ggml: aarch64: SVE kernels for q8_0_q8_0, q4_0_q8_0 vector dot (#7433 )

* Add SVE support for q4_0_q8_0 q8_0_q8_0 * remove ifdef
android : module (#7502 )
2026-02-05 13:53:23 +02:00 · 2024-05-25 11:42:31 +03:00 · 2024-05-25 11:11:33 +03:00 · 2024-05-25 13:30:59 +10:00 · 2024-05-25 11:14:42 +10:00 · 2024-05-25 11:11:48 +10:00
167 changed files with 70467 additions and 47933 deletions
--- a/.devops/nix/package.nix
+++ b/.devops/nix/package.nix
@@ -214,7 +214,6 @@ effectiveStdenv.mkDerivation (
        (cmakeBool "LLAMA_CUDA" useCuda)
        (cmakeBool "LLAMA_HIPBLAS" useRocm)
        (cmakeBool "LLAMA_METAL" useMetalKit)
-        (cmakeBool "LLAMA_MPI" useMpi)
        (cmakeBool "LLAMA_VULKAN" useVulkan)
        (cmakeBool "LLAMA_STATIC" enableStatic)
      ]
@@ -227,20 +226,20 @@ effectiveStdenv.mkDerivation (
        )
      ]
      ++ optionals useRocm [
-        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
-        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
-
-        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-        # and select the line that matches the current nixpkgs version of rocBLAS.
-        # Should likely use `rocmPackages.clr.gpuTargets`.
-        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+        (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
+        (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
      ]
      ++ optionals useMetalKit [
        (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
        (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
      ];

+    # Environment variables needed for ROCm
+    env = optionals useRocm {
+      ROCM_PATH = "${rocmPackages.clr}";
+      HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
+    };
+
    # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
    # if they haven't been added yet.
    postInstall = ''
--- a/.github/labeler.yml
+++ b/.github/labeler.yml
@@ -0,0 +1,78 @@
+# https://github.com/actions/labeler
+
+SYCL:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-sycl.h
+            - ggml-sycl.cpp
+            - README-sycl.md
+Nvidia GPU:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml-cuda/**
+Vulkan:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml_vk_generate_shaders.py
+            - ggml-vulkan*
+documentation:
+    - changed-files:
+        - any-glob-to-any-file:
+            - docs/**
+            - media/**
+testing:
+    - changed-files:
+        - any-glob-to-any-file:
+            - tests/**
+build:
+    - changed-files:
+        - any-glob-to-any-file:
+            - cmake/**
+            - CMakeLists.txt
+            - CMakePresets.json
+            - codecov.yml
+examples:
+    - changed-files:
+        - any-glob-to-any-file: examples/**
+devops:
+    - changed-files:
+        - any-glob-to-any-file:
+            - .devops/**
+            - .github/**
+            - ci/**
+python:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.py"
+            - requirements/**
+            - gguf-py/**
+            - .flake8
+script:
+    - changed-files:
+        - any-glob-to-any-file:
+            - scripts/**
+android:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/llama.android/**
+server:
+    - changed-files:
+        - any-glob-to-any-file:
+            - examples/server/**
+ggml:
+    - changed-files:
+        - any-glob-to-any-file:
+            - ggml.c
+            - ggml.h
+            - ggml-*.c
+            - ggml-*.h
+            - ggml-cuda/**
+nix:
+    - changed-files:
+        - any-glob-to-any-file:
+            - "**/*.nix"
+            - .github/workflows/nix-*.yml
+            - .devops/nix/nixpkgs-instances.nix
+embedding:
+    - changed-files:
+        - any-glob-to-any-file: examples/embedding/
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -271,49 +271,15 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-ubuntu-x64.zip
          name: llama-bin-ubuntu-x64.zip

-#  ubuntu-latest-cmake-sanitizer:
-#    runs-on: ubuntu-latest
-#
-#    continue-on-error: true
-#
-#    strategy:
-#      matrix:
-#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-#        build_type: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v4
-#
-#      - name: Dependencies
-#        id: depends
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          mkdir build
-#          cd build
-#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-#
-#      - name: Test
-#        id: cmake_test
-#        run: |
-#          cd build
-#          ctest -L main --verbose --timeout 900
-
-  ubuntu-latest-cmake-mpi:
+  ubuntu-latest-cmake-sanitizer:
    runs-on: ubuntu-latest

    continue-on-error: true

    strategy:
      matrix:
-        mpi_library: [mpich, libopenmpi-dev]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]

    steps:
      - name: Clone
@@ -324,14 +290,44 @@ jobs:
        id: depends
        run: |
          sudo apt-get update
-          sudo apt-get install build-essential ${{ matrix.mpi_library }}
+          sudo apt-get install build-essential

      - name: Build
        id: cmake_build
        run: |
          mkdir build
          cd build
-          cmake -DLLAMA_MPI=ON ..
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+  ubuntu-latest-cmake-rpc:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake -DLLAMA_RPC=ON ..
          cmake --build . --config Release -j $(nproc)

      - name: Test
@@ -362,6 +358,33 @@ jobs:
          cmake -DLLAMA_VULKAN=ON ..
          cmake --build . --config Release -j $(nproc)

+  ubuntu-22-cmake-hip:
+    runs-on: ubuntu-22.04
+    container: rocm/dev-ubuntu-22.04:6.0.2
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y build-essential git cmake rocblas-dev hipblas-dev
+
+      - name: Build with native CMake HIP support
+        id: cmake_build
+        run: |
+          cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release -j $(nproc)
+
+      - name: Build with legacy HIP support
+        id: cmake_build_legacy_hip
+        run: |
+          cmake -B build2 -S . -DCMAKE_C_COMPILER=hipcc -DCMAKE_CXX_COMPILER=hipcc -DLLAMA_HIPBLAS=ON
+          cmake --build build2 --config Release -j $(nproc)
+
  ubuntu-22-cmake-sycl:
    runs-on: ubuntu-22.04

@@ -663,24 +686,28 @@ jobs:
    strategy:
      matrix:
        include:
-          - build: 'noavx'
+          - build: 'rpc-x64'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_RPC=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'noavx-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX=OFF -DLLAMA_AVX2=OFF -DLLAMA_FMA=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx2'
+          - build: 'avx2-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx'
+          - build: 'avx-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx512'
+          - build: 'avx512-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_AVX512=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'clblast'
+          - build: 'clblast-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
-          - build: 'openblas'
+          - build: 'openblas-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
-          - build: 'kompute'
+          - build: 'kompute-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'vulkan'
+          - build: 'vulkan-x64'
            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'arm64'
-            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'llvm-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'msvc-arm64'
+            defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-msvc.cmake -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'

    steps:
      - name: Clone
@@ -691,13 +718,13 @@ jobs:

      - name: Clone Kompute submodule
        id: clone_kompute
-        if: ${{ matrix.build == 'kompute' }}
+        if: ${{ matrix.build == 'kompute-x64' }}
        run: |
          git submodule update --init kompute

      - name: Download OpenCL SDK
        id: get_opencl
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
        run: |
          curl.exe -o $env:RUNNER_TEMP/opencl.zip -L "https://github.com/KhronosGroup/OpenCL-SDK/releases/download/v${env:OPENCL_VERSION}/OpenCL-SDK-v${env:OPENCL_VERSION}-Win-x64.zip"
          mkdir $env:RUNNER_TEMP/opencl
@@ -705,7 +732,7 @@ jobs:

      - name: Download CLBlast
        id: get_clblast
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
        run: |
          curl.exe -o $env:RUNNER_TEMP/clblast.7z -L "https://github.com/CNugteren/CLBlast/releases/download/${env:CLBLAST_VERSION}/CLBlast-${env:CLBLAST_VERSION}-windows-x64.7z"
          curl.exe -o $env:RUNNER_TEMP/CLBlast.LICENSE.txt -L "https://github.com/CNugteren/CLBlast/raw/${env:CLBLAST_VERSION}/LICENSE"
@@ -718,7 +745,7 @@ jobs:

      - name: Download OpenBLAS
        id: get_openblas
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
        run: |
          curl.exe -o $env:RUNNER_TEMP/openblas.zip -L "https://github.com/xianyi/OpenBLAS/releases/download/v${env:OPENBLAS_VERSION}/OpenBLAS-${env:OPENBLAS_VERSION}-x64.zip"
          curl.exe -o $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt -L "https://github.com/xianyi/OpenBLAS/raw/v${env:OPENBLAS_VERSION}/LICENSE"
@@ -731,38 +758,41 @@ jobs:

      - name: Install Vulkan SDK
        id: get_vulkan
-        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
+        if: ${{ matrix.build == 'kompute-x64' || matrix.build == 'vulkan-x64' }}
        run: |
          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"

+      - name: Install Ninja
+        id: install_ninja
+        run: |
+          choco install ninja
+
      - name: Build
        id: cmake_build
        run: |
-          mkdir build
-          cd build
-          cmake .. ${{ matrix.defines }}
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+          cmake -S . -B build ${{ matrix.defines }}
+          cmake --build build --config Release -j ${env:NUMBER_OF_PROCESSORS}

      - name: Add clblast.dll
        id: add_clblast_dll
-        if: ${{ matrix.build == 'clblast' }}
+        if: ${{ matrix.build == 'clblast-x64' }}
        run: |
          cp $env:RUNNER_TEMP/clblast/lib/clblast.dll ./build/bin/Release
          cp $env:RUNNER_TEMP/CLBlast.LICENSE.txt ./build/bin/Release/CLBlast-${env:CLBLAST_VERSION}.txt

      - name: Add libopenblas.dll
        id: add_libopenblas_dll
-        if: ${{ matrix.build == 'openblas' }}
+        if: ${{ matrix.build == 'openblas-x64' }}
        run: |
          cp $env:RUNNER_TEMP/openblas/bin/libopenblas.dll ./build/bin/Release/openblas.dll
          cp $env:RUNNER_TEMP/OpenBLAS.LICENSE.txt ./build/bin/Release/OpenBLAS-${env:OPENBLAS_VERSION}.txt

      - name: Check AVX512F support
        id: check_avx512f
-        if: ${{ matrix.build == 'avx512' }}
+        if: ${{ matrix.build == 'avx512-x64' }}
        continue-on-error: true
        run: |
          cd build
@@ -776,14 +806,14 @@ jobs:
      - name: Test
        id: cmake_test
        # not all machines have native AVX-512
-        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'msvc-arm64' && matrix.build != 'llvm-arm64' && matrix.build != 'clblast-x64' && matrix.build != 'kompute-x64' && matrix.build != 'vulkan-x64' && (matrix.build != 'avx512-x64' || env.HAS_AVX512F == '1') }}
        run: |
          cd build
          ctest -L main -C Release --verbose --timeout 900

      - name: Test (Intel SDE)
        id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
+        if: ${{ matrix.build == 'avx512-x64' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
        run: |
          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
          # for some weird reason windows tar doesn't like sde tar.xz
@@ -811,14 +841,14 @@ jobs:
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        run: |
          Copy-Item LICENSE .\build\bin\Release\llama.cpp.txt
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip .\build\bin\Release\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip .\build\bin\Release\*

      - name: Upload artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        uses: actions/upload-artifact@v4
        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
-          name: llama-bin-win-${{ matrix.build }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}.zip
+          name: llama-bin-win-${{ matrix.build }}.zip

  windows-latest-cmake-cuda:
    runs-on: windows-latest
@@ -898,9 +928,9 @@ jobs:
        shell: bash

    env:
-      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/7dff44ba-e3af-4448-841c-0d616c8da6e7/w_BaseKit_p_2024.1.0.595_offline.exe
      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
-
+      ONEAPI_ROOT: "C:/Program Files (x86)/Intel/oneAPI"
    steps:
      - name: Clone
        id: checkout
@@ -932,6 +962,17 @@ jobs:
        id: pack_artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        run: |
+          echo "cp oneAPI running time dll files in ${{ env.ONEAPI_ROOT }} to ./build/bin"
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_sycl_blas.4.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_core.2.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/mkl/latest/bin/mkl_tbb_thread.2.dll" ./build/bin
+
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_win_proxy_loader.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/pi_level_zero.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          echo "cp oneAPI running time dll files to ./build/bin done"
          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*

      - name: Upload artifacts
@@ -941,6 +982,37 @@ jobs:
          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
          name: llama-bin-win-sycl-x64.zip

+  windows-latest-cmake-hip:
+    runs-on: windows-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DLLAMA_HIPBLAS=ON
+          cmake --build build --config Release
+
  ios-xcode-build:
    runs-on: macos-latest

--- a/.github/workflows/docker.yml
+++ b/.github/workflows/docker.yml
@@ -42,8 +42,9 @@ jobs:
          - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
          - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
+          # TODO: Disabled due to build issues https://github.com/ggerganov/llama.cpp/issues/7507
+          #- { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
+          #- { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
    steps:
      - name: Check out the repo
        uses: actions/checkout@v4
--- a/.github/workflows/labeler.yml
+++ b/.github/workflows/labeler.yml
@@ -0,0 +1,17 @@
+name: "Pull Request Labeler"
+on:
+- pull_request_target
+
+jobs:
+  labeler:
+    permissions:
+      contents: read
+      pull-requests: write
+    runs-on: ubuntu-latest
+    steps:
+    - uses: actions/checkout@v4
+      with:
+        repository: "ggerganov/llama.cpp"
+    - uses: actions/labeler@v5
+      with:
+        configuration-path: '.github/labeler.yml'
--- a/.github/workflows/server.yml
+++ b/.github/workflows/server.yml
@@ -32,10 +32,8 @@ jobs:

    strategy:
      matrix:
-        # TODO: temporary disabled due to linux kernel issues
-        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        sanitizer: [UNDEFINED]
-        build_type: [Debug]
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [RelWithDebInfo]
        include:
          - build_type: Release
            sanitizer: ""
@@ -102,10 +100,8 @@ jobs:
              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
          cmake --build build --config ${{ matrix.build_type }} -j $(nproc) --target server

-
      - name: Tests
        id: server_integration_tests
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
        run: |
          cd examples/server/tests
          PORT=8888 ./tests.sh
--- a/.github/workflows/zig-build.yml
+++ b/.github/workflows/zig-build.yml
@@ -1,29 +0,0 @@
-name: Zig CI
-
-on:
-  pull_request:
-  push:
-    branches:
-      - master
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  build:
-    strategy:
-      fail-fast: false
-      matrix:
-        runs-on: [ubuntu-latest, macos-latest, windows-latest]
-    runs-on: ${{ matrix.runs-on }}
-    steps:
-      - uses: actions/checkout@v4
-        with:
-          submodules: recursive
-          fetch-depth: 0
-      - uses: goto-bus-stop/setup-zig@v2
-        with:
-          version: 0.11.0
-      - name: Build Summary
-        run: zig build --summary all -freference-trace
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.14) # for add_link_options and implicit target directories.
 project("llama.cpp" C CXX)
 include(CheckIncludeFileCXX)

@@ -72,11 +72,13 @@ else()
    set(INS_ENB ON)
 endif()

+option(LLAMA_SVE                             "llama: enable SVE"                                OFF)
 option(LLAMA_AVX                             "llama: enable AVX"                                ${INS_ENB})
 option(LLAMA_AVX2                            "llama: enable AVX2"                               ${INS_ENB})
 option(LLAMA_AVX512                          "llama: enable AVX512"                             OFF)
 option(LLAMA_AVX512_VBMI                     "llama: enable AVX512-VBMI"                        OFF)
 option(LLAMA_AVX512_VNNI                     "llama: enable AVX512-VNNI"                        OFF)
+option(LLAMA_AVX512_BF16                     "llama: enable AVX512-BF16"                        OFF)
 option(LLAMA_FMA                             "llama: enable FMA"                                ${INS_ENB})
 # in MSVC F16C is implied with AVX2/AVX512
 if (NOT MSVC)
@@ -122,8 +124,7 @@ set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                             "llama: metal minimum macOS version")
 set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
-option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
-option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
+option(LLAMA_RPC                             "llama: use RPC"                                   OFF)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
 set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
@@ -133,6 +134,8 @@ set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeli
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
+option(LLAMA_LASX                            "llama: enable lasx"                               ON)
+option(LLAMA_LSX                             "llama: enable lsx"                                ON)

 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
@@ -296,7 +299,7 @@ if (LLAMA_BLAS)
    if (LLAMA_STATIC)
        set(BLA_STATIC ON)
    endif()
-    if ($(CMAKE_VERSION) VERSION_GREATER_EQUAL 3.22)
+    if (CMAKE_VERSION VERSION_GREATER_EQUAL 3.22)
        set(BLA_SIZEOF_INTEGER 8)
    endif()

@@ -381,10 +384,6 @@ if (LLAMA_LLAMAFILE)
    set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
 endif()

-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
-endif()
-
 if (LLAMA_CUBLAS)
    message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
    set(LLAMA_CUDA ON)
@@ -431,7 +430,7 @@ if (LLAMA_CUDA)

        if (LLAMA_STATIC)
            if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                # As of 12.3.1 CUDA Toolkit for Windows does not offer a static cublas library
                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
            else ()
                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
@@ -465,33 +464,15 @@ if (LLAMA_CUDA)
    endif()
 endif()

-if (LLAMA_MPI)
-    cmake_minimum_required(VERSION 3.10)
-    find_package(MPI)
-    if (MPI_C_FOUND)
-        message(STATUS "MPI found")
+if (LLAMA_RPC)
+    add_compile_definitions(GGML_USE_RPC)

-        set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c)
-
-        add_compile_definitions(GGML_USE_MPI)
-        add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-
-        if (NOT MSVC)
-            add_compile_options(-Wno-cast-qual)
-        endif()
-
-        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
-
-        # Even if you're only using the C header, C++ programs may bring in MPI
-        # C++ functions, so more linkage is needed
-        if (MPI_CXX_FOUND)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}     ${MPI_CXX_LIBRARIES})
-        endif()
-    else()
-        message(WARNING "MPI not found")
+    if (WIN32)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ws2_32)
    endif()
+
+    set(GGML_HEADERS_RPC ggml-rpc.h)
+    set(GGML_SOURCES_RPC ggml-rpc.cpp)
 endif()

 if (LLAMA_CLBLAST)
@@ -520,6 +501,12 @@ if (LLAMA_VULKAN)

        add_compile_definitions(GGML_USE_VULKAN)

+        # Workaround to the "can't dereference invalidated vector iterator" bug in clang-cl debug build
+        # Posssibly relevant: https://stackoverflow.com/questions/74748276/visual-studio-no-displays-the-correct-length-of-stdvector
+        if (MSVC AND CMAKE_CXX_COMPILER_ID STREQUAL "Clang")
+            add_compile_definitions(_ITERATOR_DEBUG_LEVEL=0)
+        endif()
+
        if (LLAMA_VULKAN_CHECK_RESULTS)
            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
        endif()
@@ -543,16 +530,37 @@ if (LLAMA_VULKAN)
 endif()

 if (LLAMA_HIPBLAS)
-    list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
+    if ($ENV{ROCM_PATH})
+        set(ROCM_PATH $ENV{ROCM_PATH})
+    else()
+        set(ROCM_PATH /opt/rocm)
+    endif()
+    list(APPEND CMAKE_PREFIX_PATH ${ROCM_PATH})

-    if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
+    # CMake on Windows doesn't support the HIP language yet
+    if(WIN32)
+        set(CXX_IS_HIPCC TRUE)
+    else()
+        string(REGEX MATCH "hipcc(\.bat)?$" CXX_IS_HIPCC "${CMAKE_CXX_COMPILER}")
    endif()

-    if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
-        message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
-    endif()
+    if(CXX_IS_HIPCC)
+        if(LINUX)
+            if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
+                message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
+            endif()

+            message(WARNING "Setting hipcc as the C++ compiler is legacy behavior."
+                    " Prefer setting the HIP compiler directly. See README for details.")
+        endif()
+    else()
+        # Forward AMDGPU_TARGETS to CMAKE_HIP_ARCHITECTURES.
+        if(AMDGPU_TARGETS AND NOT CMAKE_HIP_ARCHITECTURES)
+            set(CMAKE_HIP_ARCHITECTURES ${AMDGPU_TARGETS})
+        endif()
+        cmake_minimum_required(VERSION 3.21)
+        enable_language(HIP)
+    endif()
    find_package(hip     REQUIRED)
    find_package(hipblas REQUIRED)
    find_package(rocblas REQUIRED)
@@ -586,13 +594,18 @@ if (LLAMA_HIPBLAS)
    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})

-    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+    if (CXX_IS_HIPCC)
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device)
+    else()
+        set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE HIP)
+    endif()

    if (LLAMA_STATIC)
        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
    endif()

-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()

 if (LLAMA_SYCL)
@@ -995,6 +1008,11 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
        if (GGML_COMPILER_SUPPORT_DOTPROD)
            add_compile_definitions(__ARM_FEATURE_DOTPROD)
        endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vmlaq_f32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_MATMUL_INT8)
+        if (GGML_COMPILER_SUPPORT_MATMUL_INT8)
+            add_compile_definitions(__ARM_FEATURE_MATMUL_INT8)
+        endif ()
+
        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
@@ -1023,6 +1041,9 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
            # Raspberry Pi 3, 4, Zero 2 (32-bit)
            list(APPEND ARCH_FLAGS -mno-unaligned-access)
        endif()
+        if (LLAMA_SVE)
+            list(APPEND ARCH_FLAGS -march=armv8.6-a+sve)
+        endif()
    endif()
 elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LWR MATCHES "^(x86_64|i686|amd64|x64|win32)$" OR
        (NOT CMAKE_OSX_ARCHITECTURES AND NOT CMAKE_GENERATOR_PLATFORM_LWR AND
@@ -1047,6 +1068,10 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
            endif()
+            if (LLAMA_AVX512_BF16)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
+                add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+            endif()
        elseif (LLAMA_AVX2)
            list(APPEND ARCH_FLAGS /arch:AVX2)
        elseif (LLAMA_AVX)
@@ -1078,6 +1103,9 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
        if (LLAMA_AVX512_VNNI)
            list(APPEND ARCH_FLAGS -mavx512vnni)
        endif()
+        if (LLAMA_AVX512_BF16)
+            list(APPEND ARCH_FLAGS -mavx512bf16)
+        endif()
    endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
    message(STATUS "PowerPC detected")
@@ -1087,6 +1115,17 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
    endif()
+elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "loongarch64")
+    message(STATUS "loongarch64 detected")
+
+    list(APPEND ARCH_FLAGS -march=loongarch64)
+    if (LLAMA_LASX)
+        list(APPEND ARCH_FLAGS -mlasx)
+    endif()
+    if (LLAMA_LSX)
+        list(APPEND ARCH_FLAGS -mlsx)
+    endif()
+
 else()
    message(STATUS "Unknown architecture")
 endif()
@@ -1175,7 +1214,7 @@ add_library(ggml OBJECT
            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_RPC}       ${GGML_HEADERS_RPC}
            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
@@ -1262,7 +1301,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake

 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
-        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
+        "${GGML_HEADERS_METAL}" "${GGML_HEADERS_EXTRA}")

 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
 install(TARGETS ggml PUBLIC_HEADER)
@@ -1281,17 +1320,6 @@ install(
        WORLD_READ
        WORLD_EXECUTE
    DESTINATION ${CMAKE_INSTALL_BINDIR})
-install(
-    FILES convert-lora-to-ggml.py
-    PERMISSIONS
-        OWNER_READ
-        OWNER_WRITE
-        OWNER_EXECUTE
-        GROUP_READ
-        GROUP_EXECUTE
-        WORLD_READ
-        WORLD_EXECUTE
-    DESTINATION ${CMAKE_INSTALL_BINDIR})
 if (LLAMA_METAL)
    install(
        FILES ggml-metal.metal
--- a/CMakePresets.json
+++ b/CMakePresets.json
@@ -0,0 +1,45 @@
+{
+  "version": 4,
+  "configurePresets": [
+    {
+        "name":  "base",
+        "hidden": true,
+        "generator":   "Ninja",
+        "binaryDir":   "${sourceDir}/build-${presetName}",
+        "cacheVariables": {
+            "CMAKE_EXPORT_COMPILE_COMMANDS": "ON",
+            "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
+        }
+    },
+
+    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",  "hidden": true, "cacheVariables": { "LLAMA_STATIC": "ON" } },
+
+    {
+        "name": "arm64-windows-msvc", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
+        }
+    },
+
+    {
+        "name": "arm64-windows-llvm", "hidden": true,
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
+        }
+    },
+
+    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "release" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "release", "static" ] },
+
+    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "release" ] },
+    { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "release", "static" ] }
+  ]
+}
--- a/27
+++ b/27
@@ -379,15 +379,16 @@ ifneq ($(filter ppc64le%,$(UNAME_M)),)
 	CUDA_POWER_ARCH = 1
 endif

+ifneq ($(filter loongarch64%,$(UNAME_M)),)
+	MK_CFLAGS   += -mlasx
+	MK_CXXFLAGS += -mlasx
+endif
+
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif

-ifdef LLAMA_QKK_64
-	MK_CPPFLAGS += -DGGML_QKK_64
-endif
-
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
 	# `-framework Accelerate` works both with Apple Silicon and Mac Intel
@@ -399,13 +400,6 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif # LLAMA_NO_ACCELERATE

-ifdef LLAMA_MPI
-	MK_CPPFLAGS += -DGGML_USE_MPI
-	MK_CFLAGS   += -Wno-cast-qual
-	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS        += ggml-mpi.o
-endif # LLAMA_MPI
-
 ifdef LLAMA_OPENBLAS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags-only-I openblas)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other openblas)
@@ -560,10 +554,10 @@ endif # LLAMA_VULKAN
 ifdef LLAMA_HIPBLAS
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
-		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
+		AMDGPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
 	else
 		ROCM_PATH	?= /opt/rocm
-		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+		AMDGPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	endif
 	HIPCC                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
@@ -575,7 +569,7 @@ ifdef LLAMA_HIP_UMA
 endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
-	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
+	HIPFLAGS    += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
 	HIPFLAGS    += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 	HIPFLAGS    += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 	HIPFLAGS    += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
@@ -629,11 +623,6 @@ ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 endif
 endif # LLAMA_METAL

-ifdef LLAMA_MPI
-ggml-mpi.o: ggml-mpi.c ggml-mpi.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_MPI
-
 ifndef LLAMA_NO_LLAMAFILE
 sgemm.o: sgemm.cpp sgemm.h ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
--- a/README.md
+++ b/README.md
@@ -2,7 +2,7 @@

 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)

-[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT) [![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)

 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)

@@ -107,7 +107,6 @@ Typically finetunes of the base models below are supported as well.
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
- [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
 - [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
@@ -128,6 +127,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)

 (instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))

@@ -140,6 +140,8 @@ Typically finetunes of the base models below are supported as well.
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 - [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
+- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
+- [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)

 **HTTP server**

@@ -175,6 +177,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
+- [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -299,7 +302,7 @@ cd llama.cpp

 ### Build

-In order to build llama.cpp you have three different options.
+In order to build llama.cpp you have four different options.

 - Using `make`:
  - On Linux or MacOS:
@@ -380,45 +383,6 @@ To disable the Metal build at compile time use the `LLAMA_NO_METAL=1` flag or th
 When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
 argument.

-### MPI Build
-
-MPI lets you distribute the computation over a cluster of machines. Because of the serial nature of LLM prediction, this won't yield any end-to-end speed-ups, but it will let you run larger models than would otherwise fit into RAM on a single machine.
-
-First you will need MPI libraries installed on your system. The two most popular (only?) options are [MPICH](https://www.mpich.org) and [OpenMPI](https://www.open-mpi.org). Either can be installed with a package manager (`apt`, Homebrew, MacPorts, etc).
-
-Next you will need to build the project with `LLAMA_MPI` set to true on all machines; if you're building with `make`, you will also need to specify an MPI-capable compiler (when building with CMake, this is configured automatically):
-
- Using `make`:
-
-  ```bash
-  make CC=mpicc CXX=mpicxx LLAMA_MPI=1
-  ```
-
- Using `CMake`:
-
-  ```bash
-  cmake -S . -B build -DLLAMA_MPI=ON
-  ```
-
-Once the programs are built, download/convert the weights on all of the machines in your cluster. The paths to the weights and programs should be identical on all machines.
-
-Next, ensure password-less SSH access to each machine from the primary host, and create a `hostfile` with a list of the hostnames and their relative "weights" (slots). If you want to use localhost for computation, use its local subnet IP address rather than the loopback address or "localhost".
-
-Here is an example hostfile:
-
-```
-192.168.0.1:2
-malvolio.local:1
-```
-
-The above will distribute the computation across 2 processes on the first host and 1 process on the second host. Each process will use roughly an equal amount of RAM. Try to keep these numbers small, as inter-process (intra-host) communication is expensive.
-
-Finally, you're ready to run a computation using `mpirun`:
-
-```bash
-mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
-```
-
 ### BLAS Build

 Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
@@ -526,13 +490,28 @@ Building the program with BLAS support may lead to some performance improvements
    ```
  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
    ```bash
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
-        cmake -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
        && cmake --build build --config Release -- -j 16
    ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
+    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON`.
    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).

