ci, tests : use cmake to download models and remove libcurl dependency (#18791)

* ci, tests : use cmake to download models and remove libcurl dependency
* llama_dl_model -> llama_download_model
* use EXPECTED_HASH for robust model downloading
* Move llama_download_model to cmake/common.cmake

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

.devops/vulkan.Dockerfile

@@ -33,6 +33,7 @@ FROM ubuntu:$UBUNTU_VERSION AS base
 
 RUN apt-get update \
     && apt-get install -y libgomp1 curl libvulkan1 mesa-vulkan-drivers \
+    libglvnd0 libgl1 libglx0 libegl1 libgles2 \
     && apt autoremove -y \
     && apt clean -y \
     && rm -rf /tmp/* /var/tmp/* \

.github/actions/windows-setup-curl/action.yml

@@ -1,30 +0,0 @@
-name: 'Windows - Setup CURL'
-description: 'Composite action, to be reused in other workflow'
-inputs:
-  curl_version:
-    description: 'CURL version'
-    required: false
-    default: '8.6.0_6'
-  architecture:
-    description: 'Architecture of the libcurl to download'
-    required: false
-    default: 'win64'
-outputs:
-  curl_path:
-    description: "Path to the downloaded libcurl"
-    value: ${{ steps.get_libcurl.outputs.curl_path }}
-
-runs:
-  using: "composite"
-  steps:
-    - name: libCURL
-      id: get_libcurl
-      shell: powershell
-      env:
-        CURL_VERSION: ${{ inputs.curl_version }}
-        ARCHITECTURE: ${{ inputs.architecture }}
-      run: |
-        curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-${env:ARCHITECTURE}-mingw.zip"
-        mkdir $env:RUNNER_TEMP/libcurl
-        tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
-        echo "curl_path=$env:RUNNER_TEMP/libcurl" >> $env:GITHUB_OUTPUT

.github/workflows/build.yml

@@ -92,7 +92,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L 'main|curl' --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-x64:
     runs-on: macos-15-intel
@@ -152,13 +152,13 @@ jobs:
           DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.zip"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
           curl -L -o artifact.zip \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
           mkdir dawn
           unzip artifact.zip
-          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release.tar.gz -C dawn --strip-components=1
+          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -237,7 +237,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L 'main|curl' --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
       - name: Test llama2c conversion
         id: llama2c_test
@@ -532,13 +532,13 @@ jobs:
           DAWN_VERSION="v2.0.0"
           DAWN_OWNER="reeselevine"
           DAWN_REPO="dawn"
-          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.zip"
-          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
           curl -L -o artifact.zip \
-            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}"
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
           mkdir dawn
           unzip artifact.zip
-          tar -xvf Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-ubuntu-latest-Release.tar.gz -C dawn --strip-components=1
+          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
 
       - name: Build
         id: cmake_build
@@ -1418,7 +1418,6 @@ jobs:
           echo "FIXME: test on devices"
 
   openEuler-latest-cmake-cann:
-    if: ${{ github.event_name != 'pull_request' || contains(github.event.pull_request.labels.*.name, 'Ascend NPU') }}
     defaults:
       run:
         shell: bash -el {0}
@@ -1464,12 +1463,14 @@ jobs:
             "${{ steps.cann-image.outputs.image }}" \
             bash -lc '
               set -e
-              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake libcurl-devel
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
               yum clean all && rm -rf /var/cache/yum
               git config --global --add safe.directory "/workspace"
               export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
               cmake -S . -B build \
                   -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DLLAMA_CURL=OFF \
+                  -DLLAMA_OPENSSL=ON \
                   -DGGML_CANN=on \
                   -DSOC_TYPE=${SOC_TYPE}
               cmake --build build -j $(nproc)
@@ -1498,7 +1499,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Test
         id: ggml-ci
@@ -1524,7 +1525,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Test
         id: ggml-ci
@@ -1550,7 +1551,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Test
         id: ggml-ci
@@ -1576,7 +1577,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Test
         id: ggml-ci
@@ -1602,7 +1603,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential
 
       - name: Test
         id: ggml-ci
@@ -1705,6 +1706,34 @@ jobs:
         run: |
           GG_BUILD_METAL=1 bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
 
+  ggml-ci-mac-webgpu:
+    runs-on: [self-hosted, macOS, ARM64]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dawn Dependency
+        id: dawn-depends
+        run: |
+          DAWN_VERSION="v2.0.0"
+          DAWN_OWNER="reeselevine"
+          DAWN_REPO="dawn"
+          DAWN_ASSET_NAME="Dawn-5e9a4865b1635796ccc77dd30057f2b4002a1355-macos-latest-Release"
+          echo "Fetching release asset from https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          curl -L -o artifact.zip \
+            "https://github.com/${DAWN_OWNER}/${DAWN_REPO}/releases/download/${DAWN_VERSION}/${DAWN_ASSET_NAME}.zip"
+          mkdir dawn
+          unzip artifact.zip
+          tar -xvf ${DAWN_ASSET_NAME}.tar.gz -C dawn --strip-components=1
+
+      - name: Test
+        id: ggml-ci
+        run: |
+          GG_BUILD_WEBGPU=1 GG_BUILD_WEBGPU_DAWN_PREFIX="$GITHUB_WORKSPACE/dawn" \
+            bash ./ci/run.sh ~/results/llama.cpp ~/mnt/llama.cpp
+
   ggml-ci-mac-vulkan:
     runs-on: [self-hosted, macOS, ARM64]
 
@@ -1738,7 +1767,7 @@ jobs:
          id: depends
          run: |
            sudo apt-get update
-           sudo apt-get install -y build-essential libcurl4-openssl-dev
+           sudo apt-get install -y build-essential
 
        - name: Test
          id: ggml-ci
@@ -1824,7 +1853,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L 'main|curl' --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
       - name: Test llama2c conversion
         id: llama2c_test
@@ -2100,7 +2129,6 @@ jobs:
            sudo DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a \
            apt-get install -y \
             build-essential \
-            libcurl4-openssl-dev \
             python3-venv \
             gpg \
             wget \

.github/workflows/release.yml

@@ -37,13 +37,6 @@ jobs:
           key: macOS-latest-cmake-arm64
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
       - name: Build
         id: cmake_build
         run: |
@@ -52,6 +45,8 @@ jobs:
             -DCMAKE_INSTALL_RPATH='@loader_path' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DGGML_RPC=ON \
@@ -90,13 +85,6 @@ jobs:
           key: macOS-latest-cmake-x64
           evict-old-files: 1d
 
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-          brew install curl
-
       - name: Build
         id: cmake_build
         run: |
@@ -107,6 +95,8 @@ jobs:
             -DCMAKE_INSTALL_RPATH='@loader_path' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_BUILD_BORINGSSL=ON \
             -DGGML_METAL=OFF \
             -DGGML_RPC=ON \
             -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
@@ -159,7 +149,7 @@ jobs:
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential libcurl4-openssl-dev
+          sudo apt-get install build-essential libssl-dev
 
       - name: Build
         id: cmake_build
@@ -171,6 +161,8 @@ jobs:
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
             -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release -j $(nproc)
 
@@ -212,7 +204,7 @@ jobs:
           wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
           sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
           sudo apt-get update -y
-          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libcurl4-openssl-dev
+          sudo apt-get install -y build-essential mesa-vulkan-drivers vulkan-sdk libssl-dev
 
       - name: Build
         id: cmake_build
@@ -220,6 +212,8 @@ jobs:
           cmake -B build \
             -DCMAKE_INSTALL_RPATH='$ORIGIN' \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
+            -DLLAMA_CURL=OFF \
+            -DLLAMA_OPENSSL=ON \
             -DGGML_BACKEND_DL=ON \
             -DGGML_NATIVE=OFF \
             -DGGML_CPU_ALL_VARIANTS=ON \
@@ -269,34 +263,24 @@ jobs:
         run: |
           choco install ninja
 
-      - name: libCURL
-        id: get_libcurl
-        uses: ./.github/actions/windows-setup-curl
-        with:
-          architecture: ${{ matrix.arch == 'x64' && 'win64' || 'win64a' }}
-
       - name: Build
         shell: cmd
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
           call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvarsall.bat" ${{ matrix.arch == 'x64' && 'x64' || 'amd64_arm64' }}
           cmake -S . -B build -G "Ninja Multi-Config" ^
             -D CMAKE_TOOLCHAIN_FILE=cmake/${{ matrix.arch }}-windows-llvm.cmake ^
+            -DLLAMA_CURL=OFF ^
+            -DLLAMA_BUILD_BORINGSSL=ON ^
             -DGGML_NATIVE=OFF ^
             -DGGML_BACKEND_DL=ON ^
             -DGGML_CPU_ALL_VARIANTS=${{ matrix.arch == 'x64' && 'ON' || 'OFF' }} ^
             -DGGML_OPENMP=ON ^
-            -DCURL_LIBRARY="%CURL_PATH%/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="%CURL_PATH%/include" ^
             ${{ env.CMAKE_ARGS }}
           cmake --build build --config Release
 
       - name: Pack artifacts
         id: pack_artifacts
-        env:
-          CURL_PATH: ${{ steps.get_libcurl.outputs.curl_path }}
         run: |
-          Copy-Item $env:CURL_PATH\bin\libcurl-${{ matrix.arch }}.dll .\build\bin\Release\
           Copy-Item "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Redist\MSVC\14.44.35112\debug_nonredist\${{ matrix.arch }}\Microsoft.VC143.OpenMP.LLVM\libomp140.${{ matrix.arch == 'x64' && 'x86_64' || 'aarch64' }}.dll" .\build\bin\Release\
           7z a -snl llama-bin-win-cpu-${{ matrix.arch }}.zip .\build\bin\Release\*
 
@@ -744,12 +728,14 @@ jobs:
             "${{ steps.cann-image.outputs.image }}" \
             bash -lc '
               set -e
-              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake libcurl-devel
+              yum install -y --setopt=install_weak_deps=False --setopt=tsflags=nodocs git gcc gcc-c++ make cmake openssl-devel
               yum clean all && rm -rf /var/cache/yum
               git config --global --add safe.directory "/workspace"
               export LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/$(uname -m)-linux/devlib/:${LD_LIBRARY_PATH}
               cmake -S . -B build \
                   -DCMAKE_BUILD_TYPE=${BUILD_TYPE} \
+                  -DLLAMA_CURL=OFF \
+                  -DLLAMA_OPENSSL=ON \
                   -DGGML_CANN=on \
                   -DSOC_TYPE=${SOC_TYPE}
               cmake --build build -j $(nproc)

.gitignore

@@ -130,6 +130,7 @@ poetry.toml
 # Local scripts
 /run-vim.sh
 /run-chat.sh
+/run-spec.sh
 /.ccache/
 
 # IDE

CMakeLists.txt

@@ -182,6 +182,9 @@ if (NOT MSVC)
     endif()
 endif()
 
+include("cmake/license.cmake")
+license_add_file("llama.cpp" "LICENSE")
+
 #
 # 3rd-party
 #
@@ -235,6 +238,19 @@ if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS)
     add_subdirectory(tools)
 endif()
 
+# Automatically add all files from the 'licenses' directory
+file(GLOB EXTRA_LICENSES "${CMAKE_SOURCE_DIR}/licenses/LICENSE-*")
+
+foreach(FILE_PATH ${EXTRA_LICENSES})
+    get_filename_component(FILE_NAME "${FILE_PATH}" NAME)
+    string(REGEX REPLACE "^LICENSE-" "" NAME "${FILE_NAME}")
+    license_add_file("${NAME}" "${FILE_PATH}")
+endforeach()
+
+if (LLAMA_BUILD_COMMON)
+    license_generate(common)
+endif()
+
 #
 # install
 #

CONTRIBUTING.md

@@ -20,7 +20,7 @@ If AI is used to generate any portion of the code, contributors must adhere to t
 1. Explicitly disclose the manner in which AI was employed.
 2. Perform a comprehensive manual review prior to submitting the pull request.
 3. Be prepared to explain every line of code they submitted when asked about it by a maintainer.
-4. Using AI to respond to human reviewers is strictly prohibited.
+4. Using AI to write pull request descriptions or to respond to human reviewers is strictly prohibited.
 
 For more info, please refer to the [AGENTS.md](AGENTS.md) file.
 

README.md

@@ -200,6 +200,7 @@ Instructions for adding support for new models: [HOWTO-add-model.md](docs/develo
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+- [BonzAI App](https://apps.apple.com/us/app/bonzai-your-local-ai-agent/id6752847988) (proprietary)
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [Dot](https://github.com/alexpinel/Dot) (GPL)
 - [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
@@ -482,21 +483,6 @@ To learn more about model quantization, [read this documentation](tools/quantize
 
     </details>
 
-## [`llama-run`](tools/run)
-
-#### A comprehensive example for running `llama.cpp` models. Useful for inferencing. Used with RamaLama [^3].
-
-- <details>
-    <summary>Run a model with a specific prompt (by default it's pulled from Ollama registry)</summary>
-
-    ```bash
-    llama-run granite-code
-    ```
-
-    </details>
-
-[^3]: [RamaLama](https://github.com/containers/ramalama)
-
 ## [`llama-simple`](examples/simple)
 
 #### A minimal example for implementing apps with `llama.cpp`. Useful for developers.
@@ -600,7 +586,6 @@ $ echo "source ~/.llama-completion.bash" >> ~/.bashrc
 - [stb-image](https://github.com/nothings/stb) - Single-header image format decoder, used by multimodal subsystem - Public domain
 - [nlohmann/json](https://github.com/nlohmann/json) - Single-header JSON library, used by various tools/examples - MIT License
 - [minja](https://github.com/google/minja) - Minimal Jinja parser in C++, used by various tools/examples - MIT License
-- [linenoise.cpp](./tools/run/linenoise.cpp/linenoise.cpp) - C++ library that provides readline-like line editing capabilities, used by `llama-run` - BSD 2-Clause License
 - [curl](https://curl.se/) - Client-side URL transfer library, used by various tools/examples - [CURL License](https://curl.se/docs/copyright.html)
 - [miniaudio.h](https://github.com/mackron/miniaudio) - Single-header audio format decoder, used by multimodal subsystem - Public domain
 - [subprocess.h](https://github.com/sheredom/subprocess.h) - Single-header process launching solution for C and C++ - Public domain

SECURITY.md

@@ -1,12 +1,52 @@
 # Security Policy
 
+ - [**Reporting a vulnerability**](#reporting-a-vulnerability)
+ - [**Requirements**](#requirements)
+ - [**Covered Topics**](#covered-topics)
  - [**Using llama.cpp securely**](#using-llamacpp-securely)
    - [Untrusted models](#untrusted-models)
    - [Untrusted inputs](#untrusted-inputs)
    - [Data privacy](#data-privacy)
    - [Untrusted environments or networks](#untrusted-environments-or-networks)
    - [Multi-Tenant environments](#multi-tenant-environments)
- - [**Reporting a vulnerability**](#reporting-a-vulnerability)
+
+## Reporting a vulnerability
+
+If you have discovered a security vulnerability in this project that falls inside the [covered topics](#covered-topics), please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
+
+Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).
+
+A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
+
+> [!IMPORTANT]
+> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080
+
+## Requirements
+
+Before submitting your report, ensure you meet the following requirements:
+
+- You have read this policy and fully understand it.
+- AI is only permitted in an assistive capacity as stated in [AGENTS.md](AGENTS.md). We do not accept reports that are written exclusively by AI.
+- Your report must include a working Proof-of-Concept in the form of a script and/or attached files.
+
+Maintainers reserve the right to close the report if these requirements are not fulfilled.
+
+## Covered Topics
+
+Only vulnerabilities that fall within these parts of the project are considered valid. For problems falling outside of this list, please report them as issues.
+
+- `src/**/*`
+- `ggml/**/*`
+- `gguf-py/**/*`
+- `tools/server/*`, **excluding** the following topics:
+    - Web UI
+    - Features marked as experimental
+    - Features not recommended for use in untrusted environments (e.g., router, MCP)
+    - Bugs that can lead to Denial-of-Service attack
+
+Note that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities in LLaMA C++.
+
+For vulnerabilities that fall within the `vendor` directory, please report them directly to the third-party project.
 
 ## Using llama.cpp securely
 
@@ -55,19 +95,3 @@ If you intend to run multiple models in parallel with shared memory, it is your
 3. Model Sharing: In a multitenant model sharing design, tenants and users must understand the security risks of running code provided by others. Since there are no reliable methods to detect malicious models, sandboxing the model execution is the recommended approach to mitigate the risk.
 
 4. Hardware Attacks: GPUs or TPUs can also be attacked. [Researches](https://scholar.google.com/scholar?q=gpu+side+channel) has shown that side channel attacks on GPUs are possible, which can make data leak from other models or processes running on the same system at the same time.
-
-## Reporting a vulnerability
-
-Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities of LLaMA C++.
-
-<!-- normal version -->
-However, If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
-
-Please disclose it as a private [security advisory](https://github.com/ggml-org/llama.cpp/security/advisories/new).
-
-Please note that using AI to identify vulnerabilities and generate reports is permitted. However, you must (1) explicitly disclose how AI was used and (2) conduct a thorough manual review before submitting the report.
-
-A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.
-
-> [!IMPORTANT]
-> For collaborators: if you are interested in helping out with reviewing privting security disclosures, please see: https://github.com/ggml-org/llama.cpp/discussions/18080

ci/run.sh

@@ -45,7 +45,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_CURL=ON -DGGML_SCHED_NO_REALLOC=ON"
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=${LLAMA_FATAL_WARNINGS:-ON} -DLLAMA_CURL=OFF -DGGML_SCHED_NO_REALLOC=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
@@ -105,7 +105,20 @@ if [ ! -z ${GG_BUILD_VULKAN} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_WEBGPU} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_WEBGPU=1 -DGGML_METAL=OFF -DGGML_BLAS=OFF"
+
+    if [ ! -z "${GG_BUILD_WEBGPU_DAWN_PREFIX}" ]; then
+        if [ -z "${CMAKE_PREFIX_PATH}" ]; then
+            export CMAKE_PREFIX_PATH="${GG_BUILD_WEBGPU_DAWN_PREFIX}"
+        else
+            export CMAKE_PREFIX_PATH="${GG_BUILD_WEBGPU_DAWN_PREFIX}:${CMAKE_PREFIX_PATH}"
+        fi
+    fi
+
+    # For some systems, Dawn_DIR needs to be set explicitly, e.g., the lib64 path
+    if [ ! -z "${GG_BUILD_WEBGPU_DAWN_DIR}" ]; then
+        CMAKE_EXTRA="${CMAKE_EXTRA} -DDawn_DIR=${GG_BUILD_WEBGPU_DAWN_DIR}"
+    fi
 fi
 
 if [ ! -z ${GG_BUILD_MUSA} ]; then
@@ -284,7 +297,8 @@ function gg_sum_test_scripts {
 }
 
 function gg_get_model {
-    local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-f16.gguf"
+    #local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-f16.gguf"
+    local gguf_0="$MNT/models/qwen3/0.6B/ggml-model-q4_0.gguf"
     if [[ -s $gguf_0 ]]; then
         echo -n "$gguf_0"
     else

cmake/common.cmake

@@ -33,3 +33,25 @@ function(llama_add_compile_flags)
         endif()
     endif()
 endfunction()
+
+function(llama_download_model NAME HASH)
+    set(DEST "${CMAKE_BINARY_DIR}/${NAME}")
+    get_filename_component(DEST_DIR "${DEST}" DIRECTORY)
+    file(MAKE_DIRECTORY "${DEST_DIR}")
+    if(NOT EXISTS "${DEST}")
+        message(STATUS "Downloading ${NAME} from ggml-org/models...")
+    endif()
+    file(DOWNLOAD
+        "https://huggingface.co/ggml-org/models/resolve/main/${NAME}?download=true"
+        "${DEST}"
+        TLS_VERIFY ON
+        EXPECTED_HASH ${HASH}
+        STATUS status
+    )
+    list(GET status 0 code)
+    if(NOT code EQUAL 0)
+        list(GET status 1 msg)
+        message(FATAL_ERROR "Failed to download ${NAME}: ${msg}")
+    endif()
+    set(LLAMA_DOWNLOAD_MODEL "${DEST}" PARENT_SCOPE)
+endfunction()

cmake/license.cmake

@@ -0,0 +1,40 @@
+define_property(GLOBAL PROPERTY LICENSE_TEXT
+    BRIEF_DOCS "Embedded licenses"
+    FULL_DOCS  "Global string containing all aggregated licenses"
+)
+
+function(license_add_file NAME FILE)
+    if(NOT IS_ABSOLUTE "${FILE}")
+        set(FILE "${CMAKE_CURRENT_SOURCE_DIR}/${FILE}")
+    endif()
+    if(EXISTS "${FILE}")
+        set(TITLE "License for ${NAME}")
+        string(REGEX REPLACE "." "=" UNDERLINE "${TITLE}")
+        file(READ "${FILE}" TEXT)
+        get_property(TMP GLOBAL PROPERTY LICENSE_TEXT)
+        string(APPEND TMP "R\"=L=(${TITLE}\n${UNDERLINE}\n\n${TEXT})=L=\",\n")
+        set_property(GLOBAL PROPERTY LICENSE_TEXT "${TMP}")
+    else()
+        message(WARNING "License file '${FILE}' not found")
+    endif()
+endfunction()
+
+function(license_generate TARGET_NAME)
+    message(STATUS "Generating embedded license file for target: ${TARGET_NAME}")
+    get_property(TEXT GLOBAL PROPERTY LICENSE_TEXT)
+
+    set(CPP_CONTENT "// Generated by CMake\n\n")
+    string(APPEND CPP_CONTENT "const char* LICENSES[] = {\n")
+    string(APPEND CPP_CONTENT "${TEXT}")
+    string(APPEND CPP_CONTENT "nullptr\n")
+    string(APPEND CPP_CONTENT "};\n")
+
+    set(CPP_FILE "${CMAKE_BINARY_DIR}/license.cpp")
+    file(WRITE "${CPP_FILE}" "${CPP_CONTENT}")
+
+    if(TARGET ${TARGET_NAME})
+        target_sources(${TARGET_NAME} PRIVATE "${CPP_FILE}")
+    else()
+        message(FATAL_ERROR "Target '${TARGET_NAME}' does not exist")
+    endif()
+endfunction()

common/CMakeLists.txt

@@ -155,27 +155,3 @@ if (LLAMA_LLGUIDANCE)
 endif ()
 
 target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama Threads::Threads)
-
-
-#
-# copy the license files
-#
-
-# Check if running in GitHub Actions
-if (DEFINED ENV{GITHUB_ACTIONS} AND "$ENV{GITHUB_ACTIONS}" STREQUAL "true")
-    message(STATUS "Running inside GitHub Actions - copying license files")
-
-    # Copy all files from licenses/ to build/bin/
-    file(GLOB LICENSE_FILES "${CMAKE_SOURCE_DIR}/licenses/*")
-    foreach(LICENSE_FILE ${LICENSE_FILES})
-        get_filename_component(FILENAME ${LICENSE_FILE} NAME)
-        add_custom_command(
-            POST_BUILD
-            TARGET ${TARGET}
-            COMMAND ${CMAKE_COMMAND} -E copy_if_different
-                "${LICENSE_FILE}"
-                "$<TARGET_FILE_DIR:llama>/${FILENAME}"
-            COMMENT "Copying ${FILENAME} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}")
-        message(STATUS "Copying ${LICENSE_FILE} to ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${FILENAME}")
-    endforeach()
-endif()

common/arg.cpp

@@ -2,10 +2,11 @@
 
 #include "chat.h"
 #include "common.h"
+#include "download.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
 #include "sampling.h"
-#include "download.h"
+#include "preset.h"
 
 // fix problem with std::min and std::max
 #if defined(_WIN32)
@@ -47,6 +48,8 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
+extern const char * LICENSES[];
+
 using json = nlohmann::ordered_json;
 using namespace common_arg_utils;
 
@@ -268,6 +271,55 @@ static void parse_tensor_buffer_overrides(const std::string & value, std::vector
     }
 }
 
+static std::string clean_file_name(const std::string & fname) {
+    std::string clean_fname = fname;
+    string_replace_all(clean_fname, "\\", "_");
+    string_replace_all(clean_fname, "/", "_");
+    return clean_fname;
+}
+
+static bool common_params_handle_remote_preset(common_params & params, llama_example ex) {
+    GGML_ASSERT(!params.model.hf_repo.empty());
+
+    // the returned hf_repo is without tag
+    auto [hf_repo, hf_tag] = common_download_split_repo_tag(params.model.hf_repo);
+
+    // "latest" tag (default if not specified) is translated to "default" preset
+    if (hf_tag == "latest") {
+        hf_tag = "default";
+    }
+
+    const bool offline = params.offline;
+    std::string model_endpoint = get_model_endpoint();
+    auto preset_url = model_endpoint + hf_repo + "/resolve/main/preset.ini";
+
+    // prepare local path for caching
+    auto preset_fname = clean_file_name(hf_repo + "_preset.ini");
+    auto preset_path = fs_get_cache_file(preset_fname);
+    const int status = common_download_file_single(preset_url, preset_path, params.hf_token, offline);
+    const bool has_preset = status >= 200 && status < 400;
+
+    // remote preset is optional, so we don't error out if not found
+    if (has_preset) {
+        LOG_INF("applying remote preset from %s\n", preset_url.c_str());
+        common_preset_context ctx(ex, /* only_remote_allowed */ true);
+        common_preset global;
+        auto remote_presets = ctx.load_from_ini(preset_path, global);
+        remote_presets = ctx.cascade(global, remote_presets);
+        if (remote_presets.find(hf_tag) != remote_presets.end()) {
+            common_preset preset = remote_presets.at(hf_tag);
+            LOG_INF("\n%s", preset.to_ini().c_str()); // to_ini already added trailing newline
+            preset.apply_to_params(params);
+        } else {
+            throw std::runtime_error("Remote preset.ini does not contain [" + std::string(hf_tag) + "] section");
+        }
+    } else {
+        LOG_INF("%s", "no remote preset found, skipping\n");
+    }
+
+    return has_preset;
+}
+
 struct handle_model_result {
     bool found_mmproj = false;
     common_params_model mmproj;
@@ -309,9 +361,7 @@ static handle_model_result common_params_handle_model(
             // make sure model path is present (for caching purposes)
             if (model.path.empty()) {
                 // this is to avoid different repo having same file name, or same file name in different subdirs
-                std::string filename = model.hf_repo + "_" + model.hf_file;
-                // to make sure we don't have any slashes in the filename
-                string_replace_all(filename, "/", "_");
+                std::string filename = clean_file_name(model.hf_repo + "_" + model.hf_file);
                 model.path = fs_get_cache_file(filename);
             }
 
@@ -425,61 +475,87 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         }
     };
 
-    std::set<std::string> seen_args;
+    auto parse_cli_args = [&]() {
+        std::set<std::string> seen_args;
 
-    for (int i = 1; i < argc; i++) {
-        const std::string arg_prefix = "--";
+        for (int i = 1; i < argc; i++) {
+            const std::string arg_prefix = "--";
 
-        std::string arg = argv[i];
-        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
-            std::replace(arg.begin(), arg.end(), '_', '-');
-        }
-        if (arg_to_options.find(arg) == arg_to_options.end()) {
-            throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
-        }
-        if (!seen_args.insert(arg).second) {
-            LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
-        }
-        auto & tmp = arg_to_options[arg];
-        auto opt = *tmp.first;
-        bool is_positive = tmp.second;
-        if (opt.has_value_from_env()) {
-            fprintf(stderr, "warn: %s environment variable is set, but will be overwritten by command line argument %s\n", opt.env, arg.c_str());
-        }
-        try {
-            if (opt.handler_void) {
-                opt.handler_void(params);
-                continue;
+            std::string arg = argv[i];
+            if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+                std::replace(arg.begin(), arg.end(), '_', '-');
             }
-            if (opt.handler_bool) {
-                opt.handler_bool(params, is_positive);
-                continue;
+            if (arg_to_options.find(arg) == arg_to_options.end()) {
+                throw std::invalid_argument(string_format("error: invalid argument: %s", arg.c_str()));
             }
+            if (!seen_args.insert(arg).second) {
+                LOG_WRN("DEPRECATED: argument '%s' specified multiple times, use comma-separated values instead (only last value will be used)\n", arg.c_str());
+            }
+            auto & tmp = arg_to_options[arg];
+            auto opt = *tmp.first;
+            bool is_positive = tmp.second;
+            if (opt.has_value_from_env()) {
+                fprintf(stderr, "warn: %s environment variable is set, but will be overwritten by command line argument %s\n", opt.env, arg.c_str());
+            }
+            try {
+                if (opt.handler_void) {
+                    opt.handler_void(params);
+                    continue;
+                }
+                if (opt.handler_bool) {
+                    opt.handler_bool(params, is_positive);
+                    continue;
+                }
 
-            // arg with single value
-            check_arg(i);
-            std::string val = argv[++i];
-            if (opt.handler_int) {
-                opt.handler_int(params, std::stoi(val));
-                continue;
-            }
-            if (opt.handler_string) {
-                opt.handler_string(params, val);
-                continue;
-            }
+                // arg with single value
+                check_arg(i);
+                std::string val = argv[++i];
+                if (opt.handler_int) {
+                    opt.handler_int(params, std::stoi(val));
+                    continue;
+                }
+                if (opt.handler_string) {
+                    opt.handler_string(params, val);
+                    continue;
+                }
 
-            // arg with 2 values
-            check_arg(i);
-            std::string val2 = argv[++i];
-            if (opt.handler_str_str) {
-                opt.handler_str_str(params, val, val2);
-                continue;
+                // arg with 2 values
+                check_arg(i);
+                std::string val2 = argv[++i];
+                if (opt.handler_str_str) {
+                    opt.handler_str_str(params, val, val2);
+                    continue;
+                }
+            } catch (std::exception & e) {
+                throw std::invalid_argument(string_format(
+                    "error while handling argument \"%s\": %s\n\n"
+                    "usage:\n%s\n\nto show complete usage, run with -h",
+                    arg.c_str(), e.what(), opt.to_string().c_str()));
             }
-        } catch (std::exception & e) {
-            throw std::invalid_argument(string_format(
-                "error while handling argument \"%s\": %s\n\n"
-                "usage:\n%s\n\nto show complete usage, run with -h",
-                arg.c_str(), e.what(), opt.to_string().c_str()));
+        }
+    };
+
+    // parse the first time to get -hf option (used for remote preset)
+    parse_cli_args();
+
+    // maybe handle remote preset
+    if (!params.model.hf_repo.empty()) {
+        std::string cli_hf_repo = params.model.hf_repo;
+        bool has_preset = common_params_handle_remote_preset(params, ctx_arg.ex);
+
+        // special case: if hf_repo explicitly set by preset, we need to preserve it (ignore CLI value)
+        // this is useful when we have one HF repo pointing to other HF repos (one model - multiple GGUFs)
+        std::string preset_hf_repo = params.model.hf_repo;
+        bool preset_has_hf_repo = preset_hf_repo != cli_hf_repo;
+
+        if (has_preset) {
+            // re-parse CLI args to override preset values
+            parse_cli_args();
+        }
+
+        // preserve hf_repo from preset if needed
+        if (preset_has_hf_repo) {
+            params.model.hf_repo = preset_hf_repo;
         }
     }
 
@@ -679,7 +755,6 @@ static void common_params_print_completion(common_params_context & ctx_arg) {
         "llama-quantize",
         "llama-qwen2vl-cli",
         "llama-retrieval",
-        "llama-run",
         "llama-save-load-state",
         "llama-server",
         "llama-simple",
@@ -966,6 +1041,16 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             exit(0);
         }
     ));
+    add_opt(common_arg(
+        {"--license"},
+        "show source code license and dependencies",
+        [](common_params &) {
+            for (int i = 0; LICENSES[i]; ++i) {
+                printf("%s\n", LICENSES[i]);
+            }
+            exit(0);
+        }
+    ));
     add_opt(common_arg(
         {"-cl", "--cache-list"},
         "show list of models in cache",
@@ -1210,7 +1295,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.kv_unified = true;
         }
-    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_env("LLAMA_ARG_KV_UNIFIED").set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_BATCHED}));
     add_opt(common_arg(
         {"--context-shift"},
         {"--no-context-shift"},
@@ -1445,7 +1530,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, bool value) {
             params.warmup = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY}));
+    ).set_examples({LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_MTMD, LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_PERPLEXITY, LLAMA_EXAMPLE_DEBUG}));
     add_opt(common_arg(
         {"--spm-infill"},
         string_format(
@@ -1761,7 +1846,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             else if (value == "rank") { params.pooling_type = LLAMA_POOLING_TYPE_RANK; }
             else { throw std::invalid_argument("invalid value"); }
         }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_DEBUG}).set_env("LLAMA_ARG_POOLING"));
     add_opt(common_arg(
         {"--attention"}, "{causal,non-causal}",
         "attention type for embeddings, use model default if unspecified",
@@ -2089,11 +2174,22 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     add_opt(common_arg(
         {"--mmap"},
         {"--no-mmap"},
-        string_format("whether to memory-map model (if disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
+        string_format("whether to memory-map model. Explicitly enabling mmap disables direct-io. (if mmap disabled, slower load but may reduce pageouts if not using mlock) (default: %s)", params.use_mmap ? "enabled" : "disabled"),
         [](common_params & params, bool value) {
             params.use_mmap = value;
+            if (value) {
+                params.use_direct_io = false;  // disable direct io when mmap is explicitly enabled
+            }
         }
     ).set_env("LLAMA_ARG_MMAP"));
+    add_opt(common_arg(
+        {"-dio", "--direct-io"},
+        {"-ndio", "--no-direct-io"},
+        string_format("use DirectIO if available. Takes precedence over --mmap (default: %s)", params.use_direct_io ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.use_direct_io = value;
+        }
+    ).set_env("LLAMA_ARG_DIO"));
     add_opt(common_arg(
         {"--numa"}, "TYPE",
         "attempt optimizations that help on some NUMA systems\n"
@@ -2245,7 +2341,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             std::vector<std::string> split_arg{ it, {} };
             if (split_arg.size() >= llama_max_devices()) {
                 throw std::invalid_argument(
-                    string_format("got %d input configs, but system only has %d devices", (int)split_arg.size(), (int)llama_max_devices())
+                    string_format("got %zu input configs, but system only has %zu devices", split_arg.size(), llama_max_devices())
                 );
             }
             for (size_t i = 0; i < llama_max_devices(); ++i) {
@@ -2285,10 +2381,28 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_env("LLAMA_ARG_FIT"));
     add_opt(common_arg(
-        { "-fitt", "--fit-target" }, "MiB",
-        string_format("target margin per device for --fit option, default: %zu", params.fit_params_target/(1024*1024)),
-        [](common_params & params, int value) {
-            params.fit_params_target = value * size_t(1024*1024);
+        { "-fitt", "--fit-target" }, "MiB0,MiB1,MiB2,...",
+        string_format("target margin per device for --fit, comma-separated list of values, "
+            "single value is broadcast across all devices, default: %zu", params.fit_params_target[0]/(1024*1024)),
+        [](common_params & params, const std::string & value) {
+            std::string arg_next = value;
+
+            // split string by , and /
+            const std::regex regex{ R"([,/]+)" };
+            std::sregex_token_iterator it{ arg_next.begin(), arg_next.end(), regex, -1 };
+            std::vector<std::string> split_arg{ it, {} };
+            if (split_arg.size() >= llama_max_devices()) {
+                throw std::invalid_argument(
+                    string_format("got %zu input configs, but system only has %zu devices", split_arg.size(), llama_max_devices())
+                );
+            }
+            if (split_arg.size() == 1) {
+                std::fill(params.fit_params_target.begin(), params.fit_params_target.end(), std::stoul(split_arg[0]) * 1024*1024);
+                return;
+            }
+            for (size_t i = 0; i < split_arg.size(); i++) {
+                params.fit_params_target[i] = std::stoul(split_arg[i]) * 1024*1024;
+            }
         }
     ).set_env("LLAMA_ARG_FIT_TARGET"));
     add_opt(common_arg(
@@ -2609,7 +2723,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.embd_normalize = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
+    ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_DEBUG}));
     add_opt(common_arg(
         {"--embd-output-format"}, "FORMAT",
         "empty = default, \"array\" = [[],[]...], \"json\" = openai style, \"json+\" = same \"json\" + cosine similarity matrix, \"raw\" = plain whitespace-delimited output (one embedding per line)",
@@ -2687,7 +2801,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params) {
             params.embedding = true;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_EMBEDDINGS"));
+    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_DEBUG}).set_env("LLAMA_ARG_EMBEDDINGS"));
     add_opt(common_arg(
         {"--rerank", "--reranking"},
         string_format("enable reranking endpoint on server (default: %s)", "disabled"),
@@ -2763,10 +2877,18 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.n_threads_http = value;
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
+    add_opt(common_arg(
+        {"--cache-prompt"},
+        {"--no-cache-prompt"},
+        string_format("whether to enable prompt caching (default: %s)", params.cache_prompt ? "enabled" : "disabled"),
+        [](common_params & params, bool value) {
+            params.cache_prompt = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CACHE_PROMPT"));
     add_opt(common_arg(
         {"--cache-reuse"}, "N",
         string_format(
-            "min chunk size to attempt reusing from the cache via KV shifting (default: %d)\n"
+            "min chunk size to attempt reusing from the cache via KV shifting, requires prompt caching to be enabled (default: %d)\n"
             "[(card)](https://ggml.ai/f0.png)", params.n_cache_reuse
         ),
         [](common_params & params, int value) {
@@ -3378,6 +3500,27 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             }
         }
     ).set_examples({ LLAMA_EXAMPLE_FINETUNE }));
+    add_opt(common_arg(
+        {"--save-logits"},
+        string_format("save final logits to files for verification (default: %s)", params.save_logits ? "true" : "false"),
+        [](common_params & params) {
+            params.save_logits = true;
+        }
+    ).set_examples({LLAMA_EXAMPLE_DEBUG}));
+    add_opt(common_arg(
+        {"--logits-output-dir"}, "PATH",
+        string_format("directory for saving logits output files (default: %s)", params.logits_output_dir.c_str()),
+        [](common_params & params, const std::string & value) {
+            params.logits_output_dir = value;
+        }
+    ).set_examples({LLAMA_EXAMPLE_DEBUG}));
+    add_opt(common_arg(
+        {"--tensor-filter"}, "REGEX",
+        "filter tensor names for debug output (regex pattern, can be specified multiple times)",
+        [](common_params & params, const std::string & value) {
+            params.tensor_filter.push_back(value);
+        }
+    ).set_examples({LLAMA_EXAMPLE_DEBUG}));
 
