src/ggml-vulkan/vulkan-shaders/utils.comp

@@ -0,0 +1,25 @@
+#ifndef UTILS_COMP
+#define UTILS_COMP
+
+// mod and div are expensive and coordinates/dimensions are often power of 2 or equal to 1
+uint fastmod(uint a, uint b) {
+    if ((b & (b-1)) == 0) {
+        return a & (b-1);
+    }
+    return a % b;
+}
+
+uint fastdiv(uint a, uint b) {
+    return (a < b) ? 0 : (a / b);
+}
+
+void get_indices(uint idx, out uint i00, out uint i01, out uint i02, out uint i03, uint ne00, uint ne01, uint ne02, uint ne03) {
+    i03 = fastdiv(idx, (ne02*ne01*ne00));
+    const uint i03_offset = i03 * ne02*ne01*ne00;
+    i02 = fastdiv((idx - i03_offset), (ne01*ne00));
+    const uint i02_offset = i02*ne01*ne00;
+    i01 = (idx - i03_offset - i02_offset) / ne00;
+    i00 = idx - i03_offset - i02_offset - i01*ne00;
+}
+
+#endif // UTILS_COMP

ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp

@@ -657,6 +657,7 @@ void process_shaders() {
     string_to_spv("rwkv_wkv7_f32", "wkv7.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("opt_step_adamw_f32", "opt_step_adamw.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
+    string_to_spv("opt_step_sgd_f32", "opt_step_sgd.comp", merge_maps(base_dict, {{"A_TYPE", "float"}}));
 
     string_to_spv("conv2d_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
     string_to_spv("conv2d_f16_f32_unroll", "conv2d_mm.comp", {{"A_TYPE", "float16_t"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"USE_COLLECTIVES", "1"}, {"UNROLL", "[[unroll]]"}});
@@ -676,6 +677,8 @@ void process_shaders() {
 
     string_to_spv("add_id_f32", "add_id.comp", merge_maps(base_dict, {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}}));
 
+    string_to_spv("multi_add_f32", "multi_add.comp", {{"A_TYPE", "float"}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}, {"RTE16", "1"}});
+
     for (auto &c : compiles) {
         c.wait();
     }

ggml/src/ggml-zdnn/CMakeLists.txt

@@ -0,0 +1,36 @@
+if (DEFINED ZDNN_ROOT)
+    message(STATUS "zdnn: using ZDNN_ROOT override: ${ZDNN_ROOT}")
+    set(ZDNN_HINT "${ZDNN_ROOT}")
+else()
+    set(ZDNN_HINT "")
+endif()
+
+find_path(ZDNN_INCLUDE
+            NAMES zdnn.h
+            HINTS ${ZDNN_HINT} /usr /usr/local
+            PATH_SUFFIXES include)
+if (ZDNN_INCLUDE)
+    message(STATUS "zdnn: found include: ${ZDNN_INCLUDE}")
+else()
+    message(FATAL_ERROR "zdnn: include directory not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+find_library(ZDNN_LIB
+                NAMES zdnn
+                HINTS ${ZDNN_HINT} /usr /usr/local
+                PATH_SUFFIXES lib lib64)
+if (ZDNN_LIB)
+    message(STATUS "zdnn: found library: ${ZDNN_LIB}")
+else()
+    message(FATAL_ERROR "zdnn: library not found, please set ZDNN_ROOT to the proper path if necessary")
+endif()
+
+file(GLOB GGML_SOURCES_ZDNN "*.c" "*.cpp")
+file(GLOB GGML_HEADERS_ZDNN "*.h" "*.hpp")
+
+ggml_add_backend_library(ggml-zdnn ${GGML_HEADERS_ZDNN} ${GGML_SOURCES_ZDNN})
+target_link_libraries(ggml-zdnn PRIVATE ${ZDNN_LIB})
+target_include_directories(ggml-zdnn PRIVATE ${ZDNN_INCLUDE})
+target_link_directories(ggml-zdnn PRIVATE ${ZDNN_LIB})
+
+target_compile_definitions(ggml-zdnn PRIVATE GGML_USE_ZDNN)

ggml/src/ggml-zdnn/ggml-zdnn-impl.h

@@ -0,0 +1,97 @@
+#ifndef GGML_ZDNN_IMPL
+#define GGML_ZDNN_IMPL
+
+#include "zdnn.h"
+#include "ggml.h"
+#include "ggml-zdnn.h"
+
+#include <vector>
+#include <memory>
+#include <vecintrin.h>
+
+#define GGML_ZDNN_NAME    "zDNN"
+#define GGML_ZDNN_VERSION ZDNN_VERNUM
+
+#define vec_neg(a)    (-(a))                // Vector Negate
+#define vec_add(a, b) ((a) + (b))           // Vector Add
+#define vec_sub(a, b) ((a) - (b))           // Vector Subtract
+#define vec_mul(a, b) ((a) * (b))           // Vector Multiply
+#define vec_div(a, b) ((a) / (b))           // Vector Divide
+#define vec_sl(a, b)  ((a) << (b))          // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b))          // Vector Shift Right
+#define vec_sr(a, b)  ((a) >> (b))          // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b)  ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef   signed char char8x16_t  __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t   int8x16_t  __attribute__((vector_size(16)));
+typedef int16_t  int16x8_t  __attribute__((vector_size(16)));
+typedef int32_t  int32x4_t  __attribute__((vector_size(16)));
+typedef uint8_t  uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t   __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));
+
+typedef   signed long long long64x2_t  __attribute__((vector_size(16)));
+typedef unsigned long long ulong64x2_t __attribute__((vector_size(16)));
+
+#define ZDNN_CHECK(stmt)                \
+    do {                                \
+        zdnn_status status = (stmt);    \
+        GGML_ASSERT(status == ZDNN_OK); \
+    } while (0);
+
+struct ggml_backend_zdnn_device_context {
+    int zdnn_device;
+    int zdnn_device_ref_count;
+
+    bool has_parmblkformat_0;
+    bool has_parmblkformat_1;
+
+    size_t max_size;
+
+    char name[128];
+};
+
+struct ggml_backend_zdnn_context {
+    int device;
+    ggml_cgraph * gf;
+};
+
+struct ggml_backend_zdnn_buffer {
+    void * data;
+    size_t size;
+
+    zdnn_tensor_desc pre_tfm_desc;
+    zdnn_tensor_desc tfm_desc;
+    zdnn_ztensor     ztensor;
+
+    char name[GGML_MAX_NAME];
+};
+
+struct ggml_backend_zdnn_buffer_context {
+    void * all_data;
+    size_t all_size;
+    bool owned;
+
+    int n_buffers;
+    std::vector<std::unique_ptr<ggml_backend_zdnn_buffer>> buffers;
+};
+
+#endif  // GGML_ZDNN_IMPL