+    Note that if you get the following error:
+    ```
+    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+    ```
+    Try searching for a directory under `HIP_PATH` that contains the file
+    `oclc_abi_version_400.bc`. Then, add the following to the start of the
+    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
+    like:
+    ```bash
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+    HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
+        cmake -S . -B build -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build -- -j 16
+    ```
+
  - Using `make` (example for target gfx1030, build with 16 CPU threads):
    ```bash
    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
@@ -541,10 +520,8 @@ Building the program with BLAS support may lead to some performance improvements
  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
    ```bash
    set PATH=%HIP_PATH%\bin;%PATH%
-    mkdir build
-    cd build
-    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ..
-    cmake --build .
+    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+    cmake --build build
    ```
    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
@@ -710,6 +687,9 @@ Building the program with BLAS support may lead to some performance improvements

 ### Prepare and Quantize

+> [!NOTE]
+> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
+
 To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.

 Note: `convert.py` does not support LLaMA 3, you can use `convert-hf-to-gguf.py` with LLaMA 3 downloaded from Hugging Face.
--- a/build.zig
+++ b/build.zig
@@ -1,172 +0,0 @@
-// Compatible with Zig Version 0.11.0
-const std = @import("std");
-const ArrayList = std.ArrayList;
-const Compile = std.Build.Step.Compile;
-const ConfigHeader = std.Build.Step.ConfigHeader;
-const Mode = std.builtin.Mode;
-const CrossTarget = std.zig.CrossTarget;
-
-const Maker = struct {
-    builder: *std.build.Builder,
-    target: CrossTarget,
-    optimize: Mode,
-    enable_lto: bool,
-
-    include_dirs: ArrayList([]const u8),
-    cflags: ArrayList([]const u8),
-    cxxflags: ArrayList([]const u8),
-    objs: ArrayList(*Compile),
-
-    fn addInclude(m: *Maker, dir: []const u8) !void {
-        try m.include_dirs.append(dir);
-    }
-    fn addProjectInclude(m: *Maker, path: []const []const u8) !void {
-        try m.addInclude(try m.builder.build_root.join(m.builder.allocator, path));
-    }
-    fn addCFlag(m: *Maker, flag: []const u8) !void {
-        try m.cflags.append(flag);
-    }
-    fn addCxxFlag(m: *Maker, flag: []const u8) !void {
-        try m.cxxflags.append(flag);
-    }
-    fn addFlag(m: *Maker, flag: []const u8) !void {
-        try m.addCFlag(flag);
-        try m.addCxxFlag(flag);
-    }
-
-    fn init(builder: *std.build.Builder) !Maker {
-        const target = builder.standardTargetOptions(.{});
-        const zig_version = @import("builtin").zig_version_string;
-        const commit_hash = try std.ChildProcess.exec(
-            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
-        );
-        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
-            \\int LLAMA_BUILD_NUMBER = {};
-            \\char const *LLAMA_COMMIT = "{s}";
-            \\char const *LLAMA_COMPILER = "Zig {s}";
-            \\char const *LLAMA_BUILD_TARGET = "{s}";
-            \\
-        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
-        var m = Maker{
-            .builder = builder,
-            .target = target,
-            .optimize = builder.standardOptimizeOption(.{}),
-            .enable_lto = false,
-            .include_dirs = ArrayList([]const u8).init(builder.allocator),
-            .cflags = ArrayList([]const u8).init(builder.allocator),
-            .cxxflags = ArrayList([]const u8).init(builder.allocator),
-            .objs = ArrayList(*Compile).init(builder.allocator),
-        };
-
-        try m.addCFlag("-std=c11");
-        try m.addCxxFlag("-std=c++11");
-        try m.addProjectInclude(&.{});
-        try m.addProjectInclude(&.{"common"});
-        return m;
-    }
-
-    fn obj(m: *const Maker, name: []const u8, src: []const u8) *Compile {
-        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        if (o.target.getAbi() != .msvc)
-            o.defineCMacro("_GNU_SOURCE", null);
-
-        if (std.mem.endsWith(u8, src, ".c")) {
-            o.addCSourceFiles(&.{src}, m.cflags.items);
-            o.linkLibC();
-        } else {
-            o.addCSourceFiles(&.{src}, m.cxxflags.items);
-            if (o.target.getAbi() == .msvc) {
-                o.linkLibC(); // need winsdk + crt
-            } else {
-                // linkLibCpp already add (libc++ + libunwind + libc)
-                o.linkLibCpp();
-            }
-        }
-        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
-        o.want_lto = m.enable_lto;
-        return o;
-    }
-
-    fn exe(m: *const Maker, name: []const u8, src: []const u8, deps: []const *Compile) *Compile {
-        const e = m.builder.addExecutable(.{ .name = name, .target = m.target, .optimize = m.optimize });
-        e.addCSourceFiles(&.{src}, m.cxxflags.items);
-        for (deps) |d| e.addObject(d);
-        for (m.objs.items) |o| e.addObject(o);
-        for (m.include_dirs.items) |i| e.addIncludePath(.{ .path = i });
-
-        // https://github.com/ziglang/zig/issues/15448
-        if (e.target.getAbi() == .msvc) {
-            e.linkLibC(); // need winsdk + crt
-        } else {
-            // linkLibCpp already add (libc++ + libunwind + libc)
-            e.linkLibCpp();
-        }
-        m.builder.installArtifact(e);
-        e.want_lto = m.enable_lto;
-        return e;
-    }
-};
-
-pub fn build(b: *std.build.Builder) !void {
-    var make = try Maker.init(b);
-    make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
-
-    const ggml = make.obj("ggml", "ggml.c");
-    const sgemm = make.obj("sgemm", "sgemm.cpp");
-    const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
-    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
-    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
-    const unicode = make.obj("unicode", "unicode.cpp");
-    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
-    const llama = make.obj("llama", "llama.cpp");
-    const buildinfo = make.obj("common", "common/build-info.cpp");
-    const common = make.obj("common", "common/common.cpp");
-    const console = make.obj("console", "common/console.cpp");
-    const sampling = make.obj("sampling", "common/sampling.cpp");
-    const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
-    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
-    const train = make.obj("train", "common/train.cpp");
-    const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
-
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
-
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
-    if (server.target.isWindows()) {
-        server.linkSystemLibrary("ws2_32");
-    }
-
-    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
-    for (server_assets) |asset| {
-        const input_path = b.fmt("examples/server/public/{s}", .{asset});
-        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
-
-        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
-
-        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
-        defer b.allocator.free(input);
-
-        var buf = std.ArrayList(u8).init(b.allocator);
-        defer buf.deinit();
-
-        for (input) |byte| {
-            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
-        }
-
-        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
-        defer b.allocator.free(name);
-        std.mem.replaceScalar(u8, name, '.', '_');
-
-        try std.fs.cwd().writeFile(output_path, b.fmt(
-            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
-            .{ name, buf.items, name, input.len },
-        ));
-
-        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
-    }
-}
--- a/ci/run.sh
+++ b/ci/run.sh
@@ -202,12 +202,15 @@ function gg_sum_test_scripts_release {
 }

 function gg_get_model {
-    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
-    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-    if [[ -s $gguf_3b ]]; then
-        echo -n "$gguf_3b"
-    elif [[ -s $gguf_7b ]]; then
-        echo -n "$gguf_7b"
+    local gguf_0="$MNT/models/pythia/1.4B/ggml-model-f16.gguf"
+    local gguf_1="$MNT/models/pythia/2.8B/ggml-model-f16.gguf"
+    local gguf_2="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
+    if [[ -s $gguf_0 ]]; then
+        echo -n "$gguf_0"
+    elif [[ -s $gguf_1 ]]; then
+        echo -n "$gguf_1"
+    elif [[ -s $gguf_2 ]]; then
+        echo -n "$gguf_2"
    else
        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
        exit 1
@@ -256,186 +259,6 @@ function gg_sum_ctest_with_model_release {
    gg_printf '```\n'
 }

-# open_llama_3b_v2
-
-function gg_run_open_llama_3b_v2 {
-    cd ${SRC}
-
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/tokenizer.model
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/tokenizer_config.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/special_tokens_map.json
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
-
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
-    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
-    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
-
-    path_models="../models-mnt/open-llama/3B-v2"
-    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert.py ${path_models}
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
-    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
-    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
-    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
-    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
-    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
-    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
-    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
-    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
-
-    wiki_test_60="${path_wiki}/wiki.test-60.raw"
-
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
-    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
-    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
-    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
-    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
-    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
-    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
-    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
-    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
-
-    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
-
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
-    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-
-    function check_ppl {
-        qnt="$1"
-        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
-        return 0
-    }
-
-    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
-
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/3B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_3b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    set +e
-}
-
-function gg_sum_open_llama_3b_v2 {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'OpenLLaMA 3B-v2:\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
-    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
-    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
-    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
-    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
-    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
-    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
-    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
-    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
-    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
-}
-
 # open_llama_7b_v2
 # requires: GG_BUILD_CUDA

@@ -464,7 +287,7 @@ function gg_run_open_llama_7b_v2 {
    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../convert.py ${path_models}
+    python3 ../convert.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf

    model_f16="${path_models}/ggml-model-f16.gguf"
    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -549,48 +372,6 @@ function gg_run_open_llama_7b_v2 {

    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log

-    # lora
-    function compare_ppl {
-        qnt="$1"
-        ppl1=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-        ppl2=$(echo "$3" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$ppl1 < $ppl2" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: %s > %s)\n' "$qnt" "$ppl" "$ppl1" "$ppl2"
-            return 20
-        fi
-
-        printf '  - %s @ %s %s OK\n' "$qnt" "$ppl1" "$ppl2"
-        return 0
-    }
-
-    path_lora="../models-mnt/open-llama/7B-v2/lora"
-    path_shakespeare="../models-mnt/shakespeare"
-
-    shakespeare="${path_shakespeare}/shakespeare.txt"
-    lora_shakespeare="${path_lora}/ggml-adapter-model.bin"
-
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_config.json
-    gg_wget ${path_lora} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/adapter_model.bin
-    gg_wget ${path_shakespeare} https://huggingface.co/slaren/open_llama_7b_v2_shakespeare_lora/resolve/main/shakespeare.txt
-
-    python3 ../convert-lora-to-ggml.py ${path_lora}
-
-    # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
-    compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # currently not supported by the CUDA backend
-    # q8_0
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
-    #compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
-    # q8_0 + f16 lora-base
-    #(time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -t 1 -ngl 999 -c 2048 -b 512 --chunks 3 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
-    #compare_ppl "q8_0 / f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
-
    set +e
 }

@@ -601,7 +382,6 @@ function gg_sum_open_llama_7b_v2 {
    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
-    gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
@@ -614,11 +394,272 @@ function gg_sum_open_llama_7b_v2 {
    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
-    gg_printf '- shakespeare (f16):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-f16.log)"
-    gg_printf '- shakespeare (f16 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log)"
-    #gg_printf '- shakespeare (q8_0):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log)"
-    #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
+}
+
+# pythia_1.4b
+
+function gg_run_pythia_1_4b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/1.4B/ https://huggingface.co/EleutherAI/pythia-1.4b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+    head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
+
+    path_models="../models-mnt/pythia/1.4B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test_60="${path_wiki}/wiki.test-60.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_1_4b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 1.4B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
+}
+
+# pythia_2_8b
+# requires: GG_BUILD_CUDA
+
+function gg_run_pythia_2_8b {
+    cd ${SRC}
+
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/tokenizer_config.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/raw/main/special_tokens_map.json
+    gg_wget models-mnt/pythia/2.8B/ https://huggingface.co/EleutherAI/pythia-2.8b/resolve/main/pytorch_model.bin
+
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
+
+    path_models="../models-mnt/pythia/2.8B"
+    path_wiki="../models-mnt/wikitext/wikitext-2-raw"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+    model_q4_0="${path_models}/ggml-model-q4_0.gguf"
+    model_q4_1="${path_models}/ggml-model-q4_1.gguf"
+    model_q5_0="${path_models}/ggml-model-q5_0.gguf"
+    model_q5_1="${path_models}/ggml-model-q5_1.gguf"
+    model_q2_k="${path_models}/ggml-model-q2_k.gguf"
+    model_q3_k="${path_models}/ggml-model-q3_k.gguf"
+    model_q4_k="${path_models}/ggml-model-q4_k.gguf"
+    model_q5_k="${path_models}/ggml-model-q5_k.gguf"
+    model_q6_k="${path_models}/ggml-model-q6_k.gguf"
+
+    wiki_test="${path_wiki}/wiki.test.raw"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+    ./bin/quantize ${model_f16} ${model_q4_0} q4_0
+    ./bin/quantize ${model_f16} ${model_q4_1} q4_1
+    ./bin/quantize ${model_f16} ${model_q5_0} q5_0
+    ./bin/quantize ${model_f16} ${model_q5_1} q5_1
+    ./bin/quantize ${model_f16} ${model_q2_k} q2_k
+    ./bin/quantize ${model_f16} ${model_q3_k} q3_k
+    ./bin/quantize ${model_f16} ${model_q4_k} q4_k
+    ./bin/quantize ${model_f16} ${model_q5_k} q5_k
+    ./bin/quantize ${model_f16} ${model_q6_k} q6_k
+
+    (time ./bin/main --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/main --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/main --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/main --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/main --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/main --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/main --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/main --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/main --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/main --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/main --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+
+    (time ./bin/save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+
+    function check_ppl {
+        qnt="$1"
+        ppl=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$ppl > 20.0" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: ppl > 20.0)\n' "$qnt" "$ppl"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$ppl"
+        return 0
+    }
+
+    check_ppl "f16"  "$(cat $OUT/${ci}-tg-f16.log  | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q8_0" "$(cat $OUT/${ci}-tg-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_0" "$(cat $OUT/${ci}-tg-q4_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_1" "$(cat $OUT/${ci}-tg-q4_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_0" "$(cat $OUT/${ci}-tg-q5_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_1" "$(cat $OUT/${ci}-tg-q5_1.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+   #check_ppl "q2_k" "$(cat $OUT/${ci}-tg-q2_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log # note: ppl > 20.0 for this quant and model
+    check_ppl "q3_k" "$(cat $OUT/${ci}-tg-q3_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q4_k" "$(cat $OUT/${ci}-tg-q4_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+    check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
+
+    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
+
+    set +e
+}
+
+function gg_sum_pythia_2_8b {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Pythia 2.8B:\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
+    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+    gg_printf '- q4_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_0.log)"
+    gg_printf '- q4_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_1.log)"
+    gg_printf '- q5_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_0.log)"
+    gg_printf '- q5_1:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_1.log)"
+    gg_printf '- q2_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q2_k.log)"
+    gg_printf '- q3_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q3_k.log)"
+    gg_printf '- q4_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q4_k.log)"
+    gg_printf '- q5_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q5_k.log)"
+    gg_printf '- q6_k:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q6_k.log)"
+    gg_printf '- save-load-state: \n```\n%s\n```\n' "$(cat $OUT/${ci}-save-load-state.log)"
 }

 # bge-small
@@ -647,7 +688,7 @@ function gg_run_embd_bge_small {
    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log

-    python3 ../convert-hf-to-gguf.py ${path_models}
+    python3 ../convert-hf-to-gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf

    model_f16="${path_models}/ggml-model-f16.gguf"
    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
@@ -701,9 +742,10 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then

    if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
        if [ -z ${GG_BUILD_CUDA} ]; then
-            test $ret -eq 0 && gg_run open_llama_3b_v2
+            test $ret -eq 0 && gg_run pythia_1_4b
        else
-            test $ret -eq 0 && gg_run open_llama_7b_v2
+            test $ret -eq 0 && gg_run pythia_2_8b
+            #test $ret -eq 0 && gg_run open_llama_7b_v2
        fi
        test $ret -eq 0 && gg_run ctest_with_model_debug
        test $ret -eq 0 && gg_run ctest_with_model_release
--- a/cmake/arm64-windows-llvm.cmake
+++ b/cmake/arm64-windows-llvm.cmake
@@ -0,0 +1,16 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.7-a -fvectorize -ffp-model=fast" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function -Wno-gnu-zero-variadic-macro-arguments" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )
--- a/cmake/arm64-windows-msvc.cmake
+++ b/cmake/arm64-windows-msvc.cmake
@@ -0,0 +1,6 @@
+set( CMAKE_SYSTEM_NAME Windows )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-pc-windows-msvc )
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
--- a/common/common.cpp
+++ b/common/common.cpp
--- a/common/common.h
+++ b/common/common.h
@@ -27,7 +27,7 @@
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)

 #define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
+    fprintf(stderr, "%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);      \
    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)

@@ -35,14 +35,18 @@

 // build info
 extern int LLAMA_BUILD_NUMBER;
-extern char const *LLAMA_COMMIT;
-extern char const *LLAMA_COMPILER;
-extern char const *LLAMA_BUILD_TARGET;
+extern char const * LLAMA_COMMIT;
+extern char const * LLAMA_COMPILER;
+extern char const * LLAMA_BUILD_TARGET;

 struct llama_control_vector_load_info;

-int get_math_cpu_count();
-int32_t get_num_physical_cores();
+//
+// CPU utils
+//
+
+int32_t cpu_get_num_physical_cores();
+int32_t cpu_get_num_math();

 //
 // CLI argument parsing
@@ -51,7 +55,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed

-    int32_t n_threads             = get_math_cpu_count();
+    int32_t n_threads             = cpu_get_num_math();
    int32_t n_threads_draft       = -1;
    int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
    int32_t n_threads_batch_draft = -1;
@@ -82,6 +86,7 @@ struct gpt_params {
    float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
    int32_t yarn_orig_ctx         = 0;     // YaRN original context length
    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
+    std::string rpc_servers       = "";    // comma separated list of RPC servers

    ggml_backend_sched_eval_callback cb_eval = nullptr;
    void * cb_eval_user_data                 = nullptr;
@@ -140,6 +145,7 @@ struct gpt_params {
    bool random_prompt     = false; // do not randomize prompt if none provided
    bool use_color         = false; // use color to distinguish generations and inputs
    bool interactive       = false; // interactive mode
+    bool interactive_specials = false; // whether to allow special tokens from user, during interactive mode
    bool conversation      = false; // conversation mode (does not print special tokens and suffix/prefix)
    bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
    bool prompt_cache_all  = false; // save user input and generations to prompt cache
@@ -177,33 +183,34 @@ struct gpt_params {

 void gpt_params_handle_model_default(gpt_params & params);

-bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
+bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
+bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);

-bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
-
-bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
-
-void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
-
-bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
-
-std::string get_system_info(const gpt_params & params);
-
-std::string gpt_random_prompt(std::mt19937 & rng);
-
-void process_escapes(std::string& input);
-
-bool validate_file_name(const std::string & filename);
+std::string gpt_params_get_system_info(const gpt_params & params);

 //
 // String utils
 //

-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
+
 std::string string_strip(const std::string & str);
-std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
+std::string string_get_sortable_timestamp();
+std::string string_random_prompt(std::mt19937 & rng);
+
+bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+void string_process_escapes(std::string & input);
+
+//
+// Filesystem utils
+//
+
+bool fs_validate_filename(const std::string & filename);
+bool fs_create_directory_with_parents(const std::string & path);
+
+std::string fs_get_cache_directory();

 //
 // Model utils
@@ -274,29 +281,15 @@ std::string llama_detokenize_bpe(
 // defaults to true when model type is SPM, otherwise false.
 bool llama_should_add_bos_token(const llama_model * model);

-//
-// YAML utils
-//
-
-bool create_directory_with_parents(const std::string & path);
-void dump_vector_float_yaml(FILE * stream, const char * prop_name, const std::vector<float> & data);
-void dump_vector_int_yaml(FILE * stream, const char * prop_name, const std::vector<int> & data);
-void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const char * data);
-std::string get_sortable_timestamp();
-
-void dump_non_result_info_yaml(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
-    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-
 //
 // KV cache utils
 //

 // Dump the KV cache view with the number of sequences per cell.
-void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
+void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);

 // Dump the KV cache view showing individual sequences in each cell (long output).
-void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);

 //
 // Embedding utils
@@ -330,6 +323,20 @@ llama_control_vector_data llama_control_vector_load(const std::vector<llama_cont
 //
 // Split utils
 //
+
 static const char * const LLM_KV_SPLIT_NO            = "split.no";
 static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
 static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";
+
+//
+// YAML utils
+//
+
+void yaml_dump_vector_float    (FILE * stream, const char * prop_name, const std::vector<float> & data);
+void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std::vector<int> & data);
+void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
+
+void yaml_dump_non_result_info(
+    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+
--- a/common/grammar-parser.cpp
+++ b/common/grammar-parser.cpp
@@ -26,7 +26,7 @@ namespace grammar_parser {

    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
-        auto result = state.symbol_ids.insert(std::make_pair(std::string(src, len), next_id));
+        auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
        return result.first->second;
    }

@@ -142,6 +142,9 @@ namespace grammar_parser {
                pos++;
                last_sym_start = out_elements.size();
                while (*pos != '"') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
                    auto char_pair = parse_char(pos);
                         pos       = char_pair.second;
                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
@@ -156,6 +159,9 @@ namespace grammar_parser {
                }
                last_sym_start = out_elements.size();
                while (*pos != ']') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
                    auto char_pair = parse_char(pos);
                         pos       = char_pair.second;
                    enum llama_gretype type = last_sym_start < out_elements.size()
@@ -164,6 +170,9 @@ namespace grammar_parser {

                    out_elements.push_back({type, char_pair.first});
                    if (pos[0] == '-' && pos[1] != ']') {
+                        if (!pos[1]) {
+                            throw std::runtime_error("unexpected end of input");
+                        }
                        auto endchar_pair = parse_char(pos + 1);
                             pos          = endchar_pair.second;
                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -272,7 +272,7 @@ private:
                    if (literal.empty()) {
                        return false;
                    }
-                    ret.push_back(std::make_pair(literal, true));
+                    ret.emplace_back(literal, true);
                    literal.clear();
                    return true;
                };
@@ -298,7 +298,7 @@ private:
            while (i < length) {
                char c = sub_pattern[i];
                if (c == '.') {
-                    seq.push_back(std::make_pair(get_dot(), false));
+                    seq.emplace_back(get_dot(), false);
                    i++;
                } else if (c == '(') {
                    i++;
@@ -307,7 +307,7 @@ private:
                            _warnings.push_back("Unsupported pattern syntax");
                        }
                    }
-                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
+                    seq.emplace_back("(" + to_rule(transform()) + ")", false);
                } else if (c == ')') {
                    i++;
                    if (start > 0 && sub_pattern[start - 1] != '(') {
@@ -331,9 +331,9 @@ private:
                    }
                    square_brackets += ']';
                    i++;
-                    seq.push_back(std::make_pair(square_brackets, false));
+                    seq.emplace_back(square_brackets, false);
                } else if (c == '|') {
-                    seq.push_back(std::make_pair("|", false));
+                    seq.emplace_back("|", false);
                    i++;
                } else if (c == '*' || c == '+' || c == '?') {
                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
@@ -417,7 +417,7 @@ private:
                        }
                    }
                    if (!literal.empty()) {
-                        seq.push_back(std::make_pair(literal, true));
+                        seq.emplace_back(literal, true);
                    }
                }
            }
--- a/common/log.h
+++ b/common/log.h
@@ -211,7 +211,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
    #else
        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
    #endif
 #else
    #define LOG_FLF_FMT "%s"
@@ -224,7 +224,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
    #else
        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
-        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
    #endif
 #else
    #define LOG_TEE_FLF_FMT "%s"
@@ -294,7 +294,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Main LOG macro.
 //  behaves like printf, and supports arguments the exact same way.
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
@@ -308,14 +308,14 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // Secondary target can be changed just like LOG_TARGET
 //  by defining LOG_TEE_TARGET
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
 #endif

 // LOG macro variants with auto endline.
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) || defined(__clang__)
    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
--- a/common/sampling.cpp
+++ b/common/sampling.cpp
@@ -35,7 +35,7 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_

    result->prev.resize(params.n_prev);

-    result->n_considered = 0;
+    result->n_valid = 0;

    llama_sampling_set_rng_seed(result, params.seed);

@@ -66,7 +66,7 @@ void llama_sampling_reset(llama_sampling_context * ctx) {

    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
    ctx->cur.clear();
-    ctx->n_considered = 0;
+    ctx->n_valid = 0;
 }

 void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
@@ -125,7 +125,7 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
    std::string result = "CFG -> Penalties ";
    if (params.mirostat == 0) {
        for (auto sampler_type : params.samplers_sequence) {
-            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
            if (!sampler_type_name.empty()) {
                result += "-> " + sampler_type_name + " ";
            }
@@ -137,6 +137,87 @@ std::string llama_sampling_order_print(const llama_sampling_params & params) {
    return result;
 }

+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
+    switch (sampler_type) {
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
+        default : return "";
+    }
+}
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
+    // since samplers names are written multiple ways
+    // make it ready for both system names and input names
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names.size());
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
+            sampler_types.push_back(sampler_item->second);
+        }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
+    }
+    return sampler_types;
+}
+
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
+    std::unordered_map<char, llama_sampler_type> sampler_name_map {
+        {'k', llama_sampler_type::TOP_K},
+        {'p', llama_sampler_type::TOP_P},
+        {'y', llama_sampler_type::TYPICAL_P},
+        {'m', llama_sampler_type::MIN_P},
+        {'f', llama_sampler_type::TFS_Z},
+        {'t', llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names_string.size());
+    for (const auto & c : names_string) {
+        const auto sampler_item = sampler_name_map.find(c);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
+        }
+    }
+    return sampler_types;
+}
+
 // no reasons to expose this function in header
 static void sampler_queue(
                   struct llama_context * ctx_main,
@@ -179,7 +260,7 @@ static llama_token llama_sampling_sample_impl(
                  struct llama_context * ctx_main,
                  struct llama_context * ctx_cfg,
                  const int idx,
-                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
+                  bool is_resampling) {
    const llama_sampling_params & params = ctx_sampling->params;

    const float   temp            = params.temp;
@@ -188,8 +269,8 @@ static llama_token llama_sampling_sample_impl(
    const float   mirostat_eta    = params.mirostat_eta;

    std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
-    if (!is_resampling) {
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
        GGML_ASSERT(!original_logits.empty());
    }
    llama_token id = 0;
@@ -252,11 +333,11 @@ static llama_token llama_sampling_sample_impl(
            // Restore logits from the copy
            std::copy(original_logits.begin(), original_logits.end(), logits);

-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
        }
    }

-    ctx_sampling->n_considered = cur_p.size;
+    ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;

    return id;
 }
@@ -285,7 +366,8 @@ static llama_token_data_array llama_sampling_prepare_impl(
    // Get a pointer to the logits
    float * logits = llama_get_logits_ith(ctx_main, idx);

-    if (apply_grammar && original_logits != NULL) {
+    if (ctx_sampling->grammar != NULL && !apply_grammar) {
+        GGML_ASSERT(original_logits != NULL);
        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
    }
@@ -342,7 +424,7 @@ llama_token llama_sampling_sample(
                  struct llama_context * ctx_cfg,
                  const int idx) {
    // Call the implementation function with is_resampling set to false by default
-    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
 }

 llama_token_data_array llama_sampling_prepare(
--- a/common/sampling.h
+++ b/common/sampling.h
@@ -81,7 +81,7 @@ struct llama_sampling_context {
    // TODO: replace with ring-buffer
    std::vector<llama_token>      prev;
    std::vector<llama_token_data> cur;
-    size_t n_considered;
+    size_t n_valid; // Number of correct top tokens with correct probabilities.

    std::mt19937 rng;
 };
@@ -116,6 +116,11 @@ std::string llama_sampling_print(const llama_sampling_params & params);
 // Print sampling order into a string
 std::string llama_sampling_order_print(const llama_sampling_params & params);

+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
+
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call