     // presets
     add_opt(common_arg(

common/arg.h

@@ -129,11 +129,3 @@ void common_params_add_preset_options(std::vector<common_arg> & args);
 
 // initialize argument parser context - used by test-arg-parser and preset
 common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
-
-struct common_remote_params {
-    std::vector<std::string> headers;
-    long timeout = 0; // CURLOPT_TIMEOUT, in seconds ; 0 means no timeout
-    long max_size = 0; // max size of the response ; unlimited if 0 ; max is 2GB
-};
-// get remote file content, returns <http_code, raw_response_body>
-std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);

common/chat-parser.cpp

@@ -1403,6 +1403,118 @@ static void common_chat_parse_solar_open(common_chat_msg_parser & builder) {
     builder.add_content(builder.consume_rest());
 }
 
+static void common_chat_parse_exaone_moe_content(common_chat_msg_parser & builder) {
+    // 1) <tool_call>{ "name": "...", "arguments": {...} }</tool_call>
+    // 2) <tool_call>{ "id": "...", "type": "function", "function": { "name": "...", "arguments": {...} } }</tool_call>
+    static const common_regex tool_call_open(R"(<tool_call[^>]*>)");
+
+    if (!builder.syntax().parse_tool_calls) {
+        LOG_DBG("%s: not parse_tool_calls\n", __func__);
+        builder.add_content(builder.consume_rest());
+        return;
+    }
+
+    LOG_DBG("%s: parse_tool_calls\n", __func__);
+
+    // Find all <tool_call></tool_call> blocks
+    while (auto first = builder.try_find_regex(tool_call_open, std::string::npos, /* add_prelude_to_content= */ true)) {
+        builder.move_to(first->groups[0].end);
+        builder.consume_spaces();
+
+        builder.try_consume_literal("```json");
+        builder.try_consume_literal("```");
+        builder.consume_spaces();
+
+        // Consume JSON object
+        auto data = builder.consume_json();
+
+        builder.consume_spaces();
+        builder.try_consume_literal("```");
+        builder.consume_spaces();
+
+        if (!builder.try_consume_literal("</tool_call>")) {
+            throw common_chat_msg_partial_exception("incomplete tool call");
+        }
+        builder.consume_spaces();
+
+        // Extract name and arguments
+        std::string name;
+        std::string id;
+        nlohmann::ordered_json arguments;
+
+        const auto extract_args = [&](const nlohmann::ordered_json & obj) -> bool {
+            if (!obj.contains("name") || !obj.contains("arguments")) {
+                return false;
+            }
+            name = obj.at("name").get<std::string>();
+            arguments = obj.at("arguments");
+            if (obj.contains("id") && obj.at("id").is_string()) {
+                id = obj.at("id").get<std::string>();
+            }
+            return true;
+        };
+
+        if (!extract_args(data.json)) {
+            if (data.json.contains("function") && data.json.at("function").is_object()) {
+                auto fn = data.json.at("function");
+                extract_args(fn);
+                if (id.empty() && data.json.contains("id") && data.json.at("id").is_string()) {
+                    id = data.json.at("id").get<std::string>();
+                }
+            }
+        }
+
+        // If name is empty, treat the JSON object as content
+        if (name.empty()) {
+            LOG_DBG("%s: tool call missing name, treating as content\n", __func__);
+            builder.add_content(data.json.dump());
+            continue;
+        }
+
+        std::string args_str = arguments.dump();
+        if (!builder.add_tool_call(name, id, args_str)) {
+            throw common_chat_msg_partial_exception("incomplete tool call");
+        }
+    }
+
+    builder.add_content(builder.consume_rest());
+}
+
+static void common_chat_parse_exaone_moe(common_chat_msg_parser & builder) {
+    LOG_DBG("%s: parsing exaone_moe\n", __func__);
+    // EXAONE MoE outputs reasoning content between "<think>" and "</think>" tags, followed by regular content
+    // First try to parse using the standard reasoning parsing method
+    LOG_DBG("%s: thinking_forced_open: %s\n", __func__, std::to_string(builder.syntax().thinking_forced_open).c_str());
+
+    auto start_pos = builder.pos();
+    auto found_end_think = builder.try_find_literal("</think>");
+    builder.move_to(start_pos);
+
+    if (builder.syntax().thinking_forced_open && !builder.is_partial() && !found_end_think) {
+        LOG_DBG("%s: no end_think, not partial, adding content\n", __func__);
+        common_chat_parse_exaone_moe_content(builder);
+    } else if (builder.try_parse_reasoning("<think>", "</think>")) {
+        // If reasoning was parsed successfully, the remaining content is regular content
+        LOG_DBG("%s: parsed reasoning, adding content\n", __func__);
+        common_chat_parse_exaone_moe_content(builder);
+    } else {
+        if (builder.syntax().reasoning_format == COMMON_REASONING_FORMAT_NONE) {
+          LOG_DBG("%s: reasoning_format none, adding content\n", __func__);
+          common_chat_parse_exaone_moe_content(builder);
+          return;
+        }
+        // If no reasoning tags found, check if we should treat everything as reasoning
+        if (builder.syntax().thinking_forced_open) {
+            // If thinking is forced open but no tags found, treat everything as reasoning
+            LOG_DBG("%s: thinking_forced_open, adding reasoning content\n", __func__);
+            builder.add_reasoning_content(builder.consume_rest());
+        } else {
+            LOG_DBG("%s: no thinking_forced_open, adding content\n", __func__);
+            common_chat_parse_exaone_moe_content(builder);
+        }
+    }
+}
+
 static void common_chat_parse_content_only(common_chat_msg_parser & builder) {
     builder.try_parse_reasoning("<think>", "</think>");
     builder.add_content(builder.consume_rest());
@@ -1490,6 +1602,9 @@ static void common_chat_parse(common_chat_msg_parser & builder) {
         case COMMON_CHAT_FORMAT_SOLAR_OPEN:
             common_chat_parse_solar_open(builder);
             break;
+        case COMMON_CHAT_FORMAT_EXAONE_MOE:
+            common_chat_parse_exaone_moe(builder);
+            break;
         default:
             throw std::runtime_error(std::string("Unsupported format: ") + common_chat_format_name(builder.syntax().format));
     }

common/chat.cpp

@@ -670,6 +670,7 @@ const char * common_chat_format_name(common_chat_format format) {
         case COMMON_CHAT_FORMAT_APRIEL_1_5: return "Apriel 1.5";
         case COMMON_CHAT_FORMAT_XIAOMI_MIMO: return "Xiaomi MiMo";
         case COMMON_CHAT_FORMAT_SOLAR_OPEN: return "Solar Open";
+        case COMMON_CHAT_FORMAT_EXAONE_MOE: return "EXAONE MoE";
         case COMMON_CHAT_FORMAT_PEG_SIMPLE: return "peg-simple";
         case COMMON_CHAT_FORMAT_PEG_NATIVE: return "peg-native";
         case COMMON_CHAT_FORMAT_PEG_CONSTRUCTED: return "peg-constructed";
@@ -2539,6 +2540,65 @@ static common_chat_params common_chat_params_init_solar_open(const common_chat_t
     return data;
 }
 
+static common_chat_params common_chat_params_init_exaone_moe(const common_chat_template & tmpl, const struct templates_params & inputs) {
+    common_chat_params data;
+
+    data.prompt = apply(tmpl, inputs);
+    data.format = COMMON_CHAT_FORMAT_EXAONE_MOE;
+    if (string_ends_with(data.prompt, "<think>\n")) {
+        if (!inputs.enable_thinking) {
+            data.prompt += "</think>\n\n";
+        } else {
+            data.thinking_forced_open = true;
+        }
+    }
+
+    if (inputs.tools.is_array() && !inputs.tools.empty()) {
+        data.grammar_lazy = inputs.tool_choice != COMMON_CHAT_TOOL_CHOICE_REQUIRED && inputs.json_schema.is_null();
+        data.grammar = build_grammar([&](const common_grammar_builder & builder) {
+            std::vector<std::string> tool_rules;
+            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);
+                // Expect: <tool_call>{"name": "<name>", "arguments": {...}}</tool_call>
+                tool_rules.push_back(builder.add_rule(
+                    name + "-call",
+                    "\"<tool_call>\" space " +
+                        builder.add_schema(name + "-obj", json{
+                            {"type", "object"},
+                            {"properties", {
+                                {"name",      json{{"const", name}}},
+                                {"arguments", parameters},
+                            }},
+                            {"required", json::array({"name", "arguments"})},
+                        }) +
+                    " space \"</tool_call>\" space"));
+            });
+
+            auto tool_call = builder.add_rule("tool_call", string_join(tool_rules, " | "));
+            builder.add_rule("root",
+                std::string(data.thinking_forced_open ? "( \"</think>\" space )? " : "") +
+                (inputs.parallel_tool_calls ? "(" + tool_call + ")+" : tool_call));
+
+            data.grammar_triggers.push_back({
+                COMMON_GRAMMAR_TRIGGER_TYPE_PATTERN_FULL,
+                std::string(data.thinking_forced_open ? "[\\s\\S]*?(</think>\\s*)?" : "") +
+                    "(<tool_call>)[\\s\\S]*"
+            });
+            data.preserved_tokens = {
+                "<think>",
+                "</think>",
+                "<tool_call>",
+                "</tool_call>",
+            };
+        });
+    }
+
+    return data;
+}
+
 static common_chat_params common_chat_params_init_without_tools(const common_chat_template & tmpl, const struct templates_params & inputs) {
     common_chat_params data;
     data.prompt = apply(tmpl, inputs);
@@ -2709,6 +2769,13 @@ static common_chat_params common_chat_templates_apply_jinja(
         return common_chat_params_init_xiaomi_mimo(tmpl, params);
     }
 
+    // EXAONE MoE format detection
+    if (src.find("<tool_call>") != std::string::npos &&
+        src.find("<tool_result>") != std::string::npos &&
+        src.find("<|tool_declare|>") != std::string::npos) {
+        return common_chat_params_init_exaone_moe(tmpl, params);
+    }
+
     // Hermes 2/3 Pro, Qwen 2.5 Instruct (w/ tools)
     if (src.find("<tool_call>") != std::string::npos && params.json_schema.is_null()) {
         return common_chat_params_init_hermes_2_pro(tmpl, params);

common/chat.h

@@ -125,6 +125,7 @@ enum common_chat_format {
     COMMON_CHAT_FORMAT_APRIEL_1_5,
     COMMON_CHAT_FORMAT_XIAOMI_MIMO,
     COMMON_CHAT_FORMAT_SOLAR_OPEN,
+    COMMON_CHAT_FORMAT_EXAONE_MOE,
 
     // These are intended to be parsed by the PEG parser
     COMMON_CHAT_FORMAT_PEG_SIMPLE,

common/common.cpp

@@ -1097,7 +1097,7 @@ common_init_result::common_init_result(common_params & params) :
     if (params.fit_params) {
         LOG_INF("%s: fitting params to device memory, for bugs during this step try to reproduce them with -fit off, or provide --verbose logs if the bug only occurs with -fit on\n", __func__);
         llama_params_fit(params.model.path.c_str(), &mparams, &cparams,
-            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target, params.fit_params_min_ctx,
+            params.tensor_split, params.tensor_buft_overrides.data(), params.fit_params_target.data(), params.fit_params_min_ctx,
             params.verbosity >= 4 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_ERROR);
     }
 
@@ -1366,6 +1366,7 @@ struct llama_model_params common_model_params_to_llama(common_params & params) {
     mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
+    mparams.use_direct_io   = params.use_direct_io;
     mparams.use_mlock       = params.use_mlock;
     mparams.check_tensors   = params.check_tensors;
     mparams.use_extra_bufts = !params.no_extra_bufts;

common/common.h

@@ -80,6 +80,8 @@ int32_t cpu_get_num_math();
 //
 
 enum llama_example {
+    LLAMA_EXAMPLE_BATCHED,
+    LLAMA_EXAMPLE_DEBUG,
     LLAMA_EXAMPLE_COMMON,
     LLAMA_EXAMPLE_SPECULATIVE,
     LLAMA_EXAMPLE_COMPLETION,
@@ -331,12 +333,14 @@ struct common_params {
     // offload params
     std::vector<ggml_backend_dev_t> devices; // devices to use for offloading
 
-    int32_t n_gpu_layers       = -1;               // number of layers to store in VRAM, -1 is auto, <= -2 is all
-    int32_t main_gpu           = 0;                // the GPU that is used for scratch and small tensors
-    float   tensor_split[128]  = {0};              // how split tensors should be distributed across GPUs
-    bool    fit_params         = true;             // whether to fit unset model/context parameters to free device memory
-    size_t  fit_params_target  = 1024 * 1024*1024; // margin per device in bytes for fitting parameters to free memory
-    int32_t fit_params_min_ctx = 4096;             // minimum context size to set when trying to reduce memory use
+    int32_t n_gpu_layers       = -1;   // number of layers to store in VRAM, -1 is auto, <= -2 is all
+    int32_t main_gpu           = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]  = {0};  // how split tensors should be distributed across GPUs
+    bool    fit_params         = true; // whether to fit unset model/context parameters to free device memory
+    int32_t fit_params_min_ctx = 4096; // minimum context size to set when trying to reduce memory use
+
+    // margin per device in bytes for fitting parameters to free memory:
+    std::vector<size_t> fit_params_target = std::vector<size_t>(llama_max_devices(), 1024 * 1024*1024);
 
     enum llama_split_mode split_mode = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
 
@@ -372,6 +376,11 @@ struct common_params {
     std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
     std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
 
+    // llama-debug specific options
+    std::string logits_output_dir = "data"; // directory for saving logits output files                     // NOLINT
+    bool        save_logits       = false;  // whether to save logits to files                              // NOLINT
+    std::vector<std::string> tensor_filter; // filter tensor names for debug output (regex)                 // NOLINT
+
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
@@ -422,7 +431,8 @@ struct common_params {
     bool kv_unified        = false; // enable unified KV cache
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool use_mmap          = true;  // use mmap for faster loads
+    bool use_mmap          = true;  // enable mmap to use filesystem cache
+    bool use_direct_io     = true;  // read from disk without buffering for faster model loading
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
@@ -466,6 +476,7 @@ struct common_params {
     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
+    bool    cache_prompt      = true;         // whether to enable prompt caching
     int32_t n_ctx_checkpoints = 8;            // max number of context checkpoints per slot
     int32_t cache_ram_mib     = 8192;         // -1 = no limit, 0 - disable, 1 = 1 MiB, etc.
 

common/download.cpp

@@ -157,6 +157,20 @@ static std::string read_etag(const std::string & path) {
     return none;
 }
 
+static bool is_http_status_ok(int status) {
+    return status >= 200 && status < 400;
+}
+
+std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag) {
+    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
+    std::string tag = parts.size() > 1 ? parts.back() : "latest";
+    std::string hf_repo = parts[0];
+    if (string_split<std::string>(hf_repo, '/').size() != 2) {
+        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
+    }
+    return {hf_repo, tag};
+}
+
 #ifdef LLAMA_USE_CURL
 
 //
@@ -306,11 +320,14 @@ static bool common_download_head(CURL *              curl,
 }
 
 // download one single file from remote URL to local path
-static bool common_download_file_single_online(const std::string & url,
+// returns status code or -1 on error
+static int common_download_file_single_online(const std::string & url,
                                                const std::string & path,
-                                               const std::string & bearer_token) {
+                                               const std::string & bearer_token,
+                                               const common_header_list & custom_headers) {
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
+
     for (int i = 0; i < max_attempts; ++i) {
         std::string etag;
 
@@ -330,6 +347,11 @@ static bool common_download_file_single_online(const std::string & url,
         common_load_model_from_url_headers headers;
         curl_easy_setopt(curl.get(), CURLOPT_HEADERDATA, &headers);
         curl_slist_ptr http_headers;
+
+        for (const auto & h : custom_headers) {
+             std::string s = h.first + ": " + h.second;
+             http_headers.ptr = curl_slist_append(http_headers.ptr, s.c_str());
+        }
         const bool     was_perform_successful = common_download_head(curl.get(), http_headers, url, bearer_token);
         if (!was_perform_successful) {
             head_request_ok = false;
@@ -365,7 +387,7 @@ static bool common_download_file_single_online(const std::string & url,
                 LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
                 if (remove(path.c_str()) != 0) {
                     LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                    return false;
+                    return -1;
                 }
             }
 
@@ -374,14 +396,14 @@ static bool common_download_file_single_online(const std::string & url,
                 if (std::filesystem::exists(path_temporary)) {
                     if (remove(path_temporary.c_str()) != 0) {
                         LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
-                        return false;
+                        return -1;
                     }
                 }
 
                 if (std::filesystem::exists(path)) {
                     if (remove(path.c_str()) != 0) {
                         LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                        return false;
+                        return -1;
                     }
                 }
             }
@@ -408,23 +430,27 @@ static bool common_download_file_single_online(const std::string & url,
 
             long http_code = 0;
             curl_easy_getinfo(curl.get(), CURLINFO_RESPONSE_CODE, &http_code);
-            if (http_code < 200 || http_code >= 400) {
+
+            int status = static_cast<int>(http_code);
+            if (!is_http_status_ok(http_code)) {
                 LOG_ERR("%s: invalid http status code received: %ld\n", __func__, http_code);
-                return false;
+                return status; // TODO: maybe only return on certain codes
             }
 
             if (rename(path_temporary.c_str(), path.c_str()) != 0) {
                 LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-                return false;
+                return -1;
             }
+
+            return static_cast<int>(http_code);
         } else {
             LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
-        }
 
-        break;
+            return 304; // Not Modified - fake cached response
+        }
     }
 
-    return true;
+    return -1; // max attempts reached
 }
 
 std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params) {
@@ -454,8 +480,10 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string &
         curl_easy_setopt(curl.get(), CURLOPT_MAXFILESIZE, params.max_size);
     }
     http_headers.ptr = curl_slist_append(http_headers.ptr, "User-Agent: llama-cpp");
+
     for (const auto & header : params.headers) {
-        http_headers.ptr = curl_slist_append(http_headers.ptr, header.c_str());
+        std::string header_ = header.first + ": " + header.second;
+        http_headers.ptr = curl_slist_append(http_headers.ptr, header_.c_str());
     }
     curl_easy_setopt(curl.get(), CURLOPT_HTTPHEADER, http_headers.ptr);
 
@@ -617,9 +645,11 @@ static bool common_pull_file(httplib::Client & cli,
 }
 
 // download one single file from remote URL to local path
-static bool common_download_file_single_online(const std::string & url,
+// returns status code or -1 on error
+static int common_download_file_single_online(const std::string & url,
                                                const std::string & path,
-                                               const std::string & bearer_token) {
+                                               const std::string & bearer_token,
+                                               const common_header_list & custom_headers) {
     static const int max_attempts        = 3;
     static const int retry_delay_seconds = 2;
 
@@ -629,6 +659,9 @@ static bool common_download_file_single_online(const std::string & url,
     if (!bearer_token.empty()) {
         default_headers.insert({"Authorization", "Bearer " + bearer_token});
     }
+    for (const auto & h : custom_headers) {
+        default_headers.emplace(h.first, h.second);
+    }
     cli.set_default_headers(default_headers);
 
     const bool file_exists = std::filesystem::exists(path);
@@ -647,8 +680,10 @@ static bool common_download_file_single_online(const std::string & url,
             LOG_WRN("%s: HEAD invalid http status code received: %d\n", __func__, head ? head->status : -1);
             if (file_exists) {
                 LOG_INF("%s: Using cached file (HEAD failed): %s\n", __func__, path.c_str());
-                return true;
+                return 304; // 304 Not Modified - fake cached response
             }
+            return head->status; // cannot use cached file, return raw status code
+            // TODO: maybe retry only on certain codes
         }
 
         std::string etag;
@@ -680,12 +715,12 @@ static bool common_download_file_single_online(const std::string & url,
         if (file_exists) {
             if (!should_download_from_scratch) {
                 LOG_INF("%s: using cached file: %s\n", __func__, path.c_str());
-                return true;
+                return 304; // 304 Not Modified - fake cached response
             }
             LOG_WRN("%s: deleting previous downloaded file: %s\n", __func__, path.c_str());
             if (remove(path.c_str()) != 0) {
                 LOG_ERR("%s: unable to delete file: %s\n", __func__, path.c_str());
-                return false;
+                return -1;
             }
         }
 
@@ -697,7 +732,7 @@ static bool common_download_file_single_online(const std::string & url,
                 existing_size = std::filesystem::file_size(path_temporary);
             } else if (remove(path_temporary.c_str()) != 0) {
                 LOG_ERR("%s: unable to delete file: %s\n", __func__, path_temporary.c_str());
-                return false;
+                return -1;
             }
         }
 
@@ -718,15 +753,16 @@ static bool common_download_file_single_online(const std::string & url,
 
         if (std::rename(path_temporary.c_str(), path.c_str()) != 0) {
             LOG_ERR("%s: unable to rename file: %s to %s\n", __func__, path_temporary.c_str(), path.c_str());
-            return false;
+            return -1;
         }
         if (!etag.empty()) {
             write_etag(path, etag);
         }
-        break;
+
+        return head->status; // TODO: use actual GET status?
     }
 
-    return true;
+    return -1; // max attempts reached
 }
 
 std::pair<long, std::vector<char>> common_remote_get_content(const std::string          & url,
@@ -734,13 +770,9 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {{"User-Agent", "llama-cpp"}};
+
     for (const auto & header : params.headers) {
-        size_t pos = header.find(':');
-        if (pos != std::string::npos) {
-            headers.emplace(header.substr(0, pos), header.substr(pos + 1));
-        } else {
-            headers.emplace(header, "");
-        }
+        headers.emplace(header.first, header.second);
     }
 
     if (params.timeout > 0) {
@@ -769,32 +801,45 @@ std::pair<long, std::vector<char>> common_remote_get_content(const std::string
 
 #if defined(LLAMA_USE_CURL) || defined(LLAMA_USE_HTTPLIB)
 
-static bool common_download_file_single(const std::string & url,
-                                        const std::string & path,
-                                        const std::string & bearer_token,
-                                        bool                offline) {
+int common_download_file_single(const std::string & url,
+                                const std::string & path,
+                                const std::string & bearer_token,
+                                bool offline,
+                                const common_header_list & headers) {
     if (!offline) {
-        return common_download_file_single_online(url, path, bearer_token);
+        return common_download_file_single_online(url, path, bearer_token, headers);
     }
 
     if (!std::filesystem::exists(path)) {
         LOG_ERR("%s: required file is not available in cache (offline mode): %s\n", __func__, path.c_str());
-        return false;
+        return -1;
     }
 
     LOG_INF("%s: using cached file (offline mode): %s\n", __func__, path.c_str());
-    return true;
+    return 304; // Not Modified - fake cached response
 }
 
 // download multiple files from remote URLs to local paths
 // the input is a vector of pairs <url, path>
-static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls, const std::string & bearer_token, bool offline) {
+static bool common_download_file_multiple(const std::vector<std::pair<std::string, std::string>> & urls,
+                                          const std::string & bearer_token,
+                                          bool offline,
+                                          const common_header_list & headers) {
     // Prepare download in parallel
     std::vector<std::future<bool>> futures_download;
+    futures_download.reserve(urls.size());
+
     for (auto const & item : urls) {
-        futures_download.push_back(std::async(std::launch::async, [bearer_token, offline](const std::pair<std::string, std::string> & it) -> bool {
-            return common_download_file_single(it.first, it.second, bearer_token, offline);
-        }, item));
+        futures_download.push_back(
+            std::async(
+                std::launch::async,
+                [&bearer_token, offline, &headers](const std::pair<std::string, std::string> & it) -> bool {
+                    const int http_status = common_download_file_single(it.first, it.second, bearer_token, offline, headers);
+                    return is_http_status_ok(http_status);
+                },
+                item
+            )
+        );
     }
 
     // Wait for all downloads to complete
@@ -807,17 +852,18 @@ static bool common_download_file_multiple(const std::vector<std::pair<std::strin
     return true;
 }
 
-bool common_download_model(
-        const common_params_model & model,
-        const std::string & bearer_token,
-        bool offline) {
+bool common_download_model(const common_params_model & model,
+                           const std::string & bearer_token,
+                           bool offline,
+                           const common_header_list & headers) {
     // Basic validation of the model.url
     if (model.url.empty()) {
         LOG_ERR("%s: invalid model url\n", __func__);
         return false;
     }
 
-    if (!common_download_file_single(model.url, model.path, bearer_token, offline)) {
+    const int http_status = common_download_file_single(model.url, model.path, bearer_token, offline, headers);
+    if (!is_http_status_ok(http_status)) {
         return false;
     }
 
@@ -876,27 +922,26 @@ bool common_download_model(
         }
 
         // Download in parallel
-        common_download_file_multiple(urls, bearer_token, offline);
+        common_download_file_multiple(urls, bearer_token, offline, headers);
     }
 
     return true;
 }
 
-common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, const std::string & bearer_token, bool offline) {
-    auto parts = string_split<std::string>(hf_repo_with_tag, ':');
-    std::string tag = parts.size() > 1 ? parts.back() : "latest";
-    std::string hf_repo = parts[0];
-    if (string_split<std::string>(hf_repo, '/').size() != 2) {
-        throw std::invalid_argument("error: invalid HF repo format, expected <user>/<model>[:quant]\n");
-    }
+common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag,
+                                      const std::string & bearer_token,
+                                      bool offline,
+                                      const common_header_list & custom_headers) {
+    // the returned hf_repo is without tag
+    auto [hf_repo, tag] = common_download_split_repo_tag(hf_repo_with_tag);
 
     std::string url = get_model_endpoint() + "v2/" + hf_repo + "/manifests/" + tag;
 
     // headers
-    std::vector<std::string> headers;
-    headers.push_back("Accept: application/json");
+    common_header_list headers = custom_headers;
+    headers.push_back({"Accept", "application/json"});
     if (!bearer_token.empty()) {
-        headers.push_back("Authorization: Bearer " + bearer_token);
+        headers.push_back({"Authorization", "Bearer " + bearer_token});
     }
     // Important: the User-Agent must be "llama-cpp" to get the "ggufFile" field in the response
     // User-Agent header is already set in common_remote_get_content, no need to set it here
@@ -952,7 +997,7 @@ common_hf_file_res common_get_hf_file(const std::string & hf_repo_with_tag, cons
     } else if (res_code == 401) {
         throw std::runtime_error("error: model is private or does not exist; if you are accessing a gated model, please provide a valid HF token");
     } else {
-        throw std::runtime_error(string_format("error from HF API, response code: %ld, data: %s", res_code, res_str.c_str()));
+        throw std::runtime_error(string_format("error from HF API (%s), response code: %ld, data: %s", url.c_str(), res_code, res_str.c_str()));
     }
 
     // check response
@@ -1031,9 +1076,10 @@ std::string common_docker_resolve_model(const std::string & docker) {
         const std::string    url_prefix = "https://registry-1.docker.io/v2/" + repo;
         std::string          manifest_url = url_prefix + "/manifests/" + tag;
         common_remote_params manifest_params;
-        manifest_params.headers.push_back("Authorization: Bearer " + token);
-        manifest_params.headers.push_back(
-            "Accept: application/vnd.docker.distribution.manifest.v2+json,application/vnd.oci.image.manifest.v1+json");
+        manifest_params.headers.push_back({"Authorization", "Bearer " + token});
+        manifest_params.headers.push_back({"Accept",
+            "application/vnd.docker.distribution.manifest.v2+json,application/vnd.oci.image.manifest.v1+json"
+        });
         auto manifest_res = common_remote_get_content(manifest_url, manifest_params);
         if (manifest_res.first != 200) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
@@ -1070,7 +1116,8 @@ std::string common_docker_resolve_model(const std::string & docker) {
         std::string local_path = fs_get_cache_file(model_filename);
 
         const std::string blob_url = url_prefix + "/blobs/" + gguf_digest;
-        if (!common_download_file_single(blob_url, local_path, token, false)) {
+        const int http_status = common_download_file_single(blob_url, local_path, token, false, {});
+        if (!is_http_status_ok(http_status)) {
             throw std::runtime_error("Failed to download Docker Model");
         }
 
@@ -1084,11 +1131,11 @@ std::string common_docker_resolve_model(const std::string & docker) {
 
 #else
 
-common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool) {
+common_hf_file_res common_get_hf_file(const std::string &, const std::string &, bool, const common_header_list &) {
     throw std::runtime_error("download functionality is not enabled in this build");
 }
 
-bool common_download_model(const common_params_model &, const std::string &, bool) {
+bool common_download_model(const common_params_model &, const std::string &, bool, const common_header_list &) {
     throw std::runtime_error("download functionality is not enabled in this build");
 }
 
@@ -1096,6 +1143,14 @@ std::string common_docker_resolve_model(const std::string &) {
     throw std::runtime_error("download functionality is not enabled in this build");
 }
 
+int common_download_file_single(const std::string &,
+                                const std::string &,
+                                const std::string &,
+                                bool,
+                                const common_header_list &) {
+    throw std::runtime_error("download functionality is not enabled in this build");
+}
+
 #endif // LLAMA_USE_CURL || LLAMA_USE_HTTPLIB
 
 std::vector<common_cached_model_info> common_list_cached_models() {

common/download.h

@@ -1,12 +1,27 @@
 #pragma once
 
 #include <string>
+#include <vector>
 
 struct common_params_model;
 
-//
-// download functionalities
-//
+using common_header      = std::pair<std::string, std::string>;
+using common_header_list = std::vector<common_header>;
+
+struct common_remote_params {
+    common_header_list headers;
+    long timeout  = 0;           // in seconds, 0 means no timeout
+    long max_size = 0;           // unlimited if 0
+};
+
+// get remote file content, returns <http_code, raw_response_body>
+std::pair<long, std::vector<char>> common_remote_get_content(const std::string & url, const common_remote_params & params);
+
+// split HF repo with tag into <repo, tag>
+// for example: "user/model:tag" -> <"user/model", "tag">
+// if tag is not present, default to "latest"
+// example: "user/model" -> <"user/model", "latest">
+std::pair<std::string, std::string> common_download_split_repo_tag(const std::string & hf_repo_with_tag);
 
 struct common_cached_model_info {
     std::string manifest_path;
@@ -41,17 +56,29 @@ struct common_hf_file_res {
 common_hf_file_res common_get_hf_file(
     const std::string & hf_repo_with_tag,
     const std::string & bearer_token,
-    bool offline);
+    bool offline,
+    const common_header_list & headers = {}
+);
 
 // returns true if download succeeded
 bool common_download_model(
     const common_params_model & model,
     const std::string & bearer_token,
-    bool offline);
+    bool offline,
+    const common_header_list & headers = {}
+);
 
 // returns list of cached models
 std::vector<common_cached_model_info> common_list_cached_models();
 
+// download single file from url to local path
+// returns status code or -1 on error
+int common_download_file_single(const std::string & url,
+                                const std::string & path,
+                                const std::string & bearer_token,
+                                bool offline,
+                                const common_header_list & headers = {});
+
 // resolve and download model from Docker registry
 // return local path to downloaded model file
 std::string common_docker_resolve_model(const std::string & docker);

common/preset.cpp

@@ -16,6 +16,48 @@ static std::string rm_leading_dashes(const std::string & str) {
     return str.substr(pos);
 }
 
+// only allow a subset of args for remote presets for security reasons
+// do not add more args unless absolutely necessary
+// args that output to files are strictly prohibited
+static std::set<std::string> get_remote_preset_whitelist(const std::map<std::string, common_arg> & key_to_opt) {
+    static const std::set<std::string> allowed_options = {
+        "model-url",
+        "hf-repo",
+        "hf-repo-draft",
+        "hf-repo-v", // vocoder
+        "hf-file-v", // vocoder
+        "mmproj-url",
+        "pooling",
+        "jinja",
+        "batch-size",
+        "ubatch-size",
+        "cache-reuse",
+        "chat-template-kwargs",
+        "mmap",
+        // note: sampling params are automatically allowed by default
+        // negated args will be added automatically if the positive arg is specified above
+    };
+
+    std::set<std::string> allowed_keys;
+
+    for (const auto & it : key_to_opt) {
+        const std::string & key = it.first;
+        const common_arg & opt = it.second;
+        if (allowed_options.find(key) != allowed_options.end() || opt.is_sparam) {
+            allowed_keys.insert(key);
+            // also add variant keys (args without leading dashes and env vars)
+            for (const auto & arg : opt.get_args()) {
+                allowed_keys.insert(rm_leading_dashes(arg));
+            }
+            for (const auto & env : opt.get_env()) {
+                allowed_keys.insert(env);
+            }
+        }
+    }
+
+    return allowed_keys;
+}
+
 std::vector<std::string> common_preset::to_args(const std::string & bin_path) const {
     std::vector<std::string> args;
 
@@ -121,6 +163,29 @@ void common_preset::merge(const common_preset & other) {
     }
 }
 
+void common_preset::apply_to_params(common_params & params) const {
+    for (const auto & [opt, val] : options) {
+        // apply each option to params
+        if (opt.handler_string) {
+            opt.handler_string(params, val);
+        } else if (opt.handler_int) {
+            opt.handler_int(params, std::stoi(val));
+        } else if (opt.handler_bool) {
+            opt.handler_bool(params, common_arg_utils::is_truthy(val));
+        } else if (opt.handler_str_str) {
+            // not supported yet
+            throw std::runtime_error(string_format(
+                "%s: option with two values is not supported yet",
+                __func__
+            ));
+        } else if (opt.handler_void) {
+            opt.handler_void(params);
+        } else {
+            GGML_ABORT("unknown handler type");
+        }
+    }
+}
+
 static std::map<std::string, std::map<std::string, std::string>> parse_ini_from_file(const std::string & path) {
     std::map<std::string, std::map<std::string, std::string>> parsed;
 
@@ -230,10 +295,16 @@ static std::string parse_bool_arg(const common_arg & arg, const std::string & ke
     return value;
 }
 
-common_preset_context::common_preset_context(llama_example ex)
+common_preset_context::common_preset_context(llama_example ex, bool only_remote_allowed)
         : ctx_params(common_params_parser_init(default_params, ex)) {
     common_params_add_preset_options(ctx_params.options);
     key_to_opt = get_map_key_opt(ctx_params);
+
+    // setup allowed keys if only_remote_allowed is true
+    if (only_remote_allowed) {
+        filter_allowed_keys = true;
+        allowed_keys = get_remote_preset_whitelist(key_to_opt);
+    }
 }
 
 common_presets common_preset_context::load_from_ini(const std::string & path, common_preset & global) const {
@@ -249,7 +320,18 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
         }
         LOG_DBG("loading preset: %s\n", preset.name.c_str());
         for (const auto & [key, value] : section.second) {
+            if (key == "version") {
+                // skip version key (reserved for future use)
+                continue;
+            }
+
             LOG_DBG("option: %s = %s\n", key.c_str(), value.c_str());
+            if (filter_allowed_keys && allowed_keys.find(key) == allowed_keys.end()) {
+                throw std::runtime_error(string_format(
+                    "option '%s' is not allowed in remote presets",
+                    key.c_str()
+                ));
+            }
             if (key_to_opt.find(key) != key_to_opt.end()) {
                 const auto & opt = key_to_opt.at(key);
                 if (is_bool_arg(opt)) {
@@ -259,7 +341,10 @@ common_presets common_preset_context::load_from_ini(const std::string & path, co
                 }
                 LOG_DBG("accepted option: %s = %s\n", key.c_str(), preset.options[opt].c_str());
             } else {
-                // TODO: maybe warn about unknown key?
+                throw std::runtime_error(string_format(
+                    "option '%s' not recognized in preset '%s'",
+                    key.c_str(), preset.name.c_str()
+                ));
             }
         }
 

common/preset.h

@@ -6,6 +6,7 @@
 #include <string>
 #include <vector>
 #include <map>
+#include <set>
 