ggml/src/ggml-zdnn/ggml-zdnn.cpp

@@ -0,0 +1,846 @@
+#include "zdnn.h"
+#include "ggml-zdnn.h"
+#include "ggml-zdnn-impl.h"
+
+#include "ggml-impl.h"
+#include "ggml-backend-impl.h"
+
+#include <vector>
+#include <memory>
+#include <csignal>
+#include <unistd.h>
+
+inline zdnn_data_types ggml_zdnn_type_mapping(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_F32:
+            return FP32;
+        case GGML_TYPE_F16:
+            return FP16;
+        case GGML_TYPE_BF16:
+            return BFLOAT;
+        case GGML_TYPE_I8:
+            return INT8;
+        case GGML_TYPE_I32:
+            return INT32;
+        case GGML_TYPE_Q8_0:
+            return INT8;
+        default:
+            GGML_ABORT("%s: fatal: unable to determine zTensor data type",
+                       __func__);
+            break;
+    }
+}
+
+inline void ggml_zdnn_create_tensor(zdnn_tensor_desc  & pre_tfm_desc,
+                                    zdnn_tensor_desc  & tfm_desc,
+                                    zdnn_ztensor      & ztensor,
+                              const ggml_tensor       * src,
+                              const int64_t           * ne,
+                              const zdnn_data_layouts   layout) {
+    zdnn_init_pre_transformed_desc(
+        layout,
+        ggml_zdnn_type_mapping(src->type),
+        &pre_tfm_desc,
+        ne[3], ne[2], ne[1], ne[0]
+    );
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&pre_tfm_desc, &tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&pre_tfm_desc, &tfm_desc, &ztensor));
+}
+
+inline void ggml_zdnn_load_tensor(zdnn_ztensor & ztensor,
+                                          void * buffer) {
+    ZDNN_CHECK(zdnn_transform_ztensor(&ztensor, buffer));
+}
+
+inline void ggml_zdnn_init_tensor(ggml_backend_zdnn_buffer * buffer, const ggml_tensor * tensor) {
+    switch (tensor->op) {
+        case GGML_OP_MUL_MAT:
+            {
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_2D,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[1], tensor->ne[0]
+                );
+            } break;
+
+        default:
+            {
+                // For 4D tensors, GGML uses NCHW layout. However, because zDNN
+                // automatically transforms everything to NHWC, we will use it
+                // directly to avoid the performance penalty changing the
+                // layout and reshaping the tensor.
+                zdnn_init_pre_transformed_desc(
+                    ZDNN_NHWC,
+                    ggml_zdnn_type_mapping(tensor->type),
+                    &buffer->pre_tfm_desc,
+                    tensor->ne[3], tensor->ne[2], tensor->ne[1], tensor->ne[0]
+                );
+
+                // TODO: Consider adding a ggml check.
+                // TODO: If tensor = 4D, use ZDNN_NCHW by default.
+                // TODO: If tensor = 2D, use ZDNN_NHWC by default.
+            } break;
+    }
+
+    ZDNN_CHECK(zdnn_generate_transformed_desc(&buffer->pre_tfm_desc, &buffer->tfm_desc));
+    ZDNN_CHECK(zdnn_init_ztensor_with_malloc(&buffer->pre_tfm_desc, &buffer->tfm_desc, &buffer->ztensor));
+}
+
+static void ggml_zdnn_mul_mat_op(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_TENSOR_BINARY_OP_LOCALS;
+
+    const enum ggml_type type = src0->type;
+
+    GGML_ASSERT(ne0 == ne01);
+    GGML_ASSERT(ne1 == ne11);
+    GGML_ASSERT(ne2 == ne12);
+    GGML_ASSERT(ne3 == ne13);
+
+    // we don't support permuted src0 or src1
+    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb10 == ggml_type_size(src1->type));
+
+    // dst cannot be transposed or permuted
+    GGML_ASSERT(nb0 == sizeof(float));
+    GGML_ASSERT(nb0 <= nb1);
+    GGML_ASSERT(nb1 <= nb2);
+    GGML_ASSERT(nb2 <= nb3);
+
+    const ggml_tensor * weights = src0;
+    const ggml_tensor * inputs  = src1;
+          ggml_tensor * output  = dst;
+
+    ggml_backend_zdnn_buffer * weights_extra = (ggml_backend_zdnn_buffer *)weights->extra;
+    ggml_backend_zdnn_buffer * inputs_extra  = (ggml_backend_zdnn_buffer *)inputs->extra;
+    ggml_backend_zdnn_buffer * output_extra  = (ggml_backend_zdnn_buffer *)output->extra;
+
+    zdnn_tensor_desc ptd_bias, td_bias;
+    zdnn_ztensor zt_bias;
+
+    const int64_t weights_rows = ne01;
+    const int64_t weights_cols = ne00;
+    const int64_t inputs_rows  = ne11;
+    const int64_t inputs_cols  = ne10;
+
+    assert(inputs_cols == weights_cols);
+
+    const int64_t output_rows = ne1;
+    const int64_t output_cols = ne0;
+
+    const int64_t bias_dim  [GGML_MAX_DIMS]  = { 1, 1, 1, output_cols };
+    ggml_zdnn_create_tensor(ptd_bias, td_bias, zt_bias, output, bias_dim, ZDNN_1D);
+
+    void * bias_data = (void *)calloc(ne0, ggml_element_size(output));
+    if (weights_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(weights_extra->ztensor, weights->data);
+    if (inputs_extra->ztensor.is_transformed == false) ggml_zdnn_load_tensor(inputs_extra->ztensor, inputs->data);
+    ggml_zdnn_load_tensor(zt_bias, bias_data);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, weights_extra->name,
+    //               weights->ne[3], weights->ne[2], weights->ne[1], weights->ne[0],
+    //               weights_extra->pre_tfm_desc.dim1,
+    //               weights_extra->pre_tfm_desc.dim2,
+    //               weights_extra->pre_tfm_desc.dim3,
+    //               weights_extra->pre_tfm_desc.dim4);
+
+    // GGML_LOG_INFO("%s: tensor '%s' tensor dimensions: [%ld, %ld, %ld, %ld] pre_tfm_desc dimensions: [%ld, %ld, %ld, %ld]\n",
+    //               __func__, inputs_extra->name,
+    //               inputs->ne[3], inputs->ne[2], inputs->ne[1], inputs->ne[0],
+    //               inputs_extra->pre_tfm_desc.dim1,
+    //               inputs_extra->pre_tfm_desc.dim2,
+    //               inputs_extra->pre_tfm_desc.dim3,
+    //               inputs_extra->pre_tfm_desc.dim4);
+
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim1 == weights->ne[0] && "weights_extra->pre_tfm_desc.dim1 must match weights->ne[0]");
+    GGML_ASSERT(weights_extra->pre_tfm_desc.dim2 == weights->ne[1] && "weights_extra->pre_tfm_desc.dim2 must match weights->ne[1]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim1  == inputs->ne[0]  && "inputs_extra->pre_tfm_desc.dim1 must match inputs->ne[0]");
+    GGML_ASSERT(inputs_extra->pre_tfm_desc.dim2  == inputs->ne[1]  && "inputs_extra->pre_tfm_desc.dim2 must match inputs->ne[1]");
+
+    ZDNN_CHECK(zdnn_matmul_transpose_op(&inputs_extra->ztensor, &weights_extra->ztensor, &zt_bias,
+                                        false, true, MATMUL_OP_ADDITION, &output_extra->ztensor));
+    // TODO: Remove in the future as we are currently DLF16 -> FP32 then in the next op, FP32 -> DLF16 again. Inefficient.
+    ZDNN_CHECK(zdnn_transform_origtensor(&output_extra->ztensor, output->data));
+
+    ZDNN_CHECK(zdnn_free_ztensor_buffer(&zt_bias));
+    free(bias_data);
+}
+
+static void ggml_zdnn_mul_mat_dispatch(ggml_backend_zdnn_context * ctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    bool use_mul_mat_vec =
+        (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_F16)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32
+        && src0->ne[0] % 2 == 0 && src1->ne[1] == 1;
+
+    bool use_mul_mat_vec_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    bool use_mul_mat_q =
+        ggml_is_quantized(src0->type)
+        && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32;
+
+    // debug helpers
+    // GGML_LOG_INFO("%s: use_mul_mat_vec   = %d\n", __func__, use_mul_mat_vec);
+    // GGML_LOG_INFO("%s: use_mul_mat_vec_q = %d\n", __func__, use_mul_mat_vec_q);
+    // GGML_LOG_INFO("%s: use_mul_mat_q     = %d\n", __func__, use_mul_mat_q);
+    // GGML_LOG_INFO("%s: src0: %8d %8d %8d %8d\n", __func__, src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
+    // GGML_LOG_INFO("%s: src1: %8d %8d %8d %8d\n", __func__, src1->ne[0], src1->ne[1], src1->ne[2], src1->ne[3]);
+    // GGML_LOG_INFO("%s:       %8d %8d %8d %8d\n", __func__, src1->nb[0], src1->nb[1], src1->nb[2], src1->nb[3]);
+    // GGML_LOG_INFO("%s: src0 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
+    // GGML_LOG_INFO("%s: src1 is contiguous %d, transposed %d, type = %s, name = %s\n", __func__, ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
+
+    if (src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F16
+        && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)
+        && src1->ne[2] * src1->ne[3] > 1) {
+        // general KQ + KQV multi-batch
+        GGML_LOG_INFO("%s: using zdnn_mul_mat_batched for KQ + KQV multi-batch\n", __func__);
+        // ggml_zdnn_mul_mat_batched(ctx, src0, src1, dst);
+    } else if (use_mul_mat_vec) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec for vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec, nullptr);
+    } else if (use_mul_mat_vec_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_vec_q for quantized vector multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_vec_q, ggml_zdnn_quantize_row_q8_1);
+    } else if (use_mul_mat_q) {
+        GGML_LOG_INFO("%s: using zdnn_op_mul_mat_q for quantized matrix multiplication\n", __func__);
+        // ggml_zdnn_op_mul_mat(ctx, src0, src1, dst, ggml_zdnn_op_mul_mat_q, ggml_zdnn_quantize_mmq_q8_1);
+    } else {
+        // GGML_LOG_INFO("%s: using zdnn_op_mul_mat for general matrix multiplication\n", __func__);
+        ggml_zdnn_mul_mat_op(ctx, src0, src1, dst);
+    }
+}
+
+static bool ggml_zdnn_compute_forward(ggml_backend_zdnn_context * ctx, ggml_tensor * dst) {
+    switch (dst->op) {