--- a/common/train.cpp
+++ b/common/train.cpp
@@ -1380,7 +1380,7 @@ bool consume_common_train_arg(

 void finish_processing_train_args(struct train_params_common * params) {
    if (params->escape) {
-        process_escapes(params->sample_start);
+        string_process_escapes(params->sample_start);
    }
 }

--- a/convert-hf-to-gguf-update.py
+++ b/convert-hf-to-gguf-update.py
@@ -20,11 +20,13 @@
 # - Update llama.cpp with the new pre-tokenizer if necessary
 #
 # TODO: generate tokenizer tests for llama.cpp
-# TODO: automate the update of convert-hf-to-gguf.py
 #

 import logging
 import os
+import pathlib
+import re
+
 import requests
 import sys
 import json
@@ -35,6 +37,7 @@ from transformers import AutoTokenizer

 logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger("convert-hf-to-gguf-update")
+sess = requests.Session()


 class TOKENIZER_TYPE(IntEnum):
@@ -49,6 +52,10 @@ chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍

 if len(sys.argv) == 2:
    token = sys.argv[1]
+    if not token.startswith("hf_"):
+        logger.info("Huggingface token seems invalid")
+        logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
+        sys.exit(1)
 else:
    logger.info("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
    sys.exit(1)
@@ -65,70 +72,55 @@ models = [
    {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
    {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
    {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+    {"name": "stablelm2",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/stabilityai/stablelm-2-zephyr-1_6b", },
    {"name": "refact",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/smallcloudai/Refact-1_6-base", },
    {"name": "command-r",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/CohereForAI/c4ai-command-r-v01", },
    {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
    {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
    {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
+    {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
+    {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
 ]

-# make directory "models/tokenizers" if it doesn't exist
-if not os.path.exists("models/tokenizers"):
-    os.makedirs("models/tokenizers")
-

 def download_file_with_auth(url, token, save_path):
    headers = {"Authorization": f"Bearer {token}"}
-    response = requests.get(url, headers=headers)
-    if response.status_code == 200:
-        with open(save_path, 'wb') as f:
-            f.write(response.content)
-        logger.info(f"File {save_path} downloaded successfully")
-    else:
-        logger.info(f"Failed to download file. Status code: {response.status_code}")
+    response = sess.get(url, headers=headers)
+    response.raise_for_status()
+    os.makedirs(os.path.dirname(save_path), exist_ok=True)
+    with open(save_path, 'wb') as f:
+        f.write(response.content)
+    logger.info(f"File {save_path} downloaded successfully")


-# download the tokenizer models
-for model in models:
+def download_model(model):
    name = model["name"]
    repo = model["repo"]
    tokt = model["tokt"]

-    if not os.path.exists(f"models/tokenizers/{name}"):
-        os.makedirs(f"models/tokenizers/{name}")
-    else:
-        logger.info(f"Directory models/tokenizers/{name} already exists - skipping")
-        continue
-
-    logger.info(f"Downloading {name} to models/tokenizers/{name}")
-
-    url = f"{repo}/raw/main/config.json"
-    save_path = f"models/tokenizers/{name}/config.json"
-    download_file_with_auth(url, token, save_path)
-
-    url = f"{repo}/raw/main/tokenizer.json"
-    save_path = f"models/tokenizers/{name}/tokenizer.json"
-    download_file_with_auth(url, token, save_path)
-
-    # if downloaded file is less than 1KB, we likely need to download an LFS instead
-    if os.path.getsize(save_path) < 1024:
-        # remove the file
-        os.remove(save_path)
-        url = f"{repo}/resolve/main/tokenizer.json"
-        save_path = f"models/tokenizers/{name}/tokenizer.json"
-        download_file_with_auth(url, token, save_path)
+    os.makedirs(f"models/tokenizers/{name}", exist_ok=True)

+    files = ["config.json", "tokenizer.json", "tokenizer_config.json"]
    if tokt == TOKENIZER_TYPE.SPM:
-        url = f"{repo}/resolve/main/tokenizer.model"
-        save_path = f"models/tokenizers/{name}/tokenizer.model"
-        download_file_with_auth(url, token, save_path)
+        files.append("tokenizer.model")
+
+    for file in files:
+        save_path = f"models/tokenizers/{name}/{file}"
+        if os.path.isfile(save_path):
+            logger.info(f"{name}: File {save_path} already exists - skipping")
+            continue
+        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+
+
+for model in models:
+    try:
+        download_model(model)
+    except Exception as e:
+        logger.error(f"Failed to download model {model['name']}. Error: {e}")

-    url = f"{repo}/raw/main/tokenizer_config.json"
-    save_path = f"models/tokenizers/{name}/tokenizer_config.json"
-    download_file_with_auth(url, token, save_path)

 # generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function:
-# TODO: auto-update convert-hf-to-gguf.py with the generated function

 src_ifs = ""
 for model in models:
@@ -138,8 +130,17 @@ for model in models:
    if tokt == TOKENIZER_TYPE.SPM:
        continue

+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
    # create the tokenizer
-    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
+        continue  # Skip to the next model if the tokenizer can't be loaded

    chktok = tokenizer.encode(chktxt)
    chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -157,6 +158,8 @@ for model in models:
        logger.info("normalizer: " + json.dumps(normalizer, indent=4))
        pre_tokenizer = cfg["pre_tokenizer"]
        logger.info("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+        if "ignore_merges" in cfg["model"]:
+            logger.info("ignore_merges: " + json.dumps(cfg["model"]["ignore_merges"], indent=4))

    logger.info("")

@@ -206,11 +209,18 @@ src_func = f"""
        return res
 """

-print(src_func) # noqa: NP100
+convert_py_pth = pathlib.Path("convert-hf-to-gguf.py")
+convert_py = convert_py_pth.read_text()
+convert_py = re.sub(
+    r"(# Marker: Start get_vocab_base_pre)(.+?)( +# Marker: End get_vocab_base_pre)",
+    lambda m: m.group(1) + src_func + m.group(3),
+    convert_py,
+    flags=re.DOTALL | re.MULTILINE,
+)

-logger.info("\n")
-logger.info("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!")
-logger.info("\n")
+convert_py_pth.write_text(convert_py)
+
+logger.info("+++ convert-hf-to-gguf.py was updated")

 # generate tests for each tokenizer model

@@ -257,6 +267,7 @@ tests = [
    "3333333",
    "33333333",
    "333333333",
+    # "Cửa Việt", # llama-bpe fails on this
    chktxt,
 ]

@@ -277,8 +288,17 @@ for model in models:
    name = model["name"]
    tokt = model["tokt"]

+    # Skip if the tokenizer folder does not exist or there are other download issues previously
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        logger.warning(f"Directory for tokenizer {name} not found. Skipping...")
+        continue
+
    # create the tokenizer
-    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    try:
+        tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+    except OSError as e:
+        logger.error(f"Failed to load tokenizer for model {name}. Error: {e}")
+        continue  # Skip this model and continue with the next one in the loop

    with open(f"models/ggml-vocab-{name}.gguf.inp", "w", encoding="utf-8") as f:
        for text in tests:
--- a/convert-hf-to-gguf.py
+++ b/convert-hf-to-gguf.py
--- a/convert-lora-to-ggml.py
+++ b/convert-lora-to-ggml.py
@@ -1,150 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import json
-import os
-import struct
-import sys
-from pathlib import Path
-from typing import Any, BinaryIO, Sequence
-
-import numpy as np
-import torch
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
-logging.basicConfig(level=logging.DEBUG)
-logger = logging.getLogger("lora-to-gguf")
-
-NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
-
-
-def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
-    fout.write(b"ggla"[::-1])  # magic (ggml lora)
-    fout.write(struct.pack("i", 1))  # file version
-    fout.write(struct.pack("i", params["r"]))
-    # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
-    # but some models ship a float value instead
-    # let's convert to int, but fail if lossless conversion is not possible
-    assert (
-        int(params["lora_alpha"]) == params["lora_alpha"]
-    ), "cannot convert float to int losslessly"
-    fout.write(struct.pack("i", int(params["lora_alpha"])))
-
-
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
-    sname = name.encode("utf-8")
-    fout.write(
-        struct.pack(
-            "iii",
-            len(shape),
-            len(sname),
-            NUMPY_TYPE_TO_FTYPE[data_type.name],
-        )
-    )
-    fout.write(struct.pack("i" * len(shape), *shape[::-1]))
-    fout.write(sname)
-    fout.seek((fout.tell() + 31) & -32)
-
-
-if __name__ == '__main__':
-    if len(sys.argv) < 2:
-        logger.info(f"Usage: python {sys.argv[0]} <path> [arch]")
-        logger.info("Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'")
-        logger.info(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-        sys.exit(1)
-
-    input_json = os.path.join(sys.argv[1], "adapter_config.json")
-    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
-
-    if os.path.exists(input_model):
-        model = torch.load(input_model, map_location="cpu")
-    else:
-        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-        model = load_file(input_model, device="cpu")
-
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
-
-    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-        logger.error(f"Error: unsupported architecture {arch_name}")
-        sys.exit(1)
-
-    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
-
-    with open(input_json, "r") as f:
-        params = json.load(f)
-
-    if params["peft_type"] != "LORA":
-        logger.error(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-        sys.exit(1)
-
-    if params["fan_in_fan_out"] is True:
-        logger.error("Error: param fan_in_fan_out is not supported")
-        sys.exit(1)
-
-    if params["bias"] is not None and params["bias"] != "none":
-        logger.error("Error: param bias is not supported")
-        sys.exit(1)
-
-    # TODO: these seem to be layers that have been trained but without lora.
-    # doesn't seem widely used but eventually should be supported
-    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-        logger.error("Error: param modules_to_save is not supported")
-        sys.exit(1)
-
-    with open(output_path, "wb") as fout:
-        fout.truncate()
-
-        write_file_header(fout, params)
-        for k, v in model.items():
-            orig_k = k
-            if k.endswith(".default.weight"):
-                k = k.replace(".default.weight", ".weight")
-            if k in ["llama_proj.weight", "llama_proj.bias"]:
-                continue
-            if k.endswith("lora_A.weight"):
-                if v.dtype != torch.float16 and v.dtype != torch.float32:
-                    v = v.float()
-                v = v.T
-            else:
-                v = v.float()
-
-            t = v.detach().numpy()
-
-            prefix = "base_model.model."
-            if k.startswith(prefix):
-                k = k[len(prefix) :]
-
-            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-            if k.endswith(lora_suffixes):
-                suffix = k[-len(lora_suffixes[0]):]
-                k = k[: -len(lora_suffixes[0])]
-            else:
-                logger.error(f"Error: unrecognized tensor name {orig_k}")
-                sys.exit(1)
-
-            tname = name_map.get_name(k)
-            if tname is None:
-                logger.error(f"Error: could not map tensor name {orig_k}")
-                logger.error(" Note: the arch parameter must be specified if the model is not llama")
-                sys.exit(1)
-
-            if suffix == ".lora_A.weight":
-                tname += ".weight.loraA"
-            elif suffix == ".lora_B.weight":
-                tname += ".weight.loraB"
-            else:
-                assert False
-
-            logger.info(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-            write_tensor_header(fout, tname, t.shape, t.dtype)
-            t.tofile(fout)
-
-    logger.info(f"Converted {input_json} and {input_model} to {output_path}")
--- a/convert-persimmon-to-gguf.py
+++ b/convert-persimmon-to-gguf.py
@@ -1,143 +0,0 @@
-#!/usr/bin/env python3
-from __future__ import annotations
-
-import logging
-import argparse
-import os
-import sys
-from pathlib import Path
-from pprint import pprint
-
-import torch
-from sentencepiece import SentencePieceProcessor
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
-import gguf
-
-logger = logging.getLogger("persimmon-to-gguf")
-
-
-def _flatten_dict(dct, tensors, prefix=None):
-    assert isinstance(dct, dict)
-    for key in dct.keys():
-        new_prefix = prefix + '.' + key if prefix is not None else key
-        if isinstance(dct[key], torch.Tensor):
-            tensors[new_prefix] = dct[key]
-        elif isinstance(dct[key], dict):
-            _flatten_dict(dct[key], tensors, new_prefix)
-        else:
-            raise ValueError(type(dct[key]))
-    return None
-
-
-def _get_sentencepiece_tokenizer_info(dir_model: Path):
-    tokenizer_path = dir_model / 'adept_vocab.model'
-    logger.info('getting sentencepiece tokenizer from', tokenizer_path)
-    tokenizer = SentencePieceProcessor(str(tokenizer_path))
-    logger.info('adding tokens')
-    tokens: list[bytes] = []
-    scores: list[float] = []
-    toktypes: list[int] = []
-
-    for i in range(tokenizer.vocab_size()):
-        text: bytes
-        score: float
-
-        piece = tokenizer.id_to_piece(i)
-        text = piece.encode("utf-8")
-        score = tokenizer.get_score(i)
-
-        toktype = 1
-        if tokenizer.is_unknown(i):
-            toktype = 2
-        if tokenizer.is_control(i):
-            toktype = 3
-        if tokenizer.is_unused(i):
-            toktype = 5
-        if tokenizer.is_byte(i):
-            toktype = 6
-
-        tokens.append(text)
-        scores.append(score)
-        toktypes.append(toktype)
-        pass
-    return tokens, scores, toktypes
-
-
-def main():
-    parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
-    parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
-    parser.add_argument("--ckpt-path",           type=Path, help="path to persimmon checkpoint .pt file")
-    parser.add_argument("--model-dir",           type=Path, help="directory containing model e.g. 8b_chat_model_release")
-    parser.add_argument("--adept-inference-dir", type=str,  help="path to adept-inference code directory")
-    parser.add_argument("--verbose",  action="store_true",  help="increase output verbosity")
-    args = parser.parse_args()
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-    sys.path.append(str(args.adept_inference_dir))
-    persimmon_model = torch.load(args.ckpt_path)
-    hparams = persimmon_model['args']
-    pprint(hparams)
-    tensors: dict[str, torch.Tensor] = {}
-    _flatten_dict(persimmon_model['model'], tensors, None)
-
-    arch = gguf.MODEL_ARCH.PERSIMMON
-    gguf_writer = gguf.GGUFWriter(args.outfile, gguf.MODEL_ARCH_NAMES[arch])
-
-    block_count = hparams.num_layers
-    head_count = hparams.num_attention_heads
-    head_count_kv = head_count
-    ctx_length = hparams.seq_length
-    hidden_size = hparams.hidden_size
-
-    gguf_writer.add_name('persimmon-8b-chat')
-    gguf_writer.add_context_length(ctx_length)
-    gguf_writer.add_embedding_length(hidden_size)
-    gguf_writer.add_block_count(block_count)
-    gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
-    gguf_writer.add_head_count(head_count)
-    gguf_writer.add_head_count_kv(head_count_kv)
-    gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
-    gguf_writer.add_layer_norm_eps(hparams.layernorm_epsilon)
-
-    tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
-    gguf_writer.add_tokenizer_model('llama')
-    gguf_writer.add_tokenizer_pre('default')
-    gguf_writer.add_token_list(tokens)
-    gguf_writer.add_token_scores(scores)
-    gguf_writer.add_token_types(toktypes)
-    gguf_writer.add_bos_token_id(71013)
-    gguf_writer.add_eos_token_id(71013)
-
-    tensor_map = gguf.get_tensor_name_map(arch, block_count)
-    logger.info(tensor_map)
-    for name in tensors.keys():
-        data_torch = tensors[name]
-        if name.endswith(".self_attention.rotary_emb.inv_freq"):
-            continue
-        old_dtype = data_torch.dtype
-        # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data_torch.to(torch.float32).squeeze().numpy()
-        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
-        if new_name is None:
-            raise ValueError(f"Can not map tensor '{name}'")
-
-        n_dims = len(data.shape)
-        logger.debug(f"{new_name}, n_dims = {str(n_dims)}, {str(old_dtype)} --> {str(data.dtype)}")
-        gguf_writer.add_tensor(new_name, data)
-    logger.info("gguf: write header")
-    gguf_writer.write_header_to_file()
-    logger.info("gguf: write metadata")
-    gguf_writer.write_kv_data_to_file()
-    logger.info("gguf: write tensors")
-    gguf_writer.write_tensors_to_file()
-
-    gguf_writer.close()
-
-    logger.info(f"gguf: model successfully exported to '{args.outfile}'")
-
-
-if __name__ == '__main__':
-    main()
--- a/convert.py
+++ b/convert.py
@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable, Optional

 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -284,6 +284,7 @@ class Params:
        n_experts      = None
        n_experts_used = None
        f_rope_freq_base = None
+        n_ff = None

        # hack to determine LLaMA v1 vs v2 vs CodeLlama
        if config.get("moe"):
@@ -308,6 +309,8 @@ class Params:
            n_experts_used = config["moe"]["num_experts_per_tok"]
            f_rope_freq_base = 1e6

+        assert n_ff is not None
+
        return Params(
            n_vocab          = model["tok_embeddings.weight"].shape[0],
            n_embd           = config["dim"],
@@ -341,10 +344,47 @@ class Params:
        return params


+@dataclass
+class Metadata:
+    name: Optional[str] = None
+    author: Optional[str] = None
+    version: Optional[str] = None
+    url: Optional[str] = None
+    description: Optional[str] = None
+    licence: Optional[str] = None
+    source_url: Optional[str] = None
+    source_hf_repo: Optional[str] = None
+
+    @staticmethod
+    def load(metadata_path: Path) -> Metadata:
+        if metadata_path is None or not metadata_path.exists():
+            return Metadata()
+
+        with open(metadata_path, 'r') as file:
+            data = json.load(file)
+
+        # Create a new Metadata instance
+        metadata = Metadata()
+
+        # Assigning values to Metadata attributes if they exist in the JSON file
+        # This is based on LLM_KV_NAMES mapping in llama.cpp
+        metadata.name = data.get("general.name")
+        metadata.author = data.get("general.author")
+        metadata.version = data.get("general.version")
+        metadata.url = data.get("general.url")
+        metadata.description = data.get("general.description")
+        metadata.license = data.get("general.license")
+        metadata.source_url = data.get("general.source.url")
+        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
+
+        return metadata
+
+
 #
 # vocab
 #

+
@runtime_checkable
 class BaseVocab(Protocol):
    tokenizer_model: ClassVar[str]
@@ -462,7 +502,8 @@ class SentencePieceVocab(Vocab):
            # not found in alternate location either
            raise FileNotFoundError('Cannot find tokenizer.model')

-        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        self.sentencepiece_tokenizer = SentencePieceProcessor()
+        self.sentencepiece_tokenizer.LoadFromFile(str(fname_tokenizer))
        vocab_size = self.sentencepiece_tokenizer.vocab_size()

        new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
@@ -482,23 +523,23 @@ class SentencePieceVocab(Vocab):
    def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
        tokenizer = self.sentencepiece_tokenizer
        for i in range(tokenizer.vocab_size()):
-            piece = tokenizer.id_to_piece(i)
+            piece = tokenizer.IdToPiece(i)
            text         = piece.encode("utf-8")
-            score: float = tokenizer.get_score(i)
+            score: float = tokenizer.GetScore(i)

            toktype = gguf.TokenType.NORMAL
-            if tokenizer.is_unknown(i):
+            if tokenizer.IsUnknown(i):
                toktype = gguf.TokenType.UNKNOWN
-            if tokenizer.is_control(i):
+            if tokenizer.IsControl(i):
                toktype = gguf.TokenType.CONTROL

            # NOTE: I think added_tokens are user defined.
            # ref: https://github.com/google/sentencepiece/blob/master/src/sentencepiece_model.proto
            # if tokenizer.is_user_defined(i): toktype = gguf.TokenType.USER_DEFINED

-            if tokenizer.is_unused(i):
+            if tokenizer.IsUnused(i):
                toktype = gguf.TokenType.UNUSED
-            if tokenizer.is_byte(i):
+            if tokenizer.IsByte(i):
                toktype = gguf.TokenType.BYTE

            yield text, score, toktype
@@ -906,7 +947,7 @@ class LazyUnpickler(pickle.Unpickler):
    def rebuild_from_type_v2(func, new_type, args, state):
        return func(*args)

-    CLASSES = {
+    CLASSES: dict[tuple[str, str], type[LazyTensor] | LazyStorageKind] = {
        # getattr used here as a workaround for mypy not being smart enough to determine
        # the staticmethods have a __func__ attribute.
        ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
@@ -1062,21 +1103,42 @@ class OutputFile:
    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
        self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)

-    def add_meta_arch(self, params: Params) -> None:
+    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
+        # Metadata About The Model And Its Provenence
        name = "LLaMA"
-
-        # TODO: better logic to determine model name
-        if params.n_ctx == 4096:
-            name = "LLaMA v2"
+        if metadata is not None and metadata.name is not None:
+            name = metadata.name
        elif params.path_model is not None:
-            name = str(params.path_model.parent).split('/')[-1]
+            name = params.path_model.name
+        elif params.n_ctx == 4096:
+            # Heuristic detection of LLaMA v2 model
+            name = "LLaMA v2"

-        self.gguf.add_name                (name)
-        self.gguf.add_vocab_size          (params.n_vocab)
-        self.gguf.add_context_length      (params.n_ctx)
-        self.gguf.add_embedding_length    (params.n_embd)
-        self.gguf.add_block_count         (params.n_layer)
-        self.gguf.add_feed_forward_length (params.n_ff)
+        self.gguf.add_name(name)
+
+        if metadata is not None:
+            if metadata.author is not None:
+                self.gguf.add_author(metadata.author)
+            if metadata.version is not None:
+                self.gguf.add_version(metadata.version)
+            if metadata.url is not None:
+                self.gguf.add_url(metadata.url)
+            if metadata.description is not None:
+                self.gguf.add_description(metadata.description)
+            if metadata.licence is not None:
+                self.gguf.add_licence(metadata.licence)
+            if metadata.source_url is not None:
+                self.gguf.add_source_url(metadata.source_url)
+            if metadata.source_hf_repo is not None:
+                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
+
+    def add_meta_arch(self, params: Params) -> None:
+        # Metadata About The Neural Architecture Itself
+        self.gguf.add_vocab_size(params.n_vocab)
+        self.gguf.add_context_length(params.n_ctx)
+        self.gguf.add_embedding_length(params.n_embd)
+        self.gguf.add_block_count(params.n_layer)
+        self.gguf.add_feed_forward_length(params.n_ff)
        self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
        self.gguf.add_head_count          (params.n_head)
        self.gguf.add_head_count_kv       (params.n_head_kv)
@@ -1179,13 +1241,14 @@ class OutputFile:
    @staticmethod
    def write_vocab_only(
        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
    ) -> None:
        check_vocab_size(params, vocab, pad_vocab=pad_vocab)

        of = OutputFile(fname_out, endianess=endianess)

        # meta data
+        of.add_meta_model(params, metadata)
        of.add_meta_arch(params)
        of.add_meta_vocab(vocab)
        of.add_meta_special_vocab(svocab)
@@ -1212,12 +1275,14 @@ class OutputFile:
        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
        concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
        pad_vocab: bool = False,
+        metadata: Metadata = None,
    ) -> None:
        check_vocab_size(params, vocab, pad_vocab=pad_vocab)

        of = OutputFile(fname_out, endianess=endianess)

        # meta data
+        of.add_meta_model(params, metadata)
        of.add_meta_arch(params)
        if isinstance(vocab, Vocab):
            of.add_meta_vocab(vocab)
@@ -1253,6 +1318,37 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
    raise ValueError(f"Unexpected combination of types: {name_to_type}")


+def model_parameter_count(model: LazyModel) -> int:
+    total_model_parameters = 0
+    for i, (name, lazy_tensor) in enumerate(model.items()):
+        sum_weights_in_tensor = 1
+        for dim in lazy_tensor.shape:
+            sum_weights_in_tensor *= dim
+        total_model_parameters += sum_weights_in_tensor
+    return total_model_parameters
+
+
+def model_parameter_count_rounded_notation(model_params_count: int) -> str:
+    if model_params_count > 1e12 :
+        # Trillions Of Parameters
+        scaled_model_params = model_params_count * 1e-12
+        scale_suffix = "T"
+    elif model_params_count > 1e9 :
+        # Billions Of Parameters
+        scaled_model_params = model_params_count * 1e-9
+        scale_suffix = "B"
+    elif model_params_count > 1e6 :
+        # Millions Of Parameters
+        scaled_model_params = model_params_count * 1e-6
+        scale_suffix = "M"
+    else:
+        # Thousands Of Parameters
+        scaled_model_params = model_params_count * 1e-3
+        scale_suffix = "K"
+
+    return f"{round(scaled_model_params)}{scale_suffix}"
+
+
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
    return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
            for (name, tensor) in model.items()}
@@ -1432,13 +1528,35 @@ class VocabFactory:
        return vocab, special_vocab


-def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
-    namestr = {
-        GGMLFileType.AllF32:    "f32",
-        GGMLFileType.MostlyF16: "f16",
-        GGMLFileType.MostlyQ8_0:"q8_0",
+def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
+    quantization = {
+        GGMLFileType.AllF32:    "F32",
+        GGMLFileType.MostlyF16: "F16",
+        GGMLFileType.MostlyQ8_0: "Q8_0",
    }[file_type]
-    ret = model_paths[0].parent / f"ggml-model-{namestr}.gguf"
+
+    parameters = model_parameter_count_rounded_notation(model_params_count)
+
+    expert_count = ""
+    if params.n_experts is not None:
+        expert_count = f"{params.n_experts}x"
+
+    version = ""
+    if metadata is not None and metadata.version is not None:
+        version = f"-{metadata.version}"
+
+    name = "ggml-model"
+    if metadata is not None and metadata.name is not None:
+        name = metadata.name
+    elif params.path_model is not None:
+        name = params.path_model.name
+
+    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
+
+
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
+    ret = model_paths[0].parent / f"{default_filename}.gguf"
    if ret in model_paths:
        logger.error(
            f"Error: Default output path ({ret}) would overwrite the input. "
@@ -1476,17 +1594,30 @@ def main(args_in: list[str] | None = None) -> None:
    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
    parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
+    parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")

    args = parser.parse_args(args_in)

    if args.verbose:
        logging.basicConfig(level=logging.DEBUG)
-    elif args.dump_single or args.dump:
+    elif args.dump_single or args.dump or args.get_outfile:
        # Avoid printing anything besides the dump output
        logging.basicConfig(level=logging.WARNING)
    else:
        logging.basicConfig(level=logging.INFO)

+    metadata = Metadata.load(args.metadata)
+
+    if args.get_outfile:
+        model_plus = load_some_model(args.model)
+        params = Params.load(model_plus)
+        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = model_parameter_count(model_plus.model)
+        ftype   = pick_output_type(model, args.outtype)
+        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
+        return
+
    if args.no_vocab and args.vocab_only:
        raise ValueError("--vocab-only does not make sense with --no-vocab")

@@ -1500,6 +1631,9 @@ def main(args_in: list[str] | None = None) -> None:
    else:
        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)

+    model_params_count = model_parameter_count(model_plus.model)
+    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
+
    if args.dump:
        do_dump_model(model_plus)
        return
@@ -1553,7 +1687,7 @@ def main(args_in: list[str] | None = None) -> None:
                f_norm_eps = 1e-5,
            )
        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess=endianess, pad_vocab=args.pad_vocab)
+                                    endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
        logger.info(f"Wrote {outfile}")
        return

@@ -1566,13 +1700,13 @@ def main(args_in: list[str] | None = None) -> None:
    model   = convert_model_names(model, params, args.skip_unknown)
    ftype   = pick_output_type(model, args.outtype)
    model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)

    params.ftype = ftype
    logger.info(f"Writing {outfile}, format {ftype}")

    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
+                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab, metadata=metadata)
    logger.info(f"Wrote {outfile}")


--- a/docs/debugging-tests.md
+++ b/docs/debugging-tests.md
@@ -0,0 +1,104 @@
+# Debugging Tests Tips
+
+## How to run & execute or debug a specific test without anything else to keep the feedback loop short?
+
+There is a script called debug-test.sh in the scripts folder whose parameter takes a REGEX and an optional test number.
+
+For example, running the following command will output an interactive list from which you can select a test. It takes this form:
+
+`debug-test.sh [OPTION]... <test_regex> <test_number>`
+
+It will then build & run in the debugger for you.
+
+To just execute a test and get back a PASS or FAIL message run:
+
+```bash
+./scripts/debug-test.sh test-tokenizer
+```
+
+To test in GDB use the `-g` flag to enable gdb test mode.
+
+```bash
+./scripts/debug-test.sh -g test-tokenizer
+
+# Once in the debugger, i.e. at the chevrons prompt, setting a breakpoint could be as follows:
+>>> b main
+```
+
+To speed up the testing loop, if you know your test number you can just run it similar to below:
+
+```bash
+./scripts/debug-test.sh test 23
+```
+
+For further reference use `debug-test.sh -h` to print help.
+
+&nbsp;
+
+### How does the script work?
+If you want to be able to use the concepts contained in the script separately, the important ones are briefly outlined below.
+
+#### Step 1: Reset and Setup folder context
+
+From base of this repository, let's create `build-ci-debug` as our build context.
+
+```bash
+rm -rf build-ci-debug && mkdir build-ci-debug && cd build-ci-debug
+```
+
+#### Step 2: Setup Build Environment and Compile Test Binaries
+
+Setup and trigger a build under debug mode. You may adapt the arguments as needed, but in this case these are sane defaults.
+
+```bash
+cmake -DCMAKE_BUILD_TYPE=Debug -DLLAMA_CUDA=1 -DLLAMA_FATAL_WARNINGS=ON ..
+make -j
+```
+
+#### Step 3: Find all tests available that matches REGEX
+
+The output of this command will give you the command & arguments needed to run GDB.
+
+* `-R test-tokenizer` : looks for all the test files named `test-tokenizer*` (R=Regex)
+* `-N` : "show-only" disables test execution & shows test commands that you can feed to GDB.
+* `-V` : Verbose Mode
+
+```bash
+ctest -R "test-tokenizer" -V -N
+```
+
+This may return output similar to below (focusing on key lines to pay attention to):
+
+```bash
+...
+1: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+1: Working Directory: .
+Labels: main
+  Test  #1: test-tokenizer-0-llama-spm
+...
+4: Test command: ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-falcon.gguf"
+4: Working Directory: .
+Labels: main
+  Test  #4: test-tokenizer-0-falcon
+...
+```
+
+#### Step 4: Identify Test Command for Debugging
+
+So for test #1 above we can tell these two pieces of relevant information:
+* Test Binary: `~/llama.cpp/build-ci-debug/bin/test-tokenizer-0`
+* Test GGUF Model: `~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf`
+
+#### Step 5: Run GDB on test command
+
+Based on the ctest 'test command' report above we can then run a gdb session via this command below:
+
+```bash
+gdb --args ${Test Binary} ${Test GGUF Model}
+```
+
+Example:
+
+```bash
+gdb --args ~/llama.cpp/build-ci-debug/bin/test-tokenizer-0 "~/llama.cpp/tests/../models/ggml-vocab-llama-spm.gguf"
+```
--- a/examples/CMakeLists.txt
+++ b/examples/CMakeLists.txt
@@ -49,4 +49,7 @@ else()
        add_subdirectory(server)
    endif()
    add_subdirectory(export-lora)
+    if (LLAMA_RPC)
+        add_subdirectory(rpc)
+    endif()
 endif()
--- a/examples/batched/batched.cpp
+++ b/examples/batched/batched.cpp
@@ -48,7 +48,7 @@ int main(int argc, char ** argv) {
        params.prompt = "Hello my name is";
    }

-    process_escapes(params.prompt);
+    string_process_escapes(params.prompt);

    // init LLM

--- a/examples/convert-llama2c-to-ggml/README.md
+++ b/examples/convert-llama2c-to-ggml/README.md
@@ -2,7 +2,7 @@

 This example reads weights from project [llama2.c](https://github.com/karpathy/llama2.c) and saves them in ggml compatible format. The vocab that is available in `models/ggml-vocab.bin` is used by default.

-To convert the model first download the models from the [llma2.c](https://github.com/karpathy/llama2.c) repository:
+To convert the model first download the models from the [llama2.c](https://github.com/karpathy/llama2.c) repository:

 `$ make -j`

--- a/examples/embedding/embedding.cpp
+++ b/examples/embedding/embedding.cpp
@@ -49,6 +49,12 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
        }

        float * out = output + batch.seq_id[i][0] * n_embd;
+        //TODO: I would also add a parameter here to enable normalization or not.
+        /*fprintf(stdout, "unnormalized_embedding:");
+        for (int hh = 0; hh < n_embd; hh++) {
+            fprintf(stdout, "%9.6f ", embd[hh]);
+        }
+        fprintf(stdout, "\n");*/
        llama_embd_normalize(embd, out, n_embd);
    }
 }
@@ -74,7 +80,7 @@ int main(int argc, char ** argv) {

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
    }

    llama_backend_init();
@@ -101,7 +107,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
    }

    // split the prompt into lines
@@ -123,10 +129,12 @@ int main(int argc, char ** argv) {
        inputs.push_back(inp);
    }

-    // add SEP if not present
+    // check if the last token is SEP
+    // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
    for (auto & inp : inputs) {
        if (inp.empty() || inp.back() != llama_token_sep(model)) {
-            inp.push_back(llama_token_sep(model));
+            fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
+            fprintf(stderr, "%s:          'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
        }
    }

@@ -203,6 +211,7 @@ int main(int argc, char ** argv) {

    // clean up