 //
 // INI preset parser and writer
@@ -40,6 +41,9 @@ struct common_preset {
 
     // merge another preset into this one, overwriting existing options
     void merge(const common_preset & other);
+
+    // apply preset options to common_params
+    void apply_to_params(common_params & params) const;
 };
 
 // interface for multiple presets in one file
@@ -50,7 +54,12 @@ struct common_preset_context {
     common_params default_params; // unused for now
     common_params_context ctx_params;
     std::map<std::string, common_arg> key_to_opt;
-    common_preset_context(llama_example ex);
+
+    bool filter_allowed_keys = false;
+    std::set<std::string> allowed_keys;
+
+    // if only_remote_allowed is true, only accept whitelisted keys
+    common_preset_context(llama_example ex, bool only_remote_allowed = false);
 
     // load presets from INI file
     common_presets load_from_ini(const std::string & path, common_preset & global) const;

convert_hf_to_gguf.py

@@ -528,7 +528,11 @@ class ModelBase:
         return ()
 
     def prepare_tensors(self):
-        max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
+        # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
+        if self.tensor_map.mapping:
+            max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
+        else:
+            max_name_len = len("vision_encoder.weight,")  # Default reasonable length
 
         for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()):
             # we don't need these
@@ -771,8 +775,8 @@ class TextModel(ModelBase):
 
         self.rope_parameters = self.hparams.get("rope_parameters", self.hparams.get("rope_scaling")) or {}
 
-        rope_theta = self.find_hparam(["rope_theta", "global_rope_theta", "rotary_emb_base"], optional=True)
-        local_rope_theta = self.find_hparam(["local_rope_theta", "rope_local_theta", "swa_rope_theta", "rope_local_base_freq"], optional=True)
+        rope_theta = self.find_hparam(["global_rope_theta", "rope_global_theta", "rope_theta_global", "rope_theta", "rotary_emb_base"], optional=True)
+        local_rope_theta = self.find_hparam(["local_rope_theta", "rope_local_theta", "rope_theta_local", "swa_rope_theta", "rope_local_base_freq"], optional=True)
 
         # Ensure "rope_theta" and "rope_type" is mirrored in rope_parameters
         if "full_attention" not in self.rope_parameters and "sliding_attention" not in self.rope_parameters:
@@ -1248,6 +1252,9 @@ class TextModel(ModelBase):
         if chkhsh == "16389f0a1f51ee53e562ffd51c371dc508639ab0e4261502071836e50e223e91":
             # ref: https://huggingface.co/upstage/Solar-Open-100B
             res = "solar-open"
+        if chkhsh == "6c81ce329e0802883b22eabab0d3fa48357337ef1ecb45443828bf1f6254833f":
+            # ref: https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B
+            res = "exaone-moe"
 
         if res is None:
             logger.warning("\n")
@@ -4363,7 +4370,37 @@ class Qwen3NextModel(Qwen2MoeModel):
         elif name.endswith("norm.weight") and not name.endswith("linear_attn.norm.weight"):
             data_torch = data_torch + 1
 
-        yield from super().modify_tensors(data_torch, name, bid)
+        if "in_proj_qkvz.weight" in name:
+            # original order:  [q, k, v, z] * head_count
+            # corrected order: [q * head_count, k * head_count, v * head_count, z * head_count]
+            head_k_dim = self.hparams["linear_key_head_dim"]
+            head_v_dim = self.hparams["linear_value_head_dim"]
+            num_v_heads = self.hparams["linear_num_value_heads"]
+            num_k_heads = self.hparams["linear_num_key_heads"]
+            hidden_size = self.hparams["hidden_size"]
+            split_arg_list_qkvz = [
+                head_k_dim, # q partition
+                head_k_dim, # k partition
+                (num_v_heads // num_k_heads * head_v_dim), # v partition
+                (num_v_heads // num_k_heads * head_v_dim), # z partition
+            ]
+            # view as (n_embd, head_count, [q+k+v+z])
+            data_torch = data_torch.permute(1, 0).contiguous()
+            data_torch = data_torch.view(-1, num_k_heads, sum(split_arg_list_qkvz))
+            # split into q, k, v, z
+            q, k, v, z = torch.split(data_torch, split_arg_list_qkvz, dim=-1)
+            # flatten dim + head_count
+            q = q.contiguous().view(hidden_size, -1)
+            k = k.contiguous().view(hidden_size, -1)
+            v = v.contiguous().view(hidden_size, -1)
+            z = z.contiguous().view(hidden_size, -1)
+            # stack back
+            qkv = torch.cat([q, k, v], dim=-1).permute(1, 0).contiguous()
+            z = z.permute(1, 0).contiguous()
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV,  bid, ".weight"), qkv)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_GATE, bid, ".weight"), z)
+        else:
+            yield from super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("RND1")
@@ -6038,7 +6075,175 @@ class Gemma3VisionModel(MmprojModel):
         return [] # skip other tensors
 
 
+class ConformerAudioModel(MmprojModel):
+    _batch_norm_tensors: list[dict[str, Tensor]] | None = None
+
+    @staticmethod
+    def is_audio_tensor(name: str):
+        return any(p in name for p in ["audio", "codebook", "conformer", "depth_embedding", "depthformer", "depth_linear"])
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        if ConformerAudioModel.is_audio_tensor(name):
+            if ".conv" in name or "_conv" in name and ".weight" in name:
+                return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # fold running_mean, running_var and eps into weight and bias for batch_norm
+        if "batch_norm" in name:
+            if self._batch_norm_tensors is None:
+                self._batch_norm_tensors = [{} for _ in range(self.block_count)]
+            assert bid is not None
+            self._batch_norm_tensors[bid][name] = data_torch
+
+            if len(self._batch_norm_tensors[bid]) < 5:
+                return []
+
+            weight = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.weight"]
+            bias = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.bias"]
+            running_mean = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.running_mean"]
+            running_var = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.running_var"]
+            eps = 1e-5 # default value
+
+            a = weight / torch.sqrt(running_var + eps)
+            b = bias - running_mean * a
+            return [
+                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.weight"), a),
+                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.bias"), b),
+            ]
+
+        # reshape conv weights
+        if name.startswith("conformer.pre_encode.conv.") and name.endswith(".bias"):
+            data_torch = data_torch[:, None, None]
+        if "conv.depthwise_conv" in name and name.endswith(".weight"):
+            assert data_torch.shape[1] == 1
+            data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[2])
+        if "conv.pointwise_conv" in name and name.endswith(".weight"):
+            assert data_torch.shape[2] == 1
+            data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @ModelBase.register("Gemma3nForConditionalGeneration")
+class Gemma3nVisionAudioModel(ConformerAudioModel):
+    has_audio_encoder = True
+    has_vision_encoder = True
+
+    # Double indexed mapping for MobileNetV5 blocks (not supported by tensor_mapping.py)
+    # This is the only known model having this, so we prefer implementing it outside of tensor_mapping.py
+    block_tensor_mapping = {
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.conv_exp.weight":             "v.blk.{bid}.{sid}.conv_exp.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.bn1.weight":                  "v.blk.{bid}.{sid}.bn1.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.conv_pwl.weight":             "v.blk.{bid}.{sid}.conv_pwl.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.bn2.weight":                  "v.blk.{bid}.{sid}.bn2.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.dw_start.conv.weight":        "v.blk.{bid}.{sid}.dw_start.conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.dw_start.bn.weight":          "v.blk.{bid}.{sid}.dw_start.bn.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.dw_mid.conv.weight":          "v.blk.{bid}.{sid}.dw_mid.conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.dw_mid.bn.weight":            "v.blk.{bid}.{sid}.dw_mid.bn.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.pw_exp.conv.weight":          "v.blk.{bid}.{sid}.pw_exp.conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.pw_exp.bn.weight":            "v.blk.{bid}.{sid}.pw_exp.bn.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.pw_proj.conv.weight":         "v.blk.{bid}.{sid}.pw_proj.conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.pw_proj.bn.weight":           "v.blk.{bid}.{sid}.pw_proj.bn.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.layer_scale.gamma":           "v.blk.{bid}.{sid}.layer_scale.gamma",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.query.proj.weight":      "v.blk.{bid}.{sid}.attn.query.proj.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.key.proj.weight":        "v.blk.{bid}.{sid}.attn.key.proj.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.value.proj.weight":      "v.blk.{bid}.{sid}.attn.value.proj.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.output.proj.weight":     "v.blk.{bid}.{sid}.attn.output.proj.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.key.down_conv.weight":   "v.blk.{bid}.{sid}.attn.key.down_conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.key.norm.weight":        "v.blk.{bid}.{sid}.attn.key.norm.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.value.down_conv.weight": "v.blk.{bid}.{sid}.attn.value.down_conv.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.attn.value.norm.weight":      "v.blk.{bid}.{sid}.attn.value.norm.weight",
+        "model.vision_tower.timm_model.blocks.{bid}.{sid}.norm.weight":                 "v.blk.{bid}.{sid}.norm.weight",
+    }
+
+    def __init__(self, *args, **kwargs):
+        # Parent init will call find_hparam which now returns 0 for empty keys
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        self.hparams_vision["n_layers"] = 128 # fake value for audio encoder, vision encoder doesn't use it
+        self.hparams_vision["intermediate_size"] = self.hparams_vision.get("intermediate_size", 2048) * 4
+        self.hparams_vision["num_attention_heads"] = self.hparams_vision.get("num_attention_heads", 8)
+
+        # MobileNetV5 does not use image_mean/std
+        self.preprocessor_config["image_mean"] = [0.0 ,0.0 , 0.0]
+        self.preprocessor_config["image_std"] = [1.0 ,1.0 ,1.0]
+        self.hparams_vision["image_size"] = self.preprocessor_config.get(
+            "size", {"height": 768, "width": 768}
+        )["height"]
+
+        # Image sequence length (256 tokens = 16x16 for Gemma3n)
+        image_seq_length = self.preprocessor_config.get("image_seq_length", 256)
+        image_size = self.hparams_vision["image_size"]
+        self.hparams_vision["patch_size"] = image_size // image_seq_length
+
+        # remap audio hparams
+        assert self.hparams_audio is not None
+        self.hparams_audio["n_layers"] = self.hparams_audio["conf_num_hidden_layers"]
+        self.hparams_audio["num_attention_heads"] = self.hparams_audio["conf_num_attention_heads"]
+        self.hparams_audio["feat_in"] = self.hparams_audio["input_feat_size"]
+        self.hparams_audio["intermediate_size"] = self.hparams_audio.get("intermediate_size", 6144)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+
+        # vision params
+        self.gguf_writer.add_clip_vision_projector_type(gguf.VisionProjectorType.GEMMA3NV)
+        self.gguf_writer.add_vision_attention_layernorm_eps(self.hparams.get("layer_norm_eps", 1e-6))
+
+        # audio params
+        assert self.hparams_audio is not None
+        self.gguf_writer.add_clip_audio_projector_type(gguf.VisionProjectorType.GEMMA3NA)
+        self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
+        self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # Force quantization settings for specific tensor types
+        if "input_projection" in name or "input_proj" in name:
+            return gguf.GGMLQuantizationType.F16
+        if ".embeddings." in name or "stem" in name:
+            return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def custom_map(self, name: str) -> str:
+        """Parses names like model.vision_tower.timm_model.blocks.1.2.suffix and applies template mapping."""
+        parts = name.split(".")
+        # MobileNet blocks have at least 7 parts: model, vision_tower, timm_model, blocks, bid, sid, and suffix
+        if len(parts) >= 7:
+            bid, sid = parts[4], parts[5]
+            suffix = ".".join(parts[6:])
+            template = f"model.vision_tower.timm_model.blocks.{{bid}}.{{sid}}.{suffix}"
+            if template in self.block_tensor_mapping:
+                return self.block_tensor_mapping[template].format(bid=bid, sid=sid)
+
+        raise ValueError(f"Unknown name: {name}")
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if (ConformerAudioModel.is_audio_tensor(name)):
+            name = name.replace("model.audio_tower.conformer.", "conformer.layers.")
+            return super().modify_tensors(data_torch, name, bid)
+
+        # Gemma3n uses
+        # - model.embed_vision.* for projection layers
+        # - model.vision_tower.* for vision encoder
+        # Skip non-vision tensors
+        if not (name.startswith("model.embed_vision.") or name.startswith("model.vision_tower.")):
+            return []
+
+        if name.startswith("model.vision_tower.timm_model.blocks."):
+            # Double-indexed block tensors through custom logic
+            new_name = self.custom_map(name)
+        else:
+            # Route non-repeating (conv_stem, msfa, embedding, etc.) and un-catched through tensor_mapping.py
+            new_name = self.map_tensor_name(name)
+
+        if new_name.endswith("conv_stem.conv.bias") or new_name.endswith("layer_scale.gamma"):
+            data_torch = data_torch.unsqueeze(0).unsqueeze(-1).unsqueeze(-1) # [1, C, 1, 1]
+
+        return [(new_name, data_torch)]
+
+
+@ModelBase.register("Gemma3nForCausalLM", "Gemma3nForConditionalGeneration")
 class Gemma3NModel(Gemma3Model):
     model_arch = gguf.MODEL_ARCH.GEMMA3N
     norm_shift = 0.0 # same value with Gemma3p5RMSNorm scale_shift on python code
@@ -6061,8 +6266,25 @@ class Gemma3NModel(Gemma3Model):
         ]
 
     def set_vocab(self):
+        # For Gemma3n multimodal models, we need the FULL vocab_size (262400)
+        # which includes special tokens from 262144-262399 for vision/audio.
+        # The vocab_size_per_layer_input (262144) is only the embedding size per layer.
+        # Temporarily override the hparams lookup order to prioritize vocab_size.
+
+        # Store original vocab_size_per_layer_input if it exists
+        vocab_size_per_layer_input = self.hparams.get("vocab_size_per_layer_input")
+
+        # Temporarily remove vocab_size_per_layer_input to force using vocab_size
+        if vocab_size_per_layer_input is not None:
+            del self.hparams["vocab_size_per_layer_input"]
+
+        # Call parent set_vocab which will now use vocab_size (262400)
         super().set_vocab()
 
+        # Restore vocab_size_per_layer_input for later use
+        if vocab_size_per_layer_input is not None:
+            self.hparams["vocab_size_per_layer_input"] = vocab_size_per_layer_input
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_altup_active_idx(self.hparams["altup_active_idx"])
@@ -6098,8 +6320,32 @@ class Gemma3NModel(Gemma3Model):
         if "language_model." not in name:
             return [] # skip non-language model tensors
 
+        # Pad token embeddings for vision/audio special tokens (262144-262399)
+        if "embed_tokens.weight" in name or "embed_tokens_per_layer" in name:
+            # Move to CPU to avoid meta device issues during padding
+            data_torch = data_torch.to(device="cpu")
+
+            vocab_size = self.hparams.get("vocab_size", 262400)
+            current_size = data_torch.shape[0]  # First dimension is vocab_size
+
+            if current_size < vocab_size:
+                # Pad with zeros for vision/audio tokens (they get embeddings from vision tower)
+                padding_size = vocab_size - current_size
+                tensor_type = "per-layer embeddings" if "per_layer" in name else "token embeddings"
+                logger.info(f"Padding {tensor_type} shape {list(data_torch.shape)} from {current_size} to {vocab_size} (adding {padding_size} vision/audio token slots)")
+
+                # Create padding with zeros (vision tokens won't use these embeddings)
+                padding = torch.zeros((padding_size, data_torch.shape[1]), dtype=data_torch.dtype, device=data_torch.device)
+                data_torch = torch.cat([data_torch, padding], dim=0)
+
+            # Continue with normal processing
+            name = name.replace("language_model.", "")
+            return [(self.map_tensor_name(name), data_torch)]
+
         if "altup_unembed_projections" in name:
             data_torch = data_torch.to(device="cpu")
+            # altup_unembed matrices are [hidden_size, hidden_size], NOT vocab-based
+            # They should NOT be padded
             if ".0." in name:
                 self._altup_unembd[0] = data_torch
             elif ".1." in name:
@@ -8505,6 +8751,106 @@ class Exaone4Model(TextModel):
                 yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
 
+@ModelBase.register("ExaoneMoEForCausalLM")
+class ExaoneMoEModel(Exaone4Model):
+    model_arch = gguf.MODEL_ARCH.EXAONE_MOE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.block_count = self.hparams["num_hidden_layers"] + self.hparams.get("num_nextn_predict_layers", 0)
+        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_expert_count(self.hparams["num_experts"])
+        moe_intermediate_size = self.hparams["moe_intermediate_size"]
+        num_shared_experts = self.hparams["num_shared_experts"]
+        self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
+        self.gguf_writer.add_expert_shared_count(num_shared_experts)
+        self.gguf_writer.add_expert_shared_feed_forward_length(moe_intermediate_size * num_shared_experts)
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+        n_dense_layer = self.hparams.get("first_k_dense_replace", self.hparams.get("first_last_k_dense_replace", 0))
+        self.gguf_writer.add_leading_dense_block_count(n_dense_layer)
+        # For here, we hard-code the number of NextN/MTP layers to 1 for K-EXAONE,
+        # so that we can convert MTP weights to GGUF format for speculative decoding.
+        # This is because HF config of K-EXAONE does not have `num_nextn_predict_layers` at now.
+        # Will be updated when HF config is updated.
+        self.gguf_writer.add_nextn_predict_layers(self.hparams.get("num_nextn_predict_layers", 1))
+
+        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("mtp."):
+            if name.find("layers.") != -1:
+                # `mtp.layers.0.[module_name]` format
+                name = name.replace(f"mtp.layers.{bid}", f"model.layers.{bid + self.hparams['num_hidden_layers']}")
+            else:
+                # mtp fc/norm weights
+                remapper = {
+                    "mtp.fc": "model.layers.{bid}.eh_proj",
+                    "mtp.pre_fc_norm_embedding": "model.layers.{bid}.enorm",
+                    "mtp.pre_fc_norm_hidden": "model.layers.{bid}.hnorm",
+                    "mtp.norm": "model.layers.{bid}.shared_head.norm",
+                }
+                _n = Path(name)
+                new_name = remapper[_n.stem] + _n.suffix
+
+                # set shared weights for all NextN/MTP layers
+                tensors = []
+                for bid in range(self.hparams['num_hidden_layers'], self.block_count):
+                    new_name = new_name.format(bid=bid)
+                    tensors.append((self.map_tensor_name(new_name), data_torch))
+                return tensors
+
+        if name.endswith("e_score_correction_bias"):
+            name = name.replace("e_score_correction_bias", "e_score_correction.bias")
+
+        if name.find("mlp.experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    def prepare_tensors(self):
+        super().prepare_tensors()
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @ModelBase.register("GraniteForCausalLM")
 class GraniteModel(LlamaModel):
     """Conversion for IBM's GraniteForCausalLM"""
@@ -9936,7 +10282,7 @@ class LFM2Model(TextModel):
         self._add_feed_forward_length()
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-        if self._is_vision_tensor(name) or self._is_audio_tensor(name):
+        if self._is_vision_tensor(name) or ConformerAudioModel.is_audio_tensor(name):
             # skip multimodal tensors
             return []
 
@@ -9952,9 +10298,6 @@ class LFM2Model(TextModel):
     def _is_vision_tensor(self, name: str) -> bool:
         return "vision_tower" in name or "multi_modal_projector" in name
 
-    def _is_audio_tensor(self, name: str):
-        return any(p in name for p in ["audio", "codebook", "conformer", "depth_embedding", "depthformer", "depth_linear"])
-
 
 @ModelBase.register("Lfm2Model")
 class LFM2ColBertModel(LFM2Model):
@@ -10082,13 +10425,11 @@ class LFM2VLModel(MmprojModel):
 
 
 @ModelBase.register("Lfm2AudioForConditionalGeneration")
-class LFM2AudioModel(MmprojModel):
+class LFM2AudioModel(ConformerAudioModel):
     has_vision_encoder = False
     has_audio_encoder = True
     model_name = "Lfm2AudioEncoder"
 
-    _batch_norm_tensors: list[dict[str, Tensor]] | None = None
-
     def get_audio_config(self) -> dict[str, Any] | None:
         return self.global_config.get("encoder")
 
@@ -10102,12 +10443,7 @@ class LFM2AudioModel(MmprojModel):
         self.gguf_writer.add_audio_num_mel_bins(self.hparams_audio["feat_in"])
         self.gguf_writer.add_audio_attention_layernorm_eps(1e-5)
 
-    def tensor_force_quant(self, name, new_name, bid, n_dims):
-        if ".conv" in name and ".weight" in name:
-            return gguf.GGMLQuantizationType.F32
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
-    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+    def modify_tensors(self, data_torch, name, bid):
         # skip language model tensors
         if name.startswith("lfm."):
             return []
@@ -10120,40 +10456,7 @@ class LFM2AudioModel(MmprojModel):
         if any(p in name for p in ["codebook_offsets", "depth_embeddings", "depth_linear", "depthformer"]):
             return []
 
-        # fold running_mean, running_var and eps into weight and bias for batch_norm
-        if "batch_norm" in name:
-            if self._batch_norm_tensors is None:
-                self._batch_norm_tensors = [{} for _ in range(self.block_count)]
-            assert bid is not None
-            self._batch_norm_tensors[bid][name] = data_torch
-
-            if len(self._batch_norm_tensors[bid]) < 5:
-                return []
-
-            weight = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.weight"]
-            bias = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.bias"]
-            running_mean = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.running_mean"]
-            running_var = self._batch_norm_tensors[bid][f"conformer.layers.{bid}.conv.batch_norm.running_var"]
-            eps = 1e-5 # default value
-
-            a = weight / torch.sqrt(running_var + eps)
-            b = bias - running_mean * a
-            return [
-                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.weight"), a),
-                (self.map_tensor_name(f"conformer.layers.{bid}.conv.batch_norm.bias"), b),
-            ]
-
-        # reshape conv weights
-        if name.startswith("conformer.pre_encode.conv.") and name.endswith(".bias"):
-            data_torch = data_torch[:, None, None]
-        if "conv.depthwise_conv" in name and name.endswith(".weight"):
-            assert data_torch.shape[1] == 1
-            data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[2])
-        if "conv.pointwise_conv" in name and name.endswith(".weight"):
-            assert data_torch.shape[2] == 1
-            data_torch = data_torch.reshape(data_torch.shape[0], data_torch.shape[1])
-
-        return [(self.map_tensor_name(name), data_torch)]
+        return super().modify_tensors(data_torch, name, bid)
 
 
 @ModelBase.register("SmallThinkerForCausalLM")
@@ -10974,8 +11277,8 @@ def parse_args() -> argparse.Namespace:
 
     parser.add_argument(
         "--sentence-transformers-dense-modules", action="store_true",
-        help=("Whether to include sentence-transformers dense modules."
-              "It can be used for sentence-transformers models, like google/embeddinggemma-300m"
+        help=("Whether to include sentence-transformers dense modules. "
+              "It can be used for sentence-transformers models, like google/embeddinggemma-300m. "
               "Default these modules are not included.")
     )
 

convert_hf_to_gguf_update.py

@@ -147,6 +147,7 @@ models = [
     {"name": "kormo",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/KORMo-Team/KORMo-tokenizer", },
     {"name": "youtu",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tencent/Youtu-LLM-2B", },
     {"name": "solar-open",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/upstage/Solar-Open-100B", },
+    {"name": "exaone-moe",       "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/K-EXAONE-236B-A23B", },
 ]
 
 # some models are known to be broken upstream, so we will skip them as exceptions

docs/ops.md

@@ -57,7 +57,6 @@ Legend:
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|               GROUP_NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -71,10 +70,9 @@ Legend:
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                     NORM_MUL_ADD | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ |
+|                         OUT_PROD | 🟡 | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | 🟡 |
 |                              PAD | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ | ❌ |
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -99,7 +97,6 @@ Legend:
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | ❌ | ❌ | ❌ |
-|                          SOFTCAP | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |

docs/ops/BLAS.csv

@@ -965,6 +965,7 @@
 "BLAS","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,1,2560],ne_kernel=[3,3,1,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","BLAS"
 "BLAS","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[12,12,2,2560],ne_kernel=[3,3,2,2560],s0=1,s1=1,p0=1,p1=1,d0=1,d1=1,is_2D=1","support","0","no","BLAS"
 "BLAS","IM2COL","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[5,5,1,32],ne_kernel=[3,4,1,32],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","BLAS"
+"BLAS","IM2COL","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[2,2,1536,729],ne_kernel=[2,2,1536,4096],s0=1,s1=1,p0=0,p1=0,d0=1,d1=1,is_2D=1","support","0","no","BLAS"
 "BLAS","IM2COL_3D","type_input=f32,type_kernel=f32,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","BLAS"
 "BLAS","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f32,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","BLAS"
 "BLAS","IM2COL_3D","type_input=f32,type_kernel=f16,dst_type=f16,ne_input=[10,10,10,9],ne_kernel=[3,3,3,1],IC=3,s0=1,s1=1,s2=1,p0=1,p1=1,p2=1,d0=1,d1=1,d2=1,v=0","support","0","no","BLAS"
@@ -4964,6 +4965,7 @@
 "BLAS","CONV_TRANSPOSE_1D","ne_input=[2,1,1,1],ne_kernel=[3,1,1,1],s0=1,p0=0,d0=1","support","0","no","BLAS"
 "BLAS","CONV_TRANSPOSE_2D","ne_input=[3,2,3,1],ne_kernel=[2,2,1,3],stride=1","support","0","no","BLAS"
 "BLAS","CONV_TRANSPOSE_2D","ne_input=[10,10,9,1],ne_kernel=[3,3,1,9],stride=2","support","0","no","BLAS"
+"BLAS","CONV_TRANSPOSE_2D","ne_input=[129,63,35,1],ne_kernel=[3,3,48,35],stride=1","support","0","no","BLAS"
 "BLAS","COUNT_EQUAL","type=f32,ne=[4,500,1,1]","support","0","no","BLAS"
 "BLAS","COUNT_EQUAL","type=f32,ne=[4,5000,1,1]","support","0","no","BLAS"
 "BLAS","ARGMAX","type=f32,ne=[32,1,1,1]","support","0","no","BLAS"
@@ -5715,15 +5717,15 @@
 "BLAS","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","BLAS"
 "BLAS","RMS_NORM","type=f32,ne=[64,5,4,3],v=0,eps=0.000001,inplace=1","support","0","no","BLAS"
 "BLAS","L2_NORM","type=f32,ne=[64,5,4,3]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[8,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
-"BLAS","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[6,1024,1,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,1024,4,1],ne_b=[3,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[6,1536,1,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,1536,4,1],ne_b=[3,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[6,2048,1,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[3,2048,4,1],ne_b=[3,2048,1,1]","support","0","no","BLAS"
 "BLAS","SSM_CONV","type=f32,ne_a=[4,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","BLAS"
 "BLAS","SSM_CONV","type=f32,ne_a=[8,1024,1,1],ne_b=[4,1024,1,1]","support","0","no","BLAS"
 "BLAS","SSM_CONV","type=f32,ne_a=[4,1024,4,1],ne_b=[4,1024,1,1]","support","0","no","BLAS"
@@ -5733,6 +5735,15 @@
 "BLAS","SSM_CONV","type=f32,ne_a=[4,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","BLAS"
 "BLAS","SSM_CONV","type=f32,ne_a=[8,2048,1,1],ne_b=[4,2048,1,1]","support","0","no","BLAS"
 "BLAS","SSM_CONV","type=f32,ne_a=[4,2048,4,1],ne_b=[4,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,1024,1,1],ne_b=[9,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[18,1024,1,1],ne_b=[9,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,1024,4,1],ne_b=[9,1024,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,1536,1,1],ne_b=[9,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[18,1536,1,1],ne_b=[9,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,1536,4,1],ne_b=[9,1536,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,2048,1,1],ne_b=[9,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[18,2048,1,1],ne_b=[9,2048,1,1]","support","0","no","BLAS"
+"BLAS","SSM_CONV","type=f32,ne_a=[9,2048,4,1],ne_b=[9,2048,1,1]","support","0","no","BLAS"
 "BLAS","SSM_SCAN","type=f32,d_state=16,head_dim=1,n_head=1024,n_group=1,n_seq_tokens=32,n_seqs=4","support","0","no","BLAS"
 "BLAS","SSM_SCAN","type=f32,d_state=128,head_dim=64,n_head=16,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","BLAS"
 "BLAS","SSM_SCAN","type=f32,d_state=256,head_dim=64,n_head=8,n_group=2,n_seq_tokens=32,n_seqs=4","support","0","no","BLAS"
@@ -6592,6 +6603,30 @@
 "BLAS","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=67,bs=[1,1],nr=[4,1],per=[0,2,1,3],k_v=0,o=1","support","0","no","BLAS"
 "BLAS","MUL_MAT","type_a=f32,type_b=f32,m=64,n=77,k=77,bs=[12,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","BLAS"
 "BLAS","MUL_MAT","type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","1","yes","BLAS"
+"BLAS","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=2048,k=8192,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=f32,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=f16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=bf16,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q4_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q4_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q5_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q5_1,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q8_0,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=mxfp4,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q2_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q3_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q4_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q5_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=q6_K,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq2_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq2_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq2_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq3_xxs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq1_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq1_m,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq4_nl,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq3_s,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
+"BLAS","MUL_MAT","type_a=iq4_xs,type_b=f32,m=1,n=64,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1","support","0","no","BLAS"
 "BLAS","MUL_MAT","type_a=f16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","0","no","BLAS"
 "BLAS","MUL_MAT","type_a=f16,type_b=f32,m=128,n=1,k=1056,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=2112,o=1","support","0","no","BLAS"
 "BLAS","MUL_MAT","type_a=bf16,type_b=f32,m=1056,n=1,k=128,bs=[1,1],nr=[1,1],per=[0,2,1,3],k_v=0,o=1","support","0","no","BLAS"
@@ -8916,6 +8951,11 @@
 "BLAS","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=0,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=0.000000,inplace=0","support","0","no","BLAS"
 "BLAS","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","0","no","BLAS"
 "BLAS","SOFT_MAX","type=f32,ne=[32,2,32,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","0","no","BLAS"
+"BLAS","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f32,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","0","no","BLAS"
+"BLAS","SOFT_MAX","type=f32,ne=[200001,2,3,1],mask=1,sinks=1,m_prec=f16,nr23=[1,1],scale=0.100000,max_bias=8.000000,inplace=0","support","0","no","BLAS"
+"BLAS","SOFT_MAX","type=f32,ne=[200000,1,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","0","no","BLAS"
+"BLAS","SOFT_MAX","type=f32,ne=[200000,4,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","0","no","BLAS"
+"BLAS","SOFT_MAX","type=f32,ne=[643251,3,1,1],mask=0,sinks=0,m_prec=f32,nr23=[1,1],scale=1.000000,max_bias=0.000000,inplace=0","support","0","no","BLAS"
 "BLAS","SOFT_MAX_BACK","type=f32,ne=[16,16,1,1],scale=1.000000,max_bias=0.000000","support","0","no","BLAS"
 "BLAS","SOFT_MAX_BACK","type=f32,ne=[15,15,1,1],scale=1.000000,max_bias=0.000000","support","0","no","BLAS"
 "BLAS","SOFT_MAX_BACK","type=f32,ne=[16,16,2,3],scale=1.000000,max_bias=0.000000","support","0","no","BLAS"
@@ -8968,6 +9008,7 @@
 "BLAS","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -8977,6 +9018,7 @@
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -8987,11 +9029,13 @@
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
@@ -9001,6 +9045,7 @@
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
@@ -9011,11 +9056,13 @@
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
@@ -9025,6 +9072,7 @@
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
@@ -9035,11 +9083,13 @@
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
@@ -9049,6 +9099,7 @@
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
@@ -9059,6 +9110,7 @@
 "BLAS","ROPE","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -9184,6 +9236,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -9193,6 +9246,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -9203,11 +9257,13 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
@@ -9217,6 +9273,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
@@ -9227,11 +9284,13 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
@@ -9241,6 +9300,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
@@ -9251,11 +9311,13 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,40,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,52,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,64,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,1,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,71,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,8,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
@@ -9265,6 +9327,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=20,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,2,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,32,4,1],n_dims=32,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=128,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,12,2,1],n_dims=20,mode=8,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
@@ -9275,6 +9338,7 @@
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,28,2,1],n_dims=32,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[80,16,2,1],n_dims=80,mode=24,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[128,16,2,1],n_dims=128,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
+"BLAS","ROPE_BACK","type=f32,ne_a=[16,16,8192,1],n_dims=16,mode=40,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f32,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=1,v=1,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f16,ne_a=[128,32,2,1],n_dims=128,mode=0,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
 "BLAS","ROPE_BACK","type=f16,ne_a=[64,128,2,1],n_dims=64,mode=2,n_ctx=512,fs=1.000000,ef=0.000000,af=1.000000,ff=0,v=0,inplace=0","support","0","no","BLAS"
@@ -9542,333 +9606,333 @@
 "BLAS","ARGSORT","type=f32,ne=[2048,2,1,3],order=1","support","0","no","BLAS"
 "BLAS","ARGSORT","type=f32,ne=[2049,2,1,3],order=1","support","0","no","BLAS"
 "BLAS","ARGSORT","type=f32,ne=[2,8,8192,1],order=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[12,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[13,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[13,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=100","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=500","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=1023","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=9999","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=1","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=2","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=3","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=7","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=15","support","0","no","BLAS"
-"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=15","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[12,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[13,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[13,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[15,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[19,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[27,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[43,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[64,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[75,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[128,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[139,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[256,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[267,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[512,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[523,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1035,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2059,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4096,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[4107,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8192,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[8203,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16395,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32768,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[32779,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65536,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[65547,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131072,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[131083,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262144,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[262155,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=100,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=500,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=1023,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524288,1,1,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[524299,1,2,1],k=9999,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=1,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=2,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=3,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=7,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16,10,10,10],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[60,10,10,10],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1023,2,1,3],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1024,2,1,3],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[1025,2,1,3],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[16384,1,1,1],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2047,2,1,3],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2048,2,1,3],k=15,ties=0","support","0","no","BLAS"
+"BLAS","TOP_K","type=f32,ne=[2049,2,1,3],k=15,ties=0","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=0","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=nearest,transpose=1","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest,flags=none","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=nearest","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=nearest","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=0","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear,transpose=1","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=0","support","0","no","BLAS"
 "BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bicubic,transpose=1","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=0","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=513,transpose=1","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear,flags=none","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear,flags=align_corners","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear,flags=align_corners","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear,flags=align_corners","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic,flags=align_corners","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic,flags=align_corners","support","0","no","BLAS"
-"BLAS","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic,flags=align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bicubic","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=0","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[512,512,3,2],scale_factor=2,mode=bilinear|antialias,transpose=1","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|antialias","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[5,7,11,13],ne_tgt=[2,5,7,11],mode=bilinear|antialias","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bilinear|align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bilinear|align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bilinear|align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[2,5,7,11],ne_tgt=[5,7,11,13],mode=bicubic|align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[1,4,3,2],ne_tgt=[2,8,3,2],mode=bicubic|align_corners","support","0","no","BLAS"
+"BLAS","UPSCALE","type=f32,ne=[4,1,3,2],ne_tgt=[1,1,3,2],mode=bicubic|align_corners","support","0","no","BLAS"
 "BLAS","SUM","type=f32,ne=[10,5,4,3]","support","0","no","BLAS"
 "BLAS","SUM_ROWS","type=f32,ne=[10,5,4,3],permute=0,slice=0","support","0","no","BLAS"
 "BLAS","SUM","type=f32,ne=[11,5,6,3],permute=[0,2,1,3]","support","0","no","BLAS"
@@ -9891,8 +9955,9 @@
 "BLAS","GROUP_NORM","type=f32,ne=[64,64,320,1],num_groups=32,eps=0.000001","support","0","no","BLAS"
 "BLAS","GROUP_NORM","type=f32,ne=[9,9,1280,1],num_groups=32,eps=0.000001","support","0","no","BLAS"
 "BLAS","ACC","type=f32,ne_a=[256,17,1,1],ne_b=[256,16,1,1]","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,1,1],pad_0=1,pad_1=1,circular=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[33,17,2,1],pad_0=4,pad_1=3,circular=1","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,3,1],lp0=1,rp0=1,lp1=1,rp1=1,lp2=1,rp2=1,lp3=1,rp3=1,v=0,circular=0","support","0","no","BLAS"
 "BLAS","PAD_REFLECT_1D","type=f32,ne_a=[512,34,2,1],pad_0=10,pad_1=9","support","0","no","BLAS"
 "BLAS","PAD_REFLECT_1D","type=f32,ne_a=[3000,384,4,1],pad_0=10,pad_1=9","support","0","no","BLAS"
 "BLAS","ROLL","shift0=3,shift1=-2,shift3=1,shift4=-1","support","0","no","BLAS"
@@ -9914,6 +9979,7 @@
 "BLAS","CUMSUM","type=f32,ne=[2048,5,4,3]","support","0","no","BLAS"
 "BLAS","CUMSUM","type=f32,ne=[242004,1,1,1]","support","0","no","BLAS"
 "BLAS","CUMSUM","type=f32,ne=[375960,1,1,1]","support","0","no","BLAS"
+"BLAS","CUMSUM","type=f32,ne=[20481,4,1,1]","support","0","no","BLAS"
 "BLAS","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","BLAS"
 "BLAS","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","0","no","BLAS"
 "BLAS","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","0","no","BLAS"
@@ -9923,17 +9989,41 @@
 "BLAS","FILL","type=f32,ne=[303,207,11,3],c=2.000000","support","0","no","BLAS"
 "BLAS","FILL","type=f32,ne=[800,600,4,4],c=-152.000000","support","0","no","BLAS"
 "BLAS","FILL","type=f32,ne=[2048,512,2,2],c=3.500000","support","0","no","BLAS"
+"BLAS","DIAG","type=f32,ne=[10,1,4,3]","support","0","no","BLAS"
+"BLAS","DIAG","type=f32,ne=[79,1,19,13]","support","0","no","BLAS"
+"BLAS","DIAG","type=f32,ne=[256,1,8,16]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[10,10,4,3],ne_rhs=[3,10,4,3]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[11,11,1,1],ne_rhs=[5,11,1,1]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[17,17,2,4],ne_rhs=[9,17,2,4]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[30,30,7,1],ne_rhs=[8,30,7,1]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[42,42,5,2],ne_rhs=[10,42,5,2]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[10,64,2,2]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[64,64,2,2],ne_rhs=[64,64,2,2]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[79,79,5,3],ne_rhs=[417,79,5,3]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,2],ne_rhs=[32,128,4,2]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[80,80,2,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[79,80,2,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,2,8],ne_rhs=[81,80,2,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[80,80,8,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[79,80,8,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[80,80,8,8],ne_rhs=[81,80,8,8]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[84,84,4,4],ne_rhs=[32,84,4,4]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[95,95,8,8],ne_rhs=[40,95,8,8]","support","0","no","BLAS"
 "BLAS","SOLVE_TRI","type=f32,ne_lhs=[100,100,4,4],ne_rhs=[41,100,4,4]","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1","support","0","no","BLAS"
-"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[31,128,4,4]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,4],ne_rhs=[32,128,4,4]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[128,128,3,4],ne_rhs=[32,128,3,4]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[128,128,4,1],ne_rhs=[32,128,4,1]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[200,64,4,4]","support","0","no","BLAS"
+"BLAS","SOLVE_TRI","type=f32,ne_lhs=[64,64,4,4],ne_rhs=[384,64,4,4]","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=0,circular=1","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=0,circular=1","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=0","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[512,512,1,1],lp0=0,rp0=1,lp1=0,rp1=1,lp2=0,rp2=0,lp3=0,rp3=0,v=1,circular=1","support","0","no","BLAS"
+"BLAS","PAD","type=f32,ne_a=[11,22,33,44],lp0=1,rp0=2,lp1=3,rp1=4,lp2=5,rp2=6,lp3=7,rp3=8,v=1,circular=1","support","0","no","BLAS"
 "BLAS","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f32,permute=[0,1,2,3]","support","0","no","BLAS"
 "BLAS","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=f16,permute=[0,1,2,3]","support","0","no","BLAS"
 "BLAS","FLASH_ATTN_EXT","hsk=40,hsv=40,nh=4,nr23=[1,1],kv=113,nb=1,mask=1,sinks=1,max_bias=0.000000,logit_softcap=0.000000,prec=f32,type_KV=bf16,permute=[0,1,2,3]","support","0","no","BLAS"