+        case GGML_OP_MUL_MAT:
+            ggml_zdnn_mul_mat_dispatch(ctx, dst->src[0], dst->src[1], dst);
+            break;
+
+        default:
+            return false;
+    }
+
+    return true;
+}
+
+static enum ggml_status ggml_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * gf) {
+    ggml_backend_zdnn_context        * ctx     = (       ggml_backend_zdnn_context *)backend->context;
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)backend->device->context;
+
+    ctx->gf = gf;
+    for (int i = 0; i < gf->n_nodes; i++) {
+        ggml_tensor * node = gf->nodes[i];
+
+        if (ggml_is_empty(node)
+            || node->op == GGML_OP_NONE
+            || node->op == GGML_OP_RESHAPE
+            || node->op == GGML_OP_VIEW
+            || node->op == GGML_OP_PERMUTE
+            || node->op == GGML_OP_TRANSPOSE) {
+            continue;
+        }
+
+        bool ok = ggml_zdnn_compute_forward(ctx, node);
+        if (!ok) {
+            GGML_LOG_ERROR("%s: unsupported op %s (%s)\n",
+                           __func__, node->name, ggml_op_name(node->op));
+        }
+
+        GGML_ASSERT(ok);
+    }
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static bool ggml_zdnn_supports_op(const ggml_backend_zdnn_device_context * ctx_dev, const ggml_tensor * op) {
+    switch (op->op) {
+        case GGML_OP_NONE:
+        case GGML_OP_RESHAPE:
+        case GGML_OP_VIEW:
+        case GGML_OP_TRANSPOSE:
+        case GGML_OP_PERMUTE:
+            return true;
+
+        case GGML_OP_MUL_MAT:
+            {
+                const ggml_tensor * src0 = op->src[0];
+                const ggml_tensor * src1 = op->src[1];
+
+                const int64_t ne10 = src1->ne[0];
+                const int64_t ne0 = op->ne[0];
+                const int64_t ne1 = op->ne[1];
+
+                const int64_t max_batch = ctx_dev->max_size;
+
+                return ggml_is_matrix(src0) &&
+                       ggml_is_matrix(src1) &&
+                       ggml_is_contiguous(src0) &&
+                       ggml_is_contiguous(src1) &&
+                       src0->view_src == nullptr && src1->view_src == nullptr &&
+                       src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 &&
+                       (ne0 <= max_batch && ne1 <= max_batch && ne10 <= max_batch);
+            } break;
+
+        default:
+            return false;
+    }
+}
+
+////////////////////////////////////////////////////////////////////////////////
+
+//
+// globals
+//
+
+// initialised in ggml_backend_zdnn_reg
+static ggml_backend_reg    g_ggml_backend_zdnn_reg;
+static ggml_backend_device g_ggml_backend_zdnn_device;
+
+static ggml_backend_zdnn_device_context g_ggml_ctx_dev_main = {
+    /* .zdnn_device           = */ 0,
+    /* .zdnn_device_ref_count = */ 0,
+    /* .has_parmblkformat_0   = */ false,
+    /* .has_parmblkformat_1   = */ false,
+    /* .max_size              = */ 0,
+    /* .name                  = */ "",
+};
+
+static int ggml_backend_zdnn_device_acq(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+
+    if (ctx->zdnn_device == 0) {
+        ctx->zdnn_device = 1;
+    }
+
+    if (ctx->zdnn_device >= 1) {
+        ctx->has_parmblkformat_0 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0);
+        ctx->has_parmblkformat_1 = zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1);
+        ctx->max_size = zdnn_get_nnpa_max_dim_idx_size();
+        strncpy(ctx->name, GGML_ZDNN_NAME, sizeof(ctx->name) - 1);
+    }
+
+    ctx->zdnn_device_ref_count++;
+    return ctx->zdnn_device;
+}
+
+static void ggml_backend_zdnn_device_rel(ggml_backend_zdnn_device_context * ctx) {
+    assert(ctx != NULL);
+    assert(ctx->zdnn_device_ref_count > 0);
+
+    ctx->zdnn_device_ref_count--;
+    if (ctx->zdnn_device_ref_count == 0) {
+        if (ctx->zdnn_device >= 0) {
+            ctx->zdnn_device = 0;
+        }
+    }
+}
+
+static ggml_backend_zdnn_context * ggml_zdnn_init(ggml_backend_dev_t dev) {
+    GGML_LOG_INFO("%s: allocating\n", __func__);
+    GGML_LOG_INFO("%s: found 1 device\n", __func__);
+
+    #ifdef STATIC_LIB
+    zdnn_init();
+    #endif
+
+    ggml_backend_zdnn_context * ctx = new ggml_backend_zdnn_context();
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    int device = 1;
+    GGML_LOG_INFO("%s: picking default device: %s\n", __func__, ctx_dev->name);
+
+    ctx->device = device;
+    GGML_LOG_INFO("%s: NNPA name: %s\n", __func__, ctx_dev->name);
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_0 = %s\n", __func__, ctx_dev->has_parmblkformat_0 ? "true" : "false");
+    GGML_LOG_INFO("%s: NNPA_PARMBLKFORMAT_1 = %s\n", __func__, ctx_dev->has_parmblkformat_1 ? "true" : "false");
+
+    ctx->gf = nullptr;
+
+    return ctx;
+}
+
+static void ggml_zdnn_free(ggml_backend_zdnn_context * ctx) {
+    GGML_LOG_INFO("%s: deallocating\n", __func__);
+    delete ctx;
+}
+
+//
+// backend interface
+//
+
+static void ggml_backend_zdnn_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    for (int i = 0; i < ctx->n_buffers; i++) {
+        if (ctx->buffers[i]->ztensor.buffer != NULL && ctx->buffers[i]->ztensor.is_transformed) {
+            ZDNN_CHECK(zdnn_free_ztensor_buffer(&ctx->buffers[i]->ztensor));
+        }
+    }
+
+    delete ctx;
+}
+
+static void * ggml_backend_zdnn_buffer_get_base(ggml_backend_buffer_t buffer) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+    return ctx->all_data;
+}
+
+static enum ggml_status ggml_backend_zdnn_buffer_init_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor) {
+    if (tensor->view_src != NULL) {
+        assert(tensor->view_src->buffer->buft == buffer->buft);
+        return GGML_STATUS_SUCCESS;
+    }
+
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    const int64_t tsize = ggml_nbytes(tensor);
+    int buffer_idx = ctx->n_buffers;
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = tensor->data;
+    zdnn_buffer->size = tsize;
+    strncpy(zdnn_buffer->name, tensor->name, GGML_MAX_NAME - 1);
+
+    ggml_zdnn_init_tensor(zdnn_buffer.get(), tensor);
+    tensor->extra = zdnn_buffer.get();
+
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+    ctx->n_buffers++;
+
+    // GGML_LOG_INFO("%s: initialised tensor '%s' in buffer %d, size = %8.2f MiB\n",
+    //               __func__, tensor->name, buffer_idx, tsize);
+
+    return GGML_STATUS_SUCCESS;
+}
+
+static void ggml_backend_zdnn_buffer_memset_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
+    memset((char *)tensor->data + offset, value, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
+    memcpy((char *)tensor->data + offset, data, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {
+    memcpy(data, (const char *)tensor->data + offset, size);
+
+    GGML_UNUSED(buffer);
+}
+
+static void ggml_backend_zdnn_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    ggml_backend_zdnn_buffer_context * ctx = (ggml_backend_zdnn_buffer_context *)buffer->context;
+
+    memset(ctx->all_data, value, ctx->all_size);
+}
+
+static ggml_backend_buffer_i ggml_backend_zdnn_buffer_i = {
+    /* .free_buffer   = */ ggml_backend_zdnn_buffer_free_buffer,
+    /* .get_base      = */ ggml_backend_zdnn_buffer_get_base,
+    /* .init_tensor   = */ ggml_backend_zdnn_buffer_init_tensor,
+    /* .memset_tensor = */ ggml_backend_zdnn_buffer_memset_tensor,
+    /* .set_tensor    = */ ggml_backend_zdnn_buffer_set_tensor,
+    /* .get_tensor    = */ ggml_backend_zdnn_buffer_get_tensor,
+    /* .cpy_tensor    = */ NULL,
+    /* .clear         = */ ggml_backend_zdnn_buffer_clear,
+    /* .reset         = */ NULL,
+};
+
+//
+// default buffer type
+//
+
+static const char * ggml_backend_zdnn_buffer_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)buft->device->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    ctx->all_data  = ggml_aligned_malloc(size_aligned);
+    ctx->all_size  = size_aligned;
+    ctx->owned     = true;
+    ctx->n_buffers = 1;
+
+    if (ctx->all_data != NULL) {
+        std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+        zdnn_buffer->data = ctx->all_data;
+        zdnn_buffer->size = size_aligned;
+        ctx->buffers.push_back(std::move(zdnn_buffer));
+    }
+
+    if (size_aligned > 0 && (ctx->all_data == NULL)) {
+        GGML_LOG_ERROR("%s: error: failed to allocate buffer, size = %8.2f\n",
+                       __func__, size_aligned / 1024.0 / 1024.0);
+        delete ctx;
+        return NULL;
+    }
+
+    return ggml_backend_buffer_init(buft, ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static size_t ggml_backend_zdnn_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) {
+    return 256;
+
+    GGML_UNUSED(buft);
+}
+
+static bool ggml_backend_zdnn_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
+ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_type_zdnn = {
+        /* .iface   = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_type_zdnn;
+}
+
+static const char * ggml_backend_zdnn_buffer_from_ptr_type_get_name(ggml_backend_buffer_type_t buft) {
+    return GGML_ZDNN_NAME "_Mapped";
+
+    GGML_UNUSED(buft);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_buffer_from_ptr_type(void) {
+    static ggml_backend_buffer_type ggml_backend_buffer_from_ptr_type_zdnn = {
+        /* .iface = */ {
+            /* .get_name       = */ ggml_backend_zdnn_buffer_from_ptr_type_get_name,
+            /* .alloc_buffer   = */ ggml_backend_zdnn_buffer_type_alloc_buffer,
+            /* .get_alignment  = */ ggml_backend_zdnn_buffer_type_get_alignment,
+            /* .get_max_size   = */ NULL,