    llama_print_timings(ctx);
+    llama_batch_free(batch);
    llama_free(ctx);
    llama_free_model(model);
    llama_backend_free();
--- a/examples/eval-callback/eval-callback.cpp
+++ b/examples/eval-callback/eval-callback.cpp
@@ -52,15 +52,15 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
                    float v;
                    if (type == GGML_TYPE_F16) {
-                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) data + i);
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
                    } else if (type == GGML_TYPE_F32) {
-                        v = *(float *) data + i;
+                        v = *(float *) &data[i];
                    } else if (type == GGML_TYPE_I32) {
-                        v = (float) *(int32_t *) data + i;
+                        v = (float) *(int32_t *) &data[i];
                    } else if (type == GGML_TYPE_I16) {
-                        v = (float) *(int16_t *) data + i;
+                        v = (float) *(int16_t *) &data[i];
                    } else if (type == GGML_TYPE_I8) {
-                        v = (float) *(int8_t *) data + i;
+                        v = (float) *(int8_t *) &data[i];
                    } else {
                        GGML_ASSERT(false);
                    }
@@ -152,7 +152,7 @@ int main(int argc, char ** argv) {

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
    }

    llama_backend_init();
@@ -176,7 +176,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
    }

    bool OK = run(ctx, params);
--- a/examples/finetune/finetune.cpp
+++ b/examples/finetune/finetune.cpp
@@ -563,8 +563,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
        // not capturing these, to silcence warnings
        const int rope_mode = 0;

-        return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+        return ggml_rope_ext(ctx,
+            t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0,
            rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
        );
    };
@@ -643,7 +643,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
        struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                         set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd_head, n_head_kv, n_batch);
        struct ggml_tensor * t16;
        if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                         set_name(t16, "t16");     assert_shape_4d(t16, n_embd_head, N, n_head, n_batch);
        } else {
            struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                  set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
            struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);           set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
--- a/examples/imatrix/imatrix.cpp
+++ b/examples/imatrix/imatrix.cpp
@@ -598,7 +598,7 @@ int main(int argc, char ** argv) {

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
    }

    sparams.dataset = params.prompt_file;
@@ -667,7 +667,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
    }

    bool OK = compute_imatrix(ctx, params, compute_ppl, from_chunk);
--- a/examples/infill/infill.cpp
+++ b/examples/infill/infill.cpp
@@ -50,9 +50,9 @@ static void write_logfile(
        return;
    }

-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();

-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
    if (!success) {
        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                __func__, params.logdir.c_str());
@@ -70,7 +70,7 @@ static void write_logfile(
    fprintf(logfile, "binary: infill\n");
    char model_desc[128];
    llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);

    fprintf(logfile, "\n");
    fprintf(logfile, "######################\n");
@@ -78,8 +78,8 @@ static void write_logfile(
    fprintf(logfile, "######################\n");
    fprintf(logfile, "\n");

-    dump_string_yaml_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);

    llama_dump_timing_info_yaml(logfile, ctx);
    fclose(logfile);
@@ -236,7 +236,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
    }
    const bool add_bos = llama_should_add_bos_token(model);
    GGML_ASSERT(llama_add_eos_token(model) != 1);
@@ -621,8 +621,8 @@ int main(int argc, char ** argv) {

                if (params.escape) {
                    //process escape sequences, for the initial prompt this is done in common.cpp when we load the params, but for the interactive mode we need to do it here
-                    process_escapes(params.input_prefix);
-                    process_escapes(params.input_suffix);
+                    string_process_escapes(params.input_prefix);
+                    string_process_escapes(params.input_suffix);
                }
                suff_rm_leading_spc = params.escape;
                if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
--- a/examples/llama-bench/README.md
+++ b/examples/llama-bench/README.md
@@ -26,16 +26,21 @@ options:
  -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
  -p, --n-prompt <n>                  (default: 512)
  -n, --n-gen <n>                     (default: 128)
-  -b, --batch-size <n>                (default: 512)
-  -ctk <t>, --cache-type-k <t>        (default: f16)
-  -ctv <t>, --cache-type-v <t>        (default: f16)
-  -t, --threads <n>                   (default: 112)
+  -pg <pp,tg>                         (default: 512,128)
+  -b, --batch-size <n>                (default: 2048)
+  -ub, --ubatch-size <n>              (default: 512)
+  -ctk, --cache-type-k <t>            (default: f16)
+  -ctv, --cache-type-v <t>            (default: f16)
+  -t, --threads <n>                   (default: 16)
  -ngl, --n-gpu-layers <n>            (default: 99)
  -sm, --split-mode <none|layer|row>  (default: layer)
  -mg, --main-gpu <i>                 (default: 0)
  -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -fa, --flash-attn <0|1>             (default: 0)
  -mmp, --mmap <0|1>                  (default: 1)
-  -ts, --tensor_split <ts0/ts1/..>    (default: 0)
+  --numa <distribute|isolate|numactl> (default: disabled)
+  -embd, --embeddings <0|1>           (default: 0)
+  -ts, --tensor-split <ts0/ts1/..>    (default: 0)
  -r, --repetitions <n>               (default: 5)
  -o, --output <csv|json|md|sql>      (default: md)
  -v, --verbose                       (default: 0)
@@ -43,10 +48,11 @@ options:
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```

-llama-bench can perform two types of tests:
+llama-bench can perform three types of tests:

 - Prompt processing (pp): processing a prompt in batches (`-p`)
 - Text generation (tg): generating a sequence of tokens (`-n`)
+- Prompt processing + text generation (pg): processing a prompt followed by generating a sequence of tokens (`-pg`)

 With the exception of `-r`, `-o` and `-v`, all options can be specified multiple times to run multiple tests. Each pp and tg test is run with all combinations of the specified options. To specify multiple values for an option, the values can be separated by commas (e.g. `-n 16,32`), or the option can be specified multiple times (e.g. `-n 16 -n 32`).

--- a/examples/llama-bench/llama-bench.cpp
+++ b/examples/llama-bench/llama-bench.cpp
@@ -161,10 +161,17 @@ static const char * split_mode_str(llama_split_mode mode) {
    }
 }

+static std::string pair_str(const std::pair<int, int> & p) {
+    static char buf[32];
+    snprintf(buf, sizeof(buf), "%d,%d", p.first, p.second);
+    return buf;
+}
+
 struct cmd_params {
    std::vector<std::string> model;
    std::vector<int> n_prompt;
    std::vector<int> n_gen;
+    std::vector<std::pair<int, int>> n_pg;
    std::vector<int> n_batch;
    std::vector<int> n_ubatch;
    std::vector<ggml_type> type_k;
@@ -188,11 +195,12 @@ static const cmd_params cmd_params_defaults = {
    /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
    /* n_prompt      */ {512},
    /* n_gen         */ {128},
+    /* n_pg          */ {},
    /* n_batch       */ {2048},
    /* n_ubatch      */ {512},
    /* type_k        */ {GGML_TYPE_F16},
    /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {get_math_cpu_count()},
+    /* n_threads     */ {cpu_get_num_math()},
    /* n_gpu_layers  */ {99},
    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
    /* main_gpu      */ {0},
@@ -215,10 +223,11 @@ static void print_usage(int /* argc */, char ** argv) {
    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -pg <pp,tg>                         (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
-    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -ub, --ubatch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
+    printf("  -ctk, --cache-type-k <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv, --cache-type-v <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
@@ -304,6 +313,17 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
            }
            auto p = split<int>(argv[i], split_delim);
            params.n_gen.insert(params.n_gen.end(), p.begin(), p.end());
+        } else if (arg == "-pg") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<std::string>(argv[i], ',');
+            if (p.size() != 2) {
+                invalid_param = true;
+                break;
+            }
+            params.n_pg.push_back({std::stoi(p[0]), std::stoi(p[1])});
        } else if (arg == "-b" || arg == "--batch-size") {
            if (++i >= argc) {
                invalid_param = true;
@@ -493,6 +513,7 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
    if (params.model.empty())        { params.model = cmd_params_defaults.model; }
    if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
    if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
+    if (params.n_pg.empty())         { params.n_pg = cmd_params_defaults.n_pg; }
    if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
@@ -632,6 +653,31 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
            };
            instances.push_back(instance);
        }
+
+        for (const auto & n_pg : params.n_pg) {
+            if (n_pg.first == 0 && n_pg.second == 0) {
+                continue;
+            }
+            cmd_params_instance instance = {
+                /* .model        = */ m,
+                /* .n_prompt     = */ n_pg.first,
+                /* .n_gen        = */ n_pg.second,
+                /* .n_batch      = */ nb,
+                /* .n_ubatch     = */ nub,
+                /* .type_k       = */ tk,
+                /* .type_v       = */ tv,
+                /* .n_threads    = */ nt,
+                /* .n_gpu_layers = */ nl,
+                /* .split_mode   = */ sm,
+                /* .main_gpu     = */ mg,
+                /* .no_kv_offload= */ nkvo,
+                /* .flash_attn   = */ fa,
+                /* .tensor_split = */ ts,
+                /* .use_mmap     = */ mmp,
+                /* .embeddings   = */ embd,
+            };
+            instances.push_back(instance);
+        }
    }

    return instances;
@@ -965,6 +1011,9 @@ struct markdown_printer : public printer {
        if (field == "n_gpu_layers") {
            return 3;
        }
+        if (field == "test") {
+            return 13;
+        }

        int width = std::max((int)field.length(), 10);

@@ -1091,12 +1140,11 @@ struct markdown_printer : public printer {
                value = test::get_backend();
            } else if (field == "test") {
                if (t.n_prompt > 0 && t.n_gen == 0) {
-                    snprintf(buf, sizeof(buf), "pp %d", t.n_prompt);
+                    snprintf(buf, sizeof(buf), "pp%d", t.n_prompt);
                } else if (t.n_gen > 0 && t.n_prompt == 0) {
-                    snprintf(buf, sizeof(buf), "tg %d", t.n_gen);
+                    snprintf(buf, sizeof(buf), "tg%d", t.n_gen);
                } else {
-                    assert(false);
-                    exit(1);
+                    snprintf(buf, sizeof(buf), "pp%d+tg%d", t.n_prompt, t.n_gen);
                }
                value = buf;
            } else if (field == "t/s") {
@@ -1297,6 +1345,7 @@ int main(int argc, char ** argv) {
            llama_kv_cache_clear(ctx);

            uint64_t t_start = get_time_ns();
+
            if (t.n_prompt > 0) {
                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
            }
--- a/examples/llama.android/app/build.gradle.kts
+++ b/examples/llama.android/app/build.gradle.kts
@@ -7,8 +7,6 @@ android {
    namespace = "com.example.llama"
    compileSdk = 34

-    ndkVersion = "26.1.10909125"
-
    defaultConfig {
        applicationId = "com.example.llama"
        minSdk = 33
@@ -20,17 +18,6 @@ android {
        vectorDrawables {
            useSupportLibrary = true
        }
-        ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
    }

    buildTypes {
@@ -55,17 +42,6 @@ android {
    composeOptions {
        kotlinCompilerExtensionVersion = "1.5.1"
    }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
 }

 dependencies {
@@ -78,6 +54,7 @@ dependencies {
    implementation("androidx.compose.ui:ui-graphics")
    implementation("androidx.compose.ui:ui-tooling-preview")
    implementation("androidx.compose.material3:material3")
+    implementation(project(":llama"))
    testImplementation("junit:junit:4.13.2")
    androidTestImplementation("androidx.test.ext:junit:1.1.5")
    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
--- a/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt
+++ b/examples/llama.android/app/src/main/java/com/example/llama/MainViewModel.kt
@@ -1,5 +1,6 @@
 package com.example.llama

+import android.llama.cpp.LLamaAndroid
 import android.util.Log
 import androidx.compose.runtime.getValue
 import androidx.compose.runtime.mutableStateOf
@@ -9,7 +10,7 @@ import androidx.lifecycle.viewModelScope
 import kotlinx.coroutines.flow.catch
 import kotlinx.coroutines.launch

-class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
+class MainViewModel(private val llamaAndroid: LLamaAndroid = LLamaAndroid.instance()): ViewModel() {
    companion object {
        @JvmStatic
        private val NanosPerSecond = 1_000_000_000.0
@@ -28,7 +29,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {

        viewModelScope.launch {
            try {
-                llm.unload()
+                llamaAndroid.unload()
            } catch (exc: IllegalStateException) {
                messages += exc.message!!
            }
@@ -44,7 +45,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
        messages += ""

        viewModelScope.launch {
-            llm.send(text)
+            llamaAndroid.send(text)
                .catch {
                    Log.e(tag, "send() failed", it)
                    messages += it.message!!
@@ -57,7 +58,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
        viewModelScope.launch {
            try {
                val start = System.nanoTime()
-                val warmupResult = llm.bench(pp, tg, pl, nr)
+                val warmupResult = llamaAndroid.bench(pp, tg, pl, nr)
                val end = System.nanoTime()

                messages += warmupResult
@@ -70,7 +71,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
                    return@launch
                }

-                messages += llm.bench(512, 128, 1, 3)
+                messages += llamaAndroid.bench(512, 128, 1, 3)
            } catch (exc: IllegalStateException) {
                Log.e(tag, "bench() failed", exc)
                messages += exc.message!!
@@ -81,7 +82,7 @@ class MainViewModel(private val llm: Llm = Llm.instance()): ViewModel() {
    fun load(pathToModel: String) {
        viewModelScope.launch {
            try {
-                llm.load(pathToModel)
+                llamaAndroid.load(pathToModel)
                messages += "Loaded $pathToModel"
            } catch (exc: IllegalStateException) {
                Log.e(tag, "load() failed", exc)
--- a/examples/llama.android/build.gradle.kts
+++ b/examples/llama.android/build.gradle.kts
@@ -2,4 +2,5 @@
 plugins {
    id("com.android.application") version "8.2.0" apply false
    id("org.jetbrains.kotlin.android") version "1.9.0" apply false
+    id("com.android.library") version "8.2.0" apply false
 }
--- a/examples/llama.android/llama/.gitignore
+++ b/examples/llama.android/llama/.gitignore
@@ -0,0 +1 @@
+/build
--- a/examples/llama.android/llama/CMakeLists.txt
+++ b/examples/llama.android/llama/CMakeLists.txt
@@ -0,0 +1,55 @@
+
+# For more information about using CMake with Android Studio, read the
+# documentation: https://d.android.com/studio/projects/add-native-code.html.
+# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
+
+# Sets the minimum CMake version required for this project.
+cmake_minimum_required(VERSION 3.22.1)
+
+# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
+# Since this is the top level CMakeLists.txt, the project name is also accessible
+# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
+# build script scope).
+project("llama-android")
+
+## Fetch latest llama.cpp from GitHub
+#include(FetchContent)
+#FetchContent_Declare(
+#        llama
+#        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
+#        GIT_TAG        master
+#)
+#
+## Also provides "common"
+#FetchContent_MakeAvailable(llama)
+
+# llama.cpp CI uses the code from the current branch
+# ref: https://github.com/ggerganov/llama.cpp/pull/7341#issuecomment-2117617700
+add_subdirectory(../../../../../../ build-llama)
+
+# Creates and names a library, sets it as either STATIC
+# or SHARED, and provides the relative paths to its source code.
+# You can define multiple libraries, and CMake builds them for you.
+# Gradle automatically packages shared libraries with your APK.
+#
+# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
+# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
+# is preferred for the same purpose.
+#
+# In order to load a library into your app from Java/Kotlin, you must call
+# System.loadLibrary() and pass the name of the library defined here;
+# for GameActivity/NativeActivity derived applications, the same library name must be
+# used in the AndroidManifest.xml file.
+add_library(${CMAKE_PROJECT_NAME} SHARED
+    # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)
+
+# Specifies libraries CMake should link to your target library. You
+# can link libraries from various origins, such as libraries defined in this
+# build script, prebuilt third-party libraries, or Android system libraries.
+target_link_libraries(${CMAKE_PROJECT_NAME}
+    # List libraries link to the target library
+    llama
+    common
+    android
+    log)
--- a/examples/llama.android/llama/build.gradle.kts
+++ b/examples/llama.android/llama/build.gradle.kts
@@ -0,0 +1,68 @@
+plugins {
+    id("com.android.library")
+    id("org.jetbrains.kotlin.android")
+}
+
+android {
+    namespace = "android.llama.cpp"
+    compileSdk = 34
+
+    defaultConfig {
+        minSdk = 33
+
+        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
+        consumerProguardFiles("consumer-rules.pro")
+        ndk {
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
+        }
+        externalNativeBuild {
+            cmake {
+                arguments += "-DCMAKE_BUILD_TYPE=Release"
+                cppFlags += listOf()
+                arguments += listOf()
+
+                cppFlags("")
+            }
+        }
+    }
+
+    buildTypes {
+        release {
+            isMinifyEnabled = false
+            proguardFiles(
+                getDefaultProguardFile("proguard-android-optimize.txt"),
+                "proguard-rules.pro"
+            )
+        }
+    }
+    externalNativeBuild {
+        cmake {
+            path("src/main/cpp/CMakeLists.txt")
+            version = "3.22.1"
+        }
+    }
+    compileOptions {
+        sourceCompatibility = JavaVersion.VERSION_1_8
+        targetCompatibility = JavaVersion.VERSION_1_8
+    }
+    kotlinOptions {
+        jvmTarget = "1.8"
+    }
+
+    packaging {
+        resources {
+            excludes += "/META-INF/{AL2.0,LGPL2.1}"
+        }
+    }
+}
+
+dependencies {
+
+    implementation("androidx.core:core-ktx:1.12.0")
+    implementation("androidx.appcompat:appcompat:1.6.1")
+    implementation("com.google.android.material:material:1.11.0")
+    testImplementation("junit:junit:4.13.2")
+    androidTestImplementation("androidx.test.ext:junit:1.1.5")
+    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
+}
--- a/examples/llama.android/llama/consumer-rules.pro
+++ b/examples/llama.android/llama/consumer-rules.pro
--- a/examples/llama.android/llama/proguard-rules.pro
+++ b/examples/llama.android/llama/proguard-rules.pro
@@ -0,0 +1,21 @@
+# Add project specific ProGuard rules here.
+# You can control the set of applied configuration files using the
+# proguardFiles setting in build.gradle.
+#
+# For more details, see
+#   http://developer.android.com/guide/developing/tools/proguard.html
+
+# If your project uses WebView with JS, uncomment the following
+# and specify the fully qualified class name to the JavaScript interface
+# class:
+#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
+#   public *;
+#}
+
+# Uncomment this to preserve the line number information for
+# debugging stack traces.
+#-keepattributes SourceFile,LineNumberTable
+
+# If you keep the line number information, uncomment this to
+# hide the original source file name.
+#-renamesourcefileattribute SourceFile
--- a/examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt
+++ b/examples/llama.android/llama/src/androidTest/java/android/llama/cpp/ExampleInstrumentedTest.kt
@@ -0,0 +1,24 @@
+package android.llama.cpp
+
+import androidx.test.platform.app.InstrumentationRegistry
+import androidx.test.ext.junit.runners.AndroidJUnit4
+
+import org.junit.Test
+import org.junit.runner.RunWith
+
+import org.junit.Assert.*
+
+/**
+ * Instrumented test, which will execute on an Android device.
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+@RunWith(AndroidJUnit4::class)
+class ExampleInstrumentedTest {
+    @Test
+    fun useAppContext() {
+        // Context of the app under test.
+        val appContext = InstrumentationRegistry.getInstrumentation().targetContext
+        assertEquals("android.llama.cpp.test", appContext.packageName)
+    }
+}
--- a/examples/llama.android/llama/src/main/AndroidManifest.xml
+++ b/examples/llama.android/llama/src/main/AndroidManifest.xml
@@ -0,0 +1,4 @@
+<?xml version="1.0" encoding="utf-8"?>
+<manifest xmlns:android="http://schemas.android.com/apk/res/android">
+
+</manifest>
--- a/examples/llama.android/llama/src/main/cpp/CMakeLists.txt
+++ b/examples/llama.android/llama/src/main/cpp/CMakeLists.txt
@@ -1,4 +1,3 @@
-
 # For more information about using CMake with Android Studio, read the
 # documentation: https://d.android.com/studio/projects/add-native-code.html.
 # For more examples on how to use CMake, see https://github.com/android/ndk-samples.
@@ -36,15 +35,15 @@ FetchContent_MakeAvailable(llama)
 # for GameActivity/NativeActivity derived applications, the same library name must be
 # used in the AndroidManifest.xml file.
 add_library(${CMAKE_PROJECT_NAME} SHARED
-    # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
+        # List C/C++ source files with relative paths to this CMakeLists.txt.
+        llama-android.cpp)

 # Specifies libraries CMake should link to your target library. You
 # can link libraries from various origins, such as libraries defined in this
 # build script, prebuilt third-party libraries, or Android system libraries.
 target_link_libraries(${CMAKE_PROJECT_NAME}
-    # List libraries link to the target library
-    llama
-    common
-    android
-    log)
+        # List libraries link to the target library
+        llama
+        common
+        android
+        log)
--- a/examples/llama.android/llama/src/main/cpp/llama-android.cpp
+++ b/examples/llama.android/llama/src/main/cpp/llama-android.cpp
@@ -81,7 +81,7 @@ static void log_callback(ggml_log_level level, const char * fmt, void * data) {

 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
+Java_android_llama_cpp_LLamaAndroid_load_1model(JNIEnv *env, jobject, jstring filename) {
    llama_model_params model_params = llama_model_default_params();

    auto path_to_model = env->GetStringUTFChars(filename, 0);
@@ -101,13 +101,13 @@ Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
+Java_android_llama_cpp_LLamaAndroid_free_1model(JNIEnv *, jobject, jlong model) {
    llama_free_model(reinterpret_cast<llama_model *>(model));
 }

 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
+Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmodel) {
    auto model = reinterpret_cast<llama_model *>(jmodel);

    if (!model) {
@@ -139,25 +139,25 @@ Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_free_1context(JNIEnv *, jobject, jlong context) {
    llama_free(reinterpret_cast<llama_context *>(context));
 }

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1free(JNIEnv *, jobject) {
    llama_backend_free();
 }

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_log_1to_1android(JNIEnv *, jobject) {
    llama_log_set(log_callback, NULL);
 }

 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
+Java_android_llama_cpp_LLamaAndroid_bench_1model(
        JNIEnv *env,
        jobject,
        jlong context_pointer,
@@ -271,13 +271,13 @@ Java_com_example_llama_Llm_bench_1model(

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
 }

 extern "C"
 JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
+Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {

    // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.

@@ -313,19 +313,19 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
    llama_backend_init();
 }

 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
+Java_android_llama_cpp_LLamaAndroid_system_1info(JNIEnv *env, jobject) {
    return env->NewStringUTF(llama_print_system_info());
 }

 extern "C"
 JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
+Java_android_llama_cpp_LLamaAndroid_completion_1init(
        JNIEnv *env,
        jobject,
        jlong context_pointer,
@@ -376,7 +376,7 @@ Java_com_example_llama_Llm_completion_1init(

 extern "C"
 JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
+Java_android_llama_cpp_LLamaAndroid_completion_1loop(
        JNIEnv * env,
        jobject,
        jlong context_pointer,
@@ -438,6 +438,6 @@ Java_com_example_llama_Llm_completion_1loop(

 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
+Java_android_llama_cpp_LLamaAndroid_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
 }
--- a/examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt
+++ b/examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt
@@ -1,4 +1,4 @@
-package com.example.llama
+package android.llama.cpp

 import android.util.Log
 import kotlinx.coroutines.CoroutineDispatcher
@@ -10,7 +10,7 @@ import kotlinx.coroutines.withContext
 import java.util.concurrent.Executors
 import kotlin.concurrent.thread

-class Llm {
+class LLamaAndroid {
    private val tag: String? = this::class.simpleName

    private val threadLocalState: ThreadLocal<State> = ThreadLocal.withInitial { State.Idle }
@@ -165,8 +165,8 @@ class Llm {
        }

        // Enforce only one instance of Llm.
-        private val _instance: Llm = Llm()
+        private val _instance: LLamaAndroid = LLamaAndroid()

-        fun instance(): Llm = _instance
+        fun instance(): LLamaAndroid = _instance
    }
 }
--- a/examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt
+++ b/examples/llama.android/llama/src/test/java/android/llama/cpp/ExampleUnitTest.kt
@@ -0,0 +1,17 @@
+package android.llama.cpp
+
+import org.junit.Test
+
+import org.junit.Assert.*
+
+/**
+ * Example local unit test, which will execute on the development machine (host).
+ *
+ * See [testing documentation](http://d.android.com/tools/testing).
+ */
+class ExampleUnitTest {
+    @Test
+    fun addition_isCorrect() {
+        assertEquals(4, 2 + 2)
+    }
+}
--- a/examples/llama.android/settings.gradle.kts
+++ b/examples/llama.android/settings.gradle.kts
@@ -15,3 +15,4 @@ dependencyResolutionManagement {

 rootProject.name = "LlamaAndroid"
 include(":app")
+include(":llama")
--- a/examples/llava/clip.cpp
+++ b/examples/llava/clip.cpp
@@ -104,6 +104,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
@@ -425,6 +426,7 @@ struct clip_vision_model {
    // embeddings
    struct ggml_tensor * class_embedding;
    struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_bias;
    struct ggml_tensor * position_embeddings;

    struct ggml_tensor * pre_ln_w;
@@ -501,6 +503,11 @@ struct clip_ctx {
    bool use_gelu = false;
    int32_t ftype = 1;

+    bool has_class_embedding = true;
+    bool has_pre_norm = true;
+    bool has_post_norm = false;
+    bool has_patch_bias = false;
+
    struct gguf_context * ctx_gguf;
    struct ggml_context * ctx_data;

@@ -526,7 +533,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
    const int patch_size           = hparams.patch_size;
    const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
    const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
-    const int num_positions        = num_patches + 1;
+    const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
    const int hidden_size          = hparams.hidden_size;
    const int n_head               = hparams.n_head;
    const int d_head               = hidden_size / n_head;
@@ -557,16 +564,23 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
    inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
    inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));

+    if (ctx->has_patch_bias) {
+        // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
    // concat class_embeddings and patch_embeddings
-    struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-    ggml_set_name(embeddings, "embeddings");
-    ggml_set_input(embeddings);
+    struct ggml_tensor * embeddings = inp;
+    if (ctx->has_class_embedding) {
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+        ggml_set_name(embeddings, "embeddings");
+        ggml_set_input(embeddings);
+        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+        embeddings = ggml_acc(ctx0, embeddings, inp,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }

-    embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-
-    embeddings = ggml_acc(ctx0, embeddings, inp,
-            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);

    struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
    ggml_set_name(positions, "positions");
@@ -576,7 +590,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));

    // pre-layernorm
-    {
+    if (ctx->has_pre_norm) {
        embeddings = ggml_norm(ctx0, embeddings, eps);
        ggml_set_name(embeddings, "pre_ln");

@@ -664,6 +678,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
        embeddings = cur;
    }

+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
    // llava projector
    {
        embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@@ -1148,12 +1170,39 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {

        }

+        try {
+            vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+            new_clip->has_class_embedding = true;
+        } catch (const std::exception& e) {
+            new_clip->has_class_embedding = false;
+        }
+
+        try {
+            vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+            vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+            new_clip->has_pre_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_pre_norm = false;
+        }
+
+        try {
+            vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
+            vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
+            new_clip->has_post_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_post_norm = false;
+        }
+
+        try {
+            vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
+            new_clip->has_patch_bias = true;
+        } catch (std::exception & e) {
+            new_clip->has_patch_bias = false;
+        }
+
        try {
            vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
-            vision_model.class_embedding     = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
            vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
-            vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
-            vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
        } catch(const std::exception& e) {
            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
        }
@@ -1797,7 +1846,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    const int image_size    = hparams.image_size;
    const int patch_size    = hparams.patch_size;
    const int num_patches   = ((image_size / patch_size) * (image_size / patch_size));
-    const int num_positions = num_patches + 1;
+    const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);

    {
        struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
@@ -1825,12 +1874,14 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
    }

    {
-        struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
+        if (ctx->has_class_embedding) {
+            struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");

-        void* zero_mem = malloc(ggml_nbytes(embeddings));
-        memset(zero_mem, 0, ggml_nbytes(embeddings));
-        ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
-        free(zero_mem);
+            void* zero_mem = malloc(ggml_nbytes(embeddings));
+            memset(zero_mem, 0, ggml_nbytes(embeddings));
+            ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+            free(zero_mem);
+        }
    }

    {
--- a/examples/llava/llava-cli.cpp
+++ b/examples/llava/llava-cli.cpp
@@ -189,6 +189,11 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
    LOG_TEE("\n");

    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
+    if (!ctx_sampling) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
    std::string response = "";
    for (int i = 0; i < max_tgt_len; i++) {
        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
@@ -285,7 +290,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS

    if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        gpt_print_usage(argc, argv, params);
+        gpt_params_print_usage(argc, argv, params);
        show_additional_info(argc, argv);
        return 1;
    }
@@ -295,14 +300,10 @@ int main(int argc, char ** argv) {
        return 1;
    }

-    for (auto & image : params.image) {
+    if (prompt_contains_image(params.prompt)) {
        auto ctx_llava = llava_init_context(&params, model);

-        auto image_embed = load_image(ctx_llava, &params, image);
-        if (!image_embed) {
-            std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
-            return 1;
-        }
+        auto image_embed = load_image(ctx_llava, &params, "");

        // process the prompt
        process_prompt(ctx_llava, image_embed, &params, params.prompt);
@@ -311,7 +312,26 @@ int main(int argc, char ** argv) {
        llava_image_embed_free(image_embed);
        ctx_llava->model = NULL;
        llava_free(ctx_llava);
+    } else {
+        for (auto & image : params.image) {
+            auto ctx_llava = llava_init_context(&params, model);
+
+            auto image_embed = load_image(ctx_llava, &params, image);
+            if (!image_embed) {
+                std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
+                return 1;
+            }
+
+            // process the prompt
+            process_prompt(ctx_llava, image_embed, &params, params.prompt);
+
+            llama_print_timings(ctx_llava->ctx_llama);
+            llava_image_embed_free(image_embed);
+            ctx_llava->model = NULL;
+            llava_free(ctx_llava);
+        }
    }
+
    llama_free_model(model);

    return 0;
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@@ -88,7 +88,6 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
 // Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
 static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
    struct {
-        struct ggml_tensor * newline;
        struct ggml_context * ctx;
    } model;

@@ -150,20 +149,6 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>

    model.ctx = ggml_init(params);

-    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
-    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
-    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
-        if (newline_tmp->buffer == NULL) {