docs/preset.md

@@ -0,0 +1,97 @@
+# llama.cpp INI Presets
+
+## Introduction
+
+The INI preset feature, introduced in [PR#17859](https://github.com/ggml-org/llama.cpp/pull/17859), allows users to create reusable and shareable parameter configurations for llama.cpp.
+
+### Using Presets with the Server
+
+When running multiple models on the server (router mode), INI preset files can be used to configure model-specific parameters. Please refer to the [server documentation](../tools/server/README.md) for more details.
+
+### Using a Remote Preset
+
+> [!NOTE]
+>
+> This feature is currently only supported via the `-hf` option.
+
+For GGUF models hosted on Hugging Face, you can include a `preset.ini` file in the root directory of the repository to define specific configurations for that model.
+
+Example:
+
+```ini
+hf-repo-draft = username/my-draft-model-GGUF
+temp = 0.5
+top-k = 20
+top-p = 0.95
+```
+
+For security reasons, only certain options are allowed. Please refer to [preset.cpp](../common/preset.cpp) for the complete list of permitted options.
+
+Example usage:
+
+Assuming your repository `username/my-model-with-preset` contains a `preset.ini` with the configuration above:
+
+```sh
+llama-cli -hf username/my-model-with-preset
+
+# This is equivalent to:
+llama-cli -hf username/my-model-with-preset \
+  --hf-repo-draft username/my-draft-model-GGUF \
+  --temp 0.5 \
+  --top-k 20 \
+  --top-p 0.95
+```
+
+You can also override preset arguments by specifying them on the command line:
+
+```sh
+# Force temp = 0.1, overriding the preset value
+llama-cli -hf username/my-model-with-preset --temp 0.1
+```
+
+If you want to define multiple preset configurations for one or more GGUF models, you can create a blank HF repo for each preset. Each HF repo should contain a `preset.ini` file that references the actual model(s):
+
+```ini
+hf-repo = user/my-model-main
+hf-repo-draft = user/my-model-draft
+temp = 0.8
+ctx-size = 1024
+; (and other configurations)
+```
+
+### Named presets
+
+If you want to define multiple preset configurations for one or more GGUF models, you can create a blank HF repo containing a single `preset.ini` file that references the actual model(s):
+
+```ini
+[*]
+mmap = 1
+
+[gpt-oss-20b-hf]
+hf          = ggml-org/gpt-oss-20b-GGUF
+batch-size  = 2048
+ubatch-size = 2048
+top-p       = 1.0
+top-k       = 0
+min-p       = 0.01
+temp        = 1.0
+chat-template-kwargs = {"reasoning_effort": "high"}
+
+[gpt-oss-120b-hf]
+hf          = ggml-org/gpt-oss-120b-GGUF
+batch-size  = 2048
+ubatch-size = 2048
+top-p       = 1.0
+top-k       = 0
+min-p       = 0.01
+temp        = 1.0
+chat-template-kwargs = {"reasoning_effort": "high"}
+```
+
+You can then use it via `llama-cli` or `llama-server`, example:
+
+```sh
+llama-server -hf user/repo:gpt-oss-120b-hf
+```
+
+Please make sure to provide the correct `hf-repo` for each child preset. Otherwise, you may get error: `The specified tag is not a valid quantization scheme.`

examples/CMakeLists.txt

@@ -15,6 +15,7 @@ llama_add_compile_flags()
 if (EMSCRIPTEN)
 else()
     add_subdirectory(batched)
+    add_subdirectory(debug)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
 
@@ -34,7 +35,6 @@ else()
     add_subdirectory(gen-docs)
     add_subdirectory(training)
     add_subdirectory(diffusion)
-    add_subdirectory(model-conversion)
     if (NOT GGML_BACKEND_DL)
         add_subdirectory(convert-llama2c-to-ggml)
         # these examples use the backends directly and cannot be built with dynamic loading

examples/batched/batched.cpp

@@ -21,7 +21,7 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BATCHED, print_usage)) {
         return 1;
     }
 

examples/debug/CMakeLists.txt

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

examples/debug/README.md

@@ -0,0 +1,54 @@
+# llama.cpp/examples/debug
+
+This is a utility intended to help debug a model by registering a callback that
+logs GGML operations and tensor data. It can also store the generated logits or
+embeddings as well as the prompt and token ids for comparision with the original
+model.
+
+### Usage
+
+```shell
+llama-debug \
+  --hf-repo ggml-org/models \
+  --hf-file phi-2/ggml-model-q4_0.gguf \
+  --model phi-2-q4_0.gguf \
+  --prompt hello \
+  --save-logits \
+  --verbose
+```
+The tensor data is logged as debug and required the --verbose flag. The reason
+for this is that while useful for a model with many layers there can be a lot of
+output. You can filter the tensor names using the `--tensor-filter` option.
+
+A recommended approach is to first run without `--verbose` and see if the
+generated logits/embeddings are close to the original model. If they are not,
+then it might be required to inspect tensor by tensor and in that case it is
+useful to enable the `--verbose` flag along with `--tensor-filter` to focus on
+specific tensors.
+
+### Options
+This example supports all standard `llama.cpp` options and also accepts the
+following options:
+```console
+$ llama-debug --help
+...
+
+----- example-specific params -----
+
+--save-logits                           save final logits to files for verification (default: false)
+--logits-output-dir PATH                directory for saving logits output files (default: data)
+--tensor-filter REGEX                   filter tensor names for debug output (regex pattern, can be specified multiple times)
+```
+
+### Output Files
+
+When `--save-logits` is enabled, the following files are created in the output
+directory:
+
+* `llamacpp-<model>[-embeddings].bin`        - Binary output (logits or embeddings)
+* `llamacpp-<model>[-embeddings].txt`        - Text output (logits or embeddings, one per line)
+* `llamacpp-<model>[-embeddings]-prompt.txt` - Prompt text and token IDs
+* `llamacpp-<model>[-embeddings]-tokens.bin` - Binary token IDs for programmatic comparison
+
+These files can be compared against the original model's output to verify the
+converted model.

examples/debug/debug.cpp

@@ -0,0 +1,439 @@
+#include "arg.h"
+#include "common.h"
+#include "log.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdlib>
+#include <string>
+#include <vector>
+#include <filesystem>
+#include <fstream>
+#include <regex>
+
+static void print_usage(int, char ** argv) {
+    const std::string usage_template = R"(
+        example usage:
+
+          Print tensors:
+
+          {prog} -m model.gguf -p "Hello my name is" --verbose
+
+          The tensors to be printed can be filtered with --tensor-filter option.
+
+          Save logits/embeddings:
+
+          {prog} -m model.gguf -p "Hello my name is" --save-logits
+
+          Add --embedding to save embeddings)" "\n";
+
+    // Fix the source code indentation above that is introduced by the raw string literal.
+    std::string usage = std::regex_replace(usage_template, std::regex("\\n {8}"), "\n");
+    usage = std::regex_replace(usage, std::regex("\\{prog\\}"), argv[0]);
+    LOG("%s\n", usage.c_str());
+}
+
+static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data);
+
+struct callback_data {
+    std::vector<uint8_t>    data;
+    std::vector<std::regex> tensor_filters;
+
+    callback_data() = default;
+
+    callback_data(common_params & params, const std::vector<std::string> & filter_patterns) {
+        for (const auto & pattern : filter_patterns) {
+            try {
+                std::string anchored_pattern = "^" + pattern;
+                tensor_filters.emplace_back(anchored_pattern, std::regex::optimize);
+            } catch (const std::regex_error & e) {
+                throw std::runtime_error("Invalid regex pattern '" + pattern + "': " + e.what());
+            }
+        }
+        params.cb_eval           = ggml_debug;
+        params.cb_eval_user_data = this;
+    }
+};
+
+static bool has_pooling(llama_context * ctx) {
+    switch (llama_pooling_type(ctx)) {
+        case LLAMA_POOLING_TYPE_NONE:
+        case LLAMA_POOLING_TYPE_UNSPECIFIED:
+            return false;
+        default:
+            return true;
+    }
+}
+
+struct output_data {
+    float *                  data_ptr    = nullptr;
+    int                      data_size   = 0;
+    std::string              type_suffix;
+    std::vector<float>       embd_norm;
+    std::string              prompt;
+    std::vector<llama_token> tokens;
+
+    output_data(llama_context * ctx, const llama_model * model, const common_params & params) {
+        const llama_vocab * vocab = llama_model_get_vocab(model);
+        const bool add_bos = llama_vocab_get_add_bos(vocab);
+
+        tokens = common_tokenize(ctx, params.prompt, add_bos);
+        prompt = params.prompt;
+
+        if (params.embedding) {
+            const int n_embd       = llama_model_n_embd_out(model);
+            const bool pooling     = has_pooling(ctx);
+            const int n_embd_count = pooling ? 1 : tokens.size();
+            const int n_floats     = n_embd * n_embd_count;
+
+            float * embd_raw = pooling ? llama_get_embeddings_seq(ctx, 0) : llama_get_embeddings(ctx);
+            if (embd_raw == nullptr) {
+                throw std::runtime_error("failed to get embeddings from the model");
+            }
+
+            LOG_DBG("pooling_enabled: %s\n", pooling ? "true" : "false");
+            LOG_DBG("n_embd: %d\n", n_embd);
+            LOG_DBG("n_floats: %d\n", n_floats);
+            LOG_DBG("n_embd_count: %d\n", n_embd_count);
+
+            data_ptr    = embd_raw;
+            data_size   = n_floats;
+            type_suffix = "-embeddings";
+
+            if (params.embd_normalize >= 0) {
+                embd_norm.resize(n_floats);
+                for (int i = 0; i < n_embd_count; i++) {
+                    common_embd_normalize(embd_raw+i*n_embd, embd_norm.data()+i*n_embd, n_embd, params.embd_normalize);
+                }
+                data_ptr = embd_norm.data();
+            }
+        } else {
+            const float * logits = llama_get_logits_ith(ctx, tokens.size() - 1);
+            const int n_logits = llama_vocab_n_tokens(vocab);
+
+            data_ptr = const_cast<float*>(logits);
+            data_size = n_logits;
+            type_suffix = "";
+        }
+    }
+};
+
+static std::string ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
+        }
+    }
+    return str;
+}
+
+static inline float ggml_compute_bf16_to_fp32(ggml_bf16_t h) {
+    union {
+        float f;
+        uint32_t i;
+    } u;
+    u.i = (uint32_t)h.bits << 16;
+    return u.f;
+}
+
+static float ggml_get_float_value(const uint8_t * data, ggml_type type,
+        const size_t * nb, size_t i0, size_t i1, size_t i2, size_t i3) {
+    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+    switch (type) {
+        case GGML_TYPE_F16:
+            return ggml_fp16_to_fp32(*(const ggml_fp16_t *) &data[i]);
+        case GGML_TYPE_F32:
+            return *(const float *) &data[i];
+        case GGML_TYPE_I64:
+            return (float) *(const int64_t *) &data[i];
+        case GGML_TYPE_I32:
+            return (float) *(const int32_t *) &data[i];
+        case GGML_TYPE_I16:
+            return (float) *(const int16_t *) &data[i];
+        case GGML_TYPE_I8:
+            return (float) *(const int8_t *) &data[i];
+        case GGML_TYPE_BF16:
+            return ggml_compute_bf16_to_fp32(*(const ggml_bf16_t *) &data[i]);
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+    GGML_ASSERT(n > 0);
+    float sum    = 0;
+    float sum_sq = 0.0;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
+                    sum    += v;
+                    sum_sq += v * v;
+                }
+            }
+        }
+    }
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        LOG_DBG("                                     [\n");
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2*n) {
+                LOG_DBG("                                      ..., \n");
+                i2 = ne[2] - n;
+            }
+            LOG_DBG("                                      [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2*n) {
+                    LOG_DBG("                                       ..., \n");
+                    i1 = ne[1] - n;
+                }
+                LOG_DBG("                                       [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2*n) {
+                        LOG_DBG("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    const float v = ggml_get_float_value(data, type, nb, i0, i1, i2, i3);
+                    LOG_DBG("%12.4f", v);
+                    if (i0 < ne[0] - 1) {
+                        LOG_DBG(", ");
+                    }
+                }
+                LOG_DBG("],\n");
+            }
+            LOG_DBG("                                      ],\n");
+        }
+        LOG_DBG("                                     ]\n");
+        LOG_DBG("                                     sum    = %f\n", sum);
+        LOG_DBG("                                     sum_sq = %f\n", sum_sq);
+    }
+
+    if (std::isnan(sum)) {
+        LOG_ERR("encountered NaN - aborting\n");
+        exit(0);
+    }
+}
+
+/**
+ * GGML operations callback during the graph execution.
+ *
+ * @param t current tensor
+ * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
+ *            if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
+ *            see ggml_backend_sched_eval_callback
+ * @param user_data user data to pass at each call back
+ * @return true to receive data or continue the graph, false otherwise
+ */
+static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+
+    const struct ggml_tensor * src0 = t->src[0];
+    const struct ggml_tensor * src1 = t->src[1];
+
+    if (ask) {
+        return true; // Always retrieve data
+    }
+
+    bool matches_filter = cb_data->tensor_filters.empty();
+
+    if (!matches_filter) {
+        for (const auto & filter : cb_data->tensor_filters) {
+            if (std::regex_search(t->name, filter)) {
+                matches_filter = true;
+                break;
+            }
+        }
+    }
+
+    char src1_str[128] = {0};
+    if (src1) {
+        snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
+    }
+
+    if (matches_filter) {
+        LOG_DBG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
+             t->name,
+             ggml_type_name(t->type),
+             ggml_op_desc(t),
+             src0->name,
+             ggml_ne_string(src0).c_str(),
+             src1 ? src1_str : "",
+             ggml_ne_string(t).c_str());
+    }
+
+    const bool is_host = ggml_backend_buffer_is_host(t->buffer);
+
+    if (!is_host) {
+        auto n_bytes = ggml_nbytes(t);
+        cb_data->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+    }
+
+    if (!ggml_is_quantized(t->type) && matches_filter) {
+        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
+        ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
+    }
+
+    return true;
+}
+
+
+static void save_output_data(const output_data & output, const std::string & model_name, const std::string & output_dir) {
+    std::filesystem::create_directory(output_dir);
+    auto base_path = std::filesystem::path{output_dir} / ("llamacpp-" + model_name + output.type_suffix);
+
+    // Save logits/embeddings to binary file.
+    {
+        std::filesystem::path filepath{base_path.string() + ".bin"};
+        std::ofstream file{filepath, std::ios::binary};
+        if (!file) {
+            throw std::runtime_error("failed to open binary output file: " + filepath.string());
+        }
+        file.write(reinterpret_cast<const char*>(output.data_ptr), output.data_size * sizeof(float));
+        LOG("Data saved to %s\n", filepath.c_str());
+    }
+
+    // Save logits/embeddings to text file.
+    {
+        std::filesystem::path filepath{base_path.string() + ".txt"};
+        std::ofstream file{filepath};
+        if (!file) {
+            throw std::runtime_error("failed to open text output file: " + filepath.string());
+        }
+        for (int i = 0; i < output.data_size; i++) {
+            file << i << ": " << output.data_ptr[i] << '\n';
+        }
+        LOG("Data saved to %s\n", filepath.c_str());
+    }
+
+    // Save prompt and tokens to text file.
+    {
+        std::filesystem::path filepath{base_path.string() + "-prompt.txt"};
+        std::ofstream file{filepath};
+        if (!file) {
+            throw std::runtime_error("failed to open prompt output file: " + filepath.string());
+        }
+
+        file << "prompt: " << output.prompt << '\n';
+        file << "n_tokens: " << output.tokens.size() << '\n';
+
+        file << "token ids: ";
+        for (size_t i = 0; i < output.tokens.size(); i++) {
+            file << output.tokens[i];
+            if (i + 1 < output.tokens.size()) {
+                file << ", ";
+            }
+        }
+        file << '\n';
+        LOG("Prompt saved to %s\n", filepath.c_str());
+    }
+
+    // Save token ids to binary file.
+    {
+        std::filesystem::path filepath{base_path.string() + "-tokens.bin"};
+        std::ofstream file{filepath, std::ios::binary};
+        if (!file) {
+            throw std::runtime_error("failed to open tokens binary file: " + filepath.string());
+        }
+        file.write(reinterpret_cast<const char*>(output.tokens.data()), output.tokens.size() * sizeof(llama_token));
+        LOG("Tokens saved to %s\n", filepath.c_str());
+    }
+
+}
+
+static void print_tokenized_prompt(llama_context * ctx, const std::vector<llama_token> & tokens, const std::string & prompt) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    LOG("Model add_bos: %s\n", llama_vocab_get_add_bos(vocab) ? "true" : "false");
+    LOG("Input prompt: \"%s\"\n", prompt.c_str());
+    LOG("Token ids (%zu):\n", tokens.size());
+
+    for (auto id : tokens) {
+        std::string piece(128, '\0');
+        int n = llama_token_to_piece(vocab, id, piece.data(), piece.size(), 0, true);
+        if (n < 0) {
+            LOG_ERR("failed to convert token %d to piece\n", id);
+            continue;
+        }
+        piece.resize(n);
+        LOG("%s(%d) ", piece.c_str(), id);
+    }
+    LOG("\n");
+}
+
+static bool run(llama_context * ctx, const common_params & params) {
+    const llama_model * model = llama_get_model(ctx);
+    const llama_vocab * vocab = llama_model_get_vocab(model);
+
+    const bool add_bos = llama_vocab_get_add_bos(vocab);
+
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
+
+    if (tokens.empty()) {
+        LOG_ERR("%s : there are not input tokens to process - (try to provide a prompt with '-p')\n", __func__);
+        return false;
+    }
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
+        LOG_ERR("%s : failed to eval\n", __func__);
+        return false;
+    }
+
+    print_tokenized_prompt(ctx, tokens, params.prompt);
+
+    if (params.save_logits) {
+        output_data output {ctx, model, params};
+        std::filesystem::path model_path{params.model.path};
+        std::string model_name{model_path.stem().string()};
+        save_output_data(output, model_name, params.logits_output_dir);
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    common_params params;
+
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_DEBUG, print_usage)) {
+        return 1;
+    }
+
+    common_init();
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    callback_data cb_data(params, params.tensor_filter);
+
+    auto llama_init = common_init_from_params(params);
+
+    auto * model = llama_init->model();
+    auto * ctx   = llama_init->context();
+
+    if (model == nullptr || ctx == nullptr) {
+        LOG_ERR("%s : failed to init\n", __func__);
+        return 1;
+    }
+
+    {
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+    }
+
+    if (!run(ctx, params)) {
+        return 1;
+    }
+
+    LOG("\n");
+    llama_perf_context_print(ctx);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/diffusion/diffusion-cli.cpp

@@ -553,6 +553,7 @@ int main(int argc, char ** argv) {
     model_params.n_gpu_layers       = params.n_gpu_layers;
     model_params.devices            = params.devices.data();
     model_params.use_mmap           = params.use_mmap;
+    model_params.use_direct_io      = params.use_direct_io;
     model_params.use_mlock          = params.use_mlock;
     model_params.check_tensors      = params.check_tensors;
 

examples/eval-callback/CMakeLists.txt

@@ -6,10 +6,8 @@ target_compile_features(${TARGET} PRIVATE cxx_std_17)
 
 set(TEST_TARGET test-eval-callback)
 if(NOT ${CMAKE_SYSTEM_PROCESSOR} MATCHES "s390x")
-        add_test(NAME ${TEST_TARGET}
-                        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+    llama_download_model("tinyllamas/stories15M-q4_0.gguf" SHA256=66967fbece6dbe97886593fdbb73589584927e29119ec31f08090732d1861739)
 else()
-        add_test(NAME ${TEST_TARGET}
-                        COMMAND llama-eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K-be.gguf --model stories260K-be.gguf --prompt hello --seed 42 -ngl 0)
+    llama_download_model("tinyllamas/stories15M-be.Q4_0.gguf" SHA256=9aec857937849d976f30397e97eb1cabb53eb9dcb1ce4611ba8247fb5f44c65d)
 endif()
-set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)
+add_test(NAME ${TEST_TARGET} COMMAND llama-eval-callback -m "${LLAMA_DOWNLOAD_MODEL}" --prompt hello --seed 42 -ngl 0)

examples/model-conversion/Makefile

@@ -61,7 +61,7 @@ causal-run-converted-model:
 	@CONVERTED_MODEL="$(CONVERTED_MODEL)" ./scripts/causal/run-converted-model.sh
 
 causal-verify-logits: causal-run-original-model causal-run-converted-model
-	@./scripts/causal/compare-logits.py
+	@MODEL_PATH="$(MODEL_PATH)" ./scripts/causal/compare-logits.py
 	@MODEL_PATH="$(MODEL_PATH)" ./scripts/utils/check-nmse.py -m ${MODEL_PATH}
 
 causal-run-original-embeddings:
@@ -138,16 +138,13 @@ embedding-run-original-model-st: embedding-run-original-model
 embedding-run-converted-model:
 	@./scripts/embedding/run-converted-model.sh $(CONVERTED_EMBEDDING_MODEL) \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)") \
-	$(if $(USE_POOLING),--pooling)
-
-embedding-run-converted-model-st: USE_POOLING=1
-embedding-run-converted-model-st: embedding-run-converted-model
+	$(if $(EMBD_NORMALIZE),--embd-normalize "$(EMBD_NORMALIZE)")
 
 embedding-verify-logits: embedding-run-original-model embedding-run-converted-model
 	@./scripts/embedding/compare-embeddings-logits.sh \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 
-embedding-verify-logits-st: embedding-run-original-model-st embedding-run-converted-model-st
+embedding-verify-logits-st: embedding-run-original-model-st embedding-run-converted-model
 	@./scripts/embedding/compare-embeddings-logits.sh \
 	$(if $(PROMPTS_FILE),--prompts-file "$(PROMPTS_FILE)")
 

examples/model-conversion/README.md

@@ -198,14 +198,13 @@ model, and the other is a text file which allows for manual visual inspection.
 
 #### Using SentenceTransformer with numbered layers
 For models that have numbered SentenceTransformer layers (01_Pooling, 02_Dense,
-03_Dense, 04_Normalize), use the `-st` targets to apply all these layers:
+03_Dense, 04_Normalize), these will be applied automatically when running the
+converted model but currently there is a separate target to run the original
+version:
 
 ```console
 # Run original model with SentenceTransformer (applies all numbered layers)
 (venv) $ make embedding-run-original-model-st
-
-# Run converted model with pooling enabled
-(venv) $ make embedding-run-converted-model-st
 ```
 
 This will use the SentenceTransformer library to load and run the model, which
@@ -213,6 +212,17 @@ automatically applies all the numbered layers in the correct order. This is
 particularly useful when comparing with models that should include these
 additional transformation layers beyond just the base model output.
 
+The type of normalization can be specified for the converted model but is not
+strictly necessary as the verification uses cosine similarity and the magnitude
+of the output vectors does not affect this. But the normalization type can be
+specified as an argument to the target which might be useful for manual
+inspection:
+```console
+(venv) $ make embedding-verify-logits-st EMBD_NORMALIZE=1
+```
+The original model will apply the normalization according to the normalization
+layer specified in the modules.json configuration file.
+
 ### Model conversion
 After updates have been made to [gguf-py](../../gguf-py) to add support for the
 new model the model can be converted to GGUF format using the following command:

examples/model-conversion/logits.cpp

@@ -1,268 +0,0 @@
-#include "llama.h"
-#include "common.h"
-
-
-#include <cstdio>
-#include <cstring>
-#include <string>
-#include <vector>
-#include <ctype.h>
-#include <filesystem>
-
-static void print_usage(int, char ** argv) {
-    printf("\nexample usage:\n");
-    printf("\n    %s -m model.gguf [-ngl n_gpu_layers] -embd-mode [-pooling] [-embd-norm <norm>] [prompt]\n", argv[0]);
-    printf("\n");
-    printf("  -embd-norm: normalization type for pooled embeddings (default: 2)\n");
-    printf("              -1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm\n");
-    printf("\n");
-}
-
-int main(int argc, char ** argv) {
-    std::string model_path;
-    std::string prompt = "Hello, my name is";
-    int ngl = 0;
-    bool embedding_mode = false;
-    bool pooling_enabled = false;
-    int32_t embd_norm = 2;  // (-1=none, 0=max absolute int16, 1=taxicab, 2=Euclidean/L2, >2=p-norm)
-
-    {
-        int i = 1;
-        for (; i < argc; i++) {
-            if (strcmp(argv[i], "-m") == 0) {
-                if (i + 1 < argc) {
-                    model_path = argv[++i];
-                } else {
-                    print_usage(argc, argv);
-                    return 1;
-                }
-            } else if (strcmp(argv[i], "-ngl") == 0) {
-                if (i + 1 < argc) {
-                    try {
-                        ngl = std::stoi(argv[++i]);
-                    } catch (...) {
-                        print_usage(argc, argv);
-                        return 1;
-                    }
-                } else {
-                    print_usage(argc, argv);
-                    return 1;
-                }
-            } else if (strcmp(argv[i], "-embd-mode") == 0) {
-                embedding_mode = true;
-            } else if (strcmp(argv[i], "-pooling") == 0) {
-                pooling_enabled = true;
-            } else if (strcmp(argv[i], "-embd-norm") == 0) {
-                if (i + 1 < argc) {
-                    try {
-                        embd_norm = std::stoi(argv[++i]);
-                    } catch (...) {
-                        print_usage(argc, argv);
-                        return 1;
-                    }
-                } else {
-                    print_usage(argc, argv);
-                    return 1;
-                }
-            } else {
-                // prompt starts here
-                break;
-            }
-        }
-
-        if (model_path.empty()) {
-            print_usage(argc, argv);
-            return 1;
-        }
-
-        if (i < argc) {
-            prompt = argv[i++];
-            for (; i < argc; i++) {
-                prompt += " ";
-                prompt += argv[i];
-            }
-        }
-    }
-
-    ggml_backend_load_all();
-    llama_model_params model_params = llama_model_default_params();
-    model_params.n_gpu_layers = ngl;
-
-    llama_model * model = llama_model_load_from_file(model_path.c_str(), model_params);
-
-    if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
-        return 1;
-    }
-
-    // Extract basename from model_path
-    const char * basename = strrchr(model_path.c_str(), '/');
-    basename = (basename == NULL) ? model_path.c_str() : basename + 1;
-
-    char model_name[256];
-    strncpy(model_name, basename, 255);
-    model_name[255] = '\0';
-
-    char * dot = strrchr(model_name, '.');
-    if (dot != NULL && strcmp(dot, ".gguf") == 0) {
-        *dot = '\0';
-    }
-    printf("Model name: %s\n", model_name);
-
-    const llama_vocab * vocab = llama_model_get_vocab(model);
-    const int n_prompt = -llama_tokenize(vocab, prompt.c_str(), prompt.size(), NULL, 0, true, true);
-
-    std::vector<llama_token> prompt_tokens(n_prompt);
-    if (llama_tokenize(vocab, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), true, true) < 0) {
-        fprintf(stderr, "%s: error: failed to tokenize the prompt\n", __func__);
-        return 1;
-    }
-
-    llama_context_params ctx_params = llama_context_default_params();
-    ctx_params.n_ctx = n_prompt;
-    ctx_params.n_batch = n_prompt;
-    ctx_params.no_perf = false;
-    if (embedding_mode) {
-        ctx_params.embeddings = true;
-        ctx_params.pooling_type = pooling_enabled ? LLAMA_POOLING_TYPE_MEAN : LLAMA_POOLING_TYPE_NONE;
-        ctx_params.n_ubatch = ctx_params.n_batch;
-    }
-
-    llama_context * ctx = llama_init_from_model(model, ctx_params);
-    if (ctx == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
-        return 1;
-    }
-
-    printf("Input prompt: \"%s\"\n", prompt.c_str());
-    printf("Tokenized prompt (%d tokens): ", n_prompt);
-    for (auto id : prompt_tokens) {
-        char buf[128];
-        int n = llama_token_to_piece(vocab, id, buf, sizeof(buf), 0, true);
-        if (n < 0) {
-            fprintf(stderr, "%s: error: failed to convert token to piece\n", __func__);
-            return 1;
-        }
-        std::string s(buf, n);
-        printf("%s (%d)", s.c_str(), id);
-    }
-    printf("\n");
-
-    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
-
-    if (llama_decode(ctx, batch)) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
-        return 1;
-    }
-
-    float * data_ptr;
-    int data_size;
-    const char * type;
-    std::vector<float> embd_out;
-
-    if (embedding_mode) {
-        const int n_embd_out = llama_model_n_embd_out(model);
-        const int n_embd_count = pooling_enabled ? 1 : batch.n_tokens;
-        const int n_embeddings = n_embd_out * n_embd_count;
-        float * embeddings;
-        type = "-embeddings";
-
-        if (llama_pooling_type(ctx) != LLAMA_POOLING_TYPE_NONE) {
-            embeddings = llama_get_embeddings_seq(ctx, 0);
-            embd_out.resize(n_embeddings);
-            printf("Normalizing embeddings using norm: %d\n", embd_norm);
-            common_embd_normalize(embeddings, embd_out.data(), n_embeddings, embd_norm);
-            embeddings = embd_out.data();
-        } else {
-            embeddings = llama_get_embeddings(ctx);
-        }
-
-        printf("Embedding dimension: %d\n", n_embd_out);
-        printf("\n");
-
-        // Print embeddings in the specified format
-        for (int j = 0; j < n_embd_count; j++) {
-            printf("embedding %d: ", j);
-
-            // Print first 3 values
-            for (int i = 0; i < 3 && i < n_embd_out; i++) {
-                printf("%9.6f ", embeddings[j * n_embd_out + i]);
-            }
-
-            printf(" ... ");
-
-            // Print last 3 values
-            for (int i = n_embd_out - 3; i < n_embd_out; i++) {
-                if (i >= 0) {
-                    printf("%9.6f ", embeddings[j * n_embd_out + i]);
-                }
-            }
-
-            printf("\n");
-        }
-        printf("\n");
-
-        printf("Embeddings size: %d\n", n_embeddings);
-
-        data_ptr = embeddings;
-        data_size = n_embeddings;
-    } else {
-        float * logits = llama_get_logits_ith(ctx, batch.n_tokens - 1);
-        const int n_logits = llama_vocab_n_tokens(vocab);
-        type = "";
-        printf("Vocab size: %d\n", n_logits);
-
-        data_ptr = logits;
-        data_size = n_logits;
-    }
-
-    std::filesystem::create_directory("data");
-
-    // Save data to binary file
-    char bin_filename[512];
-    snprintf(bin_filename, sizeof(bin_filename), "data/llamacpp-%s%s.bin", model_name, type);
-    printf("Saving data to %s\n", bin_filename);
-
-    FILE * f = fopen(bin_filename, "wb");
-    if (f == NULL) {
-        fprintf(stderr, "%s: error: failed to open binary output file\n", __func__);
-        return 1;
-    }
-    fwrite(data_ptr, sizeof(float), data_size, f);
-    fclose(f);
-
-    // Also save as text for debugging
-    char txt_filename[512];
-    snprintf(txt_filename, sizeof(txt_filename), "data/llamacpp-%s%s.txt", model_name, type);
-    f = fopen(txt_filename, "w");
-    if (f == NULL) {
-        fprintf(stderr, "%s: error: failed to open text output file\n", __func__);
-        return 1;
-    }
-    for (int i = 0; i < data_size; i++) {
-        fprintf(f, "%d: %.6f\n", i, data_ptr[i]);
-    }
-    fclose(f);
-
-    if (!embedding_mode) {
-        printf("First 10 logits: ");
-        for (int i = 0; i < 10 && i < data_size; i++) {
-            printf("%.6f ", data_ptr[i]);
-        }
-        printf("\n");
-
-        printf("Last 10 logits: ");
-        for (int i = data_size - 10; i < data_size; i++) {
-            if (i >= 0) printf("%.6f ", data_ptr[i]);
-        }
-        printf("\n\n");
-    }
-
-    printf("Data saved to %s\n", bin_filename);
-    printf("Data saved to %s\n", txt_filename);
-
-    llama_free(ctx);
-    llama_model_free(model);
-
-    return 0;
-}

examples/model-conversion/scripts/causal/compare-logits.py

@@ -3,10 +3,11 @@
 import sys
 import numpy as np
 from pathlib import Path
+import os
 
 # Add utils directory to path for direct script execution
 sys.path.insert(0, str(Path(__file__).parent.parent / "utils"))
-from common import get_model_name_from_env_path  # type: ignore[import-not-found]
+from common import get_model_name_from_env_path, compare_tokens, exit_with_warning  # type: ignore[import-not-found]
 
 def quick_logits_check(pytorch_file, llamacpp_file):
     """Lightweight sanity check before NMSE"""
@@ -38,6 +39,7 @@ def quick_logits_check(pytorch_file, llamacpp_file):
     return True
 
 def main():
+    model_path = os.environ.get('MODEL_PATH')
     model_name = get_model_name_from_env_path('MODEL_PATH')
     data_dir = Path("data")
     pytorch_file = data_dir / f"pytorch-{model_name}.bin"
@@ -58,6 +60,12 @@ def main():
 
     print("Checked all required files were found. Proceeding...\n")
 
+    # Verify tokens as they are a prerequisite for logits comparison.
+    print("🔍 Token Comparison Check")
+    print("=" * 40)
+    if not compare_tokens(f"pytorch-{model_name}", f"llamacpp-{llamacpp_model_name}"):
+        exit_with_warning("\n❌ Token mismatch detected", model_path)
+    print()
 
     print("🔍 GGML Model Validation for model ", model_name)
     print("=" * 40)
@@ -73,8 +81,7 @@ def main():
         print("       Ok to proceed with NMSE check...")
         sys.exit(0)
     else:
-        print(f"❌ NOK: Top 10 predictions don't match - generation will differ")
-        sys.exit(1)
+        exit_with_warning(f"❌ NOK: Top 10 predictions don't match - generation will differ", model_path)
 
 if __name__ == "__main__":
     main()

examples/model-conversion/scripts/causal/modelcard.template

@@ -7,7 +7,7 @@ base_model:
 Recommended way to run this model:
 
 ```sh
-llama-server -hf {namespace}/{model_name}-GGUF -c 0
+llama-server -hf {namespace}/{model_name}-GGUF
 ```
 