+            /* .get_alloc_size = */ NULL,  // defaults to ggml_nbytes
+            /* .is_host        = */ ggml_backend_zdnn_buffer_type_is_host,
+        },
+        /* .device  = */ &g_ggml_backend_zdnn_device,
+        /* .context = */ NULL,
+    };
+
+    return &ggml_backend_buffer_from_ptr_type_zdnn;
+}
+
+//
+// backend
+//
+
+static const char * ggml_backend_zdnn_name(ggml_backend_t backend) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(backend);
+}
+
+static void ggml_backend_zdnn_free(ggml_backend_t backend) {
+    ggml_backend_zdnn_context * ctx = (ggml_backend_zdnn_context *)backend->context;
+
+    ggml_zdnn_free(ctx);
+    free(backend);
+}
+
+static enum ggml_status ggml_backend_zdnn_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+    return ggml_zdnn_graph_compute(backend, cgraph);
+}
+
+static ggml_backend_i ggml_backend_zdnn_i = {
+    /* .get_name           = */ ggml_backend_zdnn_name,
+    /* .free               = */ ggml_backend_zdnn_free,
+    /* .set_tensor_async   = */ NULL,
+    /* .get_tensor_async   = */ NULL,
+    /* .cpy_tensor_async   = */ NULL,
+    /* .synchronize        = */ NULL,
+    /* .graph_plan_create  = */ NULL,
+    /* .graph_plan_free    = */ NULL,
+    /* .graph_plan_update  = */ NULL,
+    /* .graph_plan_compute = */ NULL,
+    /* .graph_compute      = */ ggml_backend_zdnn_graph_compute,
+    /* .event_record       = */ NULL,
+    /* .event_wait         = */ NULL,
+};
+
+static ggml_guid_t ggml_backend_zdnn_guid(void) {
+    static const char * guid_str = "IBM-ZDNN-ACCELER";
+    return reinterpret_cast<ggml_guid_t>((void *)guid_str);
+}
+
+// TODO: remove in the future
+ggml_backend_t ggml_backend_zdnn_init(void) {
+    ggml_backend_dev_t dev = ggml_backend_reg_dev_get(ggml_backend_zdnn_reg(), 0);
+
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend_t)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+}
+
+bool ggml_backend_is_zdnn(ggml_backend_t backend) {
+    return backend != NULL &&
+           ggml_guid_matches(backend->guid, ggml_backend_zdnn_guid());
+
+    GGML_UNUSED(backend);
+}
+
+//
+// backend device
+//
+
+static const char * ggml_backend_zdnn_device_get_name(ggml_backend_dev_t dev) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(dev);
+}
+
+static const char * ggml_backend_zdnn_device_get_description(ggml_backend_dev_t dev) {
+    return "IBM Z Neural Network Processing Assist (NNPA)";
+}
+
+static void ggml_backend_zdnn_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
+    *free  = 0;
+    *total = 0;
+}
+
+static enum ggml_backend_dev_type ggml_backend_zdnn_device_get_type(ggml_backend_dev_t dev) {
+    return GGML_BACKEND_DEVICE_TYPE_ACCEL;
+
+    GGML_UNUSED(dev);
+}
+
+static void ggml_backend_zdnn_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
+    props->name        = ggml_backend_zdnn_device_get_name(dev);
+    props->description = ggml_backend_zdnn_device_get_description(dev);
+    props->type        = ggml_backend_zdnn_device_get_type(dev);
+    ggml_backend_zdnn_device_get_memory(dev, &props->memory_free, &props->memory_total);
+    props->caps = (ggml_backend_dev_caps) {
+        /* .async                = */ false,
+        /* .host_buffer          = */ false,
+        /* .buffer_from_host_ptr = */ true,
+        /* .events               = */ false,
+    };
+}
+
+static ggml_backend_t ggml_backend_zdnn_device_init(ggml_backend_dev_t dev, const char * params) {
+    ggml_backend_zdnn_context * ctx = ggml_zdnn_init(dev);
+    if (ctx == NULL) {
+        GGML_LOG_ERROR("%s: error: failed to allocate context\n", __func__);
+        return NULL;
+    }
+
+    ggml_backend_t backend = (ggml_backend *)malloc(sizeof(ggml_backend));
+    *backend = (ggml_backend) {
+        /* .guid       = */ ggml_backend_zdnn_guid(),
+        /* .iface      = */ ggml_backend_zdnn_i,
+        /* .device     = */ dev,
+        /* .context    = */ ctx,
+    };
+
+    return backend;
+
+    GGML_UNUSED(params);
+}
+
+static ggml_backend_buffer_type_t ggml_backend_zdnn_device_get_buffer_type(ggml_backend_dev_t dev) {
+    return ggml_backend_zdnn_buffer_type();
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_zdnn_device_buffer_from_ptr(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size) {
+    ggml_backend_zdnn_buffer_context * ctx = new ggml_backend_zdnn_buffer_context();
+
+    ctx->all_data  = ptr;
+    ctx->all_size  = size;
+    ctx->owned     = false;
+    ctx->n_buffers = 0;
+
+    const size_t size_page = sysconf(_SC_PAGESIZE);
+
+    // page-align the data ptr
+    {
+        const uintptr_t offs = (uintptr_t) ptr % size_page;
+        ptr  = (void *)((char *)ptr - offs);
+        size += offs;
+    }
+
+    size_t size_aligned = size;
+    if ((size_aligned % size_page) != 0) {
+        size_aligned += size_page - (size_aligned % size_page);
+    }
+
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *)dev->context;
+
+    GGML_ASSERT(ctx_dev->zdnn_device >= 0);
+    int device = ctx_dev->zdnn_device; GGML_UNUSED(device);
+
+    std::unique_ptr<ggml_backend_zdnn_buffer> zdnn_buffer = std::make_unique<ggml_backend_zdnn_buffer>();
+    zdnn_buffer->data = ptr;
+    zdnn_buffer->size = size;
+    ctx->buffers.push_back(std::move(zdnn_buffer));
+
+    GGML_LOG_INFO("%s: allocated buffer, size = %8.2f MiB\n",
+                  __func__, size_aligned / 1024.0 / 1024.0);
+
+    ++ctx->n_buffers;
+
+    return ggml_backend_buffer_init(ggml_backend_zdnn_buffer_from_ptr_type(), ggml_backend_zdnn_buffer_i, ctx, size);
+}
+
+static bool ggml_backend_zdnn_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_zdnn_device_context * ctx_dev = (ggml_backend_zdnn_device_context *) dev->context;
+
+    return ggml_zdnn_supports_op(ctx_dev, op);
+}
+
+static bool ggml_backend_zdnn_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
+    return
+        buft->iface.get_name == ggml_backend_zdnn_buffer_type_get_name ||
+        buft->iface.get_name == ggml_backend_zdnn_buffer_from_ptr_type_get_name;
+
+    GGML_UNUSED(dev);
+}
+
+static ggml_backend_device_i ggml_backend_zdnn_device_i = {
+    /* .get_name             = */ ggml_backend_zdnn_device_get_name,
+    /* .get_description      = */ ggml_backend_zdnn_device_get_description,
+    /* .get_memory           = */ ggml_backend_zdnn_device_get_memory,
+    /* .get_type             = */ ggml_backend_zdnn_device_get_type,
+    /* .get_props            = */ ggml_backend_zdnn_device_get_props,
+    /* .init_backend         = */ ggml_backend_zdnn_device_init,
+    /* .get_buffer_type      = */ ggml_backend_zdnn_device_get_buffer_type,
+    /* .get_host_buffer_type = */ NULL,
+    /* .buffer_from_host_ptr = */ ggml_backend_zdnn_device_buffer_from_ptr,
+    /* .supports_op          = */ ggml_backend_zdnn_device_supports_op,
+    /* .supports_buft        = */ ggml_backend_zdnn_device_supports_buft,
+    /* .offload_op           = */ NULL,
+    /* .event_new            = */ NULL,
+    /* .event_free           = */ NULL,
+    /* .event_synchronize    = */ NULL,
+};
+
+//
+// backend registry
+//
+
+static const char * ggml_backend_zdnn_reg_get_name(ggml_backend_reg_t reg) {
+    return GGML_ZDNN_NAME;
+
+    GGML_UNUSED(reg);
+}
+
+static size_t ggml_backend_zdnn_reg_device_count(ggml_backend_reg_t reg) {
+    if (!zdnn_is_nnpa_installed()) {
+        return 0;
+    }
+    return 1;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_dev_t ggml_backend_zdnn_reg_device_get(ggml_backend_reg_t reg, size_t index) {
+    GGML_ASSERT(index == 0);
+
+    return &g_ggml_backend_zdnn_device;
+
+    GGML_UNUSED(reg);
+    GGML_UNUSED(index);
+}
+
+static ggml_backend_feature g_ggml_backend_zdnn_features[] = {
+    { "NNPA", zdnn_is_nnpa_installed() ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_0", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_0) ? "1" : "0" },
+    { "NNPA_PARMBLKFORMAT_1", zdnn_is_nnpa_parmblk_fmt_installed(1, NNPA_PARMBLKFORMAT_1) ? "1" : "0" },
+    { NULL, NULL },
+};
+
+static ggml_backend_feature * ggml_backend_zdnn_get_features(ggml_backend_reg_t reg) {
+    return g_ggml_backend_zdnn_features;
+
+    GGML_UNUSED(reg);
+}
+
+static void * ggml_backend_zdnn_get_proc_address(ggml_backend_reg_t reg, const char * name) {
+    if (strcmp(name, "ggml_backend_get_features") == 0) {
+        return (void *) ggml_backend_zdnn_get_features;
+    }
+
+    return NULL;
+
+    GGML_UNUSED(reg);
+}
+
+static ggml_backend_reg_i ggml_backend_zdnn_reg_i = {
+    /* .get_name         = */ ggml_backend_zdnn_reg_get_name,
+    /* .get_device_count = */ ggml_backend_zdnn_reg_device_count,
+    /* .get_device       = */ ggml_backend_zdnn_reg_device_get,
+    /* .get_proc_address = */ ggml_backend_zdnn_get_proc_address,
+};
+
+static void ggml_zdnn_cleanup(void) {
+    ggml_backend_zdnn_device_rel(&g_ggml_ctx_dev_main);
+}
+
+// TODO: make thread-safe
+ggml_backend_reg_t ggml_backend_zdnn_reg(void) {
+    ggml_backend_zdnn_device_acq(&g_ggml_ctx_dev_main);
+
+    // register cleanup callback
+    atexit(ggml_zdnn_cleanup);
+
+    {
+        g_ggml_backend_zdnn_reg = (ggml_backend_reg) {
+            /* .api_version = */ GGML_ZDNN_VERSION,
+            /* .iface       = */ ggml_backend_zdnn_reg_i,
+            /* .context     = */ NULL,
+        };
+
+        g_ggml_backend_zdnn_device = (ggml_backend_device) {
+            /* .iface       = */ ggml_backend_zdnn_device_i,
+            /* .reg         = */ &g_ggml_backend_zdnn_reg,
+            /* .context     = */ &g_ggml_ctx_dev_main,
+        };
+
+        return &g_ggml_backend_zdnn_reg;
+    }
+}
+
+GGML_BACKEND_DL_IMPL(ggml_backend_zdnn_reg)