-            LOG_TEE("newline_tmp tensor buffer is NULL\n");
-        }
-        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
-    } else {
-        model.newline->data = newline_tmp->data;
-        if (model.newline->data == NULL) {
-            LOG_TEE("newline_tmp tensor data is NULL\n");
-        }
-    }
-
    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
    // fill it with the image embeddings, ignoring the base
--- a/examples/lookahead/lookahead.cpp
+++ b/examples/lookahead/lookahead.cpp
@@ -174,7 +174,7 @@ int main(int argc, char ** argv) {
        // debug
        if (dump_kv_cache) {
            llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
        }

        // build the mask from https://lmsys.org/blog/2023-11-21-lookahead-decoding/
--- a/examples/lookup/lookup.cpp
+++ b/examples/lookup/lookup.cpp
@@ -121,7 +121,7 @@ int main(int argc, char ** argv){
        // debug
        if (dump_kv_cache) {
            llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
        }

        // print current draft sequence
--- a/examples/main/README.md
+++ b/examples/main/README.md
@@ -325,3 +325,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
+
+-   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.
--- a/examples/main/main.cpp
+++ b/examples/main/main.cpp
@@ -60,9 +60,9 @@ static void write_logfile(
        return;
    }

-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();

-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
    if (!success) {
        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                __func__, params.logdir.c_str());
@@ -80,7 +80,7 @@ static void write_logfile(
    fprintf(logfile, "binary: main\n");
    char model_desc[128];
    llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, input_tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, input_tokens, model_desc);

    fprintf(logfile, "\n");
    fprintf(logfile, "######################\n");
@@ -88,8 +88,8 @@ static void write_logfile(
    fprintf(logfile, "######################\n");
    fprintf(logfile, "\n");

-    dump_string_yaml_multiline(logfile, "output", output.c_str());
-    dump_vector_int_yaml(logfile, "output_tokens", output_tokens);
+    yaml_dump_string_multiline(logfile, "output", output.c_str());
+    yaml_dump_vector_int(logfile, "output_tokens", output_tokens);

    llama_dump_timing_info_yaml(logfile, ctx);
    fclose(logfile);
@@ -181,7 +181,7 @@ int main(int argc, char ** argv) {

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
    }

    LOG("%s: llama backend init\n", __func__);
@@ -219,7 +219,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        LOG_TEE("\n");
-        LOG_TEE("%s\n", get_system_info(params).c_str());
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
    }

    std::string path_session = params.path_prompt_cache;
@@ -474,12 +474,12 @@ int main(int argc, char ** argv) {
    LOG_TEE("\n\n");

    if (params.interactive) {
-        const char *control_message;
+        const char * control_message;
        if (params.multiline_input) {
-            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
+            control_message = " - To return control to the AI, end your input with '\\'.\n"
                              " - To return control without starting a new line, end your input with '/'.\n";
        } else {
-            control_message = " - Press Return to return control to LLaMa.\n"
+            control_message = " - Press Return to return control to the AI.\n"
                              " - To return control without starting a new line, end your input with '/'.\n"
                              " - If you want to submit another line, end your input with '\\'.\n";
        }
@@ -523,6 +523,10 @@ int main(int argc, char ** argv) {
    }

    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+    if (!ctx_sampling) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }

    while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
        // predict
@@ -703,7 +707,7 @@ int main(int argc, char ** argv) {

            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);

-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);

            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());

@@ -724,7 +728,7 @@ int main(int argc, char ** argv) {

                // push the prompt in the sampling context in order to apply repetition penalties later
                // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);

                ++n_consumed;
                if ((int) embd.size() >= params.n_batch) {
@@ -875,11 +879,11 @@ int main(int argc, char ** argv) {
                        embd_inp.insert(embd_inp.end(), cml_pfx.begin(), cml_pfx.end());
                    }
                    if (params.escape) {
-                        process_escapes(buffer);
+                        string_process_escapes(buffer);
                    }

                    const auto line_pfx = ::llama_tokenize(ctx, params.input_prefix, false, true);
-                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, false);
+                    const auto line_inp = ::llama_tokenize(ctx, buffer,              false, params.interactive_specials);
                    const auto line_sfx = ::llama_tokenize(ctx, params.input_suffix, false, true);

                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
--- a/examples/parallel/parallel.cpp
+++ b/examples/parallel/parallel.cpp
@@ -210,7 +210,7 @@ int main(int argc, char ** argv) {
    while (true) {
        if (dump_kv_cache) {
            llama_kv_cache_view_update(ctx, &kvc_view);
-            dump_kv_cache_view_seqs(kvc_view, 40);
+            llama_kv_cache_dump_view_seqs(kvc_view, 40);
        }

        llama_batch_clear(batch);
--- a/examples/perplexity/README.md
+++ b/examples/perplexity/README.md
@@ -7,6 +7,8 @@ Also note that finetunes typically result in a higher perplexity value even thou

 Within llama.cpp the perplexity of base models is used primarily to judge the quality loss from e.g. quantized models vs. FP16.
 The convention among contributors is to use the Wikitext-2 test set for testing unless noted otherwise (can be obtained with `scripts/get-wikitext-2.sh`).
+When numbers are listed all command line arguments and compilation options are left at their defaults unless noted otherwise.
+llama.cpp numbers are **not** directly comparable to those of other projects because the exact values depend strongly on the implementation details.

 By default only the mean perplexity value and the corresponding uncertainty is calculated.
 The uncertainty is determined empirically by assuming a Gaussian distribution of the "correct" logits per and then applying error propagation.
@@ -32,12 +34,21 @@ In addition to the KL divergence the following statistics are calculated with `-

 ## LLaMA 3 8b Scoreboard

-Results are sorted by Kullback-Leibler divergence relative to FP16.
+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
+Results were generated using the CUDA backend and are sorted by Kullback-Leibler divergence relative to FP16.
 The "WT" importance matrices were created using varying numbers of Wikitext tokens and can be found [here](https://huggingface.co/JohannesGaessler/llama.cpp_importance_matrices/blob/main/imatrix-llama_3-8b-f16-2.7m_tokens.dat).
+Note: the FP16 logits used for the calculation of all metrics other than perplexity are stored in a binary file between runs.
+In order to save space this file does **not** contain the exact same FP32 logits but instead casts them to 16 bit unsigned integers (with some scaling).
+So the "f16" results are to be understood as the difference resulting only from this downcast.

 | Quantization | imatrix | Model size [GiB] | PPL                    | ΔPPL                   | KLD                   | Mean Δp           | RMS Δp           |
 |--------------|---------|------------------|------------------------|------------------------|-----------------------|-------------------|------------------|
-| f16          | None    |            14.97 | 6.233160 ±   0.037828  | -                      | -                     | -                 | -                |
+| f16          | None    |            14.97 | 6.233160 ±   0.037828  | 0.001524 ±   0.000755  | 0.000551 ±   0.000002 |  0.001 ± 0.002 %  | 0.787 ± 0.004 %  |
 | q8_0         | None    |             7.96 | 6.234284 ±   0.037878  | 0.002650 ±   0.001006  | 0.001355 ±   0.000006 | -0.019 ± 0.003 %  | 1.198 ± 0.007 %  |
 | q6_K         | None    |             6.14 | 6.253382 ±   0.038078  | 0.021748 ±   0.001852  | 0.005452 ±   0.000035 | -0.007 ± 0.006 %  | 2.295 ± 0.019 %  |
 | q5_K_M       | None    |             5.33 | 6.288607 ±   0.038338  | 0.056974 ±   0.002598  | 0.010762 ±   0.000079 | -0.114 ± 0.008 %  | 3.160 ± 0.031 %  |
@@ -89,6 +100,12 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence

 ## LLaMA 2 vs. LLaMA 3 Quantization comparison

+| Revision | f364eb6f           |
+|:---------|:-------------------|
+| Backend  | CUDA               |
+| CPU      | AMD Epyc 7742      |
+| GPU      | 1x NVIDIA RTX 4090 |
+
 | Metric          |          L2 7b q2_K |          L3 8b q2_K |        L2 7b q4_K_M |        L3 8b q4_K_M |          L2 7b q6_K |          L3 8b q6_K |          L2 7b q8_0 |          L3 8b q8_0 |
 |-----------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|---------------------|
 | Mean PPL        | 5.794552 ± 0.032298 | 9.751568 ± 0.063312 | 5.877078 ± 0.032781 | 6.407115 ± 0.039119 | 5.808494 ± 0.032425 | 6.253382 ± 0.038078 | 5.798542 ± 0.032366 | 6.234284 ± 0.037878 |
@@ -107,6 +124,50 @@ K-quants score better on mean Δp than the legacy quants than e.g. KL divergence
 | RMS Δp          |     9.762 ± 0.053 % |    21.421 ± 0.079 % |     3.252 ± 0.024 % |     5.519 ± 0.050 % |     1.339 ± 0.010 % |     2.295 ± 0.019 % |     0.618 ± 0.011 % |     1.198 ± 0.007 % |
 | Same top p      |    85.584 ± 0.086 % |    71.138 ± 0.119 % |    94.665 ± 0.055 % |    91.901 ± 0.072 % |    97.520 ± 0.038 % |    96.031 ± 0.051 % |    98.846 ± 0.026 % |    97.674 ± 0.040 % |

+## LLaMA 3 BF16 vs. FP16 comparison
+
+| Revision | 83330d8c      |
+|:---------|:--------------|
+| Backend  | CPU           |
+| CPU      | AMD Epyc 7742 |
+| GPU      | N/A           |
+
+Results were calculated with LLaMA 3 8b BF16 as `--kl-divergence-base` and LLaMA 3 8b FP16 as the `--model` for comparison.
+
+| Metric                         |                    Value |
+|--------------------------------|--------------------------|
+| Mean PPL(Q)                    |      6.227711 ± 0.037833 |
+| Mean PPL(base)                 |      6.225194 ± 0.037771 |
+| Cor(ln(PPL(Q)), ln(PPL(base))) |                  99.990% |
+| Mean ln(PPL(Q)/PPL(base))      |      0.000404 ± 0.000086 |
+| Mean PPL(Q)/PPL(base)          |      1.000404 ± 0.000086 |
+| Mean PPL(Q)-PPL(base)          |      0.002517 ± 0.000536 |
+| Mean    KLD                    |  0.00002515 ± 0.00000020 |
+| Maximum KLD                    |                 0.012206 |
+| 99.9%   KLD                    |                 0.000799 |
+| 99.0%   KLD                    |                 0.000222 |
+| 99.0%   KLD                    |                 0.000222 |
+| Median  KLD                    |                 0.000013 |
+| 10.0%   KLD                    |                -0.000002 |
+| 5.0%   KLD                     |                -0.000008 |
+| 1.0%   KLD                     |                -0.000023 |
+| Minimum KLD                    |                -0.000059 |
+| Mean    Δp                     | -0.0000745 ± 0.0003952 % |
+| Maximum Δp                     |                   4.186% |
+| 99.9%   Δp                     |                   1.049% |
+| 99.0%   Δp                     |                   0.439% |
+| 95.0%   Δp                     |                   0.207% |
+| 90.0%   Δp                     |                   0.125% |
+| 75.0%   Δp                     |                   0.029% |
+| Median  Δp                     |                   0.000% |
+| 25.0%   Δp                     |                  -0.030% |
+| 10.0%   Δp                     |                  -0.126% |
+| 5.0%   Δp                      |                  -0.207% |
+| 1.0%   Δp                      |                  -0.434% |
+| 0.1%   Δp                      |                  -1.016% |
+| Minimum Δp                     |                  -4.672% |
+| RMS Δp                         |          0.150 ± 0.001 % |
+| Same top p                     |         99.739 ± 0.013 % |

 ## Old Numbers

--- a/examples/perplexity/perplexity.cpp
+++ b/examples/perplexity/perplexity.cpp
@@ -44,9 +44,9 @@ static void write_logfile(
        return;
    }

-    const std::string timestamp = get_sortable_timestamp();
+    const std::string timestamp = string_get_sortable_timestamp();

-    const bool success = create_directory_with_parents(params.logdir);
+    const bool success = fs_create_directory_with_parents(params.logdir);
    if (!success) {
        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                __func__, params.logdir.c_str());
@@ -64,7 +64,7 @@ static void write_logfile(
    fprintf(logfile, "binary: main\n");
    char model_desc[128];
    llama_model_desc(model, model_desc, sizeof(model_desc));
-    dump_non_result_info_yaml(logfile, params, ctx, timestamp, results.tokens, model_desc);
+    yaml_dump_non_result_info(logfile, params, ctx, timestamp, results.tokens, model_desc);

    fprintf(logfile, "\n");
    fprintf(logfile, "######################\n");
@@ -72,9 +72,9 @@ static void write_logfile(
    fprintf(logfile, "######################\n");
    fprintf(logfile, "\n");

-    dump_vector_float_yaml(logfile, "logits", results.logits);
+    yaml_dump_vector_float(logfile, "logits", results.logits);
    fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
-    dump_vector_float_yaml(logfile, "probs", results.probs);
+    yaml_dump_vector_float(logfile, "probs", results.probs);

    llama_dump_timing_info_yaml(logfile, ctx);
    fclose(logfile);
@@ -1425,7 +1425,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
        // Use all tasks
        tasks.resize(n_task);
        printf("%s: reading tasks", __func__);
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
        int i = 0;
        for (auto& task : tasks) {
            ++i;
@@ -1675,7 +1675,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params

    llama_batch_free(batch);

-    if (n_done < 100) return;
+    if (n_done < 100 && (params.multiple_choice_tasks != 0 && params.multiple_choice_tasks < (size_t)n_task)) return;

    float p = 1.f*n_correct/n_done;
    float sigma = sqrt(p*(1-p)/(n_done-1));
@@ -2007,7 +2007,7 @@ int main(int argc, char ** argv) {

    std::mt19937 rng(params.seed);
    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
+        params.prompt = string_random_prompt(rng);
    }

    llama_backend_init();
@@ -2035,7 +2035,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
    }

    struct results_perplexity results;
--- a/examples/quantize/README.md
+++ b/examples/quantize/README.md
@@ -1,6 +1,8 @@
 # quantize

-TODO
+You can also use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to build your own quants without any setup.
+
+Note: It is synced from llama.cpp `main` every 6 hours.

 ## Llama 2 7B

--- a/examples/quantize/quantize.cpp
+++ b/examples/quantize/quantize.cpp
@@ -259,7 +259,7 @@ int main(int argc, char ** argv) {
                usage(argv[0]);
            }
        } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
-            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+            if (arg_idx == argc-1 || !string_parse_kv_override(argv[++arg_idx], kv_overrides)) {
                usage(argv[0]);
            }
        } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
@@ -284,7 +284,7 @@ int main(int argc, char ** argv) {
            } else {
                usage(argv[0]);
            }
-        } else if (strcmp(argv[arg_idx], "--keep-split")) {
+        } else if (strcmp(argv[arg_idx], "--keep-split") == 0) {
            params.keep_split = true;
        } else {
            usage(argv[0]);
--- a/examples/quantize/tests.sh
+++ b/examples/quantize/tests.sh
@@ -41,8 +41,8 @@ $SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/g
 echo PASS
 echo

-# 3. Requant model with '--keep_split'
-$QUANTIZE --allow-requantize --keep_split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
+# 3. Requant model with '--keep-split'
+$QUANTIZE --allow-requantize --keep-split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
 echo PASS
 echo

@@ -51,7 +51,7 @@ $MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt
 echo PASS
 echo

-# 4. Requant mode without '--keep_split'
+# 4. Requant mode without '--keep-split'
 $QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
 echo PASS
 echo
--- a/examples/retrieval/retrieval.cpp
+++ b/examples/retrieval/retrieval.cpp
@@ -11,7 +11,7 @@ struct retrieval_params {
 };

 static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
-    gpt_print_usage(argc, argv, gpt_params);
+    gpt_params_print_usage(argc, argv, gpt_params);
    printf("retrieval options:\n");
    printf("  --context-file FNAME  file containing context to embed.\n");
    printf("                        specify multiple files by providing --context-file option multiple times.\n");
@@ -226,7 +226,7 @@ int main(int argc, char ** argv) {
    // print system information
    {
        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
    }

    // max batch size
--- a/examples/rpc/CMakeLists.txt
+++ b/examples/rpc/CMakeLists.txt
@@ -0,0 +1,2 @@
+add_executable(rpc-server rpc-server.cpp)
+target_link_libraries(rpc-server PRIVATE ggml llama)
--- a/examples/rpc/README.md
+++ b/examples/rpc/README.md
@@ -0,0 +1,74 @@
+## Overview
+
+The `rpc-server` allows  running `ggml` backend on a remote host.
+The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
+This can be used for distributed LLM inference with `llama.cpp` in the following way:
+
+```mermaid
+flowchart TD
+    rpcb---|TCP|srva
+    rpcb---|TCP|srvb
+    rpcb-.-|TCP|srvn
+    subgraph hostn[Host N]
+    srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph hostb[Host B]
+    srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph hosta[Host A]
+    srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
+    end
+    subgraph host[Main Host]
+    ggml[llama.cpp]---rpcb[RPC backend]
+    end
+    style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
+```
+
+Each host can run a different backend, e.g. one with CUDA and another with Metal.
+You can also run multiple `rpc-server` instances on the same host, each with a different backend.
+
+## Usage
+
+On each host, build the corresponding backend with `cmake` and add `-DLLAMA_RPC=ON` to the build options.
+For example, to build the CUDA backend with RPC support:
+
+```bash
+mkdir build-rpc-cuda
+cd build-rpc-cuda
+cmake .. -DLLAMA_CUDA=ON -DLLAMA_RPC=ON
+cmake --build . --config Release
+```
+
+Then, start the `rpc-server` with the backend:
+
+```bash
+$ bin/rpc-server -p 50052
+create_backend: using CUDA backend
+ggml_cuda_init: GGML_CUDA_FORCE_MMQ:   no
+ggml_cuda_init: CUDA_USE_TENSOR_CORES: yes
+ggml_cuda_init: found 1 CUDA devices:
+  Device 0: NVIDIA T1200 Laptop GPU, compute capability 7.5, VMM: yes
+Starting RPC server on 0.0.0.0:50052
+```
+
+When using the CUDA backend, you can specify the device with the `CUDA_VISIBLE_DEVICES` environment variable, e.g.:
+```bash
+$ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
+```
+This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
+
+
+On the main host build `llama.cpp` only with `-DLLAMA_RPC=ON`:
+
+```bash
+mkdir build-rpc
+cd build-rpc
+cmake .. -DLLAMA_RPC=ON
+cmake --build . --config Release
+```
+
+Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
+
+```bash
+$ bin/main -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
+```
--- a/examples/rpc/rpc-server.cpp
+++ b/examples/rpc/rpc-server.cpp
@@ -0,0 +1,134 @@
+#ifdef GGML_USE_CUDA
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
+
+#include "ggml-rpc.h"
+#ifdef _WIN32
+#  include <windows.h>
+#else
+#  include <unistd.h>
+#endif
+#include <string>
+#include <stdio.h>
+
+struct rpc_server_params {
+    std::string host        = "0.0.0.0";
+    int         port        = 50052;
+    size_t      backend_mem = 0;
+};
+
+static void print_usage(int /*argc*/, char ** argv, rpc_server_params params) {
+    fprintf(stderr, "Usage: %s [options]\n\n", argv[0]);
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help            show this help message and exit\n");
+    fprintf(stderr, "  -H HOST, --host HOST  host to bind to (default: %s)\n", params.host.c_str());
+    fprintf(stderr, "  -p PORT, --port PORT  port to bind to (default: %d)\n", params.port);
+    fprintf(stderr, "  -m MEM, --mem MEM     backend memory size (in MB)\n");
+    fprintf(stderr, "\n");
+}
+
+static bool rpc_server_params_parse(int argc, char ** argv, rpc_server_params & params) {
+    std::string arg;
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg == "-H" || arg == "--host") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.host = argv[i];
+        } else if (arg == "-p" || arg == "--port") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.port = std::stoi(argv[i]);
+            if (params.port <= 0 || params.port > 65535) {
+                return false;
+            }
+        } else if (arg == "-m" || arg == "--mem") {
+            if (++i >= argc) {
+                return false;
+            }
+            params.backend_mem = std::stoul(argv[i]) * 1024 * 1024;
+        } else if (arg == "-h" || arg == "--help") {
+            print_usage(argc, argv, params);
+            exit(0);
+        } else {
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            print_usage(argc, argv, params);
+            exit(0);
+        }
+    }
+    return true;
+}
+
+static ggml_backend_t create_backend() {
+    ggml_backend_t backend = NULL;
+#ifdef GGML_USE_CUDA
+    fprintf(stderr, "%s: using CUDA backend\n", __func__);
+    backend = ggml_backend_cuda_init(0); // init device 0
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__);
+    }
+#elif GGML_USE_METAL
+    fprintf(stderr, "%s: using Metal backend\n", __func__);
+    backend = ggml_backend_metal_init();
+    if (!backend) {
+        fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
+    }
+#endif
+
+    // if there aren't GPU Backends fallback to CPU backend
+    if (!backend) {
+        fprintf(stderr, "%s: using CPU backend\n", __func__);
+        backend = ggml_backend_cpu_init();
+    }
+    return backend;
+}
+
+static void get_backend_memory(size_t * free_mem, size_t * total_mem) {
+#ifdef GGML_USE_CUDA
+    ggml_backend_cuda_get_device_memory(0, free_mem, total_mem);
+#else
+    #ifdef _WIN32
+        MEMORYSTATUSEX status;
+        status.dwLength = sizeof(status);
+        GlobalMemoryStatusEx(&status);
+        *total_mem = status.ullTotalPhys;
+        *free_mem = status.ullAvailPhys;
+    #else
+        long pages = sysconf(_SC_PHYS_PAGES);
+        long page_size = sysconf(_SC_PAGE_SIZE);
+        *total_mem = pages * page_size;
+        *free_mem = *total_mem;
+    #endif
+#endif
+}
+
+int main(int argc, char * argv[]) {
+    rpc_server_params params;
+    if (!rpc_server_params_parse(argc, argv, params)) {
+        fprintf(stderr, "Invalid parameters\n");
+        return 1;
+    }
+    ggml_backend_t backend = create_backend();
+    if (!backend) {
+        fprintf(stderr, "Failed to create backend\n");
+        return 1;
+    }
+    std::string endpoint = params.host + ":" + std::to_string(params.port);
+    size_t free_mem, total_mem;
+    if (params.backend_mem > 0) {
+        free_mem = params.backend_mem;
+        total_mem = params.backend_mem;
+    } else {
+        get_backend_memory(&free_mem, &total_mem);
+    }
+    printf("Starting RPC server on %s, backend memory: %zu MB\n", endpoint.c_str(), free_mem / (1024 * 1024));
+    start_rpc_server(backend, endpoint.c_str(), free_mem, total_mem);
+    ggml_backend_free(backend);
+    return 0;
+}
--- a/examples/server/README.md
+++ b/examples/server/README.md
@@ -17,8 +17,9 @@ The project is under active development, and we are [looking for feedback and co

 **Command line options:**

- `--threads N`, `-t N`: Set the number of threads to use during generation. Not used if model layers are offloaded to GPU. The server is using batching. This parameter is used only if one token is to be processed on CPU backend.
- `-tb N, --threads-batch N`: Set the number of threads to use during batch and prompt processing. If not specified, the number of threads will be set to the number of threads used for generation. Not used if model layers are offloaded to GPU.
+- `-v`, `--verbose`: Enable verbose server output. When using the `/completion` endpoint, this includes the tokenized prompt, the full request and the full response.
+- `-t N`, `--threads N`: Set the number of threads to use by CPU layers during generation. Not used by model layers that are offloaded to GPU. This option has no effect when using the maximum number of GPU layers. Default: `std::thread::hardware_concurrency()` (number of CPU cores).
+- `-tb N, --threads-batch N`: Set the number of threads to use by CPU layers during batch and prompt processing (>= 32 tokens). This option has no effect if a GPU is available. Default: `--threads`.
 - `--threads-http N`: Number of threads in the http server pool to process requests. Default: `max(std::thread::hardware_concurrency() - 1, --parallel N + 2)`
 - `-m FNAME`, `--model FNAME`: Specify the path to the LLaMA model file (e.g., `models/7B/ggml-model.gguf`).
 - `-mu MODEL_URL --model-url MODEL_URL`: Specify a remote http url to download the file. Default: unused
@@ -36,9 +37,7 @@ The project is under active development, and we are [looking for feedback and co
 - `--numa STRATEGY`: Attempt one of the below optimization strategies that may help on some NUMA systems
 - `--numa distribute`: Spread execution evenly over all nodes
 - `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system
-page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
-
+- `--numa numactl`: Use the CPU map provided by numactl. If run without this previously, it is recommended to drop the system page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/1437
 - `--numa`: Attempt optimizations that may help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
@@ -48,8 +47,8 @@ page cache before using this. See https://github.com/ggerganov/llama.cpp/issues/
 - `--path`: Path from which to serve static files. Default: disabled
 - `--api-key`: Set an api key for request authorization. By default, the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: Path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`s.
- `--embedding`: Enable embedding extraction. Default: disabled
- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`
+- `--embeddings`: Enable embedding vector output and the OAI compatible endpoint /v1/embeddings. Physical batch size (`--ubatch-size`) must be carefully defined. Default: disabled
+- `-np N`, `--parallel N`: Set the number of slots for process requests. Default: `1`. Values > 1 will allow for higher throughput with multiple parallel requests but the results will **not** be deterministic due to differences in rounding error.
 - `-cb`, `--cont-batching`: Enable continuous batching (a.k.a dynamic batching).  Default: disabled
 - `-spf FNAME`, `--system-prompt-file FNAME` Set a file to load a system prompt (initial prompt of all slots). This is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
--- a/examples/server/public_simplechat/index.html
+++ b/examples/server/public_simplechat/index.html
@@ -0,0 +1,52 @@
+<!DOCTYPE html>
+<html lang="en">
+    <head>
+        <title>SimpleChat (LlamaCPP, ...) </title>
+        <meta charset="UTF-8" />
+        <meta name="viewport" content="width=device-width, initial-scale=1" />
+        <meta name="message" content="Save Nature Save Earth" />
+        <meta name="description" content="SimpleChat: trigger LLM web service endpoints /chat/completions and /completions, single/multi chat sessions" />
+        <meta name="author" content="by Humans for All" />
+        <meta http-equiv="Cache-Control" content="no-cache, no-store, must-revalidate" />
+        <script src="simplechat.js" defer></script>
+        <link rel="stylesheet" href="simplechat.css" />
+    </head>
+    <body>
+        <div class="samecolumn" id="fullbody">
+
+            <div class="sameline">
+                <p class="heading flex-grow" > <b> SimpleChat </b> </p>
+                <div class="sameline">
+                    <label for="api-ep">Mode:</label>
+                    <select name="api-ep" id="api-ep">
+                    <option value="chat" selected>Chat</option>
+                    <option value="completion">Completion</option>
+                    </select>
+                </div>
+            </div>
+
+            <div id="sessions-div" class="sameline"></div>
+
+            <hr>
+            <div class="sameline">
+                <label for="system-in">System</label>
+                <input type="text" name="system" id="system-in" class="flex-grow"/>
+            </div>
+
+            <hr>
+            <div id="chat-div">
+                <p> Enter the system prompt above, before entering/submitting any user query.</p>
+                <p> Enter your text to the ai assistant below.</p>
+                <p> Use shift+enter for inserting enter.</p>
+                <p> Refresh the page to start over fresh.</p>
+            </div>
+
+            <hr>
+            <div class="sameline">
+                <textarea id="user-in" class="flex-grow" rows="3"></textarea>
+                <button id="user-btn">submit</button>
+            </div>
+
+        </div>
+    </body>
+</html>
--- a/examples/server/public_simplechat/readme.md
+++ b/examples/server/public_simplechat/readme.md
@@ -0,0 +1,81 @@
+
+# SimpleChat
+
+by Humans for All.
+
+
+## overview
+
+This simple web frontend, allows triggering/testing the server's /completions or /chat/completions endpoints
+in a simple way with minimal code from a common code base. Inturn additionally it tries to allow single or
+multiple independent back and forth chatting to an extent, with the ai llm model at a basic level, with their
+own system prompts.
+
+The UI follows a responsive web design so that the layout can adapt to available display space in a usable
+enough manner, in general.
+
+NOTE: Given that the idea is for basic minimal testing, it doesnt bother with any model context length and
+culling of old messages from the chat.
+
+NOTE: It doesnt set any parameters other than temperature for now. However if someone wants they can update
+the js file as needed.
+
+
+## usage
+
+One could run this web frontend directly using server itself or if anyone is thinking of adding a built in web
+frontend to configure the server over http(s) or so, then run this web frontend using something like python's
+http module.
+
+### running using examples/server
+
+bin/server -m path/model.gguf --path ../examples/server/public_simplechat [--port PORT]
+
+### running using python3's server module
+
+first run examples/server
+* bin/server -m path/model.gguf
+
+next run this web front end in examples/server/public_simplechat
+* cd ../examples/server/public_simplechat
+* python3 -m http.server PORT
+
+### using the front end
+
+Open this simple web front end from your local browser
+* http://127.0.0.1:PORT/index.html
+
+Once inside
+* Select between chat and completion mode. By default it is set to chat mode.
+* If you want to provide a system prompt, then ideally enter it first, before entering any user query.
+  * if chat.add_system_begin is used
+    * you cant change the system prompt, after it is has been submitted once along with user query.
+    * you cant set a system prompt, after you have submitted any user query
+  * if chat.add_system_anytime is used
+    * one can change the system prompt any time during chat, by changing the contents of system prompt.
+    * inturn the updated/changed system prompt will be inserted into the chat session.
+    * this allows for the subsequent user chatting to be driven by the new system prompt set above.
+* Enter your query and either press enter or click on the submit button.
+  If you want to insert enter (\n) as part of your chat/query to ai model, use shift+enter.
+* Wait for the logic to communicate with the server and get the response.
+  * the user is not allowed to enter any fresh query during this time.
+  * the user input box will be disabled and a working message will be shown in it.
+* just refresh the page, to reset wrt the chat history and or system prompt and start afresh.
+* Using NewChat one can start independent chat sessions.
+  * two independent chat sessions are setup by default.
+
+
+## Devel note
+
+Sometimes the browser may be stuborn with caching of the file, so your updates to html/css/js
+may not be visible. Also remember that just refreshing/reloading page in browser or for that
+matter clearing site data, dont directly override site caching in all cases. Worst case you may
+have to change port. Or in dev tools of browser, you may be able to disable caching fully.
+
+Concept of multiple chat sessions with different servers, as well as saving and restoring of
+those across browser usage sessions, can be woven around the SimpleChat/MultiChatUI class and
+its instances relatively easily, however given the current goal of keeping this simple, it has
+not been added, for now.
+
+By switching between chat.add_system_begin/anytime, one can control whether one can change
+the system prompt, anytime during the conversation or only at the beginning.
--- a/examples/server/public_simplechat/simplechat.css
+++ b/examples/server/public_simplechat/simplechat.css
@@ -0,0 +1,61 @@
+/**
+ * the styling of the simplechat web frontend
+ * by Humans for All
+ */
+
+#fullbody {
+    height: 98vh;
+}
+
+.heading {
+    background-color: lightgray;
+}
+
+.session-selected {
+    background-color: lightblue;
+}
+
+.role-system {
+    background-color: lightblue;
+}
+.role-user {
+    background-color: lightgray;
+}
+
+.flex-grow {
+    flex-grow: 1;
+}
+.float-right {
+    float: right;
+}
+
+#chat-div {
+    overflow: scroll;
+    flex-grow: 1;
+    flex-shrink: 1;
+    min-height: 40vh;
+}
+button {
+    min-width: 8vw;
+}
+
+.sameline {
+    display: flex;
+    flex-direction: row;
+}
+.samecolumn {
+    display: flex;
+    flex-direction: column;
+}
+
+* {
+    margin: 0.6vmin;
+}
+
+@media print {
+
+    #fullbody {
+        height: auto;
+    }
+
+}
--- a/examples/server/public_simplechat/simplechat.js
+++ b/examples/server/public_simplechat/simplechat.js
@@ -0,0 +1,478 @@
+// @ts-check
+// A simple completions and chat/completions test related web front end logic
+// by Humans for All
+
+class Roles {
+    static System = "system";
+    static User = "user";
+    static Assistant = "assistant";
+}
+
+class ApiEP {
+    static Chat = "chat";
+    static Completion = "completion";
+}
+
+let gUsageMsg = `
+    <p> Enter the system prompt above, before entering/submitting any user query.</p>
+    <p> Enter your text to the ai assistant below.</p>
+    <p> Use shift+enter for inserting enter.</p>
+    <p> Refresh the page to start over fresh.</p>
+`;
+
+class SimpleChat {
+
+    constructor() {
+        /**
+         * Maintain in a form suitable for common LLM web service chat/completions' messages entry
+         * @type {{role: string, content: string}[]}
+         */
+        this.xchat = [];
+        this.iLastSys = -1;
+    }
+
+    /**
+     * Add an entry into xchat
+     * @param {string} role
+     * @param {string|undefined|null} content
+     */
+    add(role, content) {
+        if ((content == undefined) || (content == null) || (content == "")) {
+            return false;
+        }
+        this.xchat.push( {role: role, content: content} );
+        if (role == Roles.System) {
+            this.iLastSys = this.xchat.length - 1;
+        }
+        return true;
+    }
+
+    /**
+     * Show the contents in the specified div
+     * @param {HTMLDivElement} div
+     * @param {boolean} bClear
+     */
+    show(div, bClear=true) {
+        if (bClear) {
+            div.replaceChildren();
+        }
+        let last = undefined;
+        for(const x of this.xchat) {
+            let entry = document.createElement("p");
+            entry.className = `role-${x.role}`;
+            entry.innerText = `${x.role}: ${x.content}`;