 Then, access http://localhost:8080

examples/model-conversion/scripts/causal/run-casual-gen-embeddings-org.py

@@ -67,7 +67,7 @@ with torch.no_grad():
     last_hidden_states = outputs.hidden_states[-1]
 
     # Get embeddings for all tokens
-    token_embeddings = last_hidden_states[0].cpu().numpy()  # Remove batch dimension
+    token_embeddings = last_hidden_states[0].float().cpu().numpy()  # Remove batch dimension
 
     print(f"Hidden states shape: {last_hidden_states.shape}")
     print(f"Token embeddings shape: {token_embeddings.shape}")

examples/model-conversion/scripts/causal/run-converted-model-embeddings-logits.sh

@@ -13,6 +13,6 @@ if [ -z "$CONVERTED_MODEL" ]; then
     exit 1
 fi
 
-cmake --build ../../build --target llama-logits -j8
+cmake --build ../../build --target llama-debug -j8
 
-../../build/bin/llama-logits -m $CONVERTED_MODEL -embd-mode "Hello world today"
+../../build/bin/llama-debug -m $CONVERTED_MODEL --embedding -p "Hello world today" --save-logits

examples/model-conversion/scripts/causal/run-converted-model.sh

@@ -21,6 +21,6 @@ fi
 echo $CONVERTED_MODEL
 echo $MODEL_TESTING_PROMPT
 
-cmake --build ../../build --target llama-logits -j8
+cmake --build ../../build --target llama-debug -j8
 
-../../build/bin/llama-logits -m "$CONVERTED_MODEL" "$MODEL_TESTING_PROMPT"
+../../build/bin/llama-debug -m "$CONVERTED_MODEL" -p "$MODEL_TESTING_PROMPT" --save-logits

examples/model-conversion/scripts/causal/run-org-model.py

@@ -7,12 +7,11 @@ import importlib
 import torch
 import numpy as np
 
-from pathlib import Path
 from transformers import AutoTokenizer, AutoModelForCausalLM, AutoModelForImageTextToText, AutoConfig
 
 # Add parent directory to path for imports
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
-from utils.common import debug_hook
+from utils.common import debug_hook, save_output_data
 
 def parse_arguments():
     parser = argparse.ArgumentParser(description="Process model with specified path")
@@ -126,6 +125,7 @@ def main():
     device = next(model.parameters()).device
     prompt = get_prompt(args)
     input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+    token_ids = input_ids[0].cpu().tolist()
 
     print(f"Input tokens: {input_ids}")
     print(f"Input text: {repr(prompt)}")
@@ -151,19 +151,6 @@ def main():
         print(f"Last token logits shape: {last_logits.shape}")
         print(f"Vocab size: {len(last_logits)}")
 
-        data_dir = Path("data")
-        data_dir.mkdir(exist_ok=True)
-        bin_filename = data_dir / f"pytorch-{model_name}.bin"
-        txt_filename = data_dir / f"pytorch-{model_name}.txt"
-
-        # Save to file for comparison
-        last_logits.astype(np.float32).tofile(bin_filename)
-
-        # Also save as text file for easy inspection
-        with open(txt_filename, "w") as f:
-            for i, logit in enumerate(last_logits):
-                f.write(f"{i}: {logit:.6f}\n")
-
         # Print some sample logits for quick verification
         print(f"First 10 logits: {last_logits[:10]}")
         print(f"Last 10 logits: {last_logits[-10:]}")
@@ -175,8 +162,7 @@ def main():
             token = tokenizer.decode([idx])
             print(f"  Token {idx} ({repr(token)}): {last_logits[idx]:.6f}")
 
-        print(f"Saved bin logits to: {bin_filename}")
-        print(f"Saved txt logist to: {txt_filename}")
+        save_output_data(last_logits, token_ids, prompt, model_name)
 
 if __name__ == "__main__":
     main()

examples/model-conversion/scripts/embedding/run-converted-model.sh

@@ -5,7 +5,7 @@ set -e
 # Parse command line arguments
 CONVERTED_MODEL=""
 PROMPTS_FILE=""
-USE_POOLING=""
+EMBD_NORMALIZE="2"
 
 while [[ $# -gt 0 ]]; do
     case $1 in
@@ -13,9 +13,9 @@ while [[ $# -gt 0 ]]; do
             PROMPTS_FILE="$2"
             shift 2
             ;;
-        --pooling)
-            USE_POOLING="1"
-            shift
+        --embd-normalize)
+            EMBD_NORMALIZE="$2"
+            shift 2
             ;;
         *)
             if [ -z "$CONVERTED_MODEL" ]; then
@@ -50,10 +50,5 @@ fi
 
 echo $CONVERTED_MODEL
 
-cmake --build ../../build --target llama-logits -j8
-# TODO: update logits.cpp to accept a --file/-f option for the prompt
-if [ -n "$USE_POOLING" ]; then
-    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode -pooling "$PROMPT"
-else
-    ../../build/bin/llama-logits -m "$CONVERTED_MODEL" -embd-mode "$PROMPT"
-fi
+cmake --build ../../build --target llama-debug -j8
+../../build/bin/llama-debug -m "$CONVERTED_MODEL" --embedding -p "$PROMPT" --save-logits --embd-normalize $EMBD_NORMALIZE

examples/model-conversion/scripts/embedding/run-original-model.py

@@ -3,13 +3,15 @@
 import argparse
 import os
 import sys
-import numpy as np
 import importlib
-from pathlib import Path
 
 from transformers import AutoTokenizer, AutoConfig, AutoModel
 import torch
 
+# Add parent directory to path for imports
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), '..'))
+from utils.common import save_output_data
+
 
 def parse_arguments():
     parser = argparse.ArgumentParser(description='Run original embedding model')
@@ -169,6 +171,7 @@ def main():
                 return_tensors="pt"
             )
             tokens = encoded['input_ids'][0]
+            token_ids = tokens.cpu().tolist()
             token_strings = tokenizer.convert_ids_to_tokens(tokens)
             for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
                 print(f"{token_id:6d} -> '{token_str}'")
@@ -185,6 +188,7 @@ def main():
             )
 
             tokens = encoded['input_ids'][0]
+            token_ids = tokens.cpu().tolist()
             token_strings = tokenizer.convert_ids_to_tokens(tokens)
             for i, (token_id, token_str) in enumerate(zip(tokens, token_strings)):
                 print(f"{token_id:6d} -> '{token_str}'")
@@ -228,24 +232,11 @@ def main():
 
         print()
 
-        data_dir = Path("data")
-        data_dir.mkdir(exist_ok=True)
-        bin_filename = data_dir / f"pytorch-{model_name}-embeddings.bin"
-        txt_filename = data_dir / f"pytorch-{model_name}-embeddings.txt"
-
         flattened_embeddings = all_embeddings.flatten()
-        flattened_embeddings.astype(np.float32).tofile(bin_filename)
-
-        with open(txt_filename, "w") as f:
-            idx = 0
-            for j in range(n_embd_count):
-                for value in all_embeddings[j]:
-                    f.write(f"{idx}: {value:.6f}\n")
-                    idx += 1
         print(f"Total values: {len(flattened_embeddings)} ({n_embd_count} embeddings × {n_embd} dimensions)")
         print("")
-        print(f"Saved bin embeddings to: {bin_filename}")
-        print(f"Saved txt embeddings to: {txt_filename}")
+
+        save_output_data(flattened_embeddings, token_ids, prompt_text, model_name, type_suffix="-embeddings")
 
 
 if __name__ == "__main__":

examples/model-conversion/scripts/utils/common.py

@@ -3,6 +3,11 @@
 import os
 import sys
 import torch
+import transformers
+import json
+import textwrap
+import numpy as np
+from pathlib import Path
 
 
 def get_model_name_from_env_path(env_path_name):
@@ -148,3 +153,147 @@ def setup_rope_debug(model_module_path: str, function_name: str = "apply_rotary_
     # Patch it
     setattr(module, function_name, debug_rope)
     print(f"RoPE debug patching applied to {model_module_path}.{function_name}")
+
+
+def save_output_data(data, tokens, prompt, model_name, type_suffix="", output_dir="data"):
+    """
+    Save output data (logits/embeddings), tokens, and prompt to files.
+
+    Args:
+        data:        numpy array of floats (logits or embeddings)
+        tokens:      list or array of token IDs
+        prompt:      string containing the input prompt
+        model_name:  name of the model
+        type_suffix: optional suffix like "-embeddings" (default: "")
+        output_dir:  directory to save files (default: "data")
+
+    Creates the following files in output_dir:
+        - pytorch-{model_name}{type_suffix}.bin
+        - pytorch-{model_name}{type_suffix}.txt
+        - pytorch-{model_name}{type_suffix}-prompt.txt
+        - pytorch-{model_name}{type_suffix}-tokens.bin
+    """
+    data_dir = Path(output_dir)
+    data_dir.mkdir(exist_ok=True)
+    base_path = data_dir / f"pytorch-{model_name}{type_suffix}"
+
+    # Convert and flatten logits/embeddings
+    data = data.cpu().numpy() if isinstance(data, torch.Tensor) else np.asarray(data)
+    data = data.flatten() if data.ndim > 1 else data
+
+    # Save logits/embedding files
+    data.astype(np.float32).tofile(f"{base_path}.bin")
+    print(f"Data saved to {base_path}.bin")
+
+    with open(f"{base_path}.txt", "w") as f:
+        f.writelines(f"{i}: {value:.6f}\n" for i, value in enumerate(data))
+    print(f"Data saved to {base_path}.txt")
+
+    # Convert and flatten tokens
+    tokens = tokens.cpu().numpy() if isinstance(tokens, torch.Tensor) else np.asarray(tokens)
+    tokens = tokens.flatten() if tokens.ndim > 1 else tokens
+
+    # Save token binary file
+    tokens.astype(np.int32).tofile(f"{base_path}-tokens.bin")
+    print(f"Tokens saved to {base_path}-tokens.bin")
+
+    # Save prompt file
+    with open(f"{base_path}-prompt.txt", "w") as f:
+        f.write(f"prompt: {prompt}\n")
+        f.write(f"n_tokens: {len(tokens)}\n")
+        f.write(f"token ids: {', '.join(str(int(tid)) for tid in tokens)}\n")
+    print(f"Prompt saved to {base_path}-prompt.txt")
+
+
+def compare_tokens(original, converted, type_suffix="", output_dir="data"):
+    data_dir = Path(output_dir)
+
+    # Read tokens from both models
+    tokens1_file = data_dir / f"{original}{type_suffix}-tokens.bin"
+    tokens2_file = data_dir / f"{converted}{type_suffix}-tokens.bin"
+
+    if not tokens1_file.exists():
+        print(f"Error: Token file not found: {tokens1_file}")
+        return False
+
+    if not tokens2_file.exists():
+        print(f"Error: Token file not found: {tokens2_file}")
+        return False
+
+    tokens1 = np.fromfile(tokens1_file, dtype=np.int32)
+    tokens2 = np.fromfile(tokens2_file, dtype=np.int32)
+
+    print(f"\nComparing tokens between:")
+    print(f"  Original : {original} ({len(tokens1)} tokens)")
+    print(f"  Converted: {converted} ({len(tokens2)} tokens)")
+
+    if len(tokens1) != len(tokens2):
+        print(f"\n❌ Token count mismatch: {len(tokens1)} vs {len(tokens2)}")
+        return False
+
+    if np.array_equal(tokens1, tokens2):
+        print(f"\n✅ All {len(tokens1)} tokens match!")
+        return True
+
+    mismatches = np.where(tokens1 != tokens2)[0]
+    print(f"\n❌ Found {len(mismatches)} mismatched tokens:")
+
+    num_to_show = min(len(mismatches), 10)
+    for idx in mismatches[:num_to_show]:
+        print(f"  Position {idx}: {tokens1[idx]} vs {tokens2[idx]}")
+
+    if len(mismatches) > num_to_show:
+        print(f"  ... and {len(mismatches) - num_to_show} more mismatches")
+
+    return False
+
+
+def show_version_warning(current_version, model_version):
+    if not model_version:
+        return False
+
+    try:
+        from packaging.version import parse, InvalidVersion
+        try:
+            return parse(current_version) < parse(model_version)
+        except InvalidVersion:
+            return current_version != model_version
+    except ImportError:
+        return current_version != model_version
+
+def get_model_transformers_version(model_path):
+    if not model_path:
+        return None
+
+    config_path = Path(model_path) / "config.json"
+    if not config_path.is_file():
+        return None
+
+    try:
+        with open(config_path, "r", encoding="utf-8") as f:
+            config = json.load(f)
+        return config.get("transformers_version")
+    except (IOError, json.JSONDecodeError) as e:
+        print(f"Warning: Could not read or parse {config_path}: {e}", file=sys.stderr)
+        return None
+
+def exit_with_warning(message, model_path):
+    print(message)
+
+    if model_path and transformers is not None:
+        model_transformers_version = get_model_transformers_version(model_path)
+        transformers_version       = transformers.__version__
+        if show_version_warning(transformers_version, model_transformers_version):
+            warning_message = f"""
+                =====================================================================
+                Verification failure might be due to a transformers version mismatch:
+
+                Current transformers version: {transformers_version}
+                Model's required version    : {model_transformers_version}
+
+                Consider installing the version specified by the model's config:
+                pip install transformers=={model_transformers_version}
+                =====================================================================
+            """
+            print(textwrap.dedent(warning_message))
+    sys.exit(1)

examples/model-conversion/scripts/utils/compare_tokens.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+
+import argparse
+import sys
+from common import compare_tokens  # type: ignore
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description='Compare tokens between two models',
+        formatter_class=argparse.RawDescriptionHelpFormatter,
+        epilog="""
+Examples:
+  %(prog)s pytorch-gemma-3-270m-it llamacpp-gemma-3-270m-it-bf16
+        """
+    )
+    parser.add_argument(
+        'original',
+        help='Original model name'
+    )
+    parser.add_argument(
+        'converted',
+        help='Converted model name'
+    )
+    parser.add_argument(
+        '-s', '--suffix',
+        default='',
+        help='Type suffix (e.g., "-embeddings")'
+    )
+    parser.add_argument(
+        '-d', '--data-dir',
+        default='data',
+        help='Directory containing token files (default: data)'
+    )
+    parser.add_argument(
+        '-v', '--verbose',
+        action='store_true',
+        help='Print prompts from both models'
+    )
+    return parser.parse_args()
+
+
+def main():
+    args = parse_arguments()
+
+    if args.verbose:
+        from pathlib import Path
+        data_dir = Path(args.data_dir)
+
+        prompt1_file = data_dir / f"{args.original}{args.suffix}-prompt.txt"
+        prompt2_file = data_dir / f"{args.converted}{args.suffix}-prompt.txt"
+
+        if prompt1_file.exists():
+            print(f"\nOriginal model prompt ({args.original}):")
+            print(f"  {prompt1_file.read_text().strip()}")
+
+        if prompt2_file.exists():
+            print(f"\nConverted model prompt ({args.converted}):")
+            print(f"  {prompt2_file.read_text().strip()}")
+
+        print()
+
+    result = compare_tokens(
+        args.original,
+        args.converted,
+        type_suffix=args.suffix,
+        output_dir=args.data_dir
+    )
+
+    # Enable the script to be used in shell scripts so that they can check
+    # the exit code for success/failure.
+    sys.exit(0 if result else 1)
+
+
+if __name__ == "__main__":
+    main()

examples/model-conversion/scripts/utils/semantic_check.py

@@ -4,8 +4,10 @@ import numpy as np
 import argparse
 import os
 import importlib
+from pathlib import Path
 
 from transformers import AutoTokenizer, AutoConfig, AutoModelForCausalLM, AutoModel
+from common import compare_tokens, exit_with_warning  # type: ignore[import-not-found]
 
 unreleased_model_name = os.getenv('UNRELEASED_MODEL_NAME')
 
@@ -157,9 +159,24 @@ def main():
     else:
         prompt = args.prompt
 
+    python_emb_path = Path(args.python_embeddings)
+    cpp_emb_path = Path(args.cpp_embeddings)
+
+    # Extract base names (e.g., "pytorch-model-name-embeddings.bin" -> "pytorch-model-name")
+    python_model_name = python_emb_path.stem.replace("-embeddings", "")
+    cpp_model_name = cpp_emb_path.stem.replace("-embeddings", "")
+
     print("Semantic Similarity Test Between Python and llama.cpp Embedding Models")
     print("=" * 70)
 
+    # First verify tokens match before comparing embeddings
+    print("\n🔍 Token Comparison Check")
+    print("=" * 70)
+    data_dir = python_emb_path.parent
+    if not compare_tokens(python_model_name, cpp_model_name, type_suffix="-embeddings", output_dir=str(data_dir)):
+        exit_with_warning("\n❌ Token mismatch detected", args.model_path)
+    print()
+
     # Single prompt detailed comparison
     print(f"\nTesting with prompt: '{prompt}'")
 
@@ -219,7 +236,7 @@ def main():
     elif avg_cross_sim > 0.70:
         print("⚠️  FAIR: Models have some differences")
     else:
-        print("❌ POOR: Models are significantly different")
+        exit_with_warning("❌ POOR: Models are significantly different", args.model_path)
 
 if __name__ == "__main__":
     main()

ggml/include/ggml.h

@@ -234,6 +234,11 @@
 
 #if UINTPTR_MAX == 0xFFFFFFFF
     #define GGML_MEM_ALIGN 4
+#elif defined(__EMSCRIPTEN__)
+// emscripten uses max_align_t == 8, so we need GGML_MEM_ALIGN == 8 for 64-bit wasm.
+// (for 32-bit wasm, the first conditional is true and GGML_MEM_ALIGN stays 4.)
+// ref: https://github.com/ggml-org/llama.cpp/pull/18628
+    #define GGML_MEM_ALIGN 8
 #else
     #define GGML_MEM_ALIGN 16
 #endif

ggml/src/ggml-backend-impl.h

@@ -144,7 +144,7 @@ extern "C" {
         // device description: short informative description of the device, could be the model name
         const char * (*get_description)(ggml_backend_dev_t dev);
 
-        // device memory in bytes
+        // device memory in bytes: 0 bytes to indicate no memory to report
         void         (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);
 
         // device type

ggml/src/ggml-blas/CMakeLists.txt

@@ -32,14 +32,12 @@ if (BLAS_FOUND)
                 pkg_check_modules(DepBLAS openblas)
             endif()
         elseif (${GGML_BLAS_VENDOR} MATCHES "FLAME")
-            add_compile_definitions(GGML_BLAS_USE_BLIS)
             pkg_check_modules(DepBLAS blis)
         elseif (${GGML_BLAS_VENDOR} MATCHES "ATLAS")
             pkg_check_modules(DepBLAS blas-atlas)
         elseif (${GGML_BLAS_VENDOR} MATCHES "FlexiBLAS")
             pkg_check_modules(DepBLAS flexiblas_api)
         elseif (${GGML_BLAS_VENDOR} MATCHES "Intel")
-            add_compile_definitions(GGML_BLAS_USE_MKL)
             # all Intel* libraries share the same include path
             pkg_check_modules(DepBLAS mkl-sdl)
         elseif (${GGML_BLAS_VENDOR} MATCHES "NVHPC")
@@ -74,10 +72,26 @@ if (BLAS_FOUND)
 
     target_compile_options(ggml-blas PRIVATE ${BLAS_LINKER_FLAGS})
 
-    if ("${BLAS_INCLUDE_DIRS}" MATCHES "mkl" AND (${GGML_BLAS_VENDOR} MATCHES "Generic" OR ${GGML_BLAS_VENDOR} MATCHES "Intel"))
+    if ("${GGML_BLAS_VENDOR}" STREQUAL "")
+        message(WARNING "GGML_BLAS_VENDOR is not set; some methods may not link properly.")
+    endif()
+
+    if ("${GGML_BLAS_VENDOR}" MATCHES "Intel" OR ("${BLAS_INCLUDE_DIRS}" MATCHES "mkl" AND "${GGML_BLAS_VENDOR}" MATCHES "Generic"))
         add_compile_definitions(GGML_BLAS_USE_MKL)
     endif()
 
+    if ("${GGML_BLAS_VENDOR}" MATCHES "OpenBLAS")
+        add_compile_definitions(GGML_BLAS_USE_OPENBLAS)
+    endif()
+
+    if ("${GGML_BLAS_VENDOR}" MATCHES "FLAME" OR "${GGML_BLAS_VENDOR}" MATCHES "AOCL" OR "${GGML_BLAS_VENDOR}" MATCHES "AOCL_mt")
+        add_compile_definitions(GGML_BLAS_USE_BLIS)
+    endif()
+
+    if ("${GGML_BLAS_VENDOR}" MATCHES "NVPL")
+        add_compile_definitions(GGML_BLAS_USE_NVPL)
+    endif()
+
     target_link_libraries     (ggml-blas PRIVATE ${BLAS_LIBRARIES})
     target_include_directories(ggml-blas PRIVATE ${BLAS_INCLUDE_DIRS})
 else()

ggml/src/ggml-blas/ggml-blas.cpp

@@ -115,15 +115,11 @@ static void ggml_backend_blas_mul_mat(ggml_backend_blas_context * ctx, struct gg
 #endif
     }
 
-#if defined(OPENBLAS_VERSION)
+#if defined(GGML_BLAS_USE_OPENBLAS)
     openblas_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_BLIS)
+#elif defined(GGML_BLAS_USE_BLIS)
     bli_thread_set_num_threads(ctx->n_threads);
-#endif
-
-#if defined(GGML_BLAS_USE_NVPL)
+#elif defined(GGML_BLAS_USE_NVPL)
     nvpl_blas_set_num_threads(ctx->n_threads);
 #endif
 
@@ -288,7 +284,7 @@ ggml_backend_t ggml_backend_blas_init(void) {
         /* .context = */ ctx,
     };
 
-#if defined(OPENBLAS_VERSION) && defined(GGML_USE_OPENMP)
+#if defined(GGML_BLAS_USE_OPENBLAS) && defined(GGML_USE_OPENMP)
     if (openblas_get_parallel() != OPENBLAS_OPENMP) {
         GGML_LOG_DEBUG("%s: warning: ggml is using OpenMP, but OpenBLAS was compiled without OpenMP support\n", __func__);
     }
@@ -329,7 +325,7 @@ static const char * ggml_backend_blas_device_get_description(ggml_backend_dev_t
         return "BLIS";
     #elif defined(GGML_BLAS_USE_NVPL)
         return "NVPL";
-    #elif defined(OPENBLAS_VERSION)
+    #elif defined(GGML_BLAS_USE_OPENBLAS)
         return "OpenBLAS";
     #else
         return "BLAS";

ggml/src/ggml-cann/aclnn_ops.cpp

@@ -1963,7 +1963,7 @@ static void ggml_cann_mat_mul_fp(ggml_backend_cann_context & ctx, ggml_tensor *
     acl_tensor_ptr acl_weight_tensor;
 
     // Only check env once.
-    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
+    static bool weight_to_nz = parse_bool(get_env_as_lowercase("GGML_CANN_WEIGHT_NZ").value_or("on"));
     if (weight_to_nz && is_matmul_weight(weight)) {
         acl_weight_tensor = ggml_cann_create_tensor(weight, transpose_ne, transpose_nb, n_dims, ACL_FORMAT_FRACTAL_NZ);
     } else {

ggml/src/ggml-cann/common.h

@@ -103,7 +103,7 @@ const ggml_cann_device_info & ggml_cann_info();
 void    ggml_cann_set_device(int32_t device);
 int32_t ggml_cann_get_device();
 
-std::optional<std::string> get_env(const std::string & name);
+std::optional<std::string> get_env_as_lowercase(const std::string & name);
 bool                       parse_bool(const std::string & value);
 int                        parse_integer(const std::string & value);
 

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

@@ -105,10 +105,10 @@ int32_t ggml_cann_get_device() {
 }
 
 /**
- * @brief Get the value of the specified environment variable (name).
+ * @brief Get the value of the specified environment variable (name) as lowercase.
  *        if not empty, return a std::string object
  */
-std::optional<std::string> get_env(const std::string & name) {
+std::optional<std::string> get_env_as_lowercase(const std::string & name) {
     const char * val = std::getenv(name.c_str());
     if (!val) {
         return std::nullopt;
@@ -259,7 +259,7 @@ struct ggml_cann_pool_buf_prio : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf_prio(int device) : device(device) {
-        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
+        disable_clean = parse_bool(get_env_as_lowercase("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -452,7 +452,7 @@ struct ggml_cann_pool_buf : public ggml_cann_pool {
      * @param device The device ID to associate with this buffer pool.
      */
     explicit ggml_cann_pool_buf(int device) : device(device) {
-        disable_clean = parse_bool(get_env("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
+        disable_clean = parse_bool(get_env_as_lowercase("GGML_CANN_DISABLE_BUF_POOL_CLEAN").value_or(""));
     }
 
     /**
@@ -764,7 +764,7 @@ struct ggml_cann_pool_vmm : public ggml_cann_pool {
  * @return A unique pointer to the created CANN pool.
  */
 std::unique_ptr<ggml_cann_pool> ggml_backend_cann_context::new_pool_for_device(int device) {
-    std::string mem_pool_type = get_env("GGML_CANN_MEM_POOL").value_or("");
+    std::string mem_pool_type = get_env_as_lowercase("GGML_CANN_MEM_POOL").value_or("");
 
     if (mem_pool_type == "prio") {
         GGML_LOG_INFO("%s: device %d use buffer pool with priority queue\n", __func__, device);
@@ -1217,7 +1217,7 @@ static void ggml_backend_cann_buffer_set_tensor(ggml_backend_buffer_t buffer,
     // Why aclrtSynchronizeDevice?
 
     // Only check env once.
-    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
+    static bool weight_to_nz = parse_bool(get_env_as_lowercase("GGML_CANN_WEIGHT_NZ").value_or("on"));
     if (!need_transform(tensor->type)) {
         ACL_CHECK(aclrtMemcpy((char *) tensor->data + offset, size, data, size, ACL_MEMCPY_HOST_TO_DEVICE));
         if (weight_to_nz && is_matmul_weight((const ggml_tensor *) tensor)) {
@@ -1442,7 +1442,7 @@ static size_t ggml_backend_cann_buffer_type_get_alloc_size(ggml_backend_buffer_t
     int64_t ne0  = tensor->ne[0];
 
     // Only check env once.
-    static bool weight_to_nz = parse_bool(get_env("GGML_CANN_WEIGHT_NZ").value_or("on"));
+    static bool weight_to_nz = parse_bool(get_env_as_lowercase("GGML_CANN_WEIGHT_NZ").value_or("on"));
 
     // last line must bigger than 32, because every single op deal at
     // least 32 bytes.
@@ -2136,7 +2136,7 @@ static void evaluate_and_capture_cann_graph(ggml_backend_cann_context * cann_ctx
 #endif  // USE_ACL_GRAPH
     // Only perform the graph execution if CANN graphs are not enabled, or we are capturing the graph.
     // With the use of CANN graphs, the execution will be performed by the graph launch.
-    static bool opt_fusion = parse_bool(get_env("GGML_CANN_OPERATOR_FUSION").value_or(""));
+    static bool opt_fusion = parse_bool(get_env_as_lowercase("GGML_CANN_OPERATOR_FUSION").value_or(""));
 
     if (!use_cann_graph || cann_graph_capture_required) {
         for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -2201,7 +2201,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
 #ifdef USE_ACL_GRAPH
     bool use_cann_graph = true;
 
-    static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
+    static bool prefill_use_graph = parse_bool(get_env_as_lowercase("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
     if (!prefill_use_graph) {
         // Do not use acl_graph for prefill.
         for (int i = 0; i < cgraph->n_nodes; i++) {
@@ -2541,27 +2541,6 @@ static bool ggml_backend_buft_is_cann(ggml_backend_buffer_type_t buft) {
     return buft->iface.get_name == ggml_backend_cann_buffer_type_name;
 }
 
-/**
- * @brief Determines if a tensor operation should be offloaded to the CANN
- * backend.
- *
- * This function checks if a given tensor operation should be offloaded to the
- * CANN backend based on the operation type and the size of the tensor. It
- * returns true if the second dimension (ne[1]) of the tensor is greater than or
- * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
- *
- * @param backend Pointer to the CANN backend.
- * @param op Pointer to the tensor operation to check.
- * @return bool Returns true if the operation should be offloaded, otherwise
- * false.
- */
-static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-    GGML_UNUSED(dev);
-
-    return op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS;
-}
-
 /**
  * @brief Records an event on the CANN backend stream.
  *
@@ -2637,6 +2616,7 @@ struct ggml_backend_cann_device_context {
     int         device;
     std::string name;
     std::string description;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_cann_device_get_name(ggml_backend_dev_t dev) {
@@ -2713,6 +2693,26 @@ static ggml_backend_buffer_type_t ggml_backend_cann_device_get_host_buffer_type(
     return ggml_backend_cann_host_buffer_type();
 }
 
+/**
+ * @brief Determines if a tensor operation should be offloaded to the CANN
+ * backend.
+ *
+ * This function checks if a given tensor operation should be offloaded to the
+ * CANN backend based on the operation type and the size of the tensor. It
+ * returns true if the second dimension (ne[1]) of the tensor is greater than or
+ * equal to the minimum batch size and the operation is not GGML_OP_GET_ROWS.
+ *
+ * @param backend Pointer to the CANN backend.
+ * @param op Pointer to the tensor operation to check.
+ * @return bool Returns true if the operation should be offloaded, otherwise
+ * false.
+ */
+static bool ggml_backend_cann_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_cann_device_context * dev_ctx = (ggml_backend_cann_device_context *)dev->context;
+
+    return op->ne[1] >= dev_ctx->op_offload_min_batch_size && op->op != GGML_OP_GET_ROWS;
+}
+
 /**
  * @brief Creates a new event for the CANN backend device.
  *
@@ -2829,12 +2829,14 @@ ggml_backend_reg_t ggml_backend_cann_reg() {
         if (!initialized) {
             aclInit(nullptr);
             ggml_backend_cann_reg_context * ctx = new ggml_backend_cann_reg_context;
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             for (int i = 0; i < ggml_cann_info().device_count; i++) {
                 ggml_backend_cann_device_context * dev_ctx = new ggml_backend_cann_device_context();
                 dev_ctx->description                       = aclrtGetSocName();
                 dev_ctx->device                            = i;
                 dev_ctx->name                              = GGML_CANN_NAME + std::to_string(i);
+                dev_ctx->op_offload_min_batch_size         = min_batch_size;
                 ggml_cann_set_device(i);
                 ggml_backend_dev_t dev = new ggml_backend_device{ /* .iface   = */ ggml_backend_cann_device_interface,
                                                                   /* .reg     = */ &reg,

ggml/src/ggml-cuda/CMakeLists.txt

@@ -47,7 +47,10 @@ if (CUDAToolkit_FOUND)
                 #     check Modules/Internal/CMakeCUDAArchitecturesValidate.cmake in the CMake git repository instead.
                 # However, the architectures 120a-real and 121a-real should work with basically any CMake version and
                 #     until the release of e.g. Rubin there is no benefit to shipping virtual architectures for Blackwell.
-                list(APPEND CMAKE_CUDA_ARCHITECTURES 120a-real 121a-real)
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 120a-real)
+            endif()
+            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.9")
+                list(APPEND CMAKE_CUDA_ARCHITECTURES 121a-real)
             endif()
         endif()
     endif()

ggml/src/ggml-cuda/common.cuh

@@ -262,6 +262,10 @@ static const char * cu_get_error_str(CUresult err) {
 #define FLASH_ATTN_AVAILABLE
 #endif // !defined(GGML_CUDA_NO_FA) && !(defined(GGML_USE_MUSA) && __MUSA_ARCH__ < 220)
 
+#if defined(TURING_MMA_AVAILABLE)
+#define LDMATRIX_TRANS_AVAILABLE
+#endif // defined(TURING_MMA_AVAILABLE)
+
 static bool fp16_available(const int cc) {
     return ggml_cuda_highest_compiled_arch(cc) >= GGML_CUDA_CC_PASCAL ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_PH1);

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

@@ -914,7 +914,7 @@ void launch_fattn(
 
         const int nblocks_stream_k = max_blocks;
 
-        const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || tiles_efficiency_percent < 75;
+        const bool use_stream_k = cc >= GGML_CUDA_CC_ADA_LOVELACE || amd_wmma_available(cc) || tiles_efficiency_percent < 75;
 
         blocks_num.x = use_stream_k ? nblocks_stream_k : ntiles_total;
         blocks_num.y = 1;

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

@@ -98,6 +98,19 @@ static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_co
     return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
 }
 
+static constexpr __host__ __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config_rdna(const int DKQ, const int DV, const int ncols) {
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 16, 128, 2,  64, 128, 128, 128, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 32, 128, 2,  64, 128, 128,  64, 2, true);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(256, 256, 64, 128, 2,  64, 128, 128,  64, 2, true);
+
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 16,  64, 4,  32,  96,  64, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 32, 128, 2,  32, 160, 128, 128, 1, false);
+    GGML_CUDA_FATTN_MMA_CONFIG_CASE(576, 512, 64, 256, 1,  32, 160, 128, 128, 1, false);
+
+    // TODO tune specifically for RDNA
+    return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
+}
+
 static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, const int DV, const int ncols, const int cc) {
     if (ampere_mma_available(cc)) {
         return ggml_cuda_fattn_mma_get_config_ampere(DKQ, DV, ncols);
@@ -105,6 +118,9 @@ static __host__ fattn_mma_config ggml_cuda_fattn_mma_get_config(const int DKQ, c
     if (turing_mma_available(cc)) {
         return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
     }
+    if (amd_wmma_available(cc)) {
+        return ggml_cuda_fattn_mma_get_config_rdna(DKQ, DV, ncols);
+    }
     GGML_ASSERT(volta_mma_available(cc));
     return ggml_cuda_fattn_mma_get_config_volta(DKQ, DV, ncols);
 }
@@ -116,6 +132,8 @@ static constexpr __device__ fattn_mma_config ggml_cuda_fattn_mma_get_config(cons
     return ggml_cuda_fattn_mma_get_config_turing(DKQ, DV, ncols);
 #elif defined(VOLTA_MMA_AVAILABLE)
     return ggml_cuda_fattn_mma_get_config_volta(DKQ, DV, ncols);
+#elif defined(AMD_WMMA_AVAILABLE)
+    return ggml_cuda_fattn_mma_get_config_rdna(DKQ, DV, ncols);
 #else
     GGML_UNUSED_VARS(DKQ, DV, ncols);
     return fattn_mma_config(32, 1, 0, 0, 0, 0, 0, false);
@@ -186,6 +204,23 @@ static constexpr __device__ bool ggml_cuda_fattn_mma_get_Q_in_reg(const int DKQ,
     return ggml_cuda_fattn_mma_get_config(DKQ, DV, ncols).Q_in_reg;
 }
 
+static constexpr __device__ int get_cols_per_thread() {
+#if defined(AMD_WMMA_AVAILABLE)
+    return 1; // RDNA has a single column.
+#else
+    return 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+#endif // defined(AMD_WMMA_AVAILABLE)
+}
+
+static __host__ int get_cols_per_warp(const int cc) {
+    if (turing_mma_available(cc) || amd_wmma_available(cc)) {
+        return 16;
+    } else {
+        // Volta
+        return 32;
+    }
+}
+
 // ------------------------------------------------------------------------------------------------------------------
 
 static __host__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, const int DV, const int ncols1, const int ncols2, const int cc) {
@@ -393,10 +428,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         const int jt,
         const int kb0,
         const int k_VKQ_sup) {
-#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
     constexpr int  ncols           = ncols1 * ncols2;
     constexpr int  cols_per_warp   = T_B_KQ::I;
-    constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+    constexpr int  cols_per_thread = get_cols_per_thread();
     constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa(DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
@@ -413,6 +448,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     const int k_VKQ_0 = kb0 * nbatch_fa;
 #if defined(TURING_MMA_AVAILABLE)
     T_C_KQ KQ_C[nbatch_fa/(np*(cols_per_warp == 8 ? T_C_KQ::I : T_C_KQ::J))];
+#elif defined(AMD_WMMA_AVAILABLE)
+    T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
 #else // Volta
     T_C_KQ KQ_C[nbatch_fa/(np*T_C_KQ::J)];
 #endif // defined(TURING_MMA_AVAILABLE)
@@ -461,8 +498,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     if constexpr (cols_per_warp == 8) {
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
                     } else {
-                        // Wide version of KQ_C is column-major => swap A and B.
+                        // Wide version of KQ_C is column-major
+#if defined(AMD_WMMA_AVAILABLE)
+                        // RDNA matrix C is column-major.
+                        mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[k_KQ_0/T_A_KQ::J]);
+#else
+                        // swap A and B for CUDA.
                         mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[k_KQ_0/T_A_KQ::J], K_A);
+#endif // defined(AMD_WMMA_AVAILABLE)
                     }
                 }
             }
@@ -479,8 +522,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                     T_A_KQ K_A;
                     load_ldmatrix(K_A, tile_K + i_KQ_0*stride_tile_K + (k_KQ_0 - k0_start), stride_tile_K);
 