ggml/src/ggml.c

@@ -1012,11 +1012,12 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
     "OPT_STEP_ADAMW",
+    "OPT_STEP_SGD",
 
     "GLU",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -1113,15 +1114,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
     "adamw(x)",
+    "sgd(x)",
 
     "glu(x)",
 };
 
-static_assert(GGML_OP_COUNT == 87, "GGML_OP_COUNT != 87");
+static_assert(GGML_OP_COUNT == 88, "GGML_OP_COUNT != 88");
 
 static_assert(GGML_OP_POOL_COUNT == 2, "GGML_OP_POOL_COUNT != 2");
 
-
 static const char * GGML_UNARY_OP_NAME[GGML_UNARY_OP_COUNT] = {
     "ABS",
     "SGN",
@@ -3885,6 +3886,7 @@ static struct ggml_tensor * ggml_rope_impl(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3898,15 +3900,19 @@ static struct ggml_tensor * ggml_rope_impl(
 
     GGML_ASSERT(ggml_is_vector(b));
     GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] == b->ne[0]);
+
+    bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
+    if (mrope_used) {
+        GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
+    } else {
+        GGML_ASSERT(a->ne[2] == b->ne[0]);
+    }
 
     if (c) {
         GGML_ASSERT(c->type == GGML_TYPE_F32);
         GGML_ASSERT(c->ne[0] >= n_dims / 2);
     }
 
-    int sections[4] = {0, 0, 0, 0};
-
     struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
 
     int32_t params[15] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
@@ -3916,7 +3922,11 @@ static struct ggml_tensor * ggml_rope_impl(
     memcpy(params +  8, &attn_factor,  sizeof(float));
     memcpy(params +  9, &beta_fast,    sizeof(float));
     memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(params + 11, &sections,     sizeof(int)*4);
+    if (mrope_used) {
+        memcpy(params + 11, sections,  sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    } else {
+        memset(params + 11, 0,         sizeof(int32_t) * GGML_MROPE_SECTIONS);
+    }
     ggml_set_op_params(result, params, sizeof(params));
 
     result->op     = GGML_OP_ROPE;
@@ -3934,7 +3944,7 @@ struct ggml_tensor * ggml_rope(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, false
     );
 }
 
@@ -3944,7 +3954,7 @@ struct ggml_tensor * ggml_rope_multi(
         struct ggml_tensor  * b,
         struct ggml_tensor  * c,
         int                   n_dims,
-        int                   sections[4],
+        int                   sections[GGML_MROPE_SECTIONS],
         int                   mode,
         int                   n_ctx_orig,
         float                 freq_base,
@@ -3953,36 +3963,31 @@ struct ggml_tensor * ggml_rope_multi(
         float                 attn_factor,
         float                 beta_fast,
         float                 beta_slow) {
-    // Multimodal Rotary Position Embedding
-    GGML_ASSERT((mode & 1) == 0 && "mode & 1 == 1 is no longer supported");
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, false
+    );
+}
 
-    GGML_ASSERT(ggml_is_vector(b));
-    GGML_ASSERT(b->type == GGML_TYPE_I32);
-    GGML_ASSERT(a->ne[2] * 4 == b->ne[0]); // mrope expecting 4 position ids per token
-
-    if (c) {
-        GGML_ASSERT(c->type == GGML_TYPE_F32);
-        GGML_ASSERT(c->ne[0] >= n_dims / 2);
-    }
-
-    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
-
-    int32_t params[11 + 4] = { /*n_past*/ 0, n_dims, mode, /*n_ctx*/ 0, n_ctx_orig };
-    memcpy(params +  5, &freq_base,    sizeof(float));
-    memcpy(params +  6, &freq_scale,   sizeof(float));
-    memcpy(params +  7, &ext_factor,   sizeof(float));
-    memcpy(params +  8, &attn_factor,  sizeof(float));
-    memcpy(params +  9, &beta_fast,    sizeof(float));
-    memcpy(params + 10, &beta_slow,    sizeof(float));
-    memcpy(&params[11], sections,      sizeof(int)*4);
-    ggml_set_op_params(result, params, sizeof(params));
-
-    result->op   = GGML_OP_ROPE;
-    result->src[0] = a;
-    result->src[1] = b;
-    result->src[2] = c;
-
-    return result;
+struct ggml_tensor * ggml_rope_multi_inplace(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b,
+        struct ggml_tensor  * c,
+        int                   n_dims,
+        int                   sections[GGML_MROPE_SECTIONS],
+        int                   mode,
+        int                   n_ctx_orig,
+        float                 freq_base,
+        float                 freq_scale,
+        float                 ext_factor,
+        float                 attn_factor,
+        float                 beta_fast,
+        float                 beta_slow) {
+    return ggml_rope_impl(
+        ctx, a, b, c, n_dims, sections, mode, n_ctx_orig, freq_base, freq_scale,
+        ext_factor, attn_factor, beta_fast, beta_slow, true
+    );
 }
 
 struct ggml_tensor * ggml_rope_inplace(
@@ -3992,7 +3997,7 @@ struct ggml_tensor * ggml_rope_inplace(
         int                   n_dims,
         int                   mode) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
+        ctx, a, b, NULL, n_dims, NULL, mode, 0, 10000.0f, 1.0f, 0.0f, 1.0f, 0.0f, 0.0f, true
     );
 }
 
@@ -4011,7 +4016,7 @@ struct ggml_tensor * ggml_rope_ext(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4031,7 +4036,7 @@ struct ggml_tensor * ggml_rope_ext_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, c, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, c, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
@@ -4050,7 +4055,7 @@ struct ggml_tensor * ggml_rope_custom(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, false
     );
 }
@@ -4069,7 +4074,7 @@ struct ggml_tensor * ggml_rope_custom_inplace(
         float                 beta_fast,
         float                 beta_slow) {
     return ggml_rope_impl(
-        ctx, a, b, NULL, n_dims, mode, n_ctx_orig, freq_base, freq_scale,
+        ctx, a, b, NULL, n_dims, NULL, mode, n_ctx_orig, freq_base, freq_scale,
         ext_factor, attn_factor, beta_fast, beta_slow, true
     );
 }
@@ -4267,14 +4272,13 @@ struct ggml_tensor * ggml_conv_1d_dw(
         int                   s0,
         int                   p0,
         int                   d0) {
-    struct ggml_tensor * new_a = ggml_reshape_4d(ctx, a, a->ne[0], 1, a->ne[1], a->ne[2]);
     struct ggml_tensor * new_b = ggml_reshape_4d(ctx, b, b->ne[0], 1, b->ne[1], b->ne[2]);
 
-    struct ggml_tensor * im2col = ggml_im2col(ctx, new_a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
+    struct ggml_tensor * im2col = ggml_im2col(ctx, a, new_b, s0, 0, p0, 0, d0, 0, false, GGML_TYPE_F16);
 
     struct ggml_tensor * result = ggml_mul_mat(ctx, im2col, a);
 
-    result = ggml_reshape_3d(ctx, result, b->ne[0], b->ne[1], 1);
+    result = ggml_reshape_3d(ctx, result, result->ne[0], result->ne[2], 1);
 
     return result;
 }
@@ -5602,6 +5606,28 @@ struct ggml_tensor * ggml_opt_step_adamw(
     return result;
 }
 
+// opt_step_sgd
+
+struct ggml_tensor * ggml_opt_step_sgd(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * grad,
+        struct ggml_tensor  * params) {
+    GGML_ASSERT(a->flags & GGML_TENSOR_FLAG_PARAM);
+    GGML_ASSERT(ggml_are_same_shape(a, grad));
+    GGML_ASSERT(params->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_nelements(params) == 2);
+
+    struct ggml_tensor * result = ggml_view_tensor(ctx, a);
+
+    result->op     = GGML_OP_OPT_STEP_SGD;
+    result->src[0] = a;
+    result->src[1] = grad;
+    result->src[2] = params;
+
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 struct ggml_hash_set ggml_hash_set_new(size_t size) {

include/llama.h

@@ -870,6 +870,29 @@ extern "C" {
                           size_t   n_token_capacity,
                           size_t * n_token_count_out);
 