+            div.appendChild(entry);
+            last = entry;
+        }
+        if (last !== undefined) {
+            last.scrollIntoView(false);
+        } else {
+            if (bClear) {
+                div.innerHTML = gUsageMsg;
+            }
+        }
+    }
+
+    /**
+     * Add needed fields wrt json object to be sent wrt LLM web services completions endpoint
+     * Convert the json into string.
+     * @param {Object} obj
+     */
+    request_jsonstr(obj) {
+        obj["temperature"] = 0.7;
+        return JSON.stringify(obj);
+    }
+
+    /**
+     * Return a string form of json object suitable for chat/completions
+     */
+    request_messages_jsonstr() {
+        let req = {
+            messages: this.xchat,
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Return a string form of json object suitable for /completions
+     */
+    request_prompt_jsonstr() {
+        let prompt = "";
+        for(const chat of this.xchat) {
+            prompt += `${chat.role}: ${chat.content}\n`;
+        }
+        let req = {
+            prompt: prompt,
+        }
+        return this.request_jsonstr(req);
+    }
+
+    /**
+     * Allow setting of system prompt, but only at begining.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_begin(sysPrompt, msgTag) {
+        if (this.xchat.length == 0) {
+            if (sysPrompt.length > 0) {
+                return this.add(Roles.System, sysPrompt);
+            }
+        } else {
+            if (sysPrompt.length > 0) {
+                if (this.xchat[0].role !== Roles.System) {
+                    console.error(`ERRR:SimpleChat:SC:${msgTag}:You need to specify system prompt before any user query, ignoring...`);
+                } else {
+                    if (this.xchat[0].content !== sysPrompt) {
+                        console.error(`ERRR:SimpleChat:SC:${msgTag}:You cant change system prompt, mid way through, ignoring...`);
+                    }
+                }
+            }
+        }
+        return false;
+    }
+
+    /**
+     * Allow setting of system prompt, at any time.
+     * @param {string} sysPrompt
+     * @param {string} msgTag
+     */
+    add_system_anytime(sysPrompt, msgTag) {
+        if (sysPrompt.length <= 0) {
+            return false;
+        }
+
+        if (this.iLastSys < 0) {
+            return this.add(Roles.System, sysPrompt);
+        }
+
+        let lastSys = this.xchat[this.iLastSys].content;
+        if (lastSys !== sysPrompt) {
+            return this.add(Roles.System, sysPrompt);
+        }
+        return false;
+    }
+
+    /**
+     * Retrieve the latest system prompt.
+     */
+    get_system_latest() {
+        if (this.iLastSys == -1) {
+            return "";
+        }
+        let sysPrompt = this.xchat[this.iLastSys].content;
+        return sysPrompt;
+    }
+
+}
+
+
+let gBaseURL = "http://127.0.0.1:8080";
+let gChatURL = {
+    'chat': `${gBaseURL}/chat/completions`,
+    'completion': `${gBaseURL}/completions`,
+}
+const gbCompletionFreshChatAlways = true;
+
+
+/**
+ * Set the class of the children, based on whether it is the idSelected or not.
+ * @param {HTMLDivElement} elBase
+ * @param {string} idSelected
+ * @param {string} classSelected
+ * @param {string} classUnSelected
+ */
+function el_children_config_class(elBase, idSelected, classSelected, classUnSelected="") {
+    for(let child of elBase.children) {
+        if (child.id == idSelected) {
+            child.className = classSelected;
+        } else {
+            child.className = classUnSelected;
+        }
+    }
+}
+
+/**
+ * Create button and set it up.
+ * @param {string} id
+ * @param {(this: HTMLButtonElement, ev: MouseEvent) => any} callback
+ * @param {string | undefined} name
+ * @param {string | undefined} innerText
+ */
+function el_create_button(id, callback, name=undefined, innerText=undefined) {
+    if (!name) {
+        name = id;
+    }
+    if (!innerText) {
+        innerText = id;
+    }
+    let btn = document.createElement("button");
+    btn.id = id;
+    btn.name = name;
+    btn.innerText = innerText;
+    btn.addEventListener("click", callback);
+    return btn;
+}
+
+
+class MultiChatUI {
+
+    constructor() {
+        /** @type {Object<string, SimpleChat>} */
+        this.simpleChats = {};
+        /** @type {string} */
+        this.curChatId = "";
+
+        // the ui elements
+        this.elInSystem = /** @type{HTMLInputElement} */(document.getElementById("system-in"));
+        this.elDivChat = /** @type{HTMLDivElement} */(document.getElementById("chat-div"));
+        this.elBtnUser = /** @type{HTMLButtonElement} */(document.getElementById("user-btn"));
+        this.elInUser = /** @type{HTMLInputElement} */(document.getElementById("user-in"));
+        this.elSelectApiEP = /** @type{HTMLSelectElement} */(document.getElementById("api-ep"));
+        this.elDivSessions = /** @type{HTMLDivElement} */(document.getElementById("sessions-div"));
+
+        this.validate_element(this.elInSystem, "system-in");
+        this.validate_element(this.elDivChat, "chat-div");
+        this.validate_element(this.elInUser, "user-in");
+        this.validate_element(this.elSelectApiEP, "api-ep");
+        this.validate_element(this.elDivChat, "sessions-div");
+    }
+
+    /**
+     * Check if the element got
+     * @param {HTMLElement | null} el
+     * @param {string} msgTag
+     */
+    validate_element(el, msgTag) {
+        if (el == null) {
+            throw Error(`ERRR:SimpleChat:MCUI:${msgTag} element missing in html...`);
+        } else {
+            console.debug(`INFO:SimpleChat:MCUI:${msgTag} Id[${el.id}] Name[${el["name"]}]`);
+        }
+    }
+
+    /**
+     * Reset user input ui.
+     * * clear user input
+     * * enable user input
+     * * set focus to user input
+     */
+    ui_reset_userinput() {
+        this.elInUser.value = "";
+        this.elInUser.disabled = false;
+        this.elInUser.focus();
+    }
+
+    /**
+     * Setup the needed callbacks wrt UI, curChatId to defaultChatId and
+     * optionally switch to specified defaultChatId.
+     * @param {string} defaultChatId
+     * @param {boolean} bSwitchSession
+     */
+    setup_ui(defaultChatId, bSwitchSession=false) {
+
+        this.curChatId = defaultChatId;
+        if (bSwitchSession) {
+            this.handle_session_switch(this.curChatId);
+        }
+
+        this.elBtnUser.addEventListener("click", (ev)=>{
+            if (this.elInUser.disabled) {
+                return;
+            }
+            this.handle_user_submit(this.curChatId, this.elSelectApiEP.value).catch((/** @type{Error} */reason)=>{
+                let msg = `ERRR:SimpleChat\nMCUI:HandleUserSubmit:${this.curChatId}\n${reason.name}:${reason.message}`;
+                console.debug(msg.replace("\n", ":"));
+                alert(msg);
+                this.ui_reset_userinput();
+            });
+        });
+
+        this.elInUser.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into their message using shift+enter.
+            // while just pressing enter key will lead to submitting.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                this.elBtnUser.click();
+                ev.preventDefault();
+            }
+        });
+
+        this.elInSystem.addEventListener("keyup", (ev)=> {
+            // allow user to insert enter into the system prompt using shift+enter.
+            // while just pressing enter key will lead to setting the system prompt.
+            if ((ev.key === "Enter") && (!ev.shiftKey)) {
+                let chat = this.simpleChats[this.curChatId];
+                chat.add_system_anytime(this.elInSystem.value, this.curChatId);
+                chat.show(this.elDivChat);
+                ev.preventDefault();
+            }
+        });
+
+    }
+
+    /**
+     * Setup a new chat session and optionally switch to it.
+     * @param {string} chatId
+     * @param {boolean} bSwitchSession
+     */
+    new_chat_session(chatId, bSwitchSession=false) {
+        this.simpleChats[chatId] = new SimpleChat();
+        if (bSwitchSession) {
+            this.handle_session_switch(chatId);
+        }
+    }
+
+    /**
+     * Handle user query submit request, wrt specified chat session.
+     * @param {string} chatId
+     * @param {string} apiEP
+     */
+    async handle_user_submit(chatId, apiEP) {
+
+        let chat = this.simpleChats[chatId];
+
+        chat.add_system_anytime(this.elInSystem.value, chatId);
+
+        let content = this.elInUser.value;
+        if (!chat.add(Roles.User, content)) {
+            console.debug(`WARN:SimpleChat:MCUI:${chatId}:HandleUserSubmit:Ignoring empty user input...`);
+            return;
+        }
+        chat.show(this.elDivChat);
+
+        let theBody;
+        let theUrl = gChatURL[apiEP]
+        if (apiEP == ApiEP.Chat) {
+            theBody = chat.request_messages_jsonstr();
+        } else {
+            theBody = chat.request_prompt_jsonstr();
+        }
+
+        this.elInUser.value = "working...";
+        this.elInUser.disabled = true;
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:${theUrl}:ReqBody:${theBody}`);
+        let resp = await fetch(theUrl, {
+            method: "POST",
+            headers: {
+                "Content-Type": "application/json",
+            },
+            body: theBody,
+        });
+
+        let respBody = await resp.json();
+        console.debug(`DBUG:SimpleChat:MCUI:${chatId}:HandleUserSubmit:RespBody:${JSON.stringify(respBody)}`);
+        let assistantMsg;
+        if (apiEP == ApiEP.Chat) {
+            assistantMsg = respBody["choices"][0]["message"]["content"];
+        } else {
+            try {
+                assistantMsg = respBody["choices"][0]["text"];
+            } catch {
+                assistantMsg = respBody["content"];
+            }
+        }
+        chat.add(Roles.Assistant, assistantMsg);
+        if (chatId == this.curChatId) {
+            chat.show(this.elDivChat);
+        } else {
+            console.debug(`DBUG:SimpleChat:MCUI:HandleUserSubmit:ChatId has changed:[${chatId}] [${this.curChatId}]`);
+        }
+        // Purposefully clear at end rather than begin of this function
+        // so that one can switch from chat to completion mode and sequece
+        // in a completion mode with multiple user-assistant chat data
+        // from before to be sent/occur once.
+        if ((apiEP == ApiEP.Completion) && (gbCompletionFreshChatAlways)) {
+            chat.xchat.length = 0;
+        }
+        this.ui_reset_userinput();
+    }
+
+    /**
+     * Show buttons for NewChat and available chat sessions, in the passed elDiv.
+     * If elDiv is undefined/null, then use this.elDivSessions.
+     * Take care of highlighting the selected chat-session's btn.
+     * @param {HTMLDivElement | undefined} elDiv
+     */
+    show_sessions(elDiv=undefined) {
+        if (!elDiv) {
+            elDiv = this.elDivSessions;
+        }
+        elDiv.replaceChildren();
+        // Btn for creating new chat session
+        let btnNew = el_create_button("New CHAT", (ev)=> {
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:NewChat:Current session [${this.curChatId}] awaiting response, ignoring request...`);
+                alert("ERRR:SimpleChat\nMCUI:NewChat\nWait for response to pending query, before starting new chat session");
+                return;
+            }
+            let chatId = `Chat${Object.keys(this.simpleChats).length}`;
+            let chatIdGot = prompt("INFO:SimpleChat\nMCUI:NewChat\nEnter id for new chat session", chatId);
+            if (!chatIdGot) {
+                console.error("ERRR:SimpleChat:MCUI:NewChat:Skipping based on user request...");
+                return;
+            }
+            this.new_chat_session(chatIdGot, true);
+            this.create_session_btn(elDiv, chatIdGot);
+            el_children_config_class(elDiv, chatIdGot, "session-selected", "");
+        });
+        elDiv.appendChild(btnNew);
+        // Btns for existing chat sessions
+        let chatIds = Object.keys(this.simpleChats);
+        for(let cid of chatIds) {
+            let btn = this.create_session_btn(elDiv, cid);
+            if (cid == this.curChatId) {
+                btn.className = "session-selected";
+            }
+        }
+    }
+
+    create_session_btn(elDiv, cid) {
+        let btn = el_create_button(cid, (ev)=>{
+            let target = /** @type{HTMLButtonElement} */(ev.target);
+            console.debug(`DBUG:SimpleChat:MCUI:SessionClick:${target.id}`);
+            if (this.elInUser.disabled) {
+                console.error(`ERRR:SimpleChat:MCUI:SessionClick:${target.id}:Current session [${this.curChatId}] awaiting response, ignoring switch...`);
+                alert("ERRR:SimpleChat\nMCUI:SessionClick\nWait for response to pending query, before switching");
+                return;
+            }
+            this.handle_session_switch(target.id);
+            el_children_config_class(elDiv, target.id, "session-selected", "");
+        });
+        elDiv.appendChild(btn);
+        return btn;
+    }
+
+    /**
+     * Switch ui to the specified chatId and set curChatId to same.
+     * @param {string} chatId
+     */
+    async handle_session_switch(chatId) {
+        let chat = this.simpleChats[chatId];
+        if (chat == undefined) {
+            console.error(`ERRR:SimpleChat:MCUI:HandleSessionSwitch:${chatId} missing...`);
+            return;
+        }
+        this.elInSystem.value = chat.get_system_latest();
+        this.elInUser.value = "";
+        chat.show(this.elDivChat);
+        this.elInUser.focus();
+        this.curChatId = chatId;
+        console.log(`INFO:SimpleChat:MCUI:HandleSessionSwitch:${chatId} entered...`);
+    }
+
+}
+
+
+let gMuitChat;
+const gChatIds = [ "Default", "Other" ];
+
+function startme() {
+    console.log("INFO:SimpleChat:StartMe:Starting...");
+    gMuitChat = new MultiChatUI();
+    for (let cid of gChatIds) {
+        gMuitChat.new_chat_session(cid);
+    }
+    gMuitChat.setup_ui(gChatIds[0]);
+    gMuitChat.show_sessions();
+}
+
+document.addEventListener("DOMContentLoaded", startme);
--- a/examples/server/server.cpp
+++ b/examples/server/server.cpp
@@ -102,7 +102,6 @@ struct slot_params {
    bool stream       = true;
    bool cache_prompt = false; // remember the prompt to avoid reprocessing all prompt

-    uint32_t seed      = -1; // RNG seed
    int32_t  n_keep    =  0; // number of tokens to keep from initial prompt
    int32_t  n_discard =  0; // number of tokens after n_keep that may be discarded when shifting context, 0 defaults to half
    int32_t  n_predict = -1; // new tokens to predict
@@ -651,9 +650,6 @@ struct server_context {
    std::string              system_prompt;
    std::vector<llama_token> system_tokens;

-    std::string name_user;      // this should be the antiprompt
-    std::string name_assistant;
-
    // slots / clients
    std::vector<server_slot> slots;
    json default_generation_settings_for_props;
@@ -673,6 +669,15 @@ struct server_context {
            llama_free_model(model);
            model = nullptr;
        }
+
+        // Clear any sampling context
+        for (server_slot & slot : slots) {
+            if (slot.ctx_sampling != nullptr) {
+                llama_sampling_free(slot.ctx_sampling);
+            }
+        }
+
+        llama_batch_free(batch);
    }

    bool load_model(const gpt_params & params_) {
@@ -1014,7 +1019,7 @@ struct server_context {
                        sampler_names.emplace_back(sampler_name);
                    }
                }
-                slot.sparams.samplers_sequence = sampler_types_from_names(sampler_names, false);
+                slot.sparams.samplers_sequence = llama_sampling_types_from_names(sampler_names, false);
            } else {
                slot.sparams.samplers_sequence = default_sparams.samplers_sequence;
            }
@@ -1098,15 +1103,11 @@ struct server_context {
        system_need_update = false;
    }

-    void system_prompt_set(const json & sys_props) {
-        system_prompt  = sys_props.value("prompt", "");
-        name_user      = sys_props.value("anti_prompt", "");
-        name_assistant = sys_props.value("assistant_name", "");
+    bool system_prompt_set(const std::string & sys_prompt) {
+        system_prompt = sys_prompt;

        LOG_VERBOSE("system prompt process", {
            {"system_prompt",  system_prompt},
-            {"name_user",      name_user},
-            {"name_assistant", name_assistant},
        });

        // release all slots
@@ -1115,6 +1116,7 @@ struct server_context {
        }

        system_need_update = true;
+        return true;
    }

    bool process_token(completion_token_output & result, server_slot & slot) {
@@ -1254,14 +1256,14 @@ struct server_context {
        std::vector<std::string> samplers_sequence;
        samplers_sequence.reserve(slot.sparams.samplers_sequence.size());
        for (const auto & sampler_type : slot.sparams.samplers_sequence) {
-            samplers_sequence.emplace_back(sampler_type_to_name_string(sampler_type));
+            samplers_sequence.emplace_back(llama_sampling_type_to_str(sampler_type));
        }

        return json {
            {"n_ctx",                     slot.n_ctx},
            {"n_predict",                 slot.n_predict},
            {"model",                     params.model_alias},
-            {"seed",                      slot.params.seed},
+            {"seed",                      slot.sparams.seed},
            {"temperature",               slot.sparams.temp},
            {"dynatemp_range",            slot.sparams.dynatemp_range},
            {"dynatemp_exponent",         slot.sparams.dynatemp_exponent},
@@ -1534,7 +1536,8 @@ struct server_context {
                    }

                    if (task.data.contains("system_prompt")) {
-                        system_prompt_set(task.data.at("system_prompt"));
+                        std::string sys_prompt = json_value(task.data, "system_prompt", std::string());
+                        system_prompt_set(sys_prompt);

                        for (server_slot & slot : slots) {
                            slot.n_past    = 0;
@@ -1978,8 +1981,7 @@ struct server_context {
                                slot.state = SLOT_STATE_PROCESSING;
                                slot.command = SLOT_COMMAND_NONE;
                                slot.release();
-                                slot.print_timings();
-                                send_final_response(slot);
+                                send_error(slot, "input is too large to process. increase the physical batch size", ERROR_TYPE_SERVER);
                                continue;
                            }
                        } else {
@@ -2270,10 +2272,10 @@ struct server_context {

                const size_t n_probs = std::min(cur_p.size, (size_t) slot.sparams.n_probs);
                if (n_probs > 0) {
-                    const size_t n_considered = slot.ctx_sampling->n_considered;
+                    const size_t n_valid = slot.ctx_sampling->n_valid;

                    // Make sure at least n_probs top tokens are at the front of the vector:
-                    if (slot.sparams.temp == 0.0f && n_probs > n_considered) {
+                    if (slot.sparams.temp == 0.0f && n_probs > n_valid) {
                        llama_sample_top_k(ctx, &cur_p, n_probs, 0);
                    }

@@ -2289,7 +2291,7 @@ struct server_context {
                        for (size_t i = 0; i < n_probs; ++i) {
                            result.probs.push_back({
                                cur_p.data[i].id,
-                                i >= n_considered ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
+                                i >= n_valid ? 0.0f : cur_p.data[i].p // Tokens filtered out due to e.g. top_k have 0 probability.
                            });
                        }
                    }
@@ -2383,6 +2385,7 @@ static void server_print_usage(const char * argv0, const gpt_params & params, co
    printf("  --lora-base FNAME         optional model to use as a base for the layers modified by the LoRA adapter\n");
    printf("  --host                    ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
    printf("  --port PORT               port to listen (default  (default: %d)\n", sparams.port);
+    printf("  --rpc SERVERS             comma separated list of RPC servers\n");
    printf("  --path PUBLIC_PATH        path from which to serve static files (default: disabled)\n");
    printf("  --api-key API_KEY         optional api key to enhance server security. If set, requests must include this key for access.\n");
    printf("  --api-key-file FNAME      path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
@@ -2435,6 +2438,12 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                break;
            }
            sparams.port = std::stoi(argv[i]);
+        } else if (arg == "--rpc") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rpc_servers = argv[i];
        } else if (arg == "--host") {
            if (++i >= argc) {
                invalid_param = true;
@@ -2843,7 +2852,7 @@ static void server_params_parse(int argc, char ** argv, server_params & sparams,
                invalid_param = true;
                break;
            }
-            if (!parse_kv_override(argv[i], params.kv_overrides)) {
+            if (!string_parse_kv_override(argv[i], params.kv_overrides)) {
                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
                invalid_param = true;
                break;
@@ -2918,7 +2927,7 @@ int main(int argc, char ** argv) {
    server_params_parse(argc, argv, sparams, params);

    if (!sparams.system_prompt.empty()) {
-        ctx_server.system_prompt_set(json::parse(sparams.system_prompt));
+        ctx_server.system_prompt_set(sparams.system_prompt);
    }

    if (params.model_alias == "unknown") {
@@ -3301,7 +3310,7 @@ int main(int argc, char ** argv) {
    const auto handle_slots_save = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
        json request_data = json::parse(req.body);
        std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
            res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
            return;
        }
@@ -3331,7 +3340,7 @@ int main(int argc, char ** argv) {
    const auto handle_slots_restore = [&ctx_server, &res_error, &sparams](const httplib::Request & req, httplib::Response & res, int id_slot) {
        json request_data = json::parse(req.body);
        std::string filename = request_data.at("filename");
-        if (!validate_file_name(filename)) {
+        if (!fs_validate_filename(filename)) {
            res_error(res, format_error_response("Invalid filename", ERROR_TYPE_INVALID_REQUEST));
            return;
        }
@@ -3407,8 +3416,7 @@ int main(int argc, char ** argv) {
    const auto handle_props = [&ctx_server](const httplib::Request & req, httplib::Response & res) {
        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
        json data = {
-            { "user_name",                   ctx_server.name_user.c_str() },
-            { "assistant_name",              ctx_server.name_assistant.c_str() },
+            { "system_prompt",               ctx_server.system_prompt.c_str() },
            { "default_generation_settings", ctx_server.default_generation_settings_for_props },
            { "total_slots",                 ctx_server.params.n_parallel }
        };
--- a/examples/server/tests/features/results.feature
+++ b/examples/server/tests/features/results.feature
@@ -13,6 +13,7 @@ Feature: Results

  Scenario Outline: consistent results with same seed
    Given <n_slots> slots
+    And   1.0 temperature
    Then  the server is starting
    Then  the server is healthy

@@ -26,10 +27,12 @@ Feature: Results
    Examples:
      | n_slots |
      | 1       |
-      | 2       |
+      # FIXME: unified KV cache nondeterminism
+      # | 2       |

  Scenario Outline: different results with different seed
    Given <n_slots> slots
+    And   1.0 temperature
    Then  the server is starting
    Then  the server is healthy

@@ -70,12 +73,46 @@ Feature: Results
    Then all predictions are equal
    Examples:
      | n_parallel | temp |
-      |  1         | 0.0  |
-      |  2         | 0.0  |
-      |  4         | 0.0  |
-      |  1         | 1.0  |
-      # FIXME: These tests fail on master. The problem seems to be the unified KV cache.
+      | 1          | 0.0  |
+      | 1          | 1.0  |
+      # FIXME: unified KV cache nondeterminism
      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
-      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574 .
-      # |  2         | 1.0  |
-      # |  4         | 1.0  |
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 2          | 0.0  |
+      # | 4          | 0.0  |
+      # | 2          | 1.0  |
+      # | 4          | 1.0  |
+
+  Scenario Outline: consistent token probs with same seed and prompt
+    Given <n_slots> slots
+    And   <n_kv> KV cache size
+    And   1.0 temperature
+    And   <n_predict> max tokens to predict
+    Then  the server is starting
+    Then  the server is healthy
+
+    Given 1 prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then  the server is idle
+    And   all slots are idle
+
+    Given <n_parallel> prompts "The meaning of life is" with seed 42
+    And   concurrent completion requests
+    # Then the server is busy # Not all slots will be utilized.
+    Then the server is idle
+    And  all slots are idle
+
+    Then all token probabilities are equal
+    Examples:
+      | n_slots | n_kv | n_predict | n_parallel |
+      | 4       | 1024 | 1         | 1          |
+      # FIXME: unified KV cache nondeterminism
+      # See https://github.com/ggerganov/whisper.cpp/issues/1941#issuecomment-1986923227
+      # and https://github.com/ggerganov/llama.cpp/pull/6122#discussion_r1531405574
+      # and https://github.com/ggerganov/llama.cpp/pull/7347 .
+      # | 4       | 1024 | 1         | 4          |
+      # | 4       | 1024 | 100       | 1          |
+      # This test still fails even the above patches; the first token probabilities are already different.
+      # | 4       | 1024 | 100       | 4          |
--- a/examples/server/tests/features/steps/steps.py
+++ b/examples/server/tests/features/steps/steps.py
@@ -23,6 +23,7 @@ from prometheus_client import parser
 def step_server_config(context, server_fqdn, server_port):
    context.server_fqdn = server_fqdn
    context.server_port = int(server_port)
+    context.n_threads = None
    context.n_gpu_layer = None
    if 'PORT' in os.environ:
        context.server_port = int(os.environ['PORT'])
@@ -109,6 +110,11 @@ def step_n_gpu_layer(context, ngl):
    context.n_gpu_layer = ngl


+@step('{n_threads:d} threads')
+def step_n_threads(context, n_threads):
+    context.n_thread = n_threads
+
+
@step('{draft:d} as draft')
 def step_draft(context, draft):
    context.draft = draft
@@ -193,7 +199,7 @@ async def step_wait_for_the_server_to_be_started(context, expecting_status):

        case 'ready' | 'idle':
            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=10,
+                                         timeout=30,
                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
                                         slots_idle=context.n_slots,
                                         slots_processing=0,
@@ -274,13 +280,22 @@ async def step_predictions_equal(context):

@step('all predictions are different')
@async_run_until_complete
-async def step_predictions_equal(context):
+async def step_predictions_different(context):
    n_completions = await gather_tasks_results(context)
    assert n_completions >= 2, "need at least 2 completions"
    assert_all_predictions_different(context.tasks_result)
    context.tasks_result = []


+@step('all token probabilities are equal')
+@async_run_until_complete
+async def step_token_probabilities_equal(context):
+    n_completions = await gather_tasks_results(context)
+    assert n_completions >= 2, "need at least 2 completions"
+    assert_all_token_probabilities_equal(context.tasks_result)
+    context.tasks_result = []
+
+
@step('the completion is  truncated')
 def step_assert_completion_truncated(context):
    step_assert_completion_truncated(context, '')
@@ -868,7 +883,8 @@ async def request_completion(prompt,
                                    "cache_prompt": cache_prompt,
                                    "id_slot": id_slot,
                                    "seed": seed if seed is not None else 42,
-                                    "temperature": temperature if temperature is not None else "0.8f",
+                                    "temperature": temperature if temperature is not None else 0.8,
+                                    "n_probs": 2,
                                },
                                headers=headers,
                                timeout=3600) as response:
@@ -887,6 +903,7 @@ async def oai_chat_completions(user_prompt,
                               base_path,
                               async_client,
                               debug=False,
+                               temperature=None,
                               model=None,
                               n_predict=None,
                               enable_streaming=None,
@@ -913,7 +930,8 @@ async def oai_chat_completions(user_prompt,
        "model": model,
        "max_tokens": n_predict,
        "stream": enable_streaming,
-        "seed": seed
+        "temperature": temperature if temperature is not None else 0.0,
+        "seed": seed,
    }
    if response_format is not None:
        payload['response_format'] = response_format
@@ -939,7 +957,7 @@ async def oai_chat_completions(user_prompt,
                    while event_received:
                        event_received = False
                        async for line_in_bytes in response.content:
-                            line = line_in_bytes.decode('utf8')
+                            line = line_in_bytes.decode('utf-8')
                            line = line.rstrip('\n').rstrip('\r')
                            if line == '':
                                continue
@@ -978,7 +996,8 @@ async def oai_chat_completions(user_prompt,
                max_tokens=n_predict,
                stream=enable_streaming,
                response_format=payload.get('response_format'),
-                seed=seed
+                seed=seed,
+                temperature=payload['temperature']
            )
        except openai.error.AuthenticationError as e:
            if expect_api_error is not None and expect_api_error:
@@ -1120,6 +1139,23 @@ def assert_all_predictions_different(completion_responses):
        assert content_i != content_j, "contents not different"


+def assert_all_token_probabilities_equal(completion_responses):
+    n_predict = len(completion_responses[0]['completion_probabilities'])
+    if 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON':
+        for pos in range(n_predict):
+            for i, response_i in enumerate(completion_responses):
+                probs_i = response_i['completion_probabilities'][pos]['probs']
+                print(f"pos {pos}, probs {i}: {probs_i}")
+    for pos in range(n_predict):
+        for i, response_i in enumerate(completion_responses):
+            probs_i = response_i['completion_probabilities'][pos]['probs']
+            for j, response_j in enumerate(completion_responses):
+                if i == j:
+                    continue
+                probs_j = response_j['completion_probabilities'][pos]['probs']
+            assert probs_i == probs_j, "contents not equal"
+
+
 async def gather_tasks_results(context):
    n_tasks = len(context.concurrent_tasks)
    if context.debug:
@@ -1258,6 +1294,8 @@ def start_server_background(context):
        server_args.extend(['--batch-size', context.n_batch])
    if context.n_ubatch:
        server_args.extend(['--ubatch-size', context.n_ubatch])
+    if context.n_threads:
+        server_args.extend(['--threads', context.threads])
    if context.n_gpu_layer:
        server_args.extend(['--n-gpu-layers', context.n_gpu_layer])
    if context.draft is not None:
--- a/examples/server/utils.hpp
+++ b/examples/server/utils.hpp
@@ -371,7 +371,7 @@ static json oaicompat_completion_params_parse(
    llama_params["presence_penalty"]  = json_value(body,   "presence_penalty",  0.0);
    llama_params["seed"]              = json_value(body,   "seed",              LLAMA_DEFAULT_SEED);
    llama_params["stream"]            = json_value(body,   "stream",            false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       0.0);
+    llama_params["temperature"]       = json_value(body,   "temperature",       1.0);
    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);

    // Apply chat template to the list of messages
--- a/examples/sycl/win-build-sycl.bat
+++ b/examples/sycl/win-build-sycl.bat
@@ -13,10 +13,10 @@ if %errorlevel% neq 0 goto ERROR

 ::  for FP16
 ::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON

 ::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx -DBUILD_SHARED_LIBS=ON -DCMAKE_BUILD_TYPE=Release
 if %errorlevel% neq 0 goto ERROR
 ::  build example/main only
 ::  make main
--- a/examples/tokenize/tokenize.cpp
+++ b/examples/tokenize/tokenize.cpp
@@ -3,40 +3,390 @@

 #include <cmath>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>

-int main(int argc, char ** argv) {
-    if (argc < 3 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH PROMPT [--ids]\n" , argv[0]);
+#if defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <shellapi.h>   // For CommandLineToArgvW
+#endif
+
+static void print_usage_information(const char * argv0, FILE * stream) {
+    fprintf(stream, "usage: %s [options]\n\n", argv0);
+    fprintf(stream, "The tokenize program tokenizes a prompt using a given model,\n");
+    fprintf(stream, "and prints the resulting tokens to standard output.\n\n");
+    fprintf(stream, "It needs a model file, a prompt, and optionally other flags\n");
+    fprintf(stream, "to control the behavior of the tokenizer.\n\n");
+    fprintf(stream, "    The possible options are:\n");
+    fprintf(stream, "\n");
+    fprintf(stream, "    -h, --help                           print this help and exit\n");
+    fprintf(stream, "    -m MODEL_PATH, --model MODEL_PATH    path to model.\n");
+    fprintf(stream, "    --ids                                if given, only print numerical token IDs, and not token strings.\n");
+    fprintf(stream, "                                         The output format looks like [1, 2, 3], i.e. parseable by Python.\n");
+    fprintf(stream, "    -f PROMPT_FNAME, --file PROMPT_FNAME read prompt from a file.\n");
+    fprintf(stream, "    -p PROMPT, --prompt PROMPT           read prompt from the argument.\n");
+    fprintf(stream, "    --stdin                              read prompt from standard input.\n");
+    fprintf(stream, "    --no-bos                             do not ever add a BOS token to the prompt, even if normally the model uses a BOS token.\n");
+    fprintf(stream, "    --log-disable                        disable logs. Makes stderr quiet when loading the model.\n");
+}
+
+static void llama_log_callback_null(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) text;
+    (void) user_data;
+}
+
+static std::string read_prompt_from_file(const char * filepath, bool & success) {
+    success = false;
+
+    std::ifstream in(filepath, std::ios::binary);
+    if (!in) {
+        fprintf(stderr, "%s: could not open file '%s' for reading: %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+    // do not assume the file is seekable (e.g. /dev/stdin)
+    std::stringstream buffer;
+    buffer << in.rdbuf();
+    if (in.fail()) {
+        fprintf(stderr, "%s: could not read the entire file '%s': %s\n", __func__, filepath, strerror(errno));
+        return std::string();
+    }
+
+    success = true;
+    return buffer.str();
+}
+
+//
+// Function: ingest_args(...) -> vector<string>
+//
+//  Takes argc and argv arguments, and converts them to a vector of UTF-8 encoded
+//  strings, as an STL vector<string>.
+//
+//  In particular, it handles character encoding shenanigans on Windows.
+//
+// Note: raw_argc and raw_argv are not actually read at all on Windows.
+//       On Windows we call GetCommandLineW to get the arguments in wchar_t
+//       format, ignoring the regular argc/argv arguments to main().
+//
+// TODO: potential opportunity to roll common stuff into common/console.cpp