-                    // Wide version of KQ_C is column-major => swap A and B.
+                    // Wide version of KQ_C is column-major
+#if defined(AMD_WMMA_AVAILABLE)
+                    // RDNA matrix C is column-major.
+                    mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], K_A, Q_B[0]);
+#else
+                    // swap A and B for CUDA.
                     mma(KQ_C[i_KQ_00/(np*T_A_KQ::I)], Q_B[0], K_A);
+#endif // defined(AMD_WMMA_AVAILABLE)
                 }
             }
         }
@@ -532,7 +581,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
-                    KQ_max_new[l % 2] = fmaxf(KQ_max_new[l % 2], KQ_C[k0/(np*T_C_KQ::I)].x[l] + FATTN_KQ_MAX_OFFSET);
+#if defined(AMD_WMMA_AVAILABLE)
+                    constexpr int KQ_idx = 0;
+#else
+                    // Turing + Volta:
+                    const int KQ_idx = l % 2;
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[k0/(np*T_C_KQ::I)].x[l] + FATTN_KQ_MAX_OFFSET);
                 }
             }
         }
@@ -552,8 +607,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::I + T_C_KQ::get_i(l) < k_VKQ_sup) {
-                    KQ_C[k0/(np*T_C_KQ::I)].x[l] = expf(KQ_C[k0/(np*T_C_KQ::I)].x[l] - KQ_max_new[l % 2]);
-                    KQ_rowsum_add[l % 2] += KQ_C[k0/(np*T_C_KQ::I)].x[l];
+#if defined(AMD_WMMA_AVAILABLE)
+                    constexpr int KQ_idx = 0;
+#else
+                    // Turing + Volta:
+                    const int KQ_idx = l % 2;
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    KQ_C[k0/(np*T_C_KQ::I)].x[l] = expf(KQ_C[k0/(np*T_C_KQ::I)].x[l] - KQ_max_new[KQ_idx]);
+                    KQ_rowsum_add[KQ_idx] += KQ_C[k0/(np*T_C_KQ::I)].x[l];
                 } else {
                     KQ_C[k0/(np*T_C_KQ::I)].x[l] = 0.0f;
                 }
@@ -584,8 +645,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
+#if defined(AMD_WMMA_AVAILABLE)
+                    constexpr int KQ_idx = 0;
+#else
                     // Turing + Volta:
-                    KQ_max_new[(l/2) % 2] = fmaxf(KQ_max_new[(l/2) % 2], KQ_C[(k0/(np*T_C_KQ::J))].x[l] + FATTN_KQ_MAX_OFFSET);
+                    const int KQ_idx = (l/2) % 2;
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    KQ_max_new[KQ_idx] = fmaxf(KQ_max_new[KQ_idx], KQ_C[(k0/(np*T_C_KQ::J))].x[l] + FATTN_KQ_MAX_OFFSET);
                 }
             }
         }
@@ -596,7 +662,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
             // Values per KQ column are spread across 4 threads:
             constexpr int offset_first = 2;
             constexpr int offset_last  = 1;
-#else
+#elif defined(AMD_WMMA_AVAILABLE)
+            // Values per KQ column are spread across 2 threads:
+            constexpr int offset_first = 16;
+            constexpr int offset_last  = 16;
+#else // Volta
             // Values per KQ column are spread across 2 threads:
             constexpr int offset_first = 2;
             constexpr int offset_last  = 2;
@@ -612,10 +682,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         for (int k0 = 0; k0 < nbatch_fa; k0 += np*T_C_KQ::J) {
 #pragma unroll
             for (int l = 0; l < T_C_KQ::ne; ++l) {
-                // Turing + Volta:
                 if (!oob_check || k0 + (threadIdx.y % np)*T_C_KQ::J + T_C_KQ::get_j(l) < k_VKQ_sup) {
-                    KQ_C[(k0/(np*T_C_KQ::J))].x[l] = expf(KQ_C[(k0/(np*T_C_KQ::J))].x[l] - KQ_max_new[(l/2) % 2]);
-                    KQ_rowsum_add[(l/2) % 2] += KQ_C[(k0/(np*T_C_KQ::J))].x[l];
+#if defined(AMD_WMMA_AVAILABLE)
+                    constexpr int KQ_idx = 0;
+#else
+                    // Turing + Volta:
+                    const int KQ_idx = (l/2) % 2;
+#endif // defined(AMD_WMMA_AVAILABLE)
+                    KQ_C[(k0/(np*T_C_KQ::J))].x[l] = expf(KQ_C[(k0/(np*T_C_KQ::J))].x[l] - KQ_max_new[KQ_idx]);
+                    KQ_rowsum_add[KQ_idx] += KQ_C[(k0/(np*T_C_KQ::J))].x[l];
                 } else {
                     KQ_C[(k0/(np*T_C_KQ::J))].x[l] = 0.0f;
                 }
@@ -639,7 +714,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
 
 #if defined(TURING_MMA_AVAILABLE)
         if constexpr (cols_per_warp == 8) {
-            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
 #pragma unroll
             for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
 #pragma unroll
@@ -660,6 +735,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 }
             }
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+        const half2 KQ_max_scale_h2 = make_half2(
+            KQ_max_scale[0], KQ_max_scale[0]);
+#pragma unroll
+        for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+#pragma unroll
+            for (int l = 0; l < T_C_VKQ::ne; ++l) {
+                VKQ_C[i].x[l] *= KQ_max_scale_h2;
+            }
+        }
 #else // Volta
         const half2 KQ_max_scale_h2 = make_half2(
             KQ_max_scale[(threadIdx.x / 2) % 2], KQ_max_scale[(threadIdx.x / 2) % 2]);
@@ -707,6 +792,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
     // Therefore, iterate over V in reverse and re-use the data if possible.
     static_assert(!mla || nstages <= 1, "combination of MLA and multi-stage loading not implemented");
     constexpr int reusable_cutoff = mla ? (DKQ - 1) - (DKQ - 1) % (2*nbatch_K2) - (DKQ - DV) : DV;
+#if defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
+    T_A_VKQ A_identity;
+    make_identity_mat(A_identity);
+#endif // defined(AMD_WMMA_AVAILABLE) && !defined(LDMATRIX_TRANS_AVAILABLE)
 
     // Calculate VKQ tile, need to use logical rather than physical elements for i0 due to transposition of V:
 #pragma unroll
@@ -727,7 +816,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         }
         const half2 * tile_V_i = i0_start < reusable_cutoff ? tile_V : tile_V + (i0_start - reusable_cutoff)/2;
 
-#if defined(TURING_MMA_AVAILABLE)
+#if defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
         constexpr int i0_stride = cols_per_warp == 8 ? T_C_VKQ::I : 2*T_C_VKQ::J;
 #pragma unroll
         for (int i_VKQ_0 = i0_start; i_VKQ_0 < i0_stop; i_VKQ_0 += i0_stride) {
@@ -737,12 +826,26 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 const int k0 = k00 + (threadIdx.y % np)*T_A_VKQ::J;
 
                 T_A_VKQ A; // Transposed in SRAM but not in registers, gets transposed on load.
+#if defined(LDMATRIX_TRANS_AVAILABLE)
                 load_ldmatrix_trans(A, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
+#else
+                // TODO: Try to transpose tile_V when loading gmem to smem.
+                // Use mma to transpose T_A_VKQ for RDNA.
+                T_A_VKQ A_trans;
+                load_ldmatrix(A_trans, tile_V_i + 2*k0*stride_tile_V + (i_VKQ_0 - i0_start)/2, stride_tile_V);
+                mma(A, A_trans, A_identity);
+#endif // defined(TURING_MMA_AVAILABLE)
                 if constexpr (T_B_KQ::I == 8) {
                     mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
                 } else {
-                    // Wide version of VKQ_C is column-major => swap A and B.
+                    // Wide version of VKQ_C is column-major.
+#if defined(AMD_WMMA_AVAILABLE)
+                    // RDNA matrix C is column-major.
+                    mma(VKQ_C[i_VKQ_0/i0_stride], A, B[k00/(np*T_A_VKQ::J)]);
+#else
+                    // swap A and B for CUDA.
                     mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::J)], A);
+#endif // defined(AMD_WMMA_AVAILABLE)
                 }
             }
         }
@@ -761,7 +864,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
                 mma(VKQ_C[i_VKQ_0/i0_stride], B[k00/(np*T_A_VKQ::I)], A);
             }
         }
-#endif // defined(TURING_MMA_AVAILABLE)
+#endif // defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)
 
         if constexpr (nstages <= 1) {
             __syncthreads(); // Only needed if tile_K == tile_V.
@@ -774,7 +877,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
         tile_Q, tile_K, tile_V, tile_mask,
         Q_B, VKQ_C, KQ_max, KQ_rowsum, kb0);
     NO_DEVICE_CODE;
-#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
 }
 
 #if defined(TURING_MMA_AVAILABLE)
@@ -794,6 +897,15 @@ template<> struct mma_tile_sizes<8> {
     using T_B_VKQ = tile< 8,  8, half2>; // column-major
     using T_C_VKQ = tile<16,  4, half2>; // row-major
 };
+#elif defined(AMD_WMMA_AVAILABLE)
+template<int ncols> struct mma_tile_sizes {
+    using T_A_KQ  = tile<16,  8, half2>; // row-major
+    using T_B_KQ  = tile<16,  8, half2>; // column-major
+    using T_C_KQ  = tile<16, 16, float>; // column-major
+    using T_A_VKQ = tile<16,  8, half2>; // row-major
+    using T_B_VKQ = tile<16,  8, half2>; // column-major
+    using T_C_VKQ = tile<16,  8, half2>; // column-major
+};
 #else // Volta
 template<int ncols> struct mma_tile_sizes {
     using T_A_KQ  = tile< 8,  4, half2, DATA_LAYOUT_I_MAJOR_MIRRORED>; // row-major
@@ -828,7 +940,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int jt,
         const int kb0_start,
         const int kb0_stop) {
-#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#if defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
     //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
 
     constexpr int ncols = ncols1 * ncols2;
@@ -840,7 +952,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     using     T_C_VKQ   = typename mma_tile_sizes<ncols>::T_C_VKQ;
 
     constexpr int  cols_per_warp   = T_B_KQ::I;
-    constexpr int  cols_per_thread = 2; // This is specifically KQ columns, Volta only has a single VKQ column.
+    constexpr int  cols_per_thread = get_cols_per_thread();
     constexpr int  np              = nwarps * (cols_per_warp/ncols2) / ncols1; // Number of parallel CUDA warps per Q column.
     constexpr int  nbatch_fa       = ggml_cuda_fattn_mma_get_nbatch_fa     (DKQ, DV, ncols);
     constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2     (DKQ, DV, ncols);
@@ -871,6 +983,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     T_B_KQ    Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
 #if defined(TURING_MMA_AVAILABLE)
     T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
+#elif defined(AMD_WMMA_AVAILABLE)
+    T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
 #else // Volta
     T_C_VKQ VKQ_C[                                     DV/(2*T_C_VKQ::J)];
 #endif // defined(TURING_MMA_AVAILABLE)
@@ -1010,6 +1124,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         // The partial sums are spread across 8/4 threads.
         constexpr int offset_first = cols_per_warp == 8 ? 16 : 2;
         constexpr int offset_last  = cols_per_warp == 8 ?  4 : 1;
+#elif defined(AMD_WMMA_AVAILABLE)
+        // The partial sums are spread across 2 threads.
+        constexpr int offset_first = 16;
+        constexpr int offset_last  = 16;
 #else // Volta
         // The partial sums are spread across 2 threads.
         constexpr int offset_first = 2;
@@ -1047,7 +1165,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 
 #if defined(TURING_MMA_AVAILABLE)
         if constexpr (cols_per_warp == 8) {
-            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[1]);
+            const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[cols_per_thread - 1]);
 #pragma unroll
             for (int i = 0; i < DV/T_C_VKQ::I; ++i) {
 #pragma unroll
@@ -1068,6 +1186,15 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
                 }
             }
         }
+#elif defined(AMD_WMMA_AVAILABLE)
+        const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[0], KQ_max_scale[0]);
+#pragma unroll
+        for (int i = 0; i < (DV/2)/T_C_VKQ::J; ++i) {
+#pragma unroll
+            for (int l = 0; l < T_C_VKQ::ne; ++l) {
+                VKQ_C[i].x[l] *= KQ_max_scale_h2;
+            }
+        }
 #else // Volta
         const int col = (threadIdx.x / 2) % 2;
         const half2 KQ_max_scale_h2 = make_half2(KQ_max_scale[col], KQ_max_scale[col]);
@@ -1119,6 +1246,10 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(threadIdx.x % 4);
         const float2 KQ_cmr = make_float2(KQ_max[threadIdx.x % cols_per_thread], KQ_rowsum[threadIdx.x % cols_per_thread]);
         const bool thread_should_write = threadIdx.x % 4 < cols_per_thread;
+#elif defined(AMD_WMMA_AVAILABLE)
+        const int jc_cwm = threadIdx.y*cols_per_warp + T_C_VKQ::get_i(0);
+        const float2 KQ_cmr = make_float2(KQ_max[0], KQ_rowsum[0]);
+        const bool thread_should_write = threadIdx.x / 16 < cols_per_thread;
 #else // Volta
         const int jc_cwm = threadIdx.y*cols_per_warp + T_C_KQ::get_i(threadIdx.x & 2);
         const float2 KQ_cmr = make_float2(KQ_max[(threadIdx.x & 2) / 2], KQ_rowsum[(threadIdx.x & 2) / 2]);
@@ -1319,7 +1450,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
         stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
         jt, kb0_start, kb0_stop);
     NO_DEVICE_CODE;
-#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE)
+#endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
 }
 
 template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool mla>
@@ -1346,7 +1477,7 @@ static __global__ void flash_attn_ext_f16(
                             const int32_t nb21, const int32_t nb22, const int64_t nb23,
                             const int32_t ne31, const int32_t ne32, const int32_t ne33,
                             const int32_t nb31, const int32_t nb32, const int64_t nb33) {
-#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
+#if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
 
     // Skip unused kernel variants for faster compilation:
     if (use_logit_softcap && !(DKQ == 128 || DKQ == 256)) {
@@ -1360,6 +1491,13 @@ static __global__ void flash_attn_ext_f16(
     }
 #endif // __CUDA_ARCH__ == GGML_CUDA_CC_TURING
 
+#if defined(AMD_WMMA_AVAILABLE)
+    if (ncols1*ncols2 > 32 || ncols1*ncols2 < 16 || DKQ > 128 || ncols2 == 1) {
+        NO_DEVICE_CODE;
+        return;
+    }
+#endif // defined(AMD_WMMA_AVAILABLE)
+
     static_assert(!mla || DKQ >= DV, "MLA needs DKQ >= DV");
 
     constexpr int ncols     = ncols1 * ncols2;
@@ -1473,7 +1611,7 @@ static __global__ void flash_attn_ext_f16(
               ne31, ne32, ne33,
               nb31, nb32, nb33);
     NO_DEVICE_CODE;
-#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE))
+#endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)))
 }
 
 template <int DKQ, int DV, int ncols1, int ncols2>
@@ -1492,7 +1630,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     const bool Q_in_reg       = ggml_cuda_fattn_mma_get_Q_in_reg      (DKQ, DV, ncols, cc);
     const int  nstages        = ggml_cuda_fattn_mma_get_nstages       (DKQ, DV, ncols1, ncols2, cc);
 
-    const int cols_per_warp = std::min(ncols, turing_mma_available(cc) ? 16 : 32);
+    const int cols_per_warp = std::min(ncols, get_cols_per_warp(cc));
     const int nwarps        = nthreads / WARP_SIZE;
 
     constexpr bool mla = DKQ == 576;
@@ -1512,29 +1650,34 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     float logit_softcap;
     memcpy(&logit_softcap, (const float *) KQV->op_params + 2, sizeof(float));
 
+#if defined(GGML_USE_HIP)
+    using fattn_kernel_ptr_t = const void*;
+#else
+    using fattn_kernel_ptr_t = fattn_kernel_t;
+#endif // defined(GGML_USE_HIP)
     fattn_kernel_t fattn_kernel;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
 
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
+            CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
     } else {
         constexpr bool use_logit_softcap = true;
         fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, mla>;
 
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
         static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
         if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(fattn_kernel, cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
+            CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
             shared_memory_limit_raised[id] = true;
         }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
     }
 
     launch_fattn<DV, ncols1, ncols2>

ggml/src/ggml-cuda/fattn.cu

@@ -18,12 +18,12 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
         }
     }
 
-    if (turing_mma_available(cc) && Q->ne[1] <= 16/ncols2) {
+    if ((turing_mma_available(cc) || amd_wmma_available(cc)) && Q->ne[1] <= 16/ncols2) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 16/ncols2, ncols2>(ctx, dst);
         return;
     }
 
-    if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || Q->ne[1] <= 32/ncols2) {
+    if (ggml_cuda_highest_compiled_arch(cc) == GGML_CUDA_CC_TURING || amd_wmma_available(cc) || Q->ne[1] <= 32/ncols2) {
         ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 32/ncols2, ncols2>(ctx, dst);
         return;
     }
@@ -230,7 +230,18 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
 
     // The effective batch size for the kernel can be increased by gqa_ratio.
     // The kernel versions without this optimization are also used for ALiBi, if there is no mask, or if the KV cache is not padded,
-    const bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    bool gqa_opt_applies = gqa_ratio % 2 == 0 && mask && max_bias == 0.0f && K->ne[1] % FATTN_KQ_STRIDE == 0;
+    for (const ggml_tensor * t : {Q, K, V, mask}) {
+        if (t == nullptr) {
+            continue;
+        }
+        for (size_t i = 1; i < GGML_MAX_DIMS; ++i) {
+            if (t->nb[i] % 16 != 0) {
+                gqa_opt_applies = false;
+                break;
+            }
+        }
+    }
 
     const int cc = ggml_cuda_info().devices[device].cc;
 
@@ -337,6 +348,31 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_WMMA_F16;
     }
 
+    if (amd_wmma_available(cc) && GGML_CUDA_CC_IS_RDNA4(cc) && gqa_opt_applies && Q->ne[0] <= 128 && Q->ne[0] != 40 && Q->ne[0] != 72) {
+        if (can_use_vector_kernel) {
+            if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {
+                if (Q->ne[1] == 1) {
+                    if (!gqa_opt_applies) {
+                        return BEST_FATTN_KERNEL_VEC;
+                    }
+                }
+            } else {
+                if (Q->ne[1] <= 2) {
+                    return BEST_FATTN_KERNEL_VEC;
+                }
+            }
+        }
+        int gqa_ratio_eff = 1;
+        const int ncols2_max = Q->ne[0] == 576 ? 16 : 8;
+        while (gqa_ratio % (2*gqa_ratio_eff) == 0 && gqa_ratio_eff < ncols2_max) {
+            gqa_ratio_eff *= 2;
+        }
+        if (Q->ne[1] * gqa_ratio_eff <= 8) {
+            return BEST_FATTN_KERNEL_TILE; // AMD WMMA is only faster if the full tile width of 16 can be utilized.
+        }
+        return BEST_FATTN_KERNEL_MMA_F16;
+    }
+
     // If there are no tensor cores available, use the generic tile kernel:
     if (can_use_vector_kernel) {
         if (!ggml_is_quantized(K->type) && !ggml_is_quantized(V->type)) {

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

@@ -3737,6 +3737,7 @@ static bool ggml_cuda_graph_set_enabled(ggml_backend_cuda_context * cuda_ctx) {
 
     return cuda_ctx->cuda_graph->is_enabled();
 #else
+    GGML_UNUSED(cuda_ctx);
     return false;
 #endif // USE_CUDA_GRAPH
 }
@@ -4122,6 +4123,7 @@ struct ggml_backend_cuda_device_context {
     std::string name;
     std::string description;
     std::string pci_bus_id;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_cuda_device_get_name(ggml_backend_dev_t dev) {
@@ -4676,11 +4678,9 @@ static int64_t get_op_batch_size(const ggml_tensor * op) {
 }
 
 static bool ggml_backend_cuda_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
+    ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
 
-    return get_op_batch_size(op) >= min_batch_size;
-
-    GGML_UNUSED(dev);
+    return get_op_batch_size(op) >= dev_ctx->op_offload_min_batch_size;
 }
 
 static ggml_backend_event_t ggml_backend_cuda_device_event_new(ggml_backend_dev_t dev) {
@@ -4848,6 +4848,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
             ggml_backend_cuda_reg_context * ctx = new ggml_backend_cuda_reg_context;
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             for (int i = 0; i < ggml_cuda_info().device_count; i++) {
                 ggml_backend_cuda_device_context * dev_ctx = new ggml_backend_cuda_device_context;
@@ -4861,6 +4862,7 @@ ggml_backend_reg_t ggml_backend_cuda_reg() {
                 char pci_bus_id[16] = {};
                 snprintf(pci_bus_id, sizeof(pci_bus_id), "%04x:%02x:%02x.0", prop.pciDomainID, prop.pciBusID, prop.pciDeviceID);
                 dev_ctx->pci_bus_id = pci_bus_id;
+                dev_ctx->op_offload_min_batch_size = min_batch_size;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface   = */ ggml_backend_cuda_device_interface,

ggml/src/ggml-cuda/mma.cuh

@@ -206,10 +206,16 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 16 && J == 16) {
-                // matrix C
 #if defined(RDNA3)
-                return 2 * l + (threadIdx.x / 16);
+                if constexpr (std::is_same_v<T, float> || std::is_same_v<T, int>) {
+                    // matrix C
+                    return 2 * l + (threadIdx.x / 16);
+                } else {
+                    // matrix A&B
+                    return l;
+                }
 #else
+                // matrix C is the transposed matrix A&B on RDNA4
                 return ne * (threadIdx.x / 16) + l;
 #endif // defined(RDNA3)
             } else if constexpr (I == 16 && J == 8) {
@@ -621,6 +627,21 @@ namespace ggml_cuda_mma {
 
         return ret;
     }
+#elif defined(AMD_WMMA_AVAILABLE)
+    template <int I, int J>
+    static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
+        tile<I, J/2, half2> ret;
+#pragma unroll
+        for (int l0 = 0; l0 < tile_float.ne; l0 += 2) {
+            ret.x[l0/2] = make_half2(tile_float.x[l0 + 0], tile_float.x[l0 + 1]);
+        }
+        return ret;
+    }
+
+    static __device__ __forceinline__ tile<8, 8, half2> get_transposed(const tile<16, 4, half2> & t) {
+        NO_DEVICE_CODE;
+        return tile<8, 8, half2>{};
+    }
 #else // Volta
     template <int I, int J>
     static __device__ __forceinline__ tile<I, J/2, half2> get_half2(const tile<I, J, float> & tile_float) {
@@ -639,6 +660,19 @@ namespace ggml_cuda_mma {
     }
 #endif // defined(TURING_MMA_AVAILABLE)
 
+    static __device__ __forceinline__ void make_identity_mat(tile<16, 8, half2> & t) {
+#if defined(RDNA4)
+        const int row = t.get_i(0);
+        const int left_right = t.get_j(0) / 4;
+        const int up_down = row / 8;
+        const int idx = row % 8;
+        reinterpret_cast<half*>(t.x)[idx] = left_right == up_down ? 1.0f : 0.0f;
+#else
+        GGML_UNUSED_VARS(t);
+        NO_DEVICE_CODE;
+#endif // defined(RDNA4)
+    }
+
     template <int I, int J, typename T, data_layout dl>
     static __device__ __forceinline__ void load_generic(tile<I, J, T, dl> & t, const T * __restrict__ xs0, const int stride) {
 #if defined(AMD_MFMA_AVAILABLE)
@@ -878,6 +912,17 @@ namespace ggml_cuda_mma {
             : "+r"(Dxi[2]), "+r"(Dxi[3])
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
+#elif defined(AMD_WMMA_AVAILABLE)
+#if defined(RDNA4)
+        using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
+        halfx8_t& acc_frag = reinterpret_cast<halfx8_t&>(D.x[0]);
+        const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
+        const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
+        acc_frag = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
+#else
+        GGML_UNUSED_VARS(D, A, B);
+        NO_DEVICE_CODE;
+#endif // defined(RDNA4)
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;

ggml/src/ggml-cuda/mmq.cu

@@ -190,7 +190,7 @@ void ggml_cuda_mul_mat_q(
     {
         const int64_t s11 = src1->nb[1] / ts_src1;
         const int64_t s12 = src1->nb[2] / ts_src1;
-        const int64_t s13 = src1->nb[2] / ts_src1;
+        const int64_t s13 = src1->nb[3] / ts_src1;
 
         if (use_native_mxfp4) {
             quantize_mmq_mxfp4_cuda(src1_d, ids_src1.get(), src1_q8_1.get(), src0->type, ne10, s11, s12, s13,
@@ -333,28 +333,31 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     }
 
     if (amd_wmma_available(cc)) {
-        // RDNA 4 is consistently worse on rocblas
-        // https://github.com/ggml-org/llama.cpp/pull/18537#issuecomment-3706422301
         if (GGML_CUDA_CC_IS_RDNA3(cc)) {
-            // High expert counts almost always better on MMQ
-            // due to a large amount of graph splits
+            // High expert counts are almost always better on MMQ due to
+            //     the synchronization overhead in the cuBLAS/hipBLAS path:
             // https://github.com/ggml-org/llama.cpp/pull/18202
             if (n_experts >= 64) {
                 return true;
             }
 
+            // For some quantization types MMQ can have lower peak TOPS than hipBLAS
+            //     so it's only faster for sufficiently small batch sizes:
             switch (type) {
-                // These quants are really bad on MMQ
                 case GGML_TYPE_Q2_K:
+                    return ne11 <= 128;
                 case GGML_TYPE_Q6_K:
-                // These quants are usually worse but not always
+                    return ne11 <= (GGML_CUDA_CC_IS_RDNA3_0(cc) ? 128 : 256);
                 case GGML_TYPE_IQ2_XS:
                 case GGML_TYPE_IQ2_S:
-                    return ne11 <= 128;
+                    return GGML_CUDA_CC_IS_RDNA3_5(cc) || ne11 <= 128;
                 default:
                     return true;
             }
         }
+
+        // For RDNA4 MMQ is consistently faster than dequantization + hipBLAS:
+        // https://github.com/ggml-org/llama.cpp/pull/18537#issuecomment-3706422301
         return true;
     }
 

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

@@ -138,6 +138,8 @@
 #define cudaStream_t hipStream_t
 #define cudaSuccess hipSuccess
 #define cudaOccupancyMaxActiveBlocksPerMultiprocessor hipOccupancyMaxActiveBlocksPerMultiprocessor
+#define cudaFuncSetAttribute hipFuncSetAttribute
+#define cudaFuncAttributeMaxDynamicSharedMemorySize hipFuncAttributeMaxDynamicSharedMemorySize
 #define __trap() do { abort(); __builtin_unreachable(); } while(0)
 #define CUBLAS_STATUS_SUCCESS HIPBLAS_STATUS_SUCCESS
 #define CUBLAS_STATUS_NOT_INITIALIZED HIPBLAS_STATUS_NOT_INITIALIZED

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

@@ -219,6 +219,8 @@ struct ggml_metal_device_props {
     bool use_shared_buffers;
 
     bool supports_gpu_family_apple7;
+
+    int op_offload_min_batch_size;
 };
 
 ggml_metal_device_t ggml_metal_device_init(void);

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

@@ -782,6 +782,8 @@ ggml_metal_device_t ggml_metal_device_init(void) {
 
             dev->props.supports_gpu_family_apple7 = [dev->mtl_device supportsFamily:MTLGPUFamilyApple7];
 
+            dev->props.op_offload_min_batch_size  = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
+
             dev->props.max_buffer_size            = dev->mtl_device.maxBufferLength;
             dev->props.max_working_set_size       = dev->mtl_device.recommendedMaxWorkingSetSize;
             dev->props.max_theadgroup_memory_size = dev->mtl_device.maxThreadgroupMemoryLength;

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

@@ -625,14 +625,11 @@ static int64_t get_op_batch_size(const ggml_tensor * op) {
 }
 
 static bool ggml_backend_metal_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
+    ggml_metal_device_t ctx_dev = (ggml_metal_device_t)dev->context;
 
     return (op->op == GGML_OP_MUL_MAT ||
             op->op == GGML_OP_MUL_MAT_ID) &&
-            get_op_batch_size(op) >= min_batch_size;
-
-    GGML_UNUSED(dev);
-    GGML_UNUSED(op);
+            get_op_batch_size(op) >= ggml_metal_device_get_props(ctx_dev)->op_offload_min_batch_size;
 }
 
 static ggml_backend_device_i ggml_backend_metal_device_i = {

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

@@ -9148,6 +9148,7 @@ typedef decltype(kernel_mul_mm_id_map0<1>) kernel_mul_mm_id_map0_t;
 template [[host_name("kernel_mul_mm_id_map0_ne20_1" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<1>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_2" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<2>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_4" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<4>;
+template [[host_name("kernel_mul_mm_id_map0_ne20_5" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<5>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_6" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<6>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_8" )]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<8>;
 template [[host_name("kernel_mul_mm_id_map0_ne20_10")]] kernel kernel_mul_mm_id_map0_t kernel_mul_mm_id_map0<10>;

ggml/src/ggml-opencl/CMakeLists.txt

@@ -57,6 +57,7 @@ set(GGML_OPENCL_KERNELS
     add
     add_id
     argsort
+    fill
     clamp
     cpy
     cvt
@@ -120,6 +121,8 @@ set(GGML_OPENCL_KERNELS
     tsembd
     upscale
     tanh
+    expm1
+    softplus
     pad
     repeat
     mul_mat_f16_f32

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

@@ -489,6 +489,7 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_gelu_quick, kernel_gelu_quick_4;
     cl_kernel kernel_relu;
     cl_kernel kernel_sigmoid_f32, kernel_sigmoid_f16;
+    cl_kernel kernel_fill;
     cl_kernel kernel_clamp;
     cl_kernel kernel_geglu, kernel_reglu, kernel_swiglu, kernel_swiglu_oai, kernel_geglu_erf, kernel_geglu_quick,
               kernel_geglu_f16, kernel_reglu_f16, kernel_swiglu_f16, kernel_geglu_erf_f16, kernel_geglu_quick_f16;
@@ -537,6 +538,10 @@ struct ggml_backend_opencl_context {
     cl_kernel kernel_pad;
     cl_kernel kernel_tanh_f32_nd;
     cl_kernel kernel_tanh_f16_nd;
+    cl_kernel kernel_expm1_f32_nd;
+    cl_kernel kernel_expm1_f16_nd;
+    cl_kernel kernel_softplus_f32_nd;
+    cl_kernel kernel_softplus_f16_nd;
     cl_kernel kernel_upscale;
     cl_kernel kernel_upscale_bilinear;
     cl_kernel kernel_concat_f32_contiguous;
@@ -787,6 +792,24 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         GGML_LOG_CONT(".");
     }
 
+    // fill
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "fill.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("fill.cl");
+#endif
+        cl_program prog =
+            build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+
+        CL_CHECK((backend_ctx->kernel_fill = clCreateKernel(prog, "kernel_fill_f32", &err), err));
+        GGML_LOG_CONT(".");
+
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
     // clamp
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -1780,6 +1803,56 @@ static void load_cl_kernels(ggml_backend_opencl_context *backend_ctx, ggml_cl_ve
         }
     }
 
+    // expm1
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "expm1.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("expm1.cl");
+#endif
+        cl_program prog;
+        if (!kernel_src.empty()) {
+            prog =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_expm1_f32_nd = clCreateKernel(prog, "kernel_expm1_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_expm1_f16_nd = clCreateKernel(prog, "kernel_expm1_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: expm1 kernel source not found or empty. Expm1 operation will not be available.\n");
+            prog = nullptr;
+            backend_ctx->kernel_expm1_f32_nd = nullptr;
+            backend_ctx->kernel_expm1_f16_nd = nullptr;
+        }
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
+    // softplus
+    {
+#ifdef GGML_OPENCL_EMBED_KERNELS
+        const std::string kernel_src {
+            #include "softplus.cl.h"
+        };
+#else
+        const std::string kernel_src = read_file("softplus.cl");
+#endif
+        cl_program prog;
+        if (!kernel_src.empty()) {
+            prog =
+                build_program_from_source(backend_ctx->context, backend_ctx->device, kernel_src.c_str(), compile_opts);
+            CL_CHECK((backend_ctx->kernel_softplus_f32_nd = clCreateKernel(prog, "kernel_softplus_f32_nd", &err), err));
+            CL_CHECK((backend_ctx->kernel_softplus_f16_nd = clCreateKernel(prog, "kernel_softplus_f16_nd", &err), err));
+            GGML_LOG_CONT(".");
+        } else {
+            GGML_LOG_WARN("ggml_opencl: softplus kernel source not found or empty. Softplus operation will not be available.\n");
+            prog = nullptr;
+            backend_ctx->kernel_softplus_f32_nd = nullptr;
+            backend_ctx->kernel_softplus_f16_nd = nullptr;
+        }
+        CL_CHECK(clReleaseProgram(prog));
+    }
+
     // upscale
     {
 #ifdef GGML_OPENCL_EMBED_KERNELS
@@ -3089,6 +3162,12 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 case GGML_UNARY_OP_TANH:
                    return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
                           (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                case GGML_UNARY_OP_EXPM1:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
+                case GGML_UNARY_OP_SOFTPLUS:
+                   return (op->src[0]->type == GGML_TYPE_F32 && op->type == GGML_TYPE_F32) ||
+                          (op->src[0]->type == GGML_TYPE_F16 && op->type == GGML_TYPE_F16);
                 default:
                     return false;
             }
@@ -3104,6 +3183,8 @@ static bool ggml_opencl_supports_op(ggml_backend_dev_t dev, const struct ggml_te
                 default:
                     return false;
             }
+        case GGML_OP_FILL:
+            return op->type == GGML_TYPE_F32 && ggml_is_contiguous(op);
         case GGML_OP_CLAMP:
             return op->src[0]->type == GGML_TYPE_F32;
         case GGML_OP_SOFT_MAX:
@@ -4266,8 +4347,8 @@ static const char * ggml_backend_opencl_device_get_description(ggml_backend_dev_
 }
 
 static void ggml_backend_opencl_device_get_memory(ggml_backend_dev_t dev, size_t * free, size_t * total) {
-    *free = 1;
-    *total = 1;
+    *free = 0;
+    *total = 0;
 
     GGML_UNUSED(dev);
 }
@@ -5860,6 +5941,36 @@ static void ggml_cl_sigmoid(ggml_backend_t backend, const ggml_tensor * src0, co
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
+static void ggml_cl_fill(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src0);
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+    cl_ulong offsetd = extrad->offset + dst->view_offs;
+
+    float v = 0.0f;
+    memcpy(&v, ((int32_t *) dst->op_params), sizeof(float));
+
+    const int64_t n = ggml_nelements(dst);
+
+    cl_kernel kernel = backend_ctx->kernel_fill;
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offsetd));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(float),    &v));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(float),    &n));
+
+    size_t local_work_size[1] = { 256 };
+    size_t global_work_size[1] = { ((size_t)n + local_work_size[0] - 1) / local_work_size[0] * local_work_size[0] };
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 1, global_work_size, local_work_size, dst);
+}
+
 static void ggml_cl_clamp(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -6413,6 +6524,210 @@ static void ggml_cl_tanh(ggml_backend_t backend, const ggml_tensor * src0, const
     backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
 }
 