+#define LLAMA_STATE_SEQ_FLAGS_SWA_ONLY 1
+
+    typedef uint32_t llama_state_seq_flags;
+
+    LLAMA_API size_t llama_state_seq_get_size_ext(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_get_data_ext(
+            struct llama_context * ctx,
+                         uint8_t * dst,
+                          size_t   size,
+                    llama_seq_id   seq_id,
+           llama_state_seq_flags   flags);
+
+    LLAMA_API size_t llama_state_seq_set_data_ext(
+            struct llama_context * ctx,
+                   const uint8_t * src,
+                          size_t   size,
+                    llama_seq_id   dest_seq_id,
+           llama_state_seq_flags   flags);
+
     //
     // Decoding
     //
@@ -1437,6 +1460,8 @@ extern "C" {
 
         ggml_opt_get_optimizer_params get_opt_pars; // callback for calculating optimizer parameters
         void * get_opt_pars_ud;                     // userdata for calculating optimizer parameters
+
+        enum ggml_opt_optimizer_type optimizer_type;
     };
 
     LLAMA_API void llama_opt_init(struct llama_context * lctx, struct llama_model * model, struct llama_opt_params lopt_params);

models/templates/openai-gpt-oss-120b.jinja

@@ -0,0 +1,331 @@
+{#-
+  In addition to the normal inputs of `messages` and `tools`, this template also accepts the
+  following kwargs:
+  - "builtin_tools": A list, can contain "browser" and/or "python".
+  - "model_identity": A string that optionally describes the model identity.
+  - "reasoning_effort": A string that describes the reasoning effort, defaults to "medium".
+ #}
+
+{#- Tool Definition Rendering ============================================== #}
+{%- macro render_typescript_type(param_spec, required_params, is_nullable=false) -%}
+    {%- if param_spec.type == "array" -%}
+        {%- if param_spec['items'] -%}
+            {%- if param_spec['items']['type'] == "string" -%}
+                {{- "string[]" }}
+            {%- elif param_spec['items']['type'] == "number" -%}
+                {{- "number[]" }}
+            {%- elif param_spec['items']['type'] == "integer" -%}
+                {{- "number[]" }}
+            {%- elif param_spec['items']['type'] == "boolean" -%}
+                {{- "boolean[]" }}
+            {%- else -%}
+                {%- set inner_type = render_typescript_type(param_spec['items'], required_params) -%}
+                {%- if inner_type == "object | object" or inner_type|length > 50 -%}
+                    {{- "any[]" }}
+                {%- else -%}
+                    {{- inner_type + "[]" }}
+                {%- endif -%}
+            {%- endif -%}
+            {%- if param_spec.nullable -%}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- else -%}
+            {{- "any[]" }}
+            {%- if param_spec.nullable -%}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- endif -%}
+    {%- elif param_spec.type is defined and param_spec.type is iterable and param_spec.type is not string and param_spec.type is not mapping and param_spec.type[0] is defined -%}
+        {#- Handle array of types like ["object", "object"] from Union[dict, list] #}
+        {%- if param_spec.type | length > 1 -%}
+            {{- param_spec.type | join(" | ") }}
+        {%- else -%}
+            {{- param_spec.type[0] }}
+        {%- endif -%}
+    {%- elif param_spec.oneOf -%}
+        {#- Handle oneOf schemas - check for complex unions and fallback to any #}
+        {%- set has_object_variants = false -%}
+        {%- for variant in param_spec.oneOf -%}
+            {%- if variant.type == "object" -%}
+                {%- set has_object_variants = true -%}
+            {%- endif -%}
+        {%- endfor -%}
+        {%- if has_object_variants and param_spec.oneOf|length > 1 -%}
+            {{- "any" }}
+        {%- else -%}
+            {%- for variant in param_spec.oneOf -%}
+                {{- render_typescript_type(variant, required_params) -}}
+                {%- if variant.description %}
+                    {{- "// " + variant.description }}
+                {%- endif -%}
+                {%- if variant.default is defined %}
+                    {{ "// default: " + variant.default|tojson }}
+                {%- endif -%}
+                {%- if not loop.last %}
+                    {{- " | " }}
+                {% endif -%}
+            {%- endfor -%}
+        {%- endif -%}
+    {%- elif param_spec.type == "string" -%}
+        {%- if param_spec.enum -%}
+            {{- '"' + param_spec.enum|join('" | "') + '"' -}}
+        {%- else -%}
+            {{- "string" }}
+            {%- if param_spec.nullable %}
+                {{- " | null" }}
+            {%- endif -%}
+        {%- endif -%}
+    {%- elif param_spec.type == "number" -%}
+        {{- "number" }}
+    {%- elif param_spec.type == "integer" -%}
+        {{- "number" }}
+    {%- elif param_spec.type == "boolean" -%}
+        {{- "boolean" }}
+
+    {%- elif param_spec.type == "object" -%}
+        {%- if param_spec.properties -%}
+            {{- "{\n" }}
+            {%- for prop_name, prop_spec in param_spec.properties.items() -%}
+                {{- prop_name -}}
+                {%- if prop_name not in (param_spec.required or []) -%}
+                    {{- "?" }}
+                {%- endif -%}
+                {{- ": " }}
+                {{ render_typescript_type(prop_spec, param_spec.required or []) }}
+                {%- if not loop.last -%}
+                    {{-", " }}
+                {%- endif -%}
+            {%- endfor -%}
+            {{- "}" }}
+        {%- else -%}
+            {{- "object" }}
+        {%- endif -%}
+    {%- else -%}
+        {{- "any" }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{%- macro render_tool_namespace(namespace_name, tools) -%}
+    {{- "## " + namespace_name + "\n\n" }}
+    {{- "namespace " + namespace_name + " {\n\n" }}
+    {%- for tool in tools %}
+        {%- set tool = tool.function %}
+        {{- "// " + tool.description + "\n" }}
+        {{- "type "+ tool.name + " = " }}
+        {%- if tool.parameters and tool.parameters.properties %}
+            {{- "(_: {\n" }}
+            {%- for param_name, param_spec in tool.parameters.properties.items() %}
+                {%- if param_spec.description %}
+                    {{- "// " + param_spec.description + "\n" }}
+                {%- endif %}
+                {{- param_name }}
+                {%- if param_name not in (tool.parameters.required or []) -%}
+                    {{- "?" }}
+                {%- endif -%}
+                {{- ": " }}
+                {{- render_typescript_type(param_spec, tool.parameters.required or []) }}
+                {%- if param_spec.default is defined -%}
+                    {%- if param_spec.enum %}
+                        {{- ", // default: " + param_spec.default }}
+                    {%- elif param_spec.oneOf %}
+                        {{- "// default: " + param_spec.default }}
+                    {%- else %}
+                        {{- ", // default: " + param_spec.default|tojson }}
+                    {%- endif -%}
+                {%- endif -%}
+                {%- if not loop.last %}
+                    {{- ",\n" }}
+                {%- else %}
+                    {{- ",\n" }}
+                {%- endif -%}
+            {%- endfor %}
+            {{- "}) => any;\n\n" }}
+        {%- else -%}
+            {{- "() => any;\n\n" }}
+        {%- endif -%}
+    {%- endfor %}
+    {{- "} // namespace " + namespace_name }}
+{%- endmacro -%}
+
+{%- macro render_builtin_tools(browser_tool, python_tool) -%}
+    {%- if browser_tool %}
+        {{- "## browser\n\n" }}
+        {{- "// Tool for browsing.\n" }}
+        {{- "// The `cursor` appears in brackets before each browsing display: `[{cursor}]`.\n" }}
+        {{- "// Cite information from the tool using the following format:\n" }}
+        {{- "// `【{cursor}†L{line_start}(-L{line_end})?】`, for example: `【6†L9-L11】` or `【8†L3】`.\n" }}
+        {{- "// Do not quote more than 10 words directly from the tool output.\n" }}
+        {{- "// sources=web (default: web)\n" }}
+        {{- "namespace browser {\n\n" }}
+        {{- "// Searches for information related to `query` and displays `topn` results.\n" }}
+        {{- "type search = (_: {\n" }}
+        {{- "query: string,\n" }}
+        {{- "topn?: number, // default: 10\n" }}
+        {{- "source?: string,\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "// Opens the link `id` from the page indicated by `cursor` starting at line number `loc`, showing `num_lines` lines.\n" }}
+        {{- "// Valid link ids are displayed with the formatting: `【{id}†.*】`.\n" }}
+        {{- "// If `cursor` is not provided, the most recent page is implied.\n" }}
+        {{- "// If `id` is a string, it is treated as a fully qualified URL associated with `source`.\n" }}
+        {{- "// If `loc` is not provided, the viewport will be positioned at the beginning of the document or centered on the most relevant passage, if available.\n" }}
+        {{- "// Use this function without `id` to scroll to a new location of an opened page.\n" }}
+        {{- "type open = (_: {\n" }}
+        {{- "id?: number | string, // default: -1\n" }}
+        {{- "cursor?: number, // default: -1\n" }}
+        {{- "loc?: number, // default: -1\n" }}
+        {{- "num_lines?: number, // default: -1\n" }}
+        {{- "view_source?: boolean, // default: false\n" }}
+        {{- "source?: string,\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "// Finds exact matches of `pattern` in the current page, or the page given by `cursor`.\n" }}
+        {{- "type find = (_: {\n" }}
+        {{- "pattern: string,\n" }}
+        {{- "cursor?: number, // default: -1\n" }}
+        {{- "}) => any;\n\n" }}
+        {{- "} // namespace browser\n\n" }}
+    {%- endif -%}
+
+    {%- if python_tool %}
+        {{- "## python\n\n" }}
+        {{- "Use this tool to execute Python code in your chain of thought. The code will not be shown to the user. This tool should be used for internal reasoning, but not for code that is intended to be visible to the user (e.g. when creating plots, tables, or files).\n\n" }}
+        {{- "When you send a message containing Python code to python, it will be executed in a stateful Jupyter notebook environment. python will respond with the output of the execution or time out after 120.0 seconds. The drive at '/mnt/data' can be used to save and persist user files. Internet access for this session is UNKNOWN. Depends on the cluster.\n\n" }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- System Message Construction ============================================ #}
+{%- macro build_system_message() -%}
+    {%- if model_identity is not defined %}
+        {%- set model_identity = "You are ChatGPT, a large language model trained by OpenAI." %}
+    {%- endif %}
+    {{- model_identity + "\n" }}
+    {{- "Knowledge cutoff: 2024-06\n" }}
+    {{- "Current date: " + strftime_now("%Y-%m-%d") + "\n\n" }}
+    {%- if reasoning_effort is not defined %}
+        {%- set reasoning_effort = "medium" %}
+    {%- endif %}
+    {{- "Reasoning: " + reasoning_effort + "\n\n" }}
+    {%- if builtin_tools %}
+        {{- "# Tools\n\n" }}
+        {%- set available_builtin_tools = namespace(browser=false, python=false) %}
+        {%- for tool in builtin_tools %}
+            {%- if tool == "browser" %}
+                {%- set available_builtin_tools.browser = true %}
+            {%- elif tool == "python" %}
+                {%- set available_builtin_tools.python = true %}
+            {%- endif %}
+        {%- endfor %}
+        {{- render_builtin_tools(available_builtin_tools.browser, available_builtin_tools.python) }}
+    {%- endif -%}
+    {{- "# Valid channels: analysis, commentary, final. Channel must be included for every message." }}
+    {%- if tools -%}
+        {{- "\nCalls to these tools must go to the commentary channel: 'functions'." }}
+    {%- endif -%}
+{%- endmacro -%}
+
+{#- Main Template Logic ================================================= #}
+{#- Set defaults #}
+
+{#- Render system message #}
+{{- "<|start|>system<|message|>" }}
+{{- build_system_message() }}
+{{- "<|end|>" }}
+
+{#- Extract developer message #}
+{%- if messages[0].role == "developer" or messages[0].role == "system" %}
+    {%- set developer_message = messages[0].content %}
+    {%- set loop_messages = messages[1:] %}
+{%- else %}
+    {%- set developer_message = "" %}
+    {%- set loop_messages = messages %}
+{%- endif %}
+
+{#- Render developer message #}
+{%- if developer_message or tools %}
+    {{- "<|start|>developer<|message|>" }}
+    {%- if developer_message %}
+        {{- "# Instructions\n\n" }}
+        {{- developer_message }}
+        {{- "\n\n" }}
+    {%- endif %}
+    {%- if tools -%}
+        {{- "# Tools\n\n" }}
+        {{- render_tool_namespace("functions", tools) }}
+    {%- endif -%}
+    {{- "<|end|>" }}
+{%- endif %}
+
+{#- Render messages #}
+{%- set last_tool_call = namespace(name=none) %}
+{%- for message in loop_messages -%}
+    {#- At this point only assistant/user/tool messages should remain #}
+    {%- if message.role == 'assistant' -%}
+        {#- Checks to ensure the messages are being passed in the format we expect #}
+        {%- if "content" in message %}
+            {%- if "<|channel|>analysis<|message|>" in message.content or "<|channel|>final<|message|>" in message.content %}
+                {{- raise_exception("You have passed a message containing <|channel|> tags in the content field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
+            {%- endif %}
+        {%- endif %}
+        {%- if "thinking" in message %}
+            {%- if "<|channel|>analysis<|message|>" in message.thinking or "<|channel|>final<|message|>" in message.thinking %}
+                {{- raise_exception("You have passed a message containing <|channel|> tags in the thinking field. Instead of doing this, you should pass analysis messages (the string between '<|message|>' and '<|end|>') in the 'thinking' field, and final messages (the string between '<|message|>' and '<|end|>') in the 'content' field.") }}
+            {%- endif %}
+        {%- endif %}
+        {%- if "tool_calls" in message %}
+            {#- We need very careful handling here - we want to drop the tool call analysis message if the model #}
+            {#- has output a later <|final|> message, but otherwise we want to retain it. This is the only case #}
+            {#- when we render CoT/analysis messages in inference. #}
+            {%- set future_final_message = namespace(found=false) %}
+            {%- for future_message in loop_messages[loop.index:] %}
+                {%- if future_message.role == 'assistant' and "tool_calls" not in future_message %}
+                    {%- set future_final_message.found = true %}
+                {%- endif %}
+            {%- endfor %}
+            {#- We assume max 1 tool call per message, and so we infer the tool call name #}
+            {#- in "tool" messages from the most recent assistant tool call name #}
+            {%- set tool_call = message.tool_calls[0] %}
+            {%- if tool_call.function %}
+                {%- set tool_call = tool_call.function %}
+            {%- endif %}
+            {%- if message.content and message.thinking %}
+                {{- raise_exception("Cannot pass both content and thinking in an assistant message with tool calls! Put the analysis message in one or the other, but not both.") }}
+            {%- elif message.content and not future_final_message.found %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.content + "<|end|>" }}
+            {%- elif message.thinking and not future_final_message.found %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
+            {%- endif %}
+            {{- "<|start|>assistant to=" }}
+            {{- "functions." + tool_call.name + "<|channel|>commentary " }}
+            {{- (tool_call.content_type if tool_call.content_type is defined else "json") + "<|message|>" }}
+            {{- tool_call.arguments|tojson }}
+            {{- "<|call|>" }}
+            {%- set last_tool_call.name = tool_call.name %}
+        {%- elif loop.last and not add_generation_prompt %}
+            {#- Only render the CoT if the final turn is an assistant turn and add_generation_prompt is false #}
+            {#- This is a situation that should only occur in training, never in inference. #}
+            {%- if "thinking" in message %}
+                {{- "<|start|>assistant<|channel|>analysis<|message|>" + message.thinking + "<|end|>" }}
+            {%- endif %}
+            {#- <|return|> indicates the end of generation, but <|end|> does not #}
+            {#- <|return|> should never be an input to the model, but we include it as the final token #}
+            {#- when training, so the model learns to emit it. #}
+            {{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|return|>" }}
+        {%- else %}
+            {#- CoT is dropped during all previous turns, so we never render it for inference #}
+            {{- "<|start|>assistant<|channel|>final<|message|>" + message.content + "<|end|>" }}
+            {%- set last_tool_call.name = none %}
+        {%- endif %}
+    {%- elif message.role == 'tool' -%}
+        {%- if last_tool_call.name is none %}
+            {{- raise_exception("Message has tool role, but there was no previous assistant message with a tool call!") }}
+        {%- endif %}
+        {{- "<|start|>functions." + last_tool_call.name }}
+        {{- " to=assistant<|channel|>commentary<|message|>" + message.content|tojson + "<|end|>" }}
+    {%- elif message.role == 'user' -%}
+        {{- "<|start|>user<|message|>" + message.content + "<|end|>" }}
+    {%- endif -%}
+{%- endfor -%}
+
+{#- Generation prompt #}
+{%- if add_generation_prompt -%}
+<|start|>assistant
+{%- endif -%}