+//       in relation to Windows wchar_t shenanigans.
+static std::vector<std::string> ingest_args(int raw_argc, char ** raw_argv) {
+    std::vector<std::string> argv;
+
+    // Handle Windows, if given non-ASCII arguments.
+    // We convert wchar_t arguments into UTF-8 char* on this platform.
+    // Lets you invoke 'tokenize' on Windows cmd.exe with non-ASCII characters
+    // without throwing tantrums.
+#if defined(_WIN32)
+    int argc;
+    const LPWSTR cmdline_wargv = GetCommandLineW();
+    LPWSTR * wargv = CommandLineToArgvW(cmdline_wargv, &argc);
+
+    // silence unused arg warnings
+    (void) raw_argc;
+    (void) raw_argv;
+
+    for (int i = 0; i < argc; ++i) {
+        int length_needed = WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), 0, 0, NULL, NULL);
+        char * output_buf = (char *) calloc(length_needed+1, sizeof(char));
+        GGML_ASSERT(output_buf);
+
+        WideCharToMultiByte(CP_UTF8, 0, wargv[i], wcslen(wargv[i]), output_buf, length_needed, NULL, NULL);
+        output_buf[length_needed] = '\0';
+
+        argv.push_back(output_buf);
+        free(output_buf);
+    }
+
+    LocalFree((HLOCAL) wargv);
+#else
+    int argc = raw_argc;
+    for (int i = 0; i < argc; ++i) {
+        argv.push_back(raw_argv[i]);
+    }
+#endif
+
+    GGML_ASSERT((unsigned int) argc == argv.size());
+
+    return argv;
+}
+
+//
+// Function: write_utf8_cstr_to_stdout(const char *) -> <writes to stdout>
+//
+// writes a string to standard output; taking into account that on Windows
+// to display correctly you have to use special handling. Works even if the
+// user has not set a unicode code page on a Windows cmd.exe.
+//
+// In case of invalid UTF-8, invalid_utf8 is set to true on Windows, and something
+// a human-readable is written instead.
+//
+// On non-Windows systems, simply printfs() the string.
+static void write_utf8_cstr_to_stdout(const char * str, bool & invalid_utf8) {
+        invalid_utf8 = false;
+
+#if defined(_WIN32)
+        // Are we in a console?
+        HANDLE hConsole = GetStdHandle(STD_OUTPUT_HANDLE);
+        DWORD dwMode = 0;
+
+        // According to Microsoft docs:
+        // "WriteConsole fails if it is used with a standard handle that is redirected to a file."
+        // Also according to the docs, you can use GetConsoleMode to check for that.
+        if (hConsole == INVALID_HANDLE_VALUE || !GetConsoleMode(hConsole, &dwMode)) {
+            printf("%s", str);
+            return;
+        }
+
+        // MultiByteToWideChar reports an error if str is empty, don't report
+        // them as invalid_utf8.
+        if (*str == 0) {
+            return;
+        }
+        int length_needed = MultiByteToWideChar(CP_UTF8, MB_ERR_INVALID_CHARS, str, strlen(str), NULL, 0);
+        if (length_needed == 0) {
+            DWORD err = GetLastError();
+            if (err == ERROR_NO_UNICODE_TRANSLATION) {
+                invalid_utf8 = true;
+                int len = strlen(str);
+                printf("<");
+                for (int i = 0; i < len; ++i) {
+                    if (i > 0) {
+                        printf(" ");
+                    }
+                    printf("%02x", (uint8_t) str[i]);
+                }
+                printf(">");
+                return;
+            }
+            GGML_ASSERT(false && "MultiByteToWideChar() failed in an unexpected way.");
+        }
+
+        LPWSTR wstr = (LPWSTR) calloc(length_needed+1, sizeof(*wstr));
+        GGML_ASSERT(wstr);
+
+        MultiByteToWideChar(CP_UTF8, 0, str, strlen(str), wstr, length_needed);
+        WriteConsoleW(hConsole, wstr, length_needed, NULL, NULL);
+
+        free(wstr);
+#else
+        // TODO: reporting invalid_utf8 would be useful on non-Windows too.
+        // printf will silently just write bad unicode.
+        printf("%s", str);
+#endif
+}
+
+int main(int raw_argc, char ** raw_argv) {
+    const std::vector<std::string> argv = ingest_args(raw_argc, raw_argv);
+    const int argc = argv.size();
+
+    if (argc <= 1) {
+        print_usage_information(argv[0].c_str(), stderr);
        return 1;
    }

-    const char * model_path = argv[1];
-    const char * prompt     = argv[2];
+    //////
+    // Read out all the command line arguments.
+    //////

-    const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
+    // variables where to put any arguments we see.
+    bool printing_ids = false;
+    bool no_bos = false;
+    bool disable_logging = false;
+    const char * model_path = NULL;
+    const char * prompt_path = NULL;
+    const char * prompt_arg = NULL;
+
+    // track which arguments were explicitly given
+    // used for sanity checking down the line
+    bool model_path_set = false;
+    bool prompt_path_set = false;
+    bool prompt_set = false;
+    bool stdin_set = false;
+
+    int iarg = 1;
+    for (; iarg < argc; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-h" || arg == "--help") {
+            print_usage_information(argv[0].c_str(), stdout);
+            return 0;
+        }
+        else if (arg == "--ids") {
+            printing_ids = true;
+        }
+        else if (arg == "-m" || arg == "--model") {
+            if (model_path_set) {
+                fprintf(stderr, "Error: -m or --model specified multiple times.\n");
+                return 1;
+            }
+            model_path = argv[++iarg].c_str();
+            model_path_set = true;
+        }
+        else if (arg == "--no-bos") {
+            no_bos = true;
+        }
+        else if (arg == "-p" || arg == "--prompt") {
+            if (prompt_set) {
+                fprintf(stderr, "Error: -p or --prompt specified multiple times.\n");
+                return 1;
+            }
+            prompt_arg = argv[++iarg].c_str();
+            prompt_set = true;
+        }
+        else if (arg == "-f" || arg == "--file") {
+            if (prompt_path_set) {
+                fprintf(stderr, "Error: -f or --file specified multiple times.\n");
+                return 1;
+            }
+            prompt_path = argv[++iarg].c_str();
+            prompt_path_set = true;
+        }
+        else if (arg == "--stdin") {
+            stdin_set = true;
+        }
+        else if (arg == "--log-disable") {
+            disable_logging = true;
+        }
+        else {
+            fprintf(stderr, "Error: unknown option '%s'\n", argv[iarg].c_str());
+            return 1;
+        }
+    }
+
+    //////
+    // Sanity check the command line arguments.
+    //////
+
+    // Check that we have the required stuff set.
+    if (model_path_set && model_path == NULL) {
+        fprintf(stderr, "Error: --model requires an argument.\n");
+        return 1;
+    }
+    if (!model_path_set) {
+        fprintf(stderr, "Error: must specify --model.\n");
+        return 1;
+    }
+    if (prompt_path_set && prompt_path == NULL) {
+        fprintf(stderr, "Error: --file requires an argument.\n");
+        return 1;
+    }
+    if (prompt_set && prompt_arg == NULL) {
+        fprintf(stderr, "Error: --prompt requires an argument.\n");
+        return 1;
+    }
+    const int prompts_set = !!(prompt_path_set) + !!(prompt_set) + !!(stdin_set);
+    if (prompts_set > 1) {
+        fprintf(stderr, "Error: --stdin, --file and --prompt are mutually exclusive.\n");
+        return 1;
+    }
+    // Must have some prompt.
+    if (prompts_set == 0) {
+        fprintf(stderr, "Error: must specify one of: --stdin, --file or --prompt.\n");
+        return 1;
+    }
+
+    GGML_ASSERT(model_path);
+    GGML_ASSERT(prompt_path || prompt_arg || stdin_set);
+
+    //////
+    // Figure out where will the prompt come from.
+    //////
+
+    std::string prompt;
+    if (prompt_path_set) {
+        bool success = false;
+        prompt = read_prompt_from_file(prompt_path, success);
+        if (!success) {
+            return 1;
+        }
+    } else if (prompt_set) {
+        prompt = prompt_arg;
+    } else {
+        GGML_ASSERT(stdin_set);
+        // we read stdin *after* loading model (early exit if model cannot
+        // be loaded, which can be a nicer user experience)
+    }
+
+    //////
+    // Start actually doing the tokenizing stuff.
+    //////
+
+#ifdef LOG_DISABLE_LOGS
+    disable_logging = true;
+#endif
+
+    if (disable_logging) {
+        llama_log_set(llama_log_callback_null, NULL);
+    }

    llama_backend_init();

    llama_model_params model_params = llama_model_default_params();
    model_params.vocab_only = true;
    llama_model * model = llama_load_model_from_file(model_path, model_params);
+    if (!model) {
+        fprintf(stderr, "Error: could not load model from file '%s'.\n", model_path);
+        return 1;
+    }

    llama_context_params ctx_params = llama_context_default_params();
    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    if (!ctx) {
+        fprintf(stderr, "Error: could not create context.\n");
+        return 1;
+    }
+
+    // read entire prompt from stdin?
+    if (stdin_set) {
+        GGML_ASSERT(!prompt_path_set && !prompt_set);
+
+        std::stringstream stdin_buffer;
+        stdin_buffer << std::cin.rdbuf();
+        if (std::cin.fail()) {
+            fprintf(stderr, "Error: could not read the entire standard input.\n");
+            return 1;
+        }
+
+        prompt = stdin_buffer.str();
+    }
+
+    const bool model_wants_add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = model_wants_add_bos && !no_bos;

    std::vector<llama_token> tokens;
+    tokens = ::llama_tokenize(model, prompt, add_bos, true);

-    tokens = ::llama_tokenize(model, prompt, true, true);
+    if (printing_ids) {
+        printf("[");
+    }

    for (int i = 0; i < (int) tokens.size(); i++) {
        if (printing_ids) {
-            printf("%d\n", tokens[i]);
+            if (i > 0) {
+                printf(", ");
+            }
+            printf("%d", tokens[i]);
        } else {
-            printf("%6d -> '%s'\n", tokens[i], llama_token_to_piece(ctx, tokens[i]).c_str());
+            bool invalid_utf8 = false;
+            printf("%6d -> '", tokens[i]);
+            write_utf8_cstr_to_stdout(llama_token_to_piece(ctx, tokens[i]).c_str(), invalid_utf8);
+            if (invalid_utf8) {
+                printf("' (utf-8 decode failure)\n");
+            } else {
+                printf("'\n");
+            }
        }
    }

+    if (printing_ids) {
+        printf("]\n");
+    }
+
+    // silence valgrind
+    llama_free(ctx);
+    llama_free_model(model);
+
    return 0;
 }
--- a/examples/train-text-from-scratch/train-text-from-scratch.cpp
+++ b/examples/train-text-from-scratch/train-text-from-scratch.cpp
@@ -301,8 +301,8 @@ static struct ggml_tensor * llama_build_train_graphs(
        // not capturing these, to silcence warnings
        const int rope_mode = 0;

-        return ggml_rope_custom(
-            ctx, t, KQ_pos, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        return ggml_rope_ext(
+            ctx, t, KQ_pos, nullptr, n_rot, rope_mode, n_ctx, 0, rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
        );
    };

@@ -341,7 +341,8 @@ static struct ggml_tensor * llama_build_train_graphs(
        struct ggml_tensor * t15 = ggml_permute      (ctx, t12, 0, 3, 1, 2);                        set_name(t15, "t15");     assert_shape_4d(t15, N, n_embd/n_head, n_head, n_batch);
        struct ggml_tensor * t16;
        if (enable_flash_attn) {
-            t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
+            GGML_ASSERT(false && "TODO: ggml_flash_attn_ext() not yet supported");
+            //t16 = ggml_flash_attn(ctx, t13, t14, t15, true);                                        set_name(t16, "t16");     assert_shape_4d(t16, n_embd/n_head, N, n_head, n_batch);
        } else {
            struct ggml_tensor * t16_0 = ggml_mul_mat              (ctx, t14, t13);                 set_name(t16_0, "t16_0"); assert_shape_4d(t16_0, N, N, n_head, n_batch);
            struct ggml_tensor * t16_1 = ggml_scale_inplace        (ctx, t16_0, kv_scale);          set_name(t16_1, "t16_1"); assert_shape_4d(t16_1, N, N, n_head, n_batch);
--- a/flake.lock
+++ b/flake.lock
@@ -5,11 +5,11 @@
        "nixpkgs-lib": "nixpkgs-lib"
      },
      "locked": {
-        "lastModified": 1714641030,
-        "narHash": "sha256-yzcRNDoyVP7+SCNX0wmuDju1NUCt8Dz9+lyUXEI0dbI=",
+        "lastModified": 1715865404,
+        "narHash": "sha256-/GJvTdTpuDjNn84j82cU6bXztE0MSkdnTWClUCRub78=",
        "owner": "hercules-ci",
        "repo": "flake-parts",
-        "rev": "e5d10a24b66c3ea8f150e47dfdb0416ab7c3390e",
+        "rev": "8dc45382d5206bd292f9c2768b8058a8fd8311d9",
        "type": "github"
      },
      "original": {
@@ -20,11 +20,11 @@
    },
    "nixpkgs": {
      "locked": {
-        "lastModified": 1714635257,
-        "narHash": "sha256-4cPymbty65RvF1DWQfc+Bc8B233A1BWxJnNULJKQ1EY=",
+        "lastModified": 1715961556,
+        "narHash": "sha256-+NpbZRCRisUHKQJZF3CT+xn14ZZQO+KjxIIanH3Pvn4=",
        "owner": "NixOS",
        "repo": "nixpkgs",
-        "rev": "63c3a29ca82437c87573e4c6919b09a24ea61b0f",
+        "rev": "4a6b83b05df1a8bd7d99095ec4b4d271f2956b64",
        "type": "github"
      },
      "original": {
--- a/ggml-backend.c
+++ b/ggml-backend.c
@@ -1182,9 +1182,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 static char * fmt_size(size_t size) {
    static char buffer[128];
    if (size >= 1024*1024) {
-        sprintf(buffer, "%zuM", size/1024/1024);
+        snprintf(buffer, sizeof(buffer), "%zuM", size/1024/1024);
    } else {
-        sprintf(buffer, "%zuK", size/1024);
+        snprintf(buffer, sizeof(buffer), "%zuK", size/1024);
    }
    return buffer;
 }
@@ -1895,7 +1895,6 @@ void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * t

    tensor->buffer = buffer;
    tensor->data = (char *)tensor->view_src->data + tensor->view_offs;
-    tensor->backend = tensor->view_src->backend;
    ggml_backend_buffer_init_tensor(buffer, tensor);
 }

--- a/ggml-common.h
+++ b/ggml-common.h
@@ -65,13 +65,8 @@ typedef sycl::half2 ggml_half2;
 // QK = number of values after dequantization
 // QK_K = super-block size

-#ifdef GGML_QKK_64
-#define QK_K 64
-#define K_SCALE_SIZE 4
-#else
 #define QK_K 256
 #define K_SCALE_SIZE 12
-#endif // GGML_QKK_64

 #if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
 // QR = QK / number of values before dequantization
@@ -131,13 +126,8 @@ typedef sycl::half2 ggml_half2;
 #define QI4_NL (QK4_NL / (4*QR4_NL))
 #define QR4_NL 2

-#if QK_K == 64
-#define QI4_XS QI4_NL
-#define QR4_XS QR4_NL
-#else
 #define QI4_XS (QK_K / (4*QR4_XS))
 #define QR4_XS 8
-#endif

 #endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP

@@ -228,15 +218,6 @@ static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wro
 // weight is represented as x = a * q
 // 16 blocks of 16 elements each
 // Effectively 3.4375 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    uint8_t hmask[QK_K/8]; // quants - high bit
-    uint8_t qs[QK_K/4];    // quants - low 2 bits
-    uint8_t scales[2];
-    ggml_half d;           // super-block scale
-} block_q3_K;
-static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
-#else
 typedef struct {
    uint8_t hmask[QK_K/8]; // quants - high bit
    uint8_t qs[QK_K/4];    // quants - low 2 bits
@@ -244,20 +225,11 @@ typedef struct {
    ggml_half d;           // super-block scale
 } block_q3_K;
 static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
-#endif

 // 4-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 4.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d[2];     // super-block scales/mins
-    uint8_t scales[2];  // 4-bit block scales/mins
-    uint8_t qs[QK_K/2]; // 4--bit quants
-} block_q4_K;
-static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
-#else
 typedef struct {
    union {
        struct {
@@ -270,21 +242,11 @@ typedef struct {
    uint8_t qs[QK_K/2];           // 4--bit quants
 } block_q4_K;
 static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
-#endif

 // 5-bit quantization
 // 8 blocks of 32 elements each
 // weight is represented as x = a * q + b
 // Effectively 5.5 bits per weight
-#ifdef GGML_QKK_64
-typedef struct {
-    ggml_half d;             // super-block scale
-    int8_t  scales[QK_K/16]; // 8-bit block scales
-    uint8_t qh[QK_K/8];      // quants, high bit
-    uint8_t qs[QK_K/2];      // quants, low 4 bits
-} block_q5_K;
-static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
-#else
 typedef struct {
    union {
        struct {
@@ -298,7 +260,6 @@ typedef struct {
    uint8_t qs[QK_K/2];           // quants, low 4 bits
 } block_q5_K;
 static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
-#endif

 // 6-bit quantization
 // weight is represented as x = a * q
@@ -356,11 +317,7 @@ typedef struct {
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");

 // 3.4375 bpw
-#if QK_K == 64
-#define IQ3S_N_SCALE 2
-#else
 #define IQ3S_N_SCALE QK_K/64
-#endif
 typedef struct {
    ggml_half d;
    uint8_t qs[QK_K/4];
@@ -381,16 +338,9 @@ static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wron
 typedef struct {
    uint8_t  qs[QK_K/8];      // grid index, low 8 bits
    uint8_t  qh[QK_K/16];     // grid index, high 3 bits + grid shift bit (for two groups of 8)
-#if QK_K == 64
-    ggml_half d;
-#endif
    uint8_t  scales[QK_K/32]; // 3-bit block scales (4-bit if QK_K == 64)
 } block_iq1_m;
-#if QK_K == 64
-static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32 + sizeof(ggml_half), "wrong iq1_m block size/padding");
-#else
 static_assert(sizeof(block_iq1_m) == QK_K/8 + QK_K/16 + QK_K/32, "wrong iq1_m block size/padding");
-#endif

 // Used by IQ1_M quants
 typedef union {
@@ -406,9 +356,6 @@ typedef struct {
 } block_iq4_nl;
 static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");

-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-#else
 typedef struct {
    ggml_half d;
    uint16_t scales_h;
@@ -416,7 +363,6 @@ typedef struct {
    uint8_t  qs[QK_K/2];
 } block_iq4_xs;
 static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif

 #endif // GGML_COMMON_DECL
 #endif // GGML_COMMON_DECL
--- a/ggml-cuda.cu
+++ b/ggml-cuda.cu
@@ -4,7 +4,6 @@

 #include "ggml-cuda/common.cuh"
 #include "ggml-cuda/acc.cuh"
-#include "ggml-cuda/alibi.cuh"
 #include "ggml-cuda/arange.cuh"
 #include "ggml-cuda/argsort.cuh"
 #include "ggml-cuda/binbcast.cuh"
@@ -44,19 +43,59 @@
 #include <mutex>
 #include <stdint.h>
 #include <stdio.h>
+#include <stdarg.h>
+#include <stdlib.h>
 #include <string>
 #include <vector>

 static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");

+static void ggml_cuda_default_log_callback(enum ggml_log_level level, const char * msg, void * user_data) {
+    GGML_UNUSED(level);
+    GGML_UNUSED(user_data);
+    fprintf(stderr, "%s", msg);
+}
+
+ggml_log_callback ggml_cuda_log_callback = ggml_cuda_default_log_callback;
+void * ggml_cuda_log_user_data = NULL;
+
+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data) {
+    ggml_cuda_log_callback = log_callback;
+    ggml_cuda_log_user_data = user_data;
+}
+
+#define GGML_CUDA_LOG_INFO(...) ggml_cuda_log(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
+#define GGML_CUDA_LOG_WARN(...) ggml_cuda_log(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
+#define GGML_CUDA_LOG_ERROR(...) ggml_cuda_log(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
+
+GGML_ATTRIBUTE_FORMAT(2, 3)
+static void ggml_cuda_log(enum ggml_log_level level, const char * format, ...) {
+    if (ggml_cuda_log_callback != NULL) {
+        va_list args;
+        va_start(args, format);
+        char buffer[128];
+        int len = vsnprintf(buffer, 128, format, args);
+        if (len < 128) {
+            ggml_cuda_log_callback(level, buffer, ggml_cuda_log_user_data);
+        } else {
+            std::vector<char> buffer2(len + 1);  // vsnprintf adds a null terminator
+            va_end(args);
+            va_start(args, format);
+            vsnprintf(&buffer2[0], buffer2.size(), format, args);
+            ggml_cuda_log_callback(level, buffer2.data(), ggml_cuda_log_user_data);
+        }
+        va_end(args);
+    }
+}
+
 [[noreturn]]
 void ggml_cuda_error(const char * stmt, const char * func, const char * file, int line, const char * msg) {
    int id = -1; // in case cudaGetDevice fails
    cudaGetDevice(&id);

-    fprintf(stderr, "CUDA error: %s\n", msg);
-    fprintf(stderr, "  current device: %d, in function %s at %s:%d\n", id, func, file, line);
-    fprintf(stderr, "  %s\n", stmt);
+    GGML_CUDA_LOG_ERROR("CUDA error: %s\n", msg);
+    GGML_CUDA_LOG_ERROR("  current device: %d, in function %s at %s:%d\n", id, func, file, line);
+    GGML_CUDA_LOG_ERROR("  %s\n", stmt);
    // abort with GGML_ASSERT to get a stack trace
    GGML_ASSERT(!"CUDA error");
 }
@@ -92,7 +131,7 @@ static ggml_cuda_device_info ggml_cuda_init() {

    cudaError_t err = cudaGetDeviceCount(&info.device_count);
    if (err != cudaSuccess) {
-        fprintf(stderr, "%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
+        GGML_CUDA_LOG_ERROR("%s: failed to initialize " GGML_CUDA_NAME ": %s\n", __func__, cudaGetErrorString(err));
        return info;
    }

@@ -100,16 +139,16 @@ static ggml_cuda_device_info ggml_cuda_init() {

    int64_t total_vram = 0;
 #if defined(GGML_CUDA_FORCE_MMQ)
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:   yes\n", __func__);
 #else
-    fprintf(stderr, "%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: GGML_CUDA_FORCE_MMQ:   no\n", __func__);
 #endif
 #if defined(CUDA_USE_TENSOR_CORES)
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: yes\n", __func__);
 #else
-    fprintf(stderr, "%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
+    GGML_CUDA_LOG_INFO("%s: CUDA_USE_TENSOR_CORES: no\n", __func__);
 #endif
-    fprintf(stderr, "%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
+    GGML_CUDA_LOG_INFO("%s: found %d " GGML_CUDA_NAME " devices:\n", __func__, info.device_count);
    for (int id = 0; id < info.device_count; ++id) {
        int device_vmm = 0;

@@ -130,7 +169,7 @@ static ggml_cuda_device_info ggml_cuda_init() {

        cudaDeviceProp prop;
        CUDA_CHECK(cudaGetDeviceProperties(&prop, id));
-        fprintf(stderr, "  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");
+        GGML_CUDA_LOG_INFO("  Device %d: %s, compute capability %d.%d, VMM: %s\n", id, prop.name, prop.major, prop.minor, device_vmm ? "yes" : "no");

        info.default_tensor_split[id] = total_vram;
        total_vram += prop.totalGlobalMem;
@@ -236,8 +275,8 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
        *actual_size = look_ahead_size;
        pool_size += look_ahead_size;
 #ifdef DEBUG_CUDA_MALLOC
-        fprintf(stderr, "%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
-                (uint32_t)(max_size/1024/1024), (uint32_t)(pool_size/1024/1024), (uint32_t)(size/1024/1024));
+        GGML_CUDA_LOG_INFO("%s[%d]: %d buffers, max_size = %u MB, pool_size = %u MB, requested %u MB\n", __func__, device, nnz,
+                           (uint32_t)(max_size / 1024 / 1024), (uint32_t)(pool_size / 1024 / 1024), (uint32_t)(size / 1024 / 1024));
 #endif
        return ptr;
    }
@@ -251,7 +290,7 @@ struct ggml_cuda_pool_leg : public ggml_cuda_pool {
                return;
            }
        }
-        fprintf(stderr, "WARNING: cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
+        GGML_CUDA_LOG_WARN("Cuda buffer pool full, increase MAX_CUDA_BUFFERS\n");
        ggml_cuda_set_device(device);
        CUDA_CHECK(cudaFree(ptr));
        pool_size -= size;
@@ -500,7 +539,9 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_cuda_buffer_type_alloc_buffe
    void * dev_ptr;
    cudaError_t err = cudaMalloc(&dev_ptr, size);
    if (err != cudaSuccess) {
-        fprintf(stderr, "%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size/1024.0/1024.0, buft_ctx->device, cudaGetErrorString(err));
+        // clear the error
+        cudaGetLastError();
+        GGML_CUDA_LOG_ERROR("%s: allocating %.2f MiB on device %d: cudaMalloc failed: %s\n", __func__, size / 1024.0 / 1024.0, buft_ctx->device, cudaGetErrorString(err));
        return nullptr;
    }

@@ -1003,8 +1044,8 @@ static void * ggml_cuda_host_malloc(size_t size) {
    if (err != cudaSuccess) {
        // clear the error
        cudaGetLastError();
-        fprintf(stderr, "%s: warning: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
-            size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to allocate %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
        return nullptr;
    }

@@ -2205,6 +2246,9 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
                case GGML_UNARY_OP_RELU:
                    ggml_cuda_op_relu(ctx, dst);
                    break;
+                case GGML_UNARY_OP_SIGMOID:
+                    ggml_cuda_op_sigmoid(ctx, dst);
+                    break;
                case GGML_UNARY_OP_HARDSIGMOID:
                    ggml_cuda_op_hardsigmoid(ctx, dst);
                    break;
@@ -2244,7 +2288,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
            break;
        case GGML_OP_MUL_MAT:
            if (dst->src[0]->ne[3] != dst->src[1]->ne[3]) {
-                fprintf(stderr, "%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
+                GGML_CUDA_LOG_ERROR("%s: cannot compute %s: src0->ne[3] = %" PRId64 ", src1->ne[3] = %" PRId64 " - fallback to CPU\n", __func__, dst->name, dst->src[0]->ne[3], dst->src[1]->ne[3]);
                return false;
            } else {
                ggml_cuda_mul_mat(ctx, dst->src[0], dst->src[1], dst);
@@ -2277,9 +2321,6 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg
        case GGML_OP_ROPE:
            ggml_cuda_op_rope(ctx, dst);
            break;
-        case GGML_OP_ALIBI:
-            ggml_cuda_op_alibi(ctx, dst);
-            break;
        case GGML_OP_IM2COL:
            ggml_cuda_op_im2col(ctx, dst);
            break;
@@ -2301,7 +2342,7 @@ static bool ggml_cuda_compute_forward(ggml_backend_cuda_context & ctx, struct gg

    cudaError_t err = cudaGetLastError();
    if (err != cudaSuccess) {
-        fprintf(stderr, "%s: %s failed\n", __func__, ggml_op_desc(dst));
+        GGML_CUDA_LOG_ERROR("%s: %s failed\n", __func__, ggml_op_desc(dst));
        CUDA_CHECK(err);
    }

@@ -2477,7 +2518,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
        if (ggml_cuda_info().devices[cuda_ctx->device].cc < CC_AMPERE) {
            cuda_ctx->cuda_graph->disable_due_to_gpu_arch = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to GPU architecture\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to GPU architecture\n", __func__);
 #endif
        }
    }
@@ -2524,14 +2565,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
            if (node->src[0] && ggml_backend_buffer_is_cuda_split(node->src[0]->buffer)) {
                use_cuda_graph = false; // Split buffers are not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to split buffer\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to split buffer\n", __func__);
 #endif
            }

            if (node->op == GGML_OP_MUL_MAT_ID) {
                use_cuda_graph = false; // This node type is not supported by CUDA graph capture
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to mul_mat_id\n", __func__);
 #endif
            }

@@ -2540,7 +2581,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                // Changes in batch size or context size can cause changes to the grid size of some kernels.
                use_cuda_graph = false;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to batch size > 1 [%s] [%ld %ld %ld %ld]\n", __func__, node->name, node->ne[0], node->ne[1], node->ne[2], node->ne[3]);
 #endif
            }

@@ -2559,7 +2600,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
        }

        // Disable CUDA graphs (from the next token) if the use-case is demanding too many consecutive graph updates.
-        if (cuda_graph_update_required) {
+        if (use_cuda_graph && cuda_graph_update_required) {
            cuda_ctx->cuda_graph->number_consecutive_updates++;
        } else {
            cuda_ctx->cuda_graph->number_consecutive_updates = 0;
@@ -2568,7 +2609,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
        if (cuda_ctx->cuda_graph->number_consecutive_updates >= 4) {
            cuda_ctx->cuda_graph->disable_due_to_too_many_updates = true;
 #ifndef NDEBUG
-            fprintf(stderr, "%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
+            GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to too many consecutive updates\n", __func__);
 #endif
        }
    }
@@ -2606,7 +2647,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t

                bool ok = ggml_cuda_compute_forward(*cuda_ctx, node);
                if (!ok) {
-                    fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
+                    GGML_CUDA_LOG_ERROR("%s: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
                }
                GGML_ASSERT(ok);
            }
@@ -2625,7 +2666,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
                use_cuda_graph = false;
                cuda_ctx->cuda_graph->disable_due_to_failed_graph_capture = true;
 #ifndef NDEBUG
-                fprintf(stderr, "%s: disabling CUDA graphs due to failed graph capture\n", __func__);
+                GGML_CUDA_LOG_WARN("%s: disabling CUDA graphs due to failed graph capture\n", __func__);
 #endif
            } else {
                graph_evaluated_or_captured = true; // CUDA graph has been captured
@@ -2692,7 +2733,7 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
        cudaError_t stat = cudaGraphExecUpdate(cuda_ctx->cuda_graph->instance, cuda_ctx->cuda_graph->graph, &result_info);
        if (stat == cudaErrorGraphExecUpdateFailure) {
 #ifndef NDEBUG
-            fprintf(stderr, "%s: CUDA graph update failed\n", __func__);
+            GGML_CUDA_LOG_ERROR("%s: CUDA graph update failed\n", __func__);
 #endif
            // The pre-existing graph exec cannot be updated due to violated constraints
            // so instead clear error and re-instantiate
@@ -2714,12 +2755,14 @@ GGML_CALL static enum ggml_status ggml_backend_cuda_graph_compute(ggml_backend_t
 }

 GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
+    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;
    switch (op->op) {
        case GGML_OP_UNARY:
            switch (ggml_get_unary_op(op)) {
                case GGML_UNARY_OP_GELU:
                case GGML_UNARY_OP_SILU:
                case GGML_UNARY_OP_RELU:
+                case GGML_UNARY_OP_SIGMOID:
                case GGML_UNARY_OP_HARDSIGMOID:
                case GGML_UNARY_OP_HARDSWISH:
                case GGML_UNARY_OP_GELU_QUICK:
@@ -2829,7 +2872,6 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
        case GGML_OP_DIAG_MASK_INF:
        case GGML_OP_SOFT_MAX:
        case GGML_OP_ROPE:
-        case GGML_OP_ALIBI:
        case GGML_OP_IM2COL:
        case GGML_OP_POOL_2D:
        case GGML_OP_SUM_ROWS:
@@ -2841,8 +2883,16 @@ GGML_CALL static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, cons
        case GGML_OP_ARANGE:
        case GGML_OP_TIMESTEP_EMBEDDING:
        case GGML_OP_LEAKY_RELU:
-        case GGML_OP_FLASH_ATTN_EXT:
            return true;
+        case GGML_OP_FLASH_ATTN_EXT:
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            return op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128;
+#else
+            if (op->src[0]->ne[0] == 64 || op->src[0]->ne[0] == 128) {
+                return true;
+            }
+            return ggml_cuda_info().devices[cuda_ctx->device].cc >= CC_VOLTA;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
        default:
            return false;
    }
@@ -2940,13 +2990,13 @@ static ggml_guid_t ggml_backend_cuda_guid() {

 GGML_CALL ggml_backend_t ggml_backend_cuda_init(int device) {
    if (device < 0 || device >= ggml_backend_cuda_get_device_count()) {
-        fprintf(stderr, "%s: error: invalid device %d\n", __func__, device);
+        GGML_CUDA_LOG_ERROR("%s: invalid device %d\n", __func__, device);
        return nullptr;
    }

    ggml_backend_cuda_context * ctx = new ggml_backend_cuda_context(device);
    if (ctx == nullptr) {
-        fprintf(stderr, "%s: error: failed to allocate context\n", __func__);
+        GGML_CUDA_LOG_ERROR("%s: failed to allocate context\n", __func__);
        return nullptr;
    }

@@ -2990,8 +3040,8 @@ GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size
        // clear the error
        cudaGetLastError();

-        fprintf(stderr, "%s: warning: failed to register %.2f MiB of pinned memory: %s\n", __func__,
-                size/1024.0/1024.0, cudaGetErrorString(err));
+        GGML_CUDA_LOG_WARN("%s: failed to register %.2f MiB of pinned memory: %s\n", __func__,
+                           size / 1024.0 / 1024.0, cudaGetErrorString(err));
        return false;
    }
    return true;
--- a/ggml-cuda.h
+++ b/ggml-cuda.h
@@ -38,6 +38,7 @@ GGML_API GGML_CALL void ggml_backend_cuda_get_device_memory(int device, size_t *
 GGML_API GGML_CALL bool ggml_backend_cuda_register_host_buffer(void * buffer, size_t size);
 GGML_API GGML_CALL void ggml_backend_cuda_unregister_host_buffer(void * buffer);

+GGML_API void ggml_backend_cuda_log_set_callback(ggml_log_callback log_callback, void * user_data);
 #ifdef  __cplusplus
 }
 #endif
--- a/ggml-cuda/alibi.cu
+++ b/ggml-cuda/alibi.cu
@@ -1,63 +0,0 @@
-#include "alibi.cuh"
-
-static __global__ void alibi_f32(const float * x, float * dst, const int ncols, const int k_rows,
-                                 const int n_heads_log2_floor, const float m0, const float m1) {
-    const int col = blockDim.x*blockIdx.x + threadIdx.x;
-
-    if (col >= ncols) {
-        return;
-    }
-
-    const int row = blockDim.y*blockIdx.y + threadIdx.y;
-    const int i = row*ncols + col;
-
-    const int k = row/k_rows;
-
-    float m_k;
-    if (k < n_heads_log2_floor) {
-        m_k = powf(m0, k + 1);
-    } else {
-        m_k = powf(m1, 2 * (k - n_heads_log2_floor) + 1);
-    }
-
-    dst[i] = col * m_k + x[i];
-}
-
-static void alibi_f32_cuda(const float * x, float * dst, const int ncols, const int nrows,
-                           const int k_rows, const int n_heads_log2_floor, const float m0,
-                           const float m1, cudaStream_t stream) {
-    const dim3 block_dims(CUDA_ALIBI_BLOCK_SIZE, 1, 1);
-    const int num_blocks_x = (ncols + CUDA_ALIBI_BLOCK_SIZE - 1) / (CUDA_ALIBI_BLOCK_SIZE);
-    const dim3 block_nums(num_blocks_x, nrows, 1);
-    alibi_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols, k_rows, n_heads_log2_floor, m0, m1);
-}
-
-void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
-    const ggml_tensor * src0 = dst->src[0];
-    const float * src0_d = (const float *)src0->data;
-    float * dst_d = (float *)dst->data;
-    cudaStream_t stream = ctx.stream();
-
-    GGML_ASSERT(src0->type == GGML_TYPE_F32);
-    GGML_ASSERT( dst->type == GGML_TYPE_F32);
-
-    const int64_t ne00 = src0->ne[0];
-    const int64_t ne01 = src0->ne[1];
-    const int64_t ne02 = src0->ne[2];
-    const int64_t nrows = ggml_nrows(src0);
-
-    //const int n_past = ((int32_t *) dst->op_params)[0];
-    const int n_head = ((int32_t *) dst->op_params)[1];
-    float max_bias;
-    memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
-
-    //GGML_ASSERT(ne01 + n_past == ne00);
-    GGML_ASSERT(n_head == ne02);
-
-    const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
-
-    const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
-    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_heads_log2_floor);
-
-    alibi_f32_cuda(src0_d, dst_d, ne00, nrows, ne01, n_heads_log2_floor, m0, m1, stream);
-}
--- a/ggml-cuda/alibi.cuh
+++ b/ggml-cuda/alibi.cuh
@@ -1,5 +0,0 @@
-#include "common.cuh"
-
-#define CUDA_ALIBI_BLOCK_SIZE 32
-
-void ggml_cuda_op_alibi(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/common.cuh
+++ b/ggml-cuda/common.cuh
@@ -234,122 +234,6 @@ typedef float dfloat; // dequantize float
 typedef float2 dfloat2;
 #endif //GGML_CUDA_F16

-[[noreturn]]
-static __device__ void no_device_code(
-    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
-
-#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
-           file_name, line, function_name, arch);
-    GGML_UNUSED(arch_list);
-#else
-    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
-           file_name, line, function_name, arch, arch_list);
-#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
-    __trap();
-
-    GGML_UNUSED(no_device_code); // suppress unused function warning