+static void ggml_cl_expm1(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_expm1_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_expm1_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_expm1");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0];
+    const int ne11 = dst->ne[1];
+    const int ne12 = dst->ne[2];
+    const int ne13 = dst->ne[3];
+
+    const cl_ulong nb10 = dst->nb[0];
+    const cl_ulong nb11 = dst->nb[1];
+    const cl_ulong nb12 = dst->nb[2];
+    const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
+static void ggml_cl_softplus(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src0);
+    GGML_ASSERT(src0->extra);
+    GGML_ASSERT(dst);
+    GGML_ASSERT(dst->extra);
+
+    UNUSED(src1);
+
+    ggml_backend_opencl_context *backend_ctx = (ggml_backend_opencl_context *)backend->context;
+
+    ggml_tensor_extra_cl * extra0 = (ggml_tensor_extra_cl *)src0->extra;
+    ggml_tensor_extra_cl * extrad = (ggml_tensor_extra_cl *)dst->extra;
+
+    cl_ulong offset0_abs = extra0->offset + src0->view_offs;
+    cl_ulong offsetd_abs = extrad->offset + dst->view_offs;
+
+    cl_kernel kernel;
+    if (dst->type == GGML_TYPE_F32) {
+        kernel = backend_ctx->kernel_softplus_f32_nd;
+    } else if (dst->type == GGML_TYPE_F16) {
+        kernel = backend_ctx->kernel_softplus_f16_nd;
+    } else {
+        GGML_ASSERT(false && "Unsupported type for ggml_cl_softplus");
+    }
+    GGML_ASSERT(kernel != nullptr);
+
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];
+
+    const cl_ulong nb00 = src0->nb[0];
+    const cl_ulong nb01 = src0->nb[1];
+    const cl_ulong nb02 = src0->nb[2];
+    const cl_ulong nb03 = src0->nb[3];
+
+    const int ne10 = dst->ne[0];
+    const int ne11 = dst->ne[1];
+    const int ne12 = dst->ne[2];
+    const int ne13 = dst->ne[3];
+
+    const cl_ulong nb10 = dst->nb[0];
+    const cl_ulong nb11 = dst->nb[1];
+    const cl_ulong nb12 = dst->nb[2];
+    const cl_ulong nb13 = dst->nb[3];
+
+    CL_CHECK(clSetKernelArg(kernel, 0, sizeof(cl_mem),   &extra0->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 1, sizeof(cl_ulong), &offset0_abs));
+    CL_CHECK(clSetKernelArg(kernel, 2, sizeof(cl_mem),   &extrad->data_device));
+    CL_CHECK(clSetKernelArg(kernel, 3, sizeof(cl_ulong), &offsetd_abs));
+
+    CL_CHECK(clSetKernelArg(kernel, 4, sizeof(int),      &ne00));
+    CL_CHECK(clSetKernelArg(kernel, 5, sizeof(int),      &ne01));
+    CL_CHECK(clSetKernelArg(kernel, 6, sizeof(int),      &ne02));
+    CL_CHECK(clSetKernelArg(kernel, 7, sizeof(int),      &ne03));
+    CL_CHECK(clSetKernelArg(kernel, 8, sizeof(cl_ulong), &nb00));
+    CL_CHECK(clSetKernelArg(kernel, 9, sizeof(cl_ulong), &nb01));
+    CL_CHECK(clSetKernelArg(kernel, 10, sizeof(cl_ulong),&nb02));
+    CL_CHECK(clSetKernelArg(kernel, 11, sizeof(cl_ulong),&nb03));
+
+    CL_CHECK(clSetKernelArg(kernel, 12, sizeof(int),     &ne10));
+    CL_CHECK(clSetKernelArg(kernel, 13, sizeof(int),     &ne11));
+    CL_CHECK(clSetKernelArg(kernel, 14, sizeof(int),     &ne12));
+    CL_CHECK(clSetKernelArg(kernel, 15, sizeof(int),     &ne13));
+    CL_CHECK(clSetKernelArg(kernel, 16, sizeof(cl_ulong),&nb10));
+    CL_CHECK(clSetKernelArg(kernel, 17, sizeof(cl_ulong),&nb11));
+    CL_CHECK(clSetKernelArg(kernel, 18, sizeof(cl_ulong),&nb12));
+    CL_CHECK(clSetKernelArg(kernel, 19, sizeof(cl_ulong),&nb13));
+
+    size_t global_work_size[3];
+    if (ne10 == 0 || ne11 == 0 || ne12 == 0 || ne13 == 0) { // Handle case of 0 elements
+        return;
+    }
+    global_work_size[0] = (size_t)ne10;
+    global_work_size[1] = (size_t)ne11;
+    global_work_size[2] = (size_t)ne12;
+
+    size_t lws0 = 16, lws1 = 4, lws2 = 1;
+    if (ne10 < 16) lws0 = ne10;
+    if (ne11 < 4) lws1 = ne11;
+    if (ne12 < 1) lws2 = ne12 > 0 ? ne12 : 1;
+
+    while (lws0 * lws1 * lws2 > 256 && lws0 > 1) lws0 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws1 > 1) lws1 /= 2;
+    while (lws0 * lws1 * lws2 > 256 && lws2 > 1) lws2 /= 2;
+
+
+    size_t local_work_size[] = {lws0, lws1, lws2};
+
+    size_t* local_work_size_ptr = local_work_size;
+    if (!backend_ctx->non_uniform_workgroups) {
+        if (global_work_size[0] % local_work_size[0] != 0 ||
+            global_work_size[1] % local_work_size[1] != 0 ||
+            global_work_size[2] % local_work_size[2] != 0) {
+            local_work_size_ptr = NULL;
+        }
+    }
+    if (global_work_size[0] == 0 || global_work_size[1] == 0 || global_work_size[2] == 0) return;
+
+    backend_ctx->enqueue_ndrange_kernel(kernel, 3, global_work_size, local_work_size_ptr, dst);
+}
+
 static void ggml_cl_repeat(ggml_backend_t backend, const ggml_tensor * src0, const ggml_tensor * src1_shape_def, ggml_tensor * dst) {
     GGML_ASSERT(src0);
     GGML_ASSERT(src0->extra);
@@ -9586,6 +9901,18 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
                     }
                     func = ggml_cl_tanh;
                     break;
+                case GGML_UNARY_OP_EXPM1:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_expm1;
+                    break;
+                case GGML_UNARY_OP_SOFTPLUS:
+                    if (!any_on_device) {
+                        return false;
+                    }
+                    func = ggml_cl_softplus;
+                    break;
                 default:
                     return false;
             } break;
@@ -9595,6 +9922,12 @@ bool ggml_cl_compute_forward(ggml_backend_t backend, struct ggml_tensor * tensor
             }
             func = ggml_cl_glu;
             break;
+        case GGML_OP_FILL:
+            if (!any_on_device) {
+                return false;
+            }
+            func = ggml_cl_fill;
+            break;
         case GGML_OP_CLAMP:
             if (!any_on_device) {
                 return false;

ggml/src/ggml-opencl/kernels/expm1.cl

@@ -0,0 +1,82 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// expm1
+//------------------------------------------------------------------------------
+kernel void kernel_expm1_f32_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}
+
+kernel void kernel_expm1_f16_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = exp(*src_val_ptr) - 1;
+        }
+    }
+}

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

@@ -0,0 +1,17 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// fill
+//------------------------------------------------------------------------------
+__kernel void kernel_fill_f32(
+        __global float *dst,
+        ulong offsetd,
+        float v,
+        int n
+
+) {
+    dst = (global float*)((global char*)dst + offsetd);
+    if(get_global_id(0) < n){
+        dst[get_global_id(0)] = v;
+    }
+}

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

@@ -0,0 +1,88 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+//------------------------------------------------------------------------------
+// softplus
+//------------------------------------------------------------------------------
+inline float softplus_f32(float x){
+    float ax = fabs(x);
+    float m = fmax(x, 0.0f);
+    return log1p(exp(-ax)) + m;
+}
+
+kernel void kernel_softplus_f32_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const float *src_val_ptr = (global const float *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global float *dst_val_ptr = (global float *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = softplus_f32(*src_val_ptr);
+        }
+    }
+}
+
+kernel void kernel_softplus_f16_nd(
+        global void * p_src0_base,
+        ulong off_src0_abs,
+        global void * p_dst_base,
+        ulong off_dst_abs,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13
+) {
+    int i0 = get_global_id(0);
+    int i1 = get_global_id(1);
+    int i2 = get_global_id(2);
+
+    if (i0 < ne10 && i1 < ne11 && i2 < ne12) {
+        for (int i3 = 0; i3 < ne13; ++i3) {
+            ulong src_offset_in_tensor = (ulong)i0*nb00 + (ulong)i1*nb01 + (ulong)i2*nb02 + (ulong)i3*nb03;
+            global const half *src_val_ptr = (global const half *)((global char *)p_src0_base + off_src0_abs + src_offset_in_tensor);
+
+            ulong dst_offset_in_tensor = (ulong)i0*nb10 + (ulong)i1*nb11 + (ulong)i2*nb12 + (ulong)i3*nb13;
+            global half *dst_val_ptr = (global half *)((global char *)p_dst_base + off_dst_abs + dst_offset_in_tensor);
+
+            *dst_val_ptr = (half)(softplus_f32((float)(*src_val_ptr)));
+        }
+    }
+}

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

@@ -4286,6 +4286,7 @@ struct ggml_backend_sycl_device_context {
     int device;
     std::string name;
     std::string description;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_sycl_device_get_name(ggml_backend_dev_t dev) {
@@ -4674,9 +4675,8 @@ static int64_t get_op_batch_size(const ggml_tensor * op) {
 }
 
 static bool ggml_backend_sycl_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
-    return get_op_batch_size(op) >= min_batch_size;
-    GGML_UNUSED(dev);
+    ggml_backend_sycl_device_context * sycl_ctx = (ggml_backend_sycl_device_context *)dev->context;
+    return get_op_batch_size(op) >= sycl_ctx->op_offload_min_batch_size;
 }
 
 static ggml_backend_event_t
@@ -4799,6 +4799,7 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
             ggml_backend_sycl_reg_context * ctx = new ggml_backend_sycl_reg_context;
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
 
             for (int i = 0; i < ggml_sycl_info().device_count; i++) {
                 ggml_backend_sycl_device_context * dev_ctx = new ggml_backend_sycl_device_context;
@@ -4812,6 +4813,7 @@ ggml_backend_reg_t ggml_backend_sycl_reg() {
                     prop, dpct::dev_mgr::instance().get_device(i))));
 
                 dev_ctx->description = prop.get_name();
+                dev_ctx->op_offload_min_batch_size = min_batch_size;
 
                 ggml_backend_dev_t dev = new ggml_backend_device {
                     /* .iface       = */ ggml_backend_sycl_device_interface,

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

@@ -119,6 +119,8 @@ struct ggml_backend_vk_context;
 // Max number of adds that can be fused without exceeding MAX_PARAMETER_COUNT.
 #define MAX_FUSED_ADDS (MAX_PARAMETER_COUNT - 3)
 
+typedef std::shared_ptr<struct vk_pipeline_struct> vk_pipeline;
+
 struct vk_pipeline_struct {
     std::string name;
     vk::ShaderModule shader_module;
@@ -136,9 +138,15 @@ struct vk_pipeline_struct {
     std::atomic<bool> compiled {};
     // number of registers used, extracted from pipeline executable properties
     uint32_t register_count {};
+
+#if defined(VK_EXT_shader_64bit_indexing)
+    bool is_64b_indexing {};
+#endif
+    // linked list of pipelines for multiple compilation variants.
+    // currently only used to compile a 64-bit indexing variant.
+    vk_pipeline next;
 };
 
-typedef std::shared_ptr<vk_pipeline_struct> vk_pipeline;
 typedef std::weak_ptr<vk_pipeline_struct> vk_pipeline_ref;
 
 static void ggml_vk_destroy_pipeline(vk::Device& device, vk_pipeline& pipeline);
@@ -230,9 +238,7 @@ static ggml_backend_buffer_type_i ggml_backend_vk_buffer_type_interface = {
     /* .is_host          = */ NULL,
 };
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
 class vk_memory_logger;
-#endif
 class vk_perf_logger;
 static void ggml_vk_destroy_buffer(vk_buffer& buf);
 static void ggml_vk_synchronize(ggml_backend_vk_context * ctx);
@@ -570,6 +576,7 @@ struct vk_device_struct {
     bool uma;
     bool prefer_host_memory;
     bool float_controls_rte_fp16;
+    bool subgroup_basic;
     bool subgroup_arithmetic;
     bool subgroup_shuffle;
     bool subgroup_ballot;
@@ -583,6 +590,8 @@ struct vk_device_struct {
     bool add_rms_fusion;
     uint32_t partials_binding_alignment;
 
+    bool shader_64b_indexing;
+
     bool integer_dot_product;
     // 0: default, 1: force mmvq, -1: disable mmvq
     int32_t mmvq_mode;
@@ -814,9 +823,7 @@ struct vk_device_struct {
     bool allow_sysmem_fallback;
     bool disable_graph_optimize;
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
     std::unique_ptr<vk_memory_logger> memory_logger;
-#endif
 
     ~vk_device_struct() {
         VK_LOG_DEBUG("destroy device " << name);
@@ -1504,6 +1511,11 @@ template <> void init_pushconst_fastdiv(vk_op_sum_rows_push_constants &p) {
     init_fastdiv_values(p.ne01,        p.ne0_1mp,  p.ne0_1L);
 }
 
+struct vk_quantize_q8_1_push_constants {
+    uint32_t ne;
+    uint32_t num_blocks;
+};
+
 // Allow pre-recording command buffers
 struct vk_staging_memcpy {
     vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
@@ -1547,8 +1559,9 @@ static void ggml_vk_preallocate_buffers(ggml_backend_vk_context * ctx, vk_contex
 static void ggml_vk_load_shaders(vk_device& device);
 static void ggml_pipeline_allocate_descriptor_sets(ggml_backend_vk_context * ctx);
 
-#if defined(GGML_VULKAN_MEMORY_DEBUG) || defined(GGML_VULKAN_DEBUG)
-#define VK_LOG_MEMORY(msg) std::cerr << "ggml_vulkan memory: " << msg << std::endl
+static bool vk_memory_logger_enabled = false;
+
+#define VK_LOG_MEMORY(msg) if (vk_memory_logger_enabled) { std::cerr << "ggml_vulkan memory: " << msg << std::endl; }
 
 static std::string format_size(size_t size) {
     const size_t kib = 1024;
@@ -1581,10 +1594,10 @@ private:
     std::map<vk::Buffer, size_t> allocations; // Track allocations
     size_t total_device;
     size_t total_host;
+    static std::mutex log_mutex;
 };
-#else
-#define VK_LOG_MEMORY(msg) ((void) 0)
-#endif // GGML_VULKAN_MEMORY_DEBUG
+
+std::mutex vk_memory_logger::log_mutex;
 
 static bool vk_perf_logger_enabled = false;
 static bool vk_perf_logger_concurrent = false;
@@ -1891,10 +1904,10 @@ struct ggml_backend_vk_buffer_context {
     }
 };
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
-static std::mutex log_mutex;
-
 void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
+    if (!vk_memory_logger_enabled) {
+        return;
+    }
     std::lock_guard<std::mutex> guard(log_mutex);
     vk_buffer buf = buf_ref.lock();
     const bool device = bool(buf->memory_property_flags & vk::MemoryPropertyFlagBits::eDeviceLocal);
@@ -1906,7 +1919,7 @@ void vk_memory_logger::log_allocation(vk_buffer_ref buf_ref, size_t size) {
 }
 
 void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
-    if (buf_ref.expired() || buf_ref.lock()->size == 0) {
+    if (buf_ref.expired() || buf_ref.lock()->size == 0 || !vk_memory_logger_enabled) {
         return;
     }
 
@@ -1924,7 +1937,6 @@ void vk_memory_logger::log_deallocation(vk_buffer_ref buf_ref) {
         VK_LOG_MEMORY("ERROR " << buf->device->name << ": Attempted to deallocate unknown " << type << " memory at " << buf->buffer);
     }
 }
-#endif // GGML_VULKAN_MEMORY_DEBUG
 
 struct vk_instance_t {
     vk::Instance instance;
@@ -2074,6 +2086,19 @@ static void ggml_vk_create_pipeline_func(vk_device& device, vk_pipeline& pipelin
         compute_pipeline_create_info.setPNext(&rci);
     }
 
+#if defined(VK_EXT_shader_64bit_indexing)
+    vk::PipelineCreateFlags2CreateInfo pipelineFlags2CreateInfo;
+    if (pipeline->is_64b_indexing)
+    {
+        pipelineFlags2CreateInfo.flags = vk::PipelineCreateFlagBits2::e64BitIndexingEXT;
+        if (device->pipeline_executable_properties_support) {
+            pipelineFlags2CreateInfo.flags |= vk::PipelineCreateFlagBits2::eCaptureStatisticsKHR;
+        }
+        pipelineFlags2CreateInfo.setPNext(compute_pipeline_create_info.pNext);
+        compute_pipeline_create_info.setPNext(&pipelineFlags2CreateInfo);
+    }
+#endif
+
     try {
         pipeline->pipeline = device->device.createComputePipeline(VK_NULL_HANDLE, compute_pipeline_create_info).value;
     } catch (const vk::SystemError& e) {
@@ -2564,9 +2589,7 @@ static vk_buffer ggml_vk_create_buffer(vk_device& device, size_t size, const std
         buf->bda_addr = device->device.getBufferAddress(addressInfo);
     }
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
     device->memory_logger->log_allocation(buf, size);
-#endif
 
     return buf;
 }
@@ -2623,11 +2646,9 @@ static void ggml_vk_destroy_buffer(vk_buffer& buf) {
         return;
     }
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
     if (buf->device != nullptr) {
         buf->device->memory_logger->log_deallocation(buf);
     }
-#endif
 
     buf.reset();
 }
@@ -2996,6 +3017,15 @@ static void ggml_vk_load_shaders(vk_device& device) {
         if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
             m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
             m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
+        } else if (device->vendor_id == VK_VENDOR_ID_AMD && device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary) {
+            // This is intentionally using tx_m values, slight performance increase
+            l_warptile = { 256, 128, 128, 16, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile_mmq = l_warptile_mmq_int = { 256, 128, 128, 32, subgroup_size_8, 64, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile_mmq_int_k = { 256, 128, 128, 32, subgroup_size_16, 64, 1, 4, 2, 1, subgroup_size_16 };
+        } else if (device->vendor_id == VK_VENDOR_ID_INTEL && device->coopmat_support && device->architecture == INTEL_XE2) {
+            // Xe2/Xe3 with coopmat enabled - warptile performance tuning
+            l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
+            l_warptile_mmq = { 512, 128, 128, 32, subgroup_size_8, 32, 2, tm_m, tn_m, tk_m, subgroup_size_8 };
         }
 
         l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
@@ -3051,7 +3081,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
     }
 
     std::vector<std::future<void>> compiles;
-    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& pipeline, const char *name, size_t spv_size, const void* spv_data, const char *entrypoint,
+    auto const &ggml_vk_create_pipeline = [&](vk_device& device, vk_pipeline& base_pipeline, const char *name, size_t spv_size, const void* spv_data, const char *entrypoint,
                                               uint32_t parameter_count, uint32_t push_constant_size, std::array<uint32_t, 3> wg_denoms, const std::vector<uint32_t>& specialization_constants,
                                               uint32_t align, bool disable_robustness = false, bool require_full_subgroups = false, uint32_t required_subgroup_size = 0) {
 
@@ -3059,35 +3089,49 @@ static void ggml_vk_load_shaders(vk_device& device) {
             required_subgroup_size = get_subgroup_size(name, device->architecture);
         }
 
-        if (!pipeline) {
-            pipeline = std::make_shared<vk_pipeline_struct>();
-        }
-        if (!pipeline->initialized) {
-            pipeline->name = name;
-            pipeline->parameter_count = parameter_count;
-            pipeline->push_constant_size = push_constant_size;
-            pipeline->wg_denoms = wg_denoms;
-            pipeline->align = align;
-            pipeline->initialized = true;
-        }
+        vk_pipeline *ptr = &base_pipeline;
 
-        if (!pipeline->needed || pipeline->compiled) {
-            return;
+        int num_pipelines = 1;
+#if defined(VK_EXT_shader_64bit_indexing)
+        if (device->shader_64b_indexing) {
+            num_pipelines = 2;
         }
-        // TODO: We're no longer benefitting from the async compiles (shaders are
-        // compiled individually, as needed) and this complexity can be removed.
-        {
-            // wait until fewer than N compiles are in progress
-            uint32_t N = std::max(1u, std::thread::hardware_concurrency());
-            std::unique_lock<std::mutex> guard(compile_count_mutex);
-            while (compile_count >= N) {
-                compile_count_cond.wait(guard);
+#endif
+        for (int i = 0; i < num_pipelines; ++i, ptr = &(*ptr)->next) {
+            vk_pipeline &pipeline = *ptr;
+            if (!pipeline) {
+                pipeline = std::make_shared<vk_pipeline_struct>();
+            }
+            if (!pipeline->initialized) {
+                pipeline->name = name;
+                pipeline->parameter_count = parameter_count;
+                pipeline->push_constant_size = push_constant_size;
+                pipeline->wg_denoms = wg_denoms;
+                pipeline->align = align;
+                pipeline->initialized = true;
+#if defined(VK_EXT_shader_64bit_indexing)
+                pipeline->is_64b_indexing = (i == 1);
+#endif
             }
-            compile_count++;
-        }
 
-        compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint,
-                                      parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
+            if (!pipeline->needed || pipeline->compiled) {
+                continue;
+            }
+            // TODO: We're no longer benefitting from the async compiles (shaders are
+            // compiled individually, as needed) and this complexity can be removed.
+            {
+                // wait until fewer than N compiles are in progress
+                uint32_t N = std::max(1u, std::thread::hardware_concurrency());
+                std::unique_lock<std::mutex> guard(compile_count_mutex);
+                while (compile_count >= N) {
+                    compile_count_cond.wait(guard);
+                }
+                compile_count++;
+            }
+
+            compiles.push_back(std::async(ggml_vk_create_pipeline_func, std::ref(device), std::ref(pipeline), spv_size, spv_data, entrypoint,
+                                          parameter_count, wg_denoms, specialization_constants, disable_robustness, require_full_subgroups, required_subgroup_size));
+        }
     };
 
     auto const &ggml_vk_create_pipeline2 = [&](vk_device& device, vk_pipeline& pipeline, const std::string &name, size_t spv_size, const void* spv_data, const char *entrypoint,
@@ -3336,12 +3380,12 @@ static void ggml_vk_load_shaders(vk_device& device) {
 
         GGML_ASSERT(device->subgroup_ballot);
 
-        CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
-        CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
-        CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
+        CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
+        CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
+        CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
 #if defined(GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT)
         if (device->coopmat_bf16_support) {
-            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id);
+            CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id);
         }
 #endif
 
@@ -3449,9 +3493,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #endif
 
         if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) {
-            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_subgroup_f16, wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_subgroup_f16_f32, wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
             CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
 
             CREATE_MM2(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0], matmul_id_subgroup_q4_0_f32, mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -3493,9 +3537,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
             }
 #endif
         } else {
-            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
-            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
-            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16, matmul_id_f16, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM2(GGML_TYPE_F16, pipeline_matmul_id_f16_f32, matmul_id_f16_f32, wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
 
             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_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -3610,9 +3654,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
 #endif
 
         if (device->subgroup_ballot && device->subgroup_require_full_support && subgroup_min_size_16) {
-            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
-            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_subgroup_f32_f32, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_subgroup_f16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
+            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_subgroup_f16_f32, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
             CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_subgroup_bf16, , wg_denoms, warptile_id, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size_16);
 
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_subgroup_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
@@ -3636,9 +3680,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
             CREATE_MM(GGML_TYPE_IQ4_NL,  pipeline_dequant_mul_mat_mat_id[GGML_TYPE_IQ4_NL].f32acc,  matmul_id_subgroup_iq4_nl_f32,  , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
             CREATE_MM(GGML_TYPE_MXFP4,   pipeline_dequant_mul_mat_mat_id[GGML_TYPE_MXFP4].f32acc,   matmul_id_subgroup_mxfp4_f32,   , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, mul_mat_subgroup_size);
         } else {
-            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
-            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
-            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM(GGML_TYPE_F32, pipeline_matmul_id_f32, matmul_id_f32_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16.f32acc, matmul_id_f16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
+            CREATE_MM(GGML_TYPE_F16, pipeline_matmul_id_f16_f32.f32acc, matmul_id_f16_f32, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
             CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
 
             CREATE_MM(GGML_TYPE_Q4_0, pipeline_dequant_mul_mat_mat_id[GGML_TYPE_Q4_0].f32acc, matmul_id_q4_0_f32, , mmq_wg_denoms, warptile_mmqid, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
@@ -3678,6 +3722,11 @@ static void ggml_vk_load_shaders(vk_device& device) {
         m_wg_denoms = { 64,  64, 1 };
         s_wg_denoms = { 32,  32, 1 };
 
+        if (device->vendor_id == VK_VENDOR_ID_INTEL && device->architecture == INTEL_XE2) {
+            // Xe2/Xe3 - bf16 warptile performance tuning
+            l_warptile = { 512, 128, 128, 16, subgroup_size_8, 32, 2, 4, 4, 1, subgroup_size_8 };
+        }
+
         CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_bf16, matmul_bf16, , wg_denoms, warptile, vk_mat_mat_push_constants, 3, , 0);
         CREATE_MM(GGML_TYPE_BF16, pipeline_matmul_id_bf16, matmul_id_bf16, , wg_denoms, warptile, vk_mat_mat_id_push_constants, mul_mat_id_param_count, _id, 0);
     }
@@ -3831,22 +3880,22 @@ static void ggml_vk_load_shaders(vk_device& device) {
             const uint32_t subgroup_size_int = (device->vendor_id == VK_VENDOR_ID_INTEL && device->subgroup_size_control) ? device->subgroup_min_size : device->subgroup_size;
             const uint32_t wg_size_subgroup_int = (w == DMMV_WG_SIZE_SUBGROUP) ? subgroup_size_int : (subgroup_size_int * 4);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_q8_1_f32", arr_dmmv_id_q4_0_q8_1_f32_len[reduc], arr_dmmv_id_q4_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_q8_1_f32", arr_dmmv_id_q4_1_q8_1_f32_len[reduc], arr_dmmv_id_q4_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_q8_1_f32", arr_dmmv_id_q5_0_q8_1_f32_len[reduc], arr_dmmv_id_q5_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_q8_1_f32", arr_dmmv_id_q5_1_q8_1_f32_len[reduc], arr_dmmv_id_q5_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_q8_1_f32", arr_dmmv_id_q8_0_q8_1_f32_len[reduc], arr_dmmv_id_q8_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_0], "mul_mat_vec_id_q4_0_q8_1_f32", arr_dmmv_id_q4_0_q8_1_f32_len[reduc], arr_dmmv_id_q4_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_1], "mul_mat_vec_id_q4_1_q8_1_f32", arr_dmmv_id_q4_1_q8_1_f32_len[reduc], arr_dmmv_id_q4_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_0], "mul_mat_vec_id_q5_0_q8_1_f32", arr_dmmv_id_q5_0_q8_1_f32_len[reduc], arr_dmmv_id_q5_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_1], "mul_mat_vec_id_q5_1_q8_1_f32", arr_dmmv_id_q5_1_q8_1_f32_len[reduc], arr_dmmv_id_q5_1_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q8_0], "mul_mat_vec_id_q8_0_q8_1_f32", arr_dmmv_id_q8_0_q8_1_f32_len[reduc], arr_dmmv_id_q8_0_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_MXFP4], "mul_mat_vec_id_mxfp4_q8_1_f32", arr_dmmv_id_mxfp4_q8_1_f32_len[reduc], arr_dmmv_id_mxfp4_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_MXFP4], "mul_mat_vec_id_mxfp4_q8_1_f32", arr_dmmv_id_mxfp4_q8_1_f32_len[reduc], arr_dmmv_id_mxfp4_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_stdq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_stdq_int}, 1, true, use_subgroups, subgroup_size_int);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_q8_1_f32", arr_dmmv_id_q2_k_q8_1_f32_len[reduc], arr_dmmv_id_q2_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {2*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_q8_1_f32", arr_dmmv_id_q3_k_q8_1_f32_len[reduc], arr_dmmv_id_q3_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_q8_1_f32", arr_dmmv_id_q4_k_q8_1_f32_len[reduc], arr_dmmv_id_q4_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_q8_1_f32", arr_dmmv_id_q5_k_q8_1_f32_len[reduc], arr_dmmv_id_q5_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_q8_1_f32", arr_dmmv_id_q6_k_q8_1_f32_len[reduc], arr_dmmv_id_q6_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q2_K], "mul_mat_vec_id_q2_k_q8_1_f32", arr_dmmv_id_q2_k_q8_1_f32_len[reduc], arr_dmmv_id_q2_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {2*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 2*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q3_K], "mul_mat_vec_id_q3_k_q8_1_f32", arr_dmmv_id_q3_k_q8_1_f32_len[reduc], arr_dmmv_id_q3_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_q8_1_f32", arr_dmmv_id_q4_k_q8_1_f32_len[reduc], arr_dmmv_id_q4_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_q8_1_f32", arr_dmmv_id_q5_k_q8_1_f32_len[reduc], arr_dmmv_id_q5_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_q8_1_f32", arr_dmmv_id_q6_k_q8_1_f32_len[reduc], arr_dmmv_id_q6_k_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_kq_int, 1, 1}, {wg_size_subgroup_int, 1*rm_kq_int}, 1, true, use_subgroups, subgroup_size_int);
 
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_IQ1_S], "mul_mat_vec_id_iq1_s_q8_1_f32", arr_dmmv_id_iq1_s_q8_1_f32_len[reduc], arr_dmmv_id_iq1_s_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_iq_int(0), 1, 1}, {wg_size_subgroup_int, 1*rm_iq_int(0)}, 1, true, use_subgroups, subgroup_size_int);
-            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_IQ1_M], "mul_mat_vec_id_iq1_m_q8_1_f32", arr_dmmv_id_iq1_m_q8_1_f32_len[reduc], arr_dmmv_id_iq1_m_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_push_constants), {1*rm_iq_int(0), 1, 1}, {wg_size_subgroup_int, 1*rm_iq_int(0)}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_IQ1_S], "mul_mat_vec_id_iq1_s_q8_1_f32", arr_dmmv_id_iq1_s_q8_1_f32_len[reduc], arr_dmmv_id_iq1_s_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_iq_int(0), 1, 1}, {wg_size_subgroup_int, 1*rm_iq_int(0)}, 1, true, use_subgroups, subgroup_size_int);
+            ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_q8_1_f32[w][GGML_TYPE_IQ1_M], "mul_mat_vec_id_iq1_m_q8_1_f32", arr_dmmv_id_iq1_m_q8_1_f32_len[reduc], arr_dmmv_id_iq1_m_q8_1_f32_data[reduc], "main", mul_mat_vec_id_num_bindings, sizeof(vk_mat_vec_id_push_constants), {1*rm_iq_int(0), 1, 1}, {wg_size_subgroup_int, 1*rm_iq_int(0)}, 1, true, use_subgroups, subgroup_size_int);
         }
 #endif // GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT
     }
@@ -3934,9 +3983,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, 5 * sizeof(uint32_t), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
 
     if (device->subgroup_clustered && device->subgroup_require_full_support) {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_subgroup_len, quantize_q8_1_x4_subgroup_data, "main", 2, sizeof(vk_quantize_q8_1_push_constants), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1, true, true);
     } else {
-        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, 1 * sizeof(uint32_t), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
+        ggml_vk_create_pipeline(device, device->pipeline_quantize_q8_1_x4, "quantize_q8_1_x4", quantize_q8_1_x4_len, quantize_q8_1_x4_data, "main", 2, sizeof(vk_quantize_q8_1_push_constants), {32 * device->subgroup_size / 8, 1, 1}, { device->subgroup_size }, 1);
     }
 
     for (uint32_t i = 0; i < p021_max_gqa_ratio; ++i) {
@@ -4144,9 +4193,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_add1_f16_f32, "add1_f16_f32", add1_f16_f32_len, add1_f16_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
     ggml_vk_create_pipeline(device, device->pipeline_add1_f32_f32, "add1_f32_f32", add1_f32_f32_len, add1_f32_f32_data, "main", 3, sizeof(vk_op_binary_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_arange_f32, "arange_f32", arange_f32_len, arange_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_arange_f32, "arange_f32", arange_f32_len, arange_f32_data, "main", 1, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
-    ggml_vk_create_pipeline(device, device->pipeline_fill_f32, "fill_f32", fill_f32_len, fill_f32_data, "main", 1, sizeof(vk_op_unary_push_constants), {512, 1, 1}, {}, 1);
+    ggml_vk_create_pipeline(device, device->pipeline_fill_f32, "fill_f32", fill_f32_len, fill_f32_data, "main", 1, sizeof(vk_op_push_constants), {512, 1, 1}, {}, 1);
 
 #define CREATE_GLU(name)  \
     if (device->float_controls_rte_fp16) {  \
@@ -4292,8 +4341,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
     ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv7_f32, "rwkv_wkv7_f32", rwkv_wkv7_f32_len, rwkv_wkv7_f32_data, "main", 8, sizeof(vk_op_rwkv_wkv7_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
 
     if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
-        ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, device->subgroup_size, 16}, 1, true, true);
-        ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, device->subgroup_size, 16}, 1, true, true);
+        ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, device->subgroup_size}, 1, true, true);
+        ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_subgroup_f32_len, ssm_scan_subgroup_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, device->subgroup_size}, 1, true, true);
     } else {
         ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d128, "ssm_scan_128_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {128, device->subgroup_size, 16}, 1, true, true);
         ggml_vk_create_pipeline(device, device->pipeline_ssm_scan_f32_d256, "ssm_scan_256_f32", ssm_scan_f32_len, ssm_scan_f32_data, "main", 8, sizeof(vk_op_ssm_scan_push_constants), {1, 1, 1}, {256, device->subgroup_size, 16}, 1, true, true);
@@ -4420,9 +4469,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         vk_device device = std::make_shared<vk_device_struct>();
         vk_instance.devices[idx] = device;
 
-#ifdef GGML_VULKAN_MEMORY_DEBUG
         device->memory_logger = std::unique_ptr<vk_memory_logger>(new vk_memory_logger());
-#endif
 
         size_t dev_num = vk_instance.device_indices[idx];
 
@@ -4460,6 +4507,7 @@ static vk_device ggml_vk_get_device(size_t idx) {
         bool pipeline_executable_properties_support = false;
         device->coopmat_support = false;
         device->integer_dot_product = false;
+        device->shader_64b_indexing = false;
         bool bfloat16_support = false;
 
         for (const auto& properties : ext_props) {
@@ -4507,6 +4555,10 @@ static vk_device ggml_vk_get_device(size_t idx) {
                 device->memory_priority = true;
             } else if (strcmp("VK_EXT_external_memory_host", properties.extensionName) == 0) {
                 device->external_memory_host = true;
+#if defined(VK_EXT_shader_64bit_indexing)
+            } else if (strcmp("VK_EXT_shader_64bit_indexing", properties.extensionName) == 0) {
+                device->shader_64b_indexing = true;
+#endif
             }
         }
 
@@ -4629,6 +4681,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
         }
         device->float_controls_rte_fp16 = vk12_props.shaderRoundingModeRTEFloat16;
 
+        device->subgroup_basic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
+                                 (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eBasic);
         device->subgroup_arithmetic = (vk11_props.subgroupSupportedStages & vk::ShaderStageFlagBits::eCompute) &&
                                       (vk11_props.subgroupSupportedOperations & vk::SubgroupFeatureFlagBits::eArithmetic);
 #ifdef __APPLE__
@@ -4795,6 +4849,16 @@ static vk_device ggml_vk_get_device(size_t idx) {
             device_extensions.push_back("VK_EXT_external_memory_host");
         }
 
+#if defined(VK_EXT_shader_64bit_indexing)
+        VkPhysicalDeviceShader64BitIndexingFeaturesEXT shader_64bit_indexing_features {};
+        shader_64bit_indexing_features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_SHADER_64_BIT_INDEXING_FEATURES_EXT;
+        if (device->shader_64b_indexing) {
+            last_struct->pNext = (VkBaseOutStructure *)&shader_64bit_indexing_features;
+            last_struct = (VkBaseOutStructure *)&shader_64bit_indexing_features;
+            device_extensions.push_back("VK_EXT_shader_64bit_indexing");
+        }
+#endif
+
         vkGetPhysicalDeviceFeatures2(device->physical_device, &device_features2);
 
         device->pipeline_executable_properties_support = pipeline_executable_properties_support;
@@ -5061,11 +5125,23 @@ static vk_device ggml_vk_get_device(size_t idx) {
             switch (device->vendor_id) {
 #ifndef GGML_VULKAN_RUN_TESTS
             case VK_VENDOR_ID_AMD:
+                device->mul_mat_l[i]    = device->coopmat_support && device->driver_id != vk::DriverId::eAmdProprietary;
+                device->mul_mat_m[i]    = true;
+                device->mul_mat_s[i]    = true;
+                device->mul_mat_id_l[i] = false;
+                device->mul_mat_id_m[i] = true;
+                device->mul_mat_id_s[i] = true;
+                break;
             case VK_VENDOR_ID_INTEL:
-                device->mul_mat_l[i] = false;
+                if (!device->coopmat_support || device->architecture != INTEL_XE2) {
+                    device->mul_mat_l[i] = false;
+                    device->mul_mat_id_l[i] = false;
+                } else {
+                    device->mul_mat_l[i] = true;  // if coopmat & XE2+, allow large matmul warptile config for Intel
+                    device->mul_mat_id_l[i] = true;
+                }
                 device->mul_mat_m[i] = true;
                 device->mul_mat_s[i] = true;
-                device->mul_mat_id_l[i] = false;
                 device->mul_mat_id_m[i] = true;
                 device->mul_mat_id_s[i] = true;
                 break;
@@ -5390,6 +5466,7 @@ static void ggml_vk_instance_init() {
     vk_perf_logger_enabled = getenv("GGML_VK_PERF_LOGGER") != nullptr;
     vk_perf_logger_concurrent = getenv("GGML_VK_PERF_LOGGER_CONCURRENT") != nullptr;
     vk_enable_sync_logger = getenv("GGML_VK_SYNC_LOGGER") != nullptr;
+    vk_memory_logger_enabled = getenv("GGML_VK_MEMORY_LOGGER") != nullptr;
     const char* GGML_VK_PERF_LOGGER_FREQUENCY = getenv("GGML_VK_PERF_LOGGER_FREQUENCY");
 
     if (GGML_VK_PERF_LOGGER_FREQUENCY != nullptr) {
@@ -6076,6 +6153,7 @@ static void ggml_vk_dispatch_pipeline(ggml_backend_vk_context* ctx, vk_context&
     GGML_ASSERT(ctx->descriptor_set_idx < ctx->descriptor_sets.size());
     GGML_ASSERT(descriptor_buffer_infos.size() <= MAX_PARAMETER_COUNT);
     GGML_ASSERT(pipeline->parameter_count == descriptor_buffer_infos.size());
+    GGML_ASSERT(pipeline->push_constant_size == push_constant_size(push_constants));
 
     vk::DescriptorSet& descriptor_set = ctx->descriptor_sets[ctx->descriptor_set_idx++];
     vk::WriteDescriptorSet write_descriptor_set{ descriptor_set, 0, 0, pipeline->parameter_count, vk::DescriptorType::eStorageBuffer, nullptr, descriptor_buffer_infos.begin() };
@@ -6858,10 +6936,29 @@ static void ggml_vk_quantize_q8_1(ggml_backend_vk_context * ctx, vk_context& sub
     const uint64_t max_elements = std::min<uint64_t>(uint64_t{ctx->device->properties.limits.maxComputeWorkGroupCount[0]} * pipeline->wg_denoms[0], std::numeric_limits<uint32_t>::max());
     const uint32_t elements = std::min(ne, static_cast<uint32_t>(max_elements));
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, std::array<uint32_t, 2>{ ne, num_blocks }, { elements, 1, 1 });
+    const vk_quantize_q8_1_push_constants pc = {
+        ne,
+        num_blocks,
+    };
+
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { in, out }, pc, { elements, 1, 1 });
     ggml_vk_sync_buffers(ctx, subctx);
 }
 