scripts/sync-ggml.last

@@ -1 +1 @@
-daf7906728036a82f20c69fcbd74b6f536c74d3f
+b141fc226b68e4af383101c39da90b54ede98850

src/llama-batch.cpp

@@ -477,7 +477,7 @@ llama_ubatch llama_batch_allocr::split_simple(uint32_t n_ubatch) {
 
 llama_ubatch llama_batch_allocr::split_equal(uint32_t n_ubatch, bool sequential) {
     if (sequential && has_cpl) {
-        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch\n", __func__);
+        LLAMA_LOG_ERROR("%s: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag)\n", __func__);
 
         return {};
     }

src/llama-context.cpp

@@ -1657,30 +1657,30 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
     }
 }
 
-size_t llama_context::state_seq_get_size(llama_seq_id seq_id) {
+size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
     llama_io_write_dummy io;
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size) {
+size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
     llama_io_write_buffer io(dst, size);
     try {
-        return state_seq_write_data(io, seq_id);
+        return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
         return 0;
     }
 }
 
-size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size) {
+size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
     llama_io_read_buffer io(src, size);
     try {
-        return state_seq_read_data(io, seq_id);
+        return state_seq_read_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
@@ -1778,7 +1778,7 @@ size_t llama_context::state_seq_load_file(llama_seq_id seq_id, const char * file
     {
         const size_t state_size = file.size() - file.tell();
         llama_io_read_file io(&file);
-        const size_t nread = state_seq_read_data(io, seq_id);
+        const size_t nread = state_seq_read_data(io, seq_id, 0);
         if (!nread) {
             LLAMA_LOG_ERROR("%s: failed to restore sequence state\n", __func__);
             return 0;
@@ -1802,7 +1802,7 @@ size_t llama_context::state_seq_save_file(llama_seq_id seq_id, const char * file
 
     // save the context state using stream saving
     llama_io_write_file io(&file);
-    state_seq_write_data(io, seq_id);
+    state_seq_write_data(io, seq_id, 0);
 
     const size_t res = file.tell();
     GGML_ASSERT(res == sizeof(uint32_t) * 3 + sizeof(llama_token) * n_token_count + io.n_bytes());
@@ -1971,21 +1971,21 @@ size_t llama_context::state_read_data(llama_io_read_i & io) {
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_write(io, seq_id);
+        memory->state_write(io, seq_id, flags);
     }
 
     return io.n_bytes();
 }
 
-size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id) {
+size_t llama_context::state_seq_read_data(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
     GGML_UNUSED(seq_id);
 
     if (memory) {
-        memory->state_read(io, seq_id);
+        memory->state_read(io, seq_id, flags);
     }
 
     return io.n_bytes();
@@ -2048,7 +2048,7 @@ void llama_context::opt_init(struct llama_model * model, struct llama_opt_params
     opt_params.opt_period      = n_batch / n_ubatch;
     opt_params.get_opt_pars    = lopt_params.get_opt_pars;
     opt_params.get_opt_pars_ud = lopt_params.get_opt_pars_ud;
-
+    opt_params.optimizer       = lopt_params.optimizer_type;
     opt_ctx = ggml_opt_init(opt_params);
 
     llama_opt_param_filter param_filter = lopt_params.param_filter;
@@ -2801,19 +2801,31 @@ bool llama_state_save_file(llama_context * ctx, const char * path_session, const
 }
 
 size_t llama_state_seq_get_size(llama_context * ctx, llama_seq_id seq_id) {
-    return ctx->state_seq_get_size(seq_id);
+    return llama_state_seq_get_size_ext(ctx, seq_id, 0);
 }
 
 size_t llama_state_seq_get_data(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id) {
-    ctx->synchronize();
-
-    return ctx->state_seq_get_data(seq_id, dst, size);
+    return llama_state_seq_get_data_ext(ctx, dst, size, seq_id, 0);
 }
 
 size_t llama_state_seq_set_data(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id) {
+    return llama_state_seq_set_data_ext(ctx, src, size, seq_id, 0);
+}
+
+size_t llama_state_seq_get_size_ext(llama_context * ctx, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    return ctx->state_seq_get_size(seq_id, flags);
+}
+
+size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 
-    return ctx->state_seq_set_data(seq_id, src, size);
+    return ctx->state_seq_get_data(seq_id, dst, size, flags);
+}
+
+size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    ctx->synchronize();
+
+    return ctx->state_seq_set_data(seq_id, src, size, flags);
 }
 
 size_t llama_state_seq_save_file(llama_context * ctx, const char * filepath, llama_seq_id seq_id, const llama_token * tokens, size_t n_token_count) {

src/llama-context.h

@@ -111,9 +111,9 @@ struct llama_context {
     size_t state_get_data(      uint8_t * dst, size_t size);
     size_t state_set_data(const uint8_t * src, size_t size);
 
-    size_t state_seq_get_size(llama_seq_id seq_id);
-    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size);
-    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size);
+    size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
+    size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);
 
     bool state_load_file(
             const char * filepath,
@@ -152,6 +152,7 @@ struct llama_context {
 
     void opt_init(struct llama_model * model, struct llama_opt_params lopt_params);
 
+    // TODO: more flexible combinations of logical/physical batch size and context size
     void opt_epoch(
             ggml_opt_dataset_t      dataset,
             ggml_opt_result_t       result_train,
@@ -212,8 +213,8 @@ private:
     size_t state_write_data(llama_io_write_i & io);
     size_t state_read_data (llama_io_read_i  & io);
 
-    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id);
-    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id);
+    size_t state_seq_write_data(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags);
+    size_t state_seq_read_data (llama_io_read_i  & io, llama_seq_id seq_id, llama_state_seq_flags flags);
 
     //
     // members

src/llama-kv-cache-unified-iswa.cpp

@@ -194,14 +194,20 @@ bool llama_kv_cache_unified_iswa::get_can_shift() const {
     return kv_base->get_size() == kv_swa->get_size();
 }
 
-void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
-    kv_base->state_write(io, seq_id);
-    kv_swa ->state_write(io, seq_id);
+void llama_kv_cache_unified_iswa::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_write(io, seq_id, flags);
+    }
+
+    kv_swa->state_write(io, seq_id, flags);
 }
 
-void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
-    kv_base->state_read(io, seq_id);
-    kv_swa ->state_read(io, seq_id);
+void llama_kv_cache_unified_iswa::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    if ((flags & LLAMA_STATE_SEQ_FLAGS_SWA_ONLY) == 0) {
+        kv_base->state_read(io, seq_id, flags);
+    }
+
+    kv_swa->state_read(io, seq_id, flags);
 }
 
 llama_kv_cache_unified * llama_kv_cache_unified_iswa::get_base() const {

src/llama-kv-cache-unified-iswa.h

@@ -56,8 +56,8 @@ public:
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
     // llama_kv_cache_unified_iswa specific API

src/llama-kv-cache-unified.cpp

@@ -1828,7 +1828,9 @@ bool llama_kv_cache_unified::is_masked_swa(llama_pos p0, llama_pos p1) const {
     return false;
 }
 
-void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     io.write(&n_stream, sizeof(n_stream));
 
     for (uint32_t s = 0; s < n_stream; ++s) {
@@ -1879,7 +1881,9 @@ void llama_kv_cache_unified::state_write(llama_io_write_i & io, llama_seq_id seq
     }
 }
 
-void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_kv_cache_unified::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;

src/llama-kv-cache-unified.h

@@ -136,8 +136,8 @@ public:
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     //
     // llama_kv_cache_unified specific API

src/llama-memory-hybrid.cpp

@@ -165,12 +165,16 @@ llama_pos llama_memory_hybrid::seq_pos_max(llama_seq_id seq_id) const {
     return std::min(mem_attn->seq_pos_max(seq_id), mem_recr->seq_pos_max(seq_id));
 }
 
-void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_hybrid::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     mem_attn->state_write(io, seq_id);
     mem_recr->state_write(io, seq_id);
 }
 
-void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_hybrid::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     mem_attn->state_read(io, seq_id);
     mem_recr->state_read(io, seq_id);
 }

src/llama-memory-hybrid.h

@@ -74,8 +74,8 @@ public:
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1)       override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0)       override;
 
     //
     // llama_memory_hybrid specific API

src/llama-memory-recurrent.cpp

@@ -680,7 +680,9 @@ size_t llama_memory_recurrent::size_s_bytes() const {
     return size_s_bytes;
 }
 
-void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id) const {
+void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) const {
+    GGML_UNUSED(flags);
+
     std::vector<std::pair<uint32_t, uint32_t>> cell_ranges; // ranges, from inclusive, to exclusive
     uint32_t cell_count = 0;
 
@@ -718,7 +720,9 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
     state_write_data(io, cell_ranges);
 }
 
-void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id) {
+void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama_state_seq_flags flags) {
+    GGML_UNUSED(flags);
+
     uint32_t cell_count;
     io.read_to(&cell_count, sizeof(cell_count));
 

src/llama-memory-recurrent.h

@@ -63,8 +63,8 @@ public:
 
     // state write/load
 
-    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const override;
-    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) override;
+    void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const override;
+    void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) override;
 
     uint32_t head = 0; // the location where the batch will be placed in the cache (see find_slot())
     uint32_t size = 0; // total number of cells, shared across all sequences

src/llama-memory.h

@@ -104,8 +104,8 @@ struct llama_memory_i {
     // state write/read
     //
 
-    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1) const = 0;
-    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1) = 0;
+    virtual void state_write(llama_io_write_i & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) const = 0;
+    virtual void state_read (llama_io_read_i  & io, llama_seq_id seq_id = -1, llama_state_seq_flags flags = 0) = 0;
 };
 
 using llama_memory_ptr = std::unique_ptr<llama_memory_i>;

src/llama-model.cpp

@@ -1095,6 +1095,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
                 switch (hparams.n_layer) {
+                    case 18: type = LLM_TYPE_537M; break;
                     case 26: type = LLM_TYPE_1B; break;
                     case 34: type = LLM_TYPE_4B; break;
                     case 48: type = LLM_TYPE_12B; break;

src/llama-model.h

@@ -39,6 +39,7 @@ enum llm_type {
     LLM_TYPE_410M,
     LLM_TYPE_450M,
     LLM_TYPE_475M,
+    LLM_TYPE_537M,
     LLM_TYPE_700M,
     LLM_TYPE_770M,
     LLM_TYPE_780M,

src/llama-vocab.cpp

@@ -2341,7 +2341,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
         // @ngxson : quick hack for gpt-oss, always render these tokens
         for (const auto & t : token_to_id) {
-            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>") {
+            if (t.first == "<|channel|>" || t.first == "<|message|>" || t.first == "<|start|>" || t.first == "<|constrain|>") {
                 id_to_token[t.second].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
             }
         }
@@ -2388,6 +2388,7 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
 
             if (has_return && has_call && has_end) {
                 special_eog_ids.erase(end_id);
+                id_to_token[end_id].attr = LLAMA_TOKEN_ATTR_USER_DEFINED;
                 LLAMA_LOG_WARN("%s: special_eog_ids contains both '<|return|>' and '<|call|>' tokens, removing '<|end|>' token from EOG list\n", __func__);
             }
         }