+static vk_pipeline ggml_vk_get_64b_indexing_pipeline(ggml_backend_vk_context * ctx, vk_pipeline &pipeline) {
+    GGML_UNUSED(ctx);
+#if defined(VK_EXT_shader_64bit_indexing)
+    vk_pipeline *ptr = &pipeline;
+    while (*ptr) {
+        if ((*ptr)->is_64b_indexing) {
+            return *ptr;
+        }
+        ptr = &(*ptr)->next;
+    }
+#endif
+    return pipeline;
+}
+
 static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool disable_split_k) {
     VK_LOG_DEBUG("ggml_vk_mul_mat_q_f16((" << src0 << ", name=" << src0->name << ", type=" << ggml_type_name(src0->type) << ", ne0=" << src0->ne[0] << ", ne1=" << src0->ne[1] << ", ne2=" << src0->ne[2] << ", ne3=" << src0->ne[3] << ", nb0=" << src0->nb[0] << ", nb1=" << src0->nb[1] << ", nb2=" << src0->nb[2] << ", nb3=" << src0->nb[3];
     std::cerr << "), (" << src1 << ", name=" << src1->name << ", type=" << ggml_type_name(src1->type) << ", ne0=" << src1->ne[0] << ", ne1=" << src1->ne[1] << ", ne2=" << src1->ne[2] << ", ne3=" << src1->ne[3] << ", nb0=" << src1->nb[0] << ", nb1=" << src1->nb[1] << ", nb2=" << src1->nb[2] << ", nb3=" << src1->nb[3];
@@ -6945,6 +7042,10 @@ static void ggml_vk_mul_mat_q_f16(ggml_backend_vk_context * ctx, vk_context& sub
 
     vk_pipeline pipeline = ggml_vk_guess_matmul_pipeline(ctx, mmp, ne01, ne11, aligned, qx_needs_dequant ? f16_type : src0->type, quantize_y ? GGML_TYPE_Q8_1 : (y_f32_kernel ? GGML_TYPE_F32 : src1->type));
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) : ne11;
     const uint64_t x_ne = ggml_nelements(src0);
@@ -7254,6 +7355,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
         to_q8_1 = ggml_vk_get_quantize_pipeline(ctx, GGML_TYPE_Q8_1);
     }
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        dmmv = ggml_vk_get_64b_indexing_pipeline(ctx, dmmv);
+    }
+
     const bool qx_needs_dequant = x_non_contig;
     const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
 
@@ -7449,9 +7554,15 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         gqa_ratio = 1;
     }
 
+    vk_pipeline pipeline = ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1];
+
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1], 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
@@ -7493,7 +7604,7 @@ static void ggml_vk_mul_mat_vec_p021_f16_f32(ggml_backend_vk_context * ctx, vk_c
         workgroups_z /= gqa_ratio;
     }
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_p021_f16_f32[gqa_ratio - 1],
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
         {
             d_Qx,
             d_Qy,
@@ -7543,9 +7654,14 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
     const uint32_t channel_stride_x = nb02 / sizeof(ggml_fp16_t);
     const uint32_t channel_stride_y = nb12 / sizeof(float);
 
+    vk_pipeline pipeline = ctx->device->pipeline_mul_mat_vec_nc_f16_f32;
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
+
     {
         // Request descriptor sets
-        ggml_pipeline_request_descriptor_sets(ctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32, 1);
+        ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
     }
 
     vk_subbuffer d_D = ggml_vk_tensor_subbuffer(ctx, cgraph->nodes[node_idx + ctx->num_additional_fused_ops], true);
@@ -7582,7 +7698,7 @@ static void ggml_vk_mul_mat_vec_nc_f16_f32(ggml_backend_vk_context * ctx, vk_con
 
     init_pushconst_tensor_offsets(ctx, pc, src0, src1, nullptr, nullptr, cgraph->nodes[node_idx + ctx->num_additional_fused_ops]);
 
-    ggml_vk_dispatch_pipeline(ctx, subctx, ctx->device->pipeline_mul_mat_vec_nc_f16_f32,
+    ggml_vk_dispatch_pipeline(ctx, subctx, pipeline,
         {
             d_Qx,
             d_Qy,
@@ -7601,8 +7717,9 @@ static void ggml_vk_mul_mat(ggml_backend_vk_context * ctx, vk_context& subctx, c
     // Handle huge A matrix by splitting the M dimensions. This works well for convolution use cases
     // where the M dimension is very large.
     // Split_k doesn't work with M splitting.
+    // This only supports batchsize == 1.
     const size_t nbytes = ggml_nbytes(src0);
-    const bool needs_split = nbytes > ctx->device->properties.limits.maxStorageBufferRange;
+    const bool needs_split = dst->ne[2] == 1 && dst->ne[3] == 1 && nbytes > ctx->device->properties.limits.maxStorageBufferRange;
     if (needs_split) {
         // Choose the number of rows that can fit (and divide by two, to allow for any additional offsets)
         const uint32_t M_split = ctx->device->properties.limits.maxStorageBufferRange / (2 * src0->nb[1]);
@@ -7744,6 +7861,9 @@ static void ggml_vk_mul_mat_id_q_f16(ggml_backend_vk_context * ctx, vk_context&
 
     vk_pipeline pipeline = ggml_vk_guess_matmul_id_pipeline(ctx, mmp, ne01, nei1, aligned, qx_needs_dequant ? f16_type : src0->type);
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        pipeline = ggml_vk_get_64b_indexing_pipeline(ctx, pipeline);
+    }
     // Reserve extra storage in the N dimension for the Y matrix, so we can avoid bounds-checking
     uint32_t padded_n = qy_needs_dequant ? ROUNDUP_POW2(ne11, pipeline->wg_denoms[1]) :ne11;
     const uint64_t x_ne = ggml_nelements(src0);
@@ -8005,6 +8125,10 @@ static void ggml_vk_mul_mat_vec_id_q_f16(ggml_backend_vk_context * ctx, vk_conte
     const bool qx_needs_dequant = x_non_contig;
     const bool qy_needs_dequant = !quantize_y && ((src1->type != GGML_TYPE_F16 && !f16_f32_kernel) || y_non_contig);
 
+    if (ggml_nbytes(src0) > ctx->device->properties.limits.maxStorageBufferRange) {
+        dmmv = ggml_vk_get_64b_indexing_pipeline(ctx, dmmv);
+    }
+
     // Not implemented
     GGML_ASSERT(y_non_contig || !qy_needs_dequant);  // NOLINT
     GGML_ASSERT(!qx_needs_dequant || to_fp16_vk_0 != nullptr);  // NOLINT
@@ -9849,8 +9973,9 @@ static void ggml_vk_ssm_scan(ggml_backend_vk_context * ctx, vk_context& subctx,
 
     std::array<uint32_t, 3> elements;
 
-    const int splitH = 16;
-    const uint32_t num_workgroups_x = CEIL_DIV(n_head * head_dim, splitH);
+    const uint32_t d_state = src0->ne[0];
+    uint32_t num_subgroups = d_state / ctx->device->subgroup_size;
+    const uint32_t num_workgroups_x = CEIL_DIV(n_head * head_dim, num_subgroups);
     const uint32_t num_workgroups_y = n_seq;
     elements = { num_workgroups_x, num_workgroups_y, 1 };
 
@@ -14228,6 +14353,7 @@ struct ggml_backend_vk_device_context {
     std::string description;
     bool is_integrated_gpu;
     std::string pci_bus_id;
+    int op_offload_min_batch_size;
 };
 
 static const char * ggml_backend_vk_device_get_name(ggml_backend_dev_t dev) {
@@ -14284,6 +14410,19 @@ static ggml_backend_t ggml_backend_vk_device_init(ggml_backend_dev_t dev, const
 }
 
 static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
+    ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
+    const vk_device& device = ggml_vk_get_device(ctx->device);
+
+    // reject any tensors larger than the max buffer size
+    for (int i = 0; i < GGML_MAX_SRC; i++) {
+        if (op->src[i] && ggml_nbytes(op->src[i]) > device->max_buffer_size) {
+            return false;
+        }
+    }
+    if (ggml_nbytes(op) > device->max_buffer_size) {
+        return false;
+    }
+
     switch (op->op) {
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(op)) {
@@ -14332,8 +14471,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
         case GGML_OP_MUL_MAT_ID:
             {
                 ggml_type src0_type = op->src[0]->type;
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                const vk_device& device = ggml_vk_get_device(ctx->device);
                 if (op->op == GGML_OP_MUL_MAT_ID) {
                     if (!device->mul_mat_id_s[src0_type] && !device->mul_mat_id_m[src0_type] && !device->mul_mat_id_l[src0_type]) {
                         // If there's not enough shared memory for row_ids and the result tile, fallback to CPU
@@ -14394,8 +14531,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             }
         case GGML_OP_FLASH_ATTN_EXT:
             {
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                auto device = ggml_vk_get_device(ctx->device);
                 bool coopmat2 = device->coopmat2;
                 uint32_t HSK = op->src[1]->ne[0];
                 uint32_t HSV = op->src[2]->ne[0];
@@ -14617,8 +14752,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (!ggml_is_contiguous(op) || !ggml_is_contiguous(op->src[0])) {
                     return false;
                 }
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                auto device = ggml_vk_get_device(ctx->device);
                 // pipeline_argsort_large_f32 requires vulkan memory model.
                 if (device->vulkan_memory_model) {
                     return true;
@@ -14631,8 +14764,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                 if (!ggml_is_contiguous(op) || !ggml_is_contiguous(op->src[0])) {
                     return false;
                 }
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                auto device = ggml_vk_get_device(ctx->device);
                 // We could potentially support larger, using argsort to sort the
                 // whole thing. Not clear if this is needed.
                 uint32_t min_pipeline = (uint32_t)log2f(float(op->ne[0])) + 1;
@@ -14679,8 +14810,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
         case GGML_OP_CUMSUM:
             {
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                auto device = ggml_vk_get_device(ctx->device);
                 if (device->subgroup_arithmetic && device->subgroup_require_full_support) {
                     return op->src[0]->type == GGML_TYPE_F32 && ggml_is_contiguous_rows(op->src[0]);
                 }
@@ -14688,9 +14817,6 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
             }
         case GGML_OP_SOLVE_TRI:
             {
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                const vk_device& device = ggml_vk_get_device(ctx->device);
-
                 if (op->type != GGML_TYPE_F32 || op->src[0]->type != GGML_TYPE_F32) {
                     return false;
                 }
@@ -14755,14 +14881,13 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     return false;
                 }
 
-                ggml_backend_vk_device_context * ctx = (ggml_backend_vk_device_context *)dev->context;
-                const vk_device& device = ggml_vk_get_device(ctx->device);
+                size_t shmem_size = d_state * sizeof(float);
 
-                const uint32_t SPLIT_H = 16;
+                if (shmem_size > device->properties.limits.maxComputeSharedMemorySize) {
+                    return false;
+                }
 
-                size_t stateC_size = SPLIT_H * d_state * sizeof(float);
-
-                if (stateC_size > device->properties.limits.maxComputeSharedMemorySize) {
+                if (!device->subgroup_basic) {
                     return false;
                 }
 
@@ -14802,12 +14927,10 @@ static bool ggml_backend_vk_device_supports_buft(ggml_backend_dev_t dev, ggml_ba
 }
 
 static bool ggml_backend_vk_device_offload_op(ggml_backend_dev_t dev, const ggml_tensor * op) {
-    const int min_batch_size = 32;
+    ggml_backend_vk_device_context * dev_ctx = (ggml_backend_vk_device_context *)dev->context;
 
-    return (op->ne[1] >= min_batch_size && op->op != GGML_OP_GET_ROWS) ||
-           (op->ne[2] >= min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
-
-    UNUSED(dev);
+    return (op->ne[1] >= dev_ctx->op_offload_min_batch_size && op->op != GGML_OP_GET_ROWS) ||
+           (op->ne[2] >= dev_ctx->op_offload_min_batch_size && op->op == GGML_OP_MUL_MAT_ID);
 }
 
 static ggml_backend_event_t ggml_backend_vk_device_event_new(ggml_backend_dev_t dev) {
@@ -14933,6 +15056,7 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
         static std::mutex mutex;
         std::lock_guard<std::mutex> lock(mutex);
         if (!initialized) {
+            const int min_batch_size = getenv("GGML_OP_OFFLOAD_MIN_BATCH") ? atoi(getenv("GGML_OP_OFFLOAD_MIN_BATCH")) : 32;
             for (int i = 0; i < ggml_backend_vk_get_device_count(); i++) {
                 ggml_backend_vk_device_context * ctx = new ggml_backend_vk_device_context;
                 char desc[256];
@@ -14942,6 +15066,7 @@ static ggml_backend_dev_t ggml_backend_vk_reg_get_device(ggml_backend_reg_t reg,
                 ctx->description = desc;
                 ctx->is_integrated_gpu = ggml_backend_vk_get_device_type(i) == vk::PhysicalDeviceType::eIntegratedGpu;
                 ctx->pci_bus_id = ggml_backend_vk_get_device_pci_id(i);
+                ctx->op_offload_min_batch_size = min_batch_size;
                 devices.push_back(new ggml_backend_device {
                     /* .iface   = */ ggml_backend_vk_device_i,
                     /* .reg     = */ reg,

ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs.glsl

@@ -462,7 +462,8 @@ vec2 get_dm(uint ib, uint a_offset) {
 
 #if defined(DATA_A_Q4_1) || defined(DATA_A_Q5_1)
 vec2 get_dm(uint ib, uint a_offset) {
-    return vec2(float(data_a[a_offset + ib].d), float(data_a[a_offset + ib].m));
+    const vec2 dm = vec2(data_a_packed32[a_offset + ib].dm);
+    return dm;
 }
 #endif
 

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

@@ -87,7 +87,6 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K;
 
     y_offset = QUANT_R == 1 ? 1 : QUANT_K/2;
 

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_base.glsl

@@ -65,9 +65,9 @@ void get_offsets(out uint a_offset, out uint b_offset, out uint d_offset) {
 
     a_offset =
 #ifdef MUL_MAT_ID
-            expert_id * p.batch_stride_a;
+            expert_id * (p.batch_stride_a / QUANT_K);
 #else
-            batch_idx_a * p.batch_stride_a;
+            batch_idx_a * (p.batch_stride_a / QUANT_K);
 #endif
     b_offset =
 #ifdef MUL_MAT_ID

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_m.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
                                const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
     // Compute starting index in matrix B for this superblock
     const uint y_idx = i * QUANT_K + 32 * ib32;
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
 
     // Precompute indices for quantization lookup tables
     const uint qh_base = 2 * ib32;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq1_s.comp

@@ -17,7 +17,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
             const vec4 b_val_1 = vec4(data_b_v4[base_b_idx + 2 * l + 1]);
 
             // index for data_a
-            uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+            uint ibi = a_offset + first_row * num_blocks_per_row + i;
 
             [[unroll]] for (uint n = 0; n < num_rows; ++n) {
                 const float d = float(data_a[ibi].d);

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_s.comp

@@ -12,7 +12,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint scale = (data_a[ibi].scales[ib32] >> nibble_shift) & 0xF;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint nibble_shift = 4 * (itid & 1);
     const uint ib32 = itid / 2; // 0..7
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     // Precompute db multiplication factors
     float db_vals[NUM_ROWS];
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
@@ -22,7 +22,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
         db_vals[n] = d * (0.125f + float(scale) * 0.25f);
         ibi += num_blocks_per_row;
     }
-    ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         // Preload grid and sign data for all l values
         vec4 grid0_vals[2], grid1_vals[2];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq2_xxs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint signscale = pack32(u16vec2(

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp

@@ -10,7 +10,7 @@ FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
     const uint y_idx = i * QUANT_K + 32 * ib32;
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_xxs.comp

@@ -11,7 +11,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + 16 * itid;
     const uint ib32 = itid / 2; // 0..7
 
-    uint ibi = a_offset / QUANT_K + first_row * num_blocks_per_row + i;
+    uint ibi = a_offset + first_row * num_blocks_per_row + i;
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
         const float d = float(data_a[ibi].d);
         const uint signscale = pack32(u16vec2(

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q2_k.comp

@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q3_k.comp

@@ -14,7 +14,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ix, co
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q4_k.comp

@@ -13,7 +13,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q5_k.comp

@@ -13,7 +13,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint v_im,
     const uint y2_idx = y1_idx + 128;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         const FLOAT_TYPE_VEC2 dm = FLOAT_TYPE_VEC2(data_a[ib0 + i].dm);
 
         const uint32_t scale0_u32 = data_a_packed16[ib0 + i].scales[v_im    ];

ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp

@@ -15,7 +15,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint itid,
     const uint y_idx = i * QUANT_K + y_offset;
 
     [[unroll]] for (uint n = 0; n < num_rows; ++n) {
-        const uint ib0 = a_offset / QUANT_K + (first_row+n)*num_blocks_per_row;
+        const uint ib0 = a_offset + (first_row+n)*num_blocks_per_row;
         csel ^= 1;
 
         if (!all_threads) { // when we don't have enough blocks to use all threads

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

@@ -79,7 +79,7 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
     const uint tid = gl_LocalInvocationID.x;
 
     get_offsets(a_offset, b_offset, d_offset);
-    a_offset /= QUANT_K_Q8_1;
+    a_offset *= QUANT_K / QUANT_K_Q8_1;
     b_offset /= QUANT_K_Q8_1;
 
     FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];

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

@@ -234,13 +234,13 @@ void main() {
     const uint end_k = min(p.K, (ik + 1) * p.k_split);
 #endif
 
-    uint pos_a = (
+    uint pos_a =
 #ifdef MUL_MAT_ID
-        expert_idx * p.batch_stride_a +
+        expert_idx * (p.batch_stride_a / LOAD_VEC_A) +
 #else
-        batch_idx_a * p.batch_stride_a +
+        batch_idx_a * (p.batch_stride_a / LOAD_VEC_A) +
 #endif
-        ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
+        (ir * BM * p.stride_a + start_k) / LOAD_VEC_A;
 #ifdef MUL_MAT_ID
     uint pos_b = 0;
 #else

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp

@@ -250,10 +250,10 @@ void main() {
 #endif
 
 #ifdef MUL_MAT_ID
-    uint pos_a = (expert_idx * p.batch_stride_a) / QUANT_K;
+    uint pos_a = expert_idx * (p.batch_stride_a / QUANT_K);
     uint pos_b = 0;
 #else
-    uint pos_a = (batch_idx_a * p.batch_stride_a) / QUANT_K;
+    uint pos_a = batch_idx_a * (p.batch_stride_a / QUANT_K);
     uint pos_b = batch_idx * p.batch_stride_b;
     uint pos_d = batch_idx * p.batch_stride_d + ik * p.batch_stride_d * gl_NumWorkGroups.z;
 #endif

ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.glsl

@@ -47,7 +47,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
 #endif
 #elif defined(DATA_A_Q4_0)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
             const uint ib = idx / 4;
             const uint iqs = idx & 0x03;
@@ -63,16 +63,15 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             buf_a[buf_idx + 9] = FLOAT_TYPE_VEC2(v1.zw);
 #elif defined(DATA_A_Q4_1)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + 2 * row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
             const uint ib = idx / 4;
             const uint iqs = idx & 0x03;
 
-            const float d = float(data_a_packed16[ib].d);
-            const float m = float(data_a_packed16[ib].m);
-            const uint vui = uint(data_a_packed16[ib].qs[2*iqs]) | (uint(data_a_packed16[ib].qs[2*iqs + 1]) << 16);
-            const vec4 v0 = vec4(unpack8(vui & 0x0F0F0F0F)) * d + m;
-            const vec4 v1 = vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) * d + m;
+            const vec2 dm = vec2(data_a_packed32[ib].dm);
+            const uint vui = data_a_packed32[ib].qs[iqs];
+            const vec4 v0 = vec4(unpack8(vui & 0x0F0F0F0F)) * dm.x + dm.y;
+            const vec4 v1 = vec4(unpack8((vui >> 4) & 0x0F0F0F0F)) * dm.x + dm.y;
 
             buf_a[buf_idx     ] = FLOAT_TYPE_VEC2(v0.xy);
             buf_a[buf_idx + 1 ] = FLOAT_TYPE_VEC2(v0.zw);
@@ -80,7 +79,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             buf_a[buf_idx + 9 ] = FLOAT_TYPE_VEC2(v1.zw);
 #elif defined(DATA_A_Q5_0)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
             const uint ib = idx / 8;
             const uint iqs = idx & 0x07;
@@ -97,22 +96,26 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v.yw);
 #elif defined(DATA_A_Q5_1)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
-            const uint ib = idx / 8;
-            const uint iqs = idx & 0x07;
+            const uint ib = idx / 4;
+            const uint iqs = idx & 0x03;
 
-            const float d = float(data_a_packed16[ib].d);
-            const float m = float(data_a_packed16[ib].m);
-            const uint uint_qh = data_a_packed16[ib].qh;
-            const ivec2 qh0 = ivec2(((uint_qh >> 2*iqs) << 4) & 0x10, (uint_qh >> (2*iqs + 12)) & 0x10);
-            const ivec2 qh1 = ivec2(((uint_qh >> (2*iqs + 1)) << 4) & 0x10, (uint_qh >> (2*iqs + 13)) & 0x10);
+            const vec2 dm = vec2(data_a_packed32[ib].dm);
+            const uint uint_qh = data_a_packed32[ib].qh;
+            const uvec2 qh0 = uvec2(((uint_qh >> 4*iqs) << 4) & 0x10, (uint_qh >> (4*iqs + 12)) & 0x10);
+            const uvec2 qh1 = uvec2(((uint_qh >> (4*iqs + 1)) << 4) & 0x10, (uint_qh >> (4*iqs + 13)) & 0x10);
+            const uvec2 qh2 = uvec2(((uint_qh >> (4*iqs + 2)) << 4) & 0x10, (uint_qh >> (4*iqs + 14)) & 0x10);
+            const uvec2 qh3 = uvec2(((uint_qh >> (4*iqs + 3)) << 4) & 0x10, (uint_qh >> (4*iqs + 15)) & 0x10);
 
-            const uint vui = uint(data_a_packed16[ib].qs[iqs]);
-            const vec4 v = vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, (vui >> 12) | qh1.y) * d + m;
+            const uint vui = data_a_packed32[ib].qs[iqs];
+            const vec4 v0 = vec4((vui & 0xF) | qh0.x, ((vui >> 4) & 0xF) | qh0.y, ((vui >> 8) & 0xF) | qh1.x, ((vui >> 12) & 0xF) | qh1.y) * dm.x + dm.y;
+            const vec4 v1 = vec4(((vui >> 16) & 0xF) | qh2.x, ((vui >> 20) & 0xF) | qh2.y, ((vui >> 24) & 0xF) | qh3.x, ((vui >> 28) & 0xF) | qh3.y) * dm.x + dm.y;
 
-            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xz);
-            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v.yw);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v0.xz);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(v1.xz);
+            buf_a[buf_idx + 8] = FLOAT_TYPE_VEC2(v0.yw);
+            buf_a[buf_idx + 9] = FLOAT_TYPE_VEC2(v1.yw);
 #elif defined(DATA_A_Q8_0)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -131,20 +134,21 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
 
-            const uint ib = idx / 128;                         // 2 values per idx
-            const uint iqs = idx % 128;                        // 0..127
+            const uint ib = idx / 64;                          // 4 values per idx
+            const uint iqs = (idx % 64) * 2;                   // 0,2,4..126
 
             const uint qsi = (iqs / 64) * 16 + (iqs % 16);     // 0..15
             const uint scalesi = iqs / 8;                      // 0..15
             const uint qsshift = ((iqs % 64) / 16) * 2;        // 0,2,4,6
 
-            const uvec2 qs = uvec2(unpack8(data_a_packed16[ib].qs[qsi]));
+            const vec4 qs = vec4(unpack8((data_a_packed32[ib].qs[qsi / 2] >> qsshift) & 0x03030303));
             const uint scales = data_a[ib].scales[scalesi];
             const vec2 dm = vec2(data_a[ib].dm);
 
-            const vec2 v = dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
+            const vec4 v = dm.x * float(scales & 0xF) * qs - dm.y * float(scales >> 4);
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xy);
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(v.zw);
 #elif defined(DATA_A_Q3_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -173,8 +177,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
 
-            const uint ib = idx / 128;                 // 2 values per idx
-            const uint iqs = idx % 128;                // 0..127
+            const uint ib = idx / 64;                  // 4 values per idx
+            const uint iqs = (idx % 64) * 2;           // 0,2,4..126
 
             const uint n = iqs / 32;                   // 0,1,2,3
             const uint b = (iqs % 32) / 16;            // 0,1
@@ -200,16 +204,16 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
-            const vec2 q = vec2(unpack8((uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F).xy);
+            const vec4 q = vec4(unpack8((data_a_packed32[ib].qs[qsi / 4] >> (b * 4)) & 0x0F0F0F0F));
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
-                                             fma(d, q.y, m));
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(fma(d, q.x, m), fma(d, q.y, m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(fma(d, q.z, m), fma(d, q.w, m));
 #elif defined(DATA_A_Q5_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
 
-            const uint ib = idx / 128;                 // 2 values per idx
-            const uint iqs = idx % 128;                // 0..127
+            const uint ib = idx / 64;                  // 4 values per idx
+            const uint iqs = (idx % 64) * 2;           // 0,2,4..126
 
             const uint n = iqs / 32;                   // 0,1,2,3
             const uint b = (iqs % 32) / 16;            // 0,1
@@ -236,12 +240,12 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             const float d = loadd.x * sc;
             const float m = -loadd.y * mbyte;
 
-            const uint qs = (uint(data_a_packed16[ib].qs[qsi / 2]) >> (b * 4)) & 0x0F0F;
-            const uint qh = ((uint(data_a_packed16[ib].qh[qhi / 2]) >> (iqs / 16)) & 0x0101) << 4;
-            const vec2 q = vec2(unpack8(qs | qh).xy);
+            const uint qs = (data_a_packed32[ib].qs[qsi / 4] >> (b * 4)) & 0x0F0F0F0F;
+            const uint qh = ((data_a_packed32[ib].qh[qhi / 4] >> (iqs / 16)) & 0x01010101) << 4;
+            const vec4 q = vec4(unpack8(qs | qh));
 
-            buf_a[buf_idx] = FLOAT_TYPE_VEC2(fma(d, q.x, m),
-                                             fma(d, q.y, m));
+            buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(fma(d, q.x, m), fma(d, q.y, m));
+            buf_a[buf_idx + 1] = FLOAT_TYPE_VEC2(fma(d, q.z, m), fma(d, q.w, m));
 #elif defined(DATA_A_Q6_K)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
             const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 2;
@@ -455,7 +459,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
             buf_a[buf_idx    ] = FLOAT_TYPE_VEC2(v.xy);
 #elif defined(DATA_A_IQ4_NL)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
             const uint ib = idx / 8;
             const uint iqs = idx & 0x07;
@@ -469,7 +473,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
                                                      kvalues_iq4nl[vui >> 12]);
 #elif defined(DATA_A_MXFP4)
             const uint idx = pos_a + col * p.stride_a / LOAD_VEC_A + row;
-            const uint buf_idx = col * SHMEM_STRIDE + row;
+            const uint buf_idx = col * SHMEM_STRIDE + row * LOAD_VEC_A / 4;
 
             const uint ib = idx / 8;
             const uint iqs = (idx & 0x07) * 2;

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

@@ -189,13 +189,13 @@ void main() {
     const uint end_k = min(p.K, (ik + 1) * p.k_split);
 #endif
 
-    uint pos_a_ib = (
+    uint pos_a_ib =
 #ifdef MUL_MAT_ID
-        expert_idx * p.batch_stride_a +
+        expert_idx * (p.batch_stride_a / BK) +
 #else
-        batch_idx_a * p.batch_stride_a +
+        batch_idx_a * (p.batch_stride_a / BK) +
 #endif
-        ir * BM * p.stride_a + start_k) / BK;
+        (ir * BM * p.stride_a + start_k) / BK;
 #ifdef MUL_MAT_ID
     uint pos_b_ib = 0;
 #else

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

@@ -1,6 +1,7 @@
 #version 450
 
 #extension GL_EXT_control_flow_attributes : require
+#extension GL_KHR_shader_subgroup_basic : enable
 #if USE_SUBGROUP_ADD
 #extension GL_KHR_shader_subgroup_arithmetic : enable
 #endif
@@ -9,7 +10,8 @@
 
 layout(constant_id = 0) const uint D_STATE = 128;
 layout(constant_id = 1) const uint SUBGROUP_SIZE = 32;
-layout(constant_id = 2) const uint SPLIT_H = 16;
+
+const uint32_t c_factor = D_STATE / SUBGROUP_SIZE;
 
 layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
@@ -41,22 +43,28 @@ float softplus(float x) {
     }
 }
 
-shared float stateC[SPLIT_H * D_STATE];
+#if !USE_SUBGROUP_ADD
+shared float temp[D_STATE];
+#endif
 
 void main() {
-    const uint tid = gl_LocalInvocationID.x;
-    const uint head_idx = (gl_WorkGroupID.x * SPLIT_H) / d_head;
-    const uint head_off = ((gl_WorkGroupID.x * SPLIT_H) % d_head) * 4;
-    const uint seq_idx = gl_WorkGroupID.y;
+    const uint subgroup = gl_SubgroupID;
+    const uint lane     = gl_SubgroupInvocationID;
+    const uint tid      = gl_SubgroupID * SUBGROUP_SIZE + lane;
+    const uint subgroup_idx = gl_WorkGroupID.x  * c_factor + subgroup;
+
+    const uint head_idx =  subgroup_idx / d_head;
+    const uint head_off = (subgroup_idx % d_head) * 4;
+    const uint seq_idx  = gl_WorkGroupID.y;
 
     const uint group_off = (head_idx / (n_head / n_group)) * D_STATE * 4;
     const uint s0_base_idx = (uint(ids[seq_idx]) * nb03 + head_idx * nb02 + head_off * D_STATE) / 4;
-    const uint x_base_idx = (seq_idx * nb13 + gl_WorkGroupID.x * SPLIT_H * 4) / 4;
+    const uint x_base_idx = (seq_idx * nb13 + subgroup_idx * 4) / 4;
     const uint dt_base_idx = (seq_idx * nb22 + head_idx * 4) / 4;
     const uint A_base_idx = (head_idx * nb31) / 4;
     const uint B_base_idx = (seq_idx * nb43 + group_off) / 4;
     const uint C_base_idx = (seq_idx * nb53 + group_off) / 4;
-    const uint y_base_idx = seq_idx * n_tok * n_head * d_head + gl_WorkGroupID.x * SPLIT_H;
+    const uint y_base_idx = seq_idx * n_tok * n_head * d_head + subgroup_idx;
     const uint s_base_idx = (s_off + seq_idx * nb03 + head_idx * nb02 + head_off * D_STATE) / 4;
 
     const uint stride_x = nb12 / 4;
@@ -65,76 +73,52 @@ void main() {
     const uint stride_C = nb52 / 4;
     const uint stride_y = n_head * d_head;
 
-    float state[SPLIT_H];
-    [[unroll]] for (uint j = 0; j < SPLIT_H; j++) {
-        state[j] = s0[s0_base_idx + j * D_STATE + tid];
+    float state[c_factor];
+
+    [[unroll]] for (uint j = 0; j < c_factor; j++) {
+        state[j] = s0[s0_base_idx + SUBGROUP_SIZE * j + lane];
     }
 
+    float a = A[A_base_idx];
+
     for (uint i = 0; i < n_tok; i++) {
-        const float dt_soft_plus = softplus(dt[dt_base_idx + i * stride_dt]);
+        float dt_soft_plus = softplus(dt[dt_base_idx + i * stride_dt]);
 
-        const float dA = exp(dt_soft_plus * A[A_base_idx]);
-
-        const float B_val = B[B_base_idx + i * stride_B + tid];
-        const float C_val = C[C_base_idx + i * stride_C + tid];
-
-        [[unroll]] for (uint j = 0; j < SPLIT_H; j++) {
-            const float x_dt = x[x_base_idx + i * stride_x + j] * dt_soft_plus;
+        float state_sum = 0.0f;
 
+        const float dA   = exp(dt_soft_plus * a);
+        const float x_dt = x[x_base_idx + i * stride_x] * dt_soft_plus;
+        [[unroll]] for (uint j = 0; j < c_factor; j++) {
+            float B_val = B[B_base_idx + i * stride_B + SUBGROUP_SIZE * j + lane];
+            float C_val = C[C_base_idx + i * stride_C + SUBGROUP_SIZE * j + lane];
             state[j] = (state[j] * dA) + (B_val * x_dt);
-
-            stateC[j * D_STATE + tid] = state[j] * C_val;
+            state_sum += state[j] * C_val;
         }
 
+#if USE_SUBGROUP_ADD
+        state_sum = subgroupAdd(state_sum);
+#else
+        temp[tid] = state_sum;
         barrier();
-        [[unroll]]
-        for (uint w = D_STATE / 2; w >= SUBGROUP_SIZE; w >>= 1) {
-            [[unroll]] for (uint j = 0; j < (w * SPLIT_H + D_STATE - 1) / D_STATE; j++) {
-                const uint k = (tid % w) + (D_STATE * (tid / w)) + j * D_STATE * (D_STATE / w);
-                if (k < SPLIT_H * D_STATE && (k + w) < SPLIT_H * D_STATE) {
-                    stateC[k] += stateC[k + w];
-                }
+        [[unroll]] for (uint s = SUBGROUP_SIZE / 2; s > 0; s >>= 1) {
+            if (lane < s) {
+                temp[tid] += temp[tid + s];
             }
             barrier();
         }
-
-        [[unroll]] for (uint j = 0; j < max(1, SPLIT_H / (D_STATE / SUBGROUP_SIZE)); j++) {
-            const uint idx = (tid % SUBGROUP_SIZE) +
-                            D_STATE * (tid / SUBGROUP_SIZE) +
-                            j * D_STATE * (D_STATE / SUBGROUP_SIZE);
-            const uint max_idx = SUBGROUP_SIZE - 1 +
-                            D_STATE * ((D_STATE - 1) / SUBGROUP_SIZE) +
-                            j * D_STATE * (D_STATE / SUBGROUP_SIZE);
-
-            if (idx < SPLIT_H * D_STATE ||
-                max_idx < SPLIT_H * D_STATE) {
-                float sc;
-#if USE_SUBGROUP_ADD
-                sc = stateC[idx];
-                sc = subgroupAdd(sc);
-#else
-                [[unroll]] for (uint offset = SUBGROUP_SIZE / 2; offset > 0; offset >>= 1) {
-                    if (idx + offset < SPLIT_H * D_STATE) {
-                        stateC[idx] += stateC[idx + offset];
-                    }
-                    barrier();
-                }
-                if (tid % SUBGROUP_SIZE == 0) {
-                    sc = stateC[idx];
-                }
+        // get the value from lane 0
+        state_sum = temp[subgroup * SUBGROUP_SIZE];
+        barrier();
 #endif
 
-                if (tid % SUBGROUP_SIZE == 0) {
-                    const uint k = tid / SUBGROUP_SIZE + j * (D_STATE / SUBGROUP_SIZE);
-                    d[y_base_idx + i * stride_y + k] = sc;
-                }
-            }
+        if (lane == 0) {
+            d[y_base_idx + i * stride_y] = state_sum;
         }
-
-        barrier();
     }
 
-    [[unroll]] for (uint j = 0; j < SPLIT_H; j++) {
-        d[s_base_idx + j * D_STATE + tid] = state[j];
+    // write back the state
+    [[unroll]]
+    for (int j = 0; j < c_factor; j++) {
+        d[s_base_idx + SUBGROUP_SIZE * j + lane] = state[j];
     }
 }

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

@@ -552,9 +552,9 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
 
     for (const auto& tname : type_names) {
         std::string load_vec_quant = "2";
-        if ((tname == "q4_0") || (tname == "q4_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
+        if ((tname == "q4_0") || (tname == "q4_1") || (tname == "q5_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
             load_vec_quant = "8";
-        else if ((tname == "q5_0") || (tname == "q5_1") || (tname == "q8_0") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
+        else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
             load_vec_quant = "4";
 
         if (tname == "bf16") {