-}
-
-#ifdef __CUDA_ARCH__
-#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
-#else
-#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
-#endif // __CUDA_ARCH__
-
-static __device__ __forceinline__ float warp_reduce_sum(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
-    }
-    return x;
-}
-
-static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
-        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
-    }
-    return a;
-}
-
-static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
-   }
-   return a;
-#else
-   GGML_UNUSED(a);
-   NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
-
-static __device__ __forceinline__ float warp_reduce_max(float x) {
-#pragma unroll
-    for (int mask = 16; mask > 0; mask >>= 1) {
-        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-    }
-    return x;
-}
-
-static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-
-#if CUDART_VERSION >= CUDART_HMAX
-    return __hmax(a, b);
-#else
-    return __half2float(a) > __half2float(b) ? a : b;
-#endif // CUDART_VERSION >= CUDART_HMAX
-
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-}
-static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
-
-#if CUDART_VERSION >= CUDART_HMAX
-    return __hmax2(a, b);
-#else
-    half2 ret;
-    reinterpret_cast<half&>(ret.x) =  __low2float(a) >  __low2float(b) ?  __low2half(a) :  __low2half(b);
-    reinterpret_cast<half&>(ret.y) = __high2float(a) > __high2float(b) ? __high2half(a) : __high2half(b);
-    return ret;
-#endif // CUDART_VERSION >= CUDART_HMAX
-
-#else
-    GGML_UNUSED(a);
-    GGML_UNUSED(b);
-    NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
-}
-
-static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
-#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-#pragma unroll
-   for (int mask = 16; mask > 0; mask >>= 1) {
-       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
-   }
-   return x;
-#else
-   GGML_UNUSED(x);
-   NO_DEVICE_CODE;
-#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
-}
-
-#if CUDART_VERSION < CUDART_HMASK
-static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
-    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
-    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
-    return mask_low | mask_high;
-}
-#endif // CUDART_VERSION < 12000
-
 #if defined(GGML_USE_HIPBLAS)
 #define __CUDA_ARCH__ 1300

@@ -431,18 +315,179 @@ static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #endif
    return c;
 }
+
+#if defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
+// __shfl_xor() for half2 was added in ROCm 5.6
+static __device__ __forceinline__ half2 __shfl_xor(half2 var, int laneMask, int width) {
+    typedef union half2_b32 {
+        half2 val;
+        int   b32;
+    } half2_b32_t;
+    half2_b32_t tmp;
+    tmp.val = var;
+    tmp.b32 = __shfl_xor(tmp.b32, laneMask, width);
+    return tmp.val;
+}
+#endif // defined(__HIP_PLATFORM_AMD__) && HIP_VERSION < 50600000
 #endif // defined(GGML_USE_HIPBLAS)

-#define FP16_AVAILABLE     defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__) ? \
-    defined(RDNA1) || defined(RDNA2) || defined(RDNA3) : __CUDA_ARCH__ >= CC_PASCAL
+#define FP16_AVAILABLE (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) || __CUDA_ARCH__ >= CC_PASCAL

 #define FP16_MMA_AVAILABLE !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_VOLTA

+static bool fast_fp16_available(const int cc) {
+    return cc >= CC_PASCAL && cc != 610;
+}
+
+static bool fp16_mma_available(const int cc) {
+    return cc < CC_OFFSET_AMD && cc >= CC_VOLTA;
+}
+
+[[noreturn]]
+static __device__ void no_device_code(
+    const char * file_name, const int line, const char * function_name, const int arch, const char * arch_list) {
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    printf("%s:%d: ERROR: HIP kernel %s has no device code compatible with HIP arch %d.\n",
+           file_name, line, function_name, arch);
+    GGML_UNUSED(arch_list);
+#else
+    printf("%s:%d: ERROR: CUDA kernel %s has no device code compatible with CUDA arch %d. ggml-cuda.cu was compiled for: %s\n",
+           file_name, line, function_name, arch, arch_list);
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    __trap();
+
+    GGML_UNUSED(no_device_code); // suppress unused function warning
+}
+
+#ifdef __CUDA_ARCH__
+#define NO_DEVICE_CODE no_device_code(__FILE__, __LINE__, __FUNCTION__, __CUDA_ARCH__, STRINGIZE(__CUDA_ARCH_LIST__))
+#else
+#define NO_DEVICE_CODE //GGML_ASSERT(false && "NO_DEVICE_CODE not valid in host code.")
+#endif // __CUDA_ARCH__
+
+static __device__ __forceinline__ float warp_reduce_sum(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x += __shfl_xor_sync(0xffffffff, x, mask, 32);
+    }
+    return x;
+}
+
+static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a.x += __shfl_xor_sync(0xffffffff, a.x, mask, 32);
+        a.y += __shfl_xor_sync(0xffffffff, a.y, mask, 32);
+    }
+    return a;
+}
+
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if FP16_AVAILABLE
+
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        const half2 a_other = __shfl_xor_sync(0xffffffff, a, mask, 32);
+        reinterpret_cast<half&>(a.x) +=  __low2half(a_other);
+        reinterpret_cast<half&>(a.y) += __high2half(a_other);
+    }
+    return a;
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+#else
+    NO_DEVICE_CODE;
+    return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ float warp_reduce_max(float x) {
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = fmaxf(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+}
+
+static __device__ __forceinline__ half ggml_cuda_hmax(const half a, const half b) {
+#if FP16_AVAILABLE
+
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+    return __float2half(fmaxf(__half2float(a), __half2float(b)));
+#else
+    return __hmax(a, b);
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && CUDART_VERSION < CUDART_HMAX
+
+#else
+   NO_DEVICE_CODE;
+   GGML_UNUSED(b);
+   return a;
+#endif // FP16_AVAILABLE
+}
+
+static __device__ __forceinline__ half2 ggml_cuda_hmax2(const half2 a, const half2 b) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+
+#if CUDART_VERSION >= CUDART_HMAX
+    return __hmax2(a, b);
+#else
+    half2 ret;
+    reinterpret_cast<half&>(ret.x) = __float2half(fmaxf( __low2float(a),  __low2float(b)));
+    reinterpret_cast<half&>(ret.y) = __float2half(fmaxf(__high2float(a), __high2float(b)));
+    return ret;
+#endif // CUDART_VERSION >= CUDART_HMAX
+
+#else
+    GGML_UNUSED(a);
+    GGML_UNUSED(b);
+    NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+#pragma unroll
+   for (int mask = 16; mask > 0; mask >>= 1) {
+       x = ggml_cuda_hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+   }
+   return x;
+#else
+   GGML_UNUSED(x);
+   NO_DEVICE_CODE;
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+
+#if CUDART_VERSION < CUDART_HMASK
+static __device__ __forceinline__ uint32_t __hgt2_mask(const half2 a, const half2 b) {
+    const uint32_t mask_low  = 0x0000FFFF * (float( __low2half(a)) > float( __low2half(b)));
+    const uint32_t mask_high = 0xFFFF0000 * (float(__high2half(a)) > float(__high2half(b)));
+    return mask_low | mask_high;
+}
+#endif // CUDART_VERSION < 12000
+
 // TODO: move to ggml-common.h
 static const __device__ int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};

 typedef void (*dequantize_kernel_t)(const void * vx, const int64_t ib, const int iqs, dfloat2 & v);

+static __device__ __forceinline__ float get_alibi_slope(
+    const float max_bias, const uint32_t h, const uint32_t n_head_log2, const float m0, const float m1
+) {
+    if (max_bias <= 0.0f) {
+        return 1.0f;
+    }
+    const float base = h < n_head_log2 ? m0 : m1;
+    const int   exph = h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1;
+
+    return powf(base, exph);
+}

 //////////////////////

--- a/ggml-cuda/convert.cu
+++ b/ggml-cuda/convert.cu
@@ -131,7 +131,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
    const block_q2_K * x = (const block_q2_K *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t n   = tid/32;
    const int64_t l   = tid - 32*n;
    const int64_t is  = 8*n + l/16;
@@ -145,17 +144,6 @@ static __global__ void dequantize_block_q2_K(const void * __restrict__ vx, dst_t
    y[l+32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 2) & 3) - dmin * (x[i].scales[is+2] >> 4);
    y[l+64] = dall * (x[i].scales[is+4] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+4] >> 4);
    y[l+96] = dall * (x[i].scales[is+6] & 0xF) * ((q >> 6) & 3) - dmin * (x[i].scales[is+6] >> 4);
-#else
-    const int64_t is = tid/16;  // 0 or 1
-    const int64_t il = tid%16;  // 0...15
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    dst_t * y = yy + i*QK_K + 16*is + il;
-    float dall = __low2half(x[i].dm);
-    float dmin = __high2half(x[i].dm);
-    y[ 0] = dall * (x[i].scales[is+0] & 0xF) * ((q >> 0) & 3) - dmin * (x[i].scales[is+0] >> 4);
-    y[32] = dall * (x[i].scales[is+2] & 0xF) * ((q >> 4) & 3) - dmin * (x[i].scales[is+2] >> 4);
-#endif
-
 }

 template<typename dst_t>
@@ -164,7 +152,6 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
    const int64_t i = blockIdx.x;
    const block_q3_K * x = (const block_q3_K *) vx;

-#if QK_K == 256
    const int64_t r = threadIdx.x/4;
    const int64_t tid = r/2;
    const int64_t is0 = r%2;
@@ -188,31 +175,8 @@ static __global__ void dequantize_block_q3_K(const void * __restrict__ vx, dst_t
    const uint8_t * hm = x[i].hmask;

    for (int l = l0; l < l0+4; ++l) y[l] = dl * ((int8_t)((q[l] >> shift) & 3) - ((hm[l] & m) ? 0 : 4));
-#else
-    const int64_t tid = threadIdx.x;
-    const int64_t is  = tid/16;  // 0 or 1
-    const int64_t il  = tid%16;  // 0...15
-    const int64_t im  = il/8;    // 0...1
-    const int64_t in  = il%8;    // 0...7
-
-    dst_t * y = yy + i*QK_K + 16*is + il;
-
-    const uint8_t q = x[i].qs[il] >> (2*is);
-    const uint8_t h = x[i].hmask[in] >> (2*is + im);
-    const float   d = (float)x[i].d;
-
-    if (is == 0) {
-        y[ 0] = d * ((x[i].scales[0] & 0xF) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] & 0xF) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    } else {
-        y[ 0] = d * ((x[i].scales[0] >>  4) - 8) * ((int8_t)((q >> 0) & 3) - ((h >> 0) & 1 ? 0 : 4));
-        y[32] = d * ((x[i].scales[1] >>  4) - 8) * ((int8_t)((q >> 4) & 3) - ((h >> 4) & 1 ? 0 : 4));
-    }
-#endif
-
 }

-#if QK_K == 256
 static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t & d, uint8_t & m) {
    if (j < 4) {
        d = q[j] & 63; m = q[j + 4] & 63;
@@ -221,7 +185,6 @@ static inline __device__ void get_scale_min_k4(int j, const uint8_t * q, uint8_t
        m = (q[j+4] >>  4) | ((q[j-0] >> 6) << 4);
    }
 }
-#endif

 template<typename dst_t>
 static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t * __restrict__ yy) {
@@ -229,7 +192,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t

    const int64_t i = blockIdx.x;

-#if QK_K == 256
    // assume 32 threads
    const int64_t tid = threadIdx.x;
    const int64_t il  = tid/8;
@@ -253,15 +215,6 @@ static __global__ void dequantize_block_q4_K(const void * __restrict__ vx, dst_t
        y[l + 0] = d1 * (q[l] & 0xF) - m1;
        y[l +32] = d2 * (q[l] >>  4) - m2;
    }
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t * q = x[i].qs;
-    dst_t * y = yy + i*QK_K;
-    const float d = (float)x[i].dm[0];
-    const float m = (float)x[i].dm[1];
-    y[tid+ 0] = d * (x[i].scales[0] & 0xF) * (q[tid] & 0xF) - m * (x[i].scales[0] >> 4);
-    y[tid+32] = d * (x[i].scales[1] & 0xF) * (q[tid] >>  4) - m * (x[i].scales[1] >> 4);
-#endif
 }

 template<typename dst_t>
@@ -270,7 +223,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t

    const int64_t i = blockIdx.x;

-#if QK_K == 256
    // assume 64 threads - this is very slightly better than the one below
    const int64_t tid = threadIdx.x;
    const int64_t il  = tid/16;   // il is in 0...3
@@ -297,18 +249,6 @@ static __global__ void dequantize_block_q5_K(const void * __restrict__ vx, dst_t
    hm <<= 1;
    y[32] = d2 * ((ql[ 0] >>  4) + (qh[ 0] & hm ? 16 : 0)) - m2;
    y[33] = d2 * ((ql[ 1] >>  4) + (qh[ 1] & hm ? 16 : 0)) - m2;
-#else
-    const int64_t tid = threadIdx.x;
-    const uint8_t q = x[i].qs[tid];
-    const int64_t im = tid/8;  // 0...3
-    const int64_t in = tid%8;  // 0...7
-    const int64_t is = tid/16; // 0 or 1
-    const uint8_t h = x[i].qh[in] >> im;
-    const float d = x[i].d;
-    dst_t * y = yy + i*QK_K + tid;
-    y[ 0] = d * x[i].scales[is+0] * ((q & 0xF) - ((h >> 0) & 1 ? 0 : 16));
-    y[32] = d * x[i].scales[is+2] * ((q >>  4) - ((h >> 4) & 1 ? 0 : 16));
-#endif
 }

 template<typename dst_t>
@@ -316,7 +256,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
    const block_q6_K * x = (const block_q6_K *) vx;

    const int64_t i = blockIdx.x;
-#if QK_K == 256

    // assume 64 threads - this is very slightly better than the one below
    const int64_t tid = threadIdx.x;
@@ -336,24 +275,6 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
    y[32] = d * sc[2] * ((int8_t)((ql[32] & 0xF) | (((qh >> 2) & 3) << 4)) - 32);
    y[64] = d * sc[4] * ((int8_t)((ql[ 0]  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
    y[96] = d * sc[6] * ((int8_t)((ql[32]  >> 4) | (((qh >> 6) & 3) << 4)) - 32);
-#else
-
-    // assume 32 threads
-    const int64_t tid = threadIdx.x;
-    const int64_t ip  = tid/16;         // 0 or 1
-    const int64_t il  = tid - 16*ip;    // 0...15
-
-    dst_t * y = yy + i*QK_K + 16*ip + il;
-
-    const float d = x[i].d;
-
-    const uint8_t   ql = x[i].ql[16*ip + il];
-    const uint8_t   qh = x[i].qh[il] >> (2*ip);
-    const int8_t  * sc = x[i].scales;
-
-    y[ 0] = d * sc[ip+0] * ((int8_t)((ql & 0xF) | (((qh >> 0) & 3) << 4)) - 32);
-    y[32] = d * sc[ip+2] * ((int8_t)((ql  >> 4) | (((qh >> 4) & 3) << 4)) - 32);
-#endif
 }

 template<typename dst_t>
@@ -363,7 +284,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -374,10 +294,6 @@ static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, ds
    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -387,7 +303,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
    const block_iq2_xs * x = (const block_iq2_xs *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -396,10 +311,6 @@ static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst
    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -409,7 +320,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
    const block_iq2_s * x = (const block_iq2_s *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -417,10 +327,6 @@ static __global__ void dequantize_block_iq2_s(const void * __restrict__ vx, dst_
    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
    const uint8_t signs = x[i].qs[QK_K/8+4*ib+il];
    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -430,7 +336,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
    const block_iq3_xxs * x = (const block_iq3_xxs  *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -445,10 +350,6 @@ static __global__ void dequantize_block_iq3_xxs(const void * __restrict__ vx, ds
        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
    }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -458,7 +359,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
    const block_iq3_s * x = (const block_iq3_s *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -471,10 +371,6 @@ static __global__ void dequantize_block_iq3_s(const void * __restrict__ vx, dst_
        y[j+0] = d * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
        y[j+4] = d * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
    }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -484,7 +380,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
    const block_iq1_s * x = (const block_iq1_s  *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -497,10 +392,6 @@ static __global__ void dequantize_block_iq1_s(const void * __restrict__ vx, dst_
    for (int j = 0; j < 8; ++j) {
        y[j] = d * (q[j] + delta);
    }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

 template<typename dst_t>
@@ -510,7 +401,6 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
    const block_iq1_m * x = (const block_iq1_m  *) vx;

    const int64_t tid = threadIdx.x;
-#if QK_K == 256
    const int64_t il = tid/8; // 0...3
    const int64_t ib = tid%8; // 0...7
    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
@@ -527,13 +417,8 @@ static __global__ void dequantize_block_iq1_m(const void * __restrict__ vx, dst_
    for (int j = 0; j < 8; ++j) {
        y[j] = d * (q[j] + delta);
    }
-#else
-    NO_DEVICE_CODE;
-#endif
-
 }

-
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst_t * __restrict__ yy) {

@@ -550,10 +435,8 @@ static __global__ void dequantize_block_iq4_nl(const void * __restrict__ vx, dst
        y[j+ 0] = d * kvalues_iq4nl[q4[j] & 0xf];
        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
    }
-
 }

-#if QK_K != 64
 template<typename dst_t>
 static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
    const int64_t i   = blockIdx.x;
@@ -570,7 +453,6 @@ static __global__ void dequantize_block_iq4_xs(const void * __restrict__ vx, dst
        y[j+16] = d * kvalues_iq4nl[q4[j] >>  4];
    }
 }
-#endif

 template <int qk, int qr, dequantize_kernel_t dequantize_kernel, typename dst_t>
 static void dequantize_block_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int64_t k, cudaStream_t stream) {
@@ -592,21 +474,13 @@ static void dequantize_block_q8_0_f16_cuda(const void * __restrict__ vx, half *
 template<typename dst_t>
 static void dequantize_row_q2_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
    const int nb = k / QK_K;
-#if QK_K == 256
    dequantize_block_q2_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q2_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }

 template<typename dst_t>
 static void dequantize_row_q3_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
    const int nb = k / QK_K;
-#if QK_K == 256
    dequantize_block_q3_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q3_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }

 template<typename dst_t>
@@ -632,21 +506,13 @@ static void dequantize_row_q4_K_cuda(const void * vx, dst_t * y, const int64_t k
 template<typename dst_t>
 static void dequantize_row_q5_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
    const int nb = k / QK_K;
-#if QK_K == 256
    dequantize_block_q5_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q5_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }

 template<typename dst_t>
 static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
    const int nb = k / QK_K;
-#if QK_K == 256
    dequantize_block_q6_K<<<nb, 64, 0, stream>>>(vx, y);
-#else
-    dequantize_block_q6_K<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }

 template<typename dst_t>
@@ -700,11 +566,7 @@ static void dequantize_row_iq1_m_cuda(const void * vx, dst_t * y, const int64_t
 template<typename dst_t>
 static void dequantize_row_iq4_xs_cuda(const void * vx, dst_t * y, const int64_t k, cudaStream_t stream) {
    const int nb = (k + QK_K - 1) / QK_K;
-#if QK_K == 64
-    dequantize_block_iq4_nl<<<nb, 32, 0, stream>>>(vx, y);
-#else
    dequantize_block_iq4_xs<<<nb, 32, 0, stream>>>(vx, y);
-#endif
 }

 template <typename src_t, typename dst_t>
--- a/ggml-cuda/dmmv.cu
+++ b/ggml-cuda/dmmv.cu
@@ -22,7 +22,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,

    float tmp = 0; // partial sum for thread in warp

-#if QK_K == 256
    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...15
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0,1

@@ -71,37 +70,6 @@ static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx,
        tmp += dall * sum1 - dmin * sum2;

    }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;
-
-    uint32_t uaux[2];
-    const uint8_t * d = (const uint8_t *)uaux;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint32_t * s = (const uint32_t *)x[i].scales;
-
-        uaux[0] = s[0] & 0x0f0f0f0f;
-        uaux[1] = (s[0] >> 4) & 0x0f0f0f0f;
-
-        const float2 dall = __half22float2(x[i].dm);
-
-        float sum1 = 0, sum2 = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t ql = q[l];
-            sum1 += y[l+ 0] * d[0] * ((ql >> 0) & 3)
-                  + y[l+16] * d[1] * ((ql >> 2) & 3)
-                  + y[l+32] * d[2] * ((ql >> 4) & 3)
-                  + y[l+48] * d[3] * ((ql >> 6) & 3);
-            sum2 += y[l+0] * d[4] + y[l+16] * d[5] + y[l+32] * d[6] + y[l+48] * d[7];
-        }
-        tmp += dall.x * sum1 - dall.y * sum2;
-    }
-#endif

    // sum up partial sums and write back result
    tmp = warp_reduce_sum(tmp);
@@ -123,8 +91,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,

    float tmp = 0; // partial sum for thread in warp

-#if QK_K == 256
-
    const uint16_t kmask1 = 0x0303;
    const uint16_t kmask2 = 0x0f0f;

@@ -175,34 +141,6 @@ static __global__ void dequantize_mul_mat_vec_q3_k(const void * __restrict__ vx,
        tmp += d * sum;

    }
-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15 or 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0....1 or 0...3
-    const int offset = tid * K_QUANTS_PER_ITERATION;         // 0...15 or 0...14
-    const int in = offset/8;                                 // 0 or 1
-    const int im = offset%8;                                 // 0...7
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y = yy + i * QK_K + offset;
-        const uint8_t * q = x[i].qs + offset;
-        const uint8_t * s = x[i].scales;
-
-        const float dall = (float)x[i].d;
-
-        float sum = 0;
-        for (int l = 0; l < K_QUANTS_PER_ITERATION; ++l) {
-            const uint8_t hl = x[i].hmask[im+l] >> in;
-            const uint8_t ql = q[l];
-            sum += y[l+ 0] * dall * ((s[0] & 0xF) - 8) * ((int8_t)((ql >> 0) & 3) - ((hl >> 0) & 1 ? 0 : 4))
-                 + y[l+16] * dall * ((s[0] >>  4) - 8) * ((int8_t)((ql >> 2) & 3) - ((hl >> 2) & 1 ? 0 : 4))
-                 + y[l+32] * dall * ((s[1] & 0xF) - 8) * ((int8_t)((ql >> 4) & 3) - ((hl >> 4) & 1 ? 0 : 4))
-                 + y[l+48] * dall * ((s[1] >>  4) - 8) * ((int8_t)((ql >> 6) & 3) - ((hl >> 6) & 1 ? 0 : 4));
-        }
-        tmp += sum;
-    }
-#endif

    // sum up partial sums and write back result
    tmp = warp_reduce_sum(tmp);
@@ -221,7 +159,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,

    const block_q4_K * x = (const block_q4_K *)vx + ib0;

-#if QK_K == 256
    const uint16_t kmask1 = 0x3f3f;
    const uint16_t kmask2 = 0x0f0f;
    const uint16_t kmask3 = 0xc0c0;
@@ -306,36 +243,6 @@ static __global__ void dequantize_mul_mat_vec_q4_k(const void * __restrict__ vx,
 #endif

    }
-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    uint16_t aux16[2];
-    const uint8_t * s = (const uint8_t *)aux16;
-
-    float tmp = 0;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const float   * y = yy + i*QK_K + step;
-        const uint16_t * a = (const uint16_t *)x[i].scales;
-        aux16[0] = a[0] & 0x0f0f;
-        aux16[1] = (a[0] >> 4) & 0x0f0f;
-        const float d = (float)x[i].dm[0];
-        const float m = (float)x[i].dm[1];
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * (d * s[0] * (q[j+ 0] & 0xF) - m * s[2])
-                 + y[j+16] * (d * s[0] * (q[j+16] & 0xF) - m * s[2])
-                 + y[j+32] * (d * s[1] * (q[j+ 0] >>  4) - m * s[3])
-                 + y[j+48] * (d * s[1] * (q[j+16] >>  4) - m * s[3]);
-        }
-        tmp += sum;
-    }
-
-#endif

    // sum up partial sums and write back result
    tmp = warp_reduce_sum(tmp);
@@ -355,7 +262,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,

    float tmp = 0; // partial sum for thread in warp

-#if QK_K == 256
    const uint16_t kmask1 = 0x3f3f;
    const uint16_t kmask2 = 0x0f0f;
    const uint16_t kmask3 = 0xc0c0;
@@ -426,30 +332,6 @@ static __global__ void dequantize_mul_mat_vec_q5_k(const void * __restrict__ vx,
        tmp += dall * (sum.x * sc[0] + sum.y * sc[1] + sum.z * sc[4] + sum.w * sc[5]) - dmin * smin;
    }

-#else
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...15
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);
-    const int step = tid * K_QUANTS_PER_ITERATION;
-    const int im = step/8;
-    const int in = step%8;
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-        const uint8_t * q = x[i].qs + step;
-        const int8_t  * s = x[i].scales;
-        const float   * y = yy + i*QK_K + step;
-        const float     d = x[i].d;
-        float sum = 0.f;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            const uint8_t h = x[i].qh[in+j] >> im;
-            sum += y[j+ 0] * d * s[0] * ((q[j+ 0] & 0xF) - ((h >> 0) & 1 ? 0 : 16))
-                 + y[j+16] * d * s[1] * ((q[j+16] & 0xF) - ((h >> 2) & 1 ? 0 : 16))
-                 + y[j+32] * d * s[2] * ((q[j+ 0] >>  4) - ((h >> 4) & 1 ? 0 : 16))
-                 + y[j+48] * d * s[3] * ((q[j+16] >>  4) - ((h >> 6) & 1 ? 0 : 16));
-        }
-        tmp += sum;
-    }
-#endif
-
    // sum up partial sums and write back result
    tmp = warp_reduce_sum(tmp);

@@ -470,8 +352,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,

    const block_q6_K * x = (const block_q6_K *)vx + ib0;

-#if QK_K == 256
-
    const int tid = threadIdx.x/K_QUANTS_PER_ITERATION;  // 0...31 or 0...16
    const int ix  = threadIdx.x%K_QUANTS_PER_ITERATION;  // 0 or 0, 1

@@ -526,37 +406,6 @@ static __global__ void dequantize_mul_mat_vec_q6_k(const void * __restrict__ vx,

    }

-#else
-
-    const int tid = threadIdx.x/(2*K_QUANTS_PER_ITERATION);  // 0...7
-    const int ix  = threadIdx.x%(2*K_QUANTS_PER_ITERATION);  // 0...3
-
-    const int step = tid * K_QUANTS_PER_ITERATION;
-
-    float tmp = 0; // partial sum for thread in warp
-
-    for (int i = ix; i < num_blocks_per_row; i += 2*K_QUANTS_PER_ITERATION) {
-
-        const float   * y  = yy + i * QK_K + step;
-        const uint8_t * ql = x[i].ql + step;
-        const uint8_t * qh = x[i].qh + step;
-        const int8_t  * s  = x[i].scales;
-
-        const float d = x[i+0].d;
-
-        float sum = 0;
-        for (int j = 0; j < K_QUANTS_PER_ITERATION; ++j) {
-            sum += y[j+ 0] * s[0] * d * ((int8_t)((ql[j+ 0] & 0xF) | ((qh[j] & 0x03) << 4)) - 32)
-                 + y[j+16] * s[1] * d * ((int8_t)((ql[j+16] & 0xF) | ((qh[j] & 0x0c) << 2)) - 32)
-                 + y[j+32] * s[2] * d * ((int8_t)((ql[j+ 0] >>  4) | ((qh[j] & 0x30) >> 0)) - 32)
-                 + y[j+48] * s[3] * d * ((int8_t)((ql[j+16] >>  4) | ((qh[j] & 0xc0) >> 2)) - 32);
-        }
-        tmp += sum;
-
-    }
-
-#endif
-
    // sum up partial sums and write back result
    tmp = warp_reduce_sum(tmp);

--- a/ggml-cuda/fattn-common.cuh
+++ b/ggml-cuda/fattn-common.cuh
@@ -0,0 +1,162 @@
+#include "common.cuh"
+
+#include <cstdint>
+
+#define FATTN_KQ_STRIDE       256
+#define HALF_MAX_HALF         __float2half(65504.0f/2) // Use neg. of this instead of -INFINITY to initialize KQ max vals to avoid NaN upon subtraction.
+#define SOFTMAX_FTZ_THRESHOLD -20.0f                   // Softmax exp. of values smaller than this are flushed to zero to avoid NaNs.
+
+typedef void (* fattn_kernel_t)(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3);
+
+template<int D, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_combine_results(
+        const float  * __restrict__ VKQ_parts,
+        const float2 * __restrict__ VKQ_meta,
+        float * __restrict__ dst) {
+    VKQ_parts += parallel_blocks*D * gridDim.y*blockIdx.x;
+    VKQ_meta  += parallel_blocks   * gridDim.y*blockIdx.x;
+    dst       +=                 D * gridDim.y*blockIdx.x;
+
+    const int tid = threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ float2 meta[parallel_blocks];
+    if (tid < 2*parallel_blocks) {
+        ((float *) meta)[threadIdx.x] = ((const float *)VKQ_meta) [blockIdx.y*(2*parallel_blocks) + tid];
+    }
+
+    __syncthreads();
+
+    float kqmax = meta[0].x;
+#pragma unroll
+    for (int l = 1; l < parallel_blocks; ++l) {
+        kqmax = max(kqmax, meta[l].x);
+    }
+
+    float VKQ_numerator   = 0.0f;
+    float VKQ_denominator = 0.0f;
+#pragma unroll
+    for (int l = 0; l < parallel_blocks; ++l) {
+        const float diff = meta[l].x - kqmax;
+        const float KQ_max_scale = expf(diff);
+        const uint32_t ftz_mask = 0xFFFFFFFF * (diff > SOFTMAX_FTZ_THRESHOLD);
+        *((uint32_t *) &KQ_max_scale) &= ftz_mask;
+
+        VKQ_numerator   += KQ_max_scale * VKQ_parts[l*gridDim.y*D + blockIdx.y*D + tid];
+        VKQ_denominator += KQ_max_scale * meta[l].y;
+    }
+
+    dst[blockIdx.y*D + tid] = VKQ_numerator / VKQ_denominator;
+}
+
+template <int D, int parallel_blocks>
+void launch_fattn(ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, int nwarps, int cols_per_block) {
+    const ggml_tensor * Q = dst->src[0];
+    const ggml_tensor * K = dst->src[1];
+    const ggml_tensor * V = dst->src[2];
+
+    const ggml_tensor * mask = dst->src[3];
+
+    ggml_tensor * KQV = dst;
+
+    GGML_ASSERT(Q->type == GGML_TYPE_F32);
+    GGML_ASSERT(K->type == GGML_TYPE_F16);
+    GGML_ASSERT(V->type == GGML_TYPE_F16);
+    GGML_ASSERT(KQV->type == GGML_TYPE_F32);
+
+    GGML_ASSERT(!mask || mask->type == GGML_TYPE_F16);
+    GGML_ASSERT(!mask || mask->ne[1] >= GGML_PAD(Q->ne[1], 16) &&
+                                "the Flash-Attention CUDA kernel requires the mask to be padded to 16 and at least n_queries big");
+
+    GGML_ASSERT(K->ne[1] % FATTN_KQ_STRIDE == 0 && "Incorrect KV cache padding.");
+
+    ggml_cuda_pool & pool = ctx.pool();
+    cudaStream_t main_stream = ctx.stream();
+
+    ggml_cuda_pool_alloc<float>  dst_tmp(pool);
+    ggml_cuda_pool_alloc<float2> dst_tmp_meta(pool);
+
+    if (parallel_blocks > 1) {
+        dst_tmp.alloc(parallel_blocks*ggml_nelements(KQV));
+        dst_tmp_meta.alloc(parallel_blocks*ggml_nrows(KQV));
+    }
+
+    const dim3 block_dim(WARP_SIZE, nwarps, 1);
+    const dim3 blocks_num(parallel_blocks*((Q->ne[1] + cols_per_block - 1) / cols_per_block), Q->ne[2], Q->ne[3]);
+    const int  shmem = 0;
+
+    float scale    = 1.0f;
+    float max_bias = 0.0f;
+
+    memcpy(&scale,    (float *) KQV->op_params + 0, sizeof(float));
+    memcpy(&max_bias, (float *) KQV->op_params + 1, sizeof(float));
+
+    const uint32_t n_head      = Q->ne[2];
+    const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head));
+
+    const float m0 = powf(2.0f, -(max_bias       ) / n_head_log2);
+    const float m1 = powf(2.0f, -(max_bias / 2.0f) / n_head_log2);
+
+    fattn_kernel<<<blocks_num, block_dim, shmem, main_stream>>>(
+        (const char *) Q->data,
+        (const char *) K->data,
+        (const char *) V->data,
+        mask ? ((const char *) mask->data) : nullptr,
+        (parallel_blocks) == 1 ? (float *) KQV->data : dst_tmp.ptr, dst_tmp_meta.ptr,
+        scale, max_bias, m0, m1, n_head_log2,
+        Q->ne[0], Q->ne[1], Q->ne[2], Q->ne[3],
+        K->ne[0], K->ne[1], K->ne[2], K->ne[3],
+        mask ? mask->ne[1] : 0, mask ?  mask->nb[1] : 0,
+        Q->nb[1], Q->nb[2], Q->nb[3],
+        K->nb[1], K->nb[2], K->nb[3],
+        KQV->ne[0], KQV->ne[1], KQV->ne[2], KQV->ne[3]
+    );
+    CUDA_CHECK(cudaGetLastError());
+
+    if ((parallel_blocks) == 1) {
+        return;
+    }
+
+    const dim3 block_dim_combine(D, 1, 1);
+    const dim3 blocks_num_combine(Q->ne[1], blocks_num.y, blocks_num.z);
+    const int  shmem_combine = 0;
+
+    flash_attn_combine_results<D, parallel_blocks>
+        <<<blocks_num_combine, block_dim_combine, shmem_combine, main_stream>>>
+        (dst_tmp.ptr, dst_tmp_meta.ptr, (float *) KQV->data);
+    CUDA_CHECK(cudaGetLastError());
+}
--- a/ggml-cuda/fattn-tile-f16.cu
+++ b/ggml-cuda/fattn-tile-f16.cu
@@ -0,0 +1,316 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f16.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F16 64
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ half KQ[ncols*FATTN_KQ_STRIDE_TILE_F16];
+    half2 * KQ2 = (half2 *) KQ;
+
+    __shared__ half2 KV_tmp[FATTN_KQ_STRIDE_TILE_F16][D/2 + 1]; // Pad D to avoid memory bank conflicts.
+
+    half kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -HALF_MAX_HALF;
+    }
+    half2 kqsum[ncols/nwarps] = {{0.0f, 0.0f}};
+
+    half2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ half2 Q_h2[ncols][D/2];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
+            Q_h2[j][i] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F16;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F16) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        half kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                KV_tmp[i_KQ][k_KQ] = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+            }
+        }
+
+        __syncthreads();
+
+        half2 sum2[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE][ncols/nwarps] = {{{0.0f, 0.0f}}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D/2; ++k_KQ) {
+            half2 K_k[FATTN_KQ_STRIDE_TILE_F16/WARP_SIZE];
+            half2 Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_h2[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE]*Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F16; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                half sum = __low2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]) + __high2half(sum2[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+                sum += mask ? slopeh*maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                kqmax_new[j_KQ_0/nwarps] = ggml_cuda_hmax(kqmax_new[j_KQ_0/nwarps], sum);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F16 + i_KQ] = sum;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const half2 KQ_max_scale = __half2half2(hexp(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]));
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F16/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const half2 diff = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] - __half2half2(kqmax[j0/nwarps]);
+                const half2 val = h2exp(diff);
+                kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + val;
+                KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + i] = val;
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE] *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp[k][i] = V_h2[(k_VKQ_0 + k)*stride_KV2 + i];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F16; k0 += 2) {
+            half2  V_k[(D/2)/WARP_SIZE][2];
+            half2 KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE][0] = KV_tmp[k0 + 0][i];
+                V_k[i0/WARP_SIZE][1] = KV_tmp[k0 + 1][i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ2[j*(FATTN_KQ_STRIDE_TILE_F16/2) + k0/2];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][0]* __low2half2(KQ_k[j0/nwarps]);
+                    VKQ[j0/nwarps][i0/WARP_SIZE] += V_k[i0/WARP_SIZE][1]*__high2half2(KQ_k[j0/nwarps]);
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        half kqsum_j = __low2half(kqsum[j_VKQ_0/nwarps]) + __high2half(kqsum[j_VKQ_0/nwarps]);
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            half2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val /= __half2half2(kqsum_j);
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] =  __low2float(dst_val);
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = __high2float(dst_val);
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f16<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}
--- a/ggml-cuda/fattn-tile-f16.cuh
+++ b/ggml-cuda/fattn-tile-f16.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/fattn-tile-f32.cu
+++ b/ggml-cuda/fattn-tile-f32.cu
@@ -0,0 +1,309 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-tile-f32.cuh"
+
+#define FATTN_KQ_STRIDE_TILE_F32 32
+
+template<int D, int ncols, int nwarps, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(nwarps*WARP_SIZE, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_tile_ext_f32(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half2  * V_h2  = (const half2  *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slope = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+
+    __shared__ float KQ[ncols*FATTN_KQ_STRIDE_TILE_F32];
+
+    __shared__ float KV_tmp[FATTN_KQ_STRIDE_TILE_F32][D + 1]; // Pad D to avoid memory bank conflicts.
+    float2 * KV_tmp2 = (float2 *) KV_tmp;
+
+    float kqmax[ncols/nwarps];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        kqmax[j0/nwarps] = -FLT_MAX/2.0f;
+    }
+    float kqsum[ncols/nwarps] = {0.0f};
+
+    float2 VKQ[ncols/nwarps][(D/2)/WARP_SIZE] = {{{0.0f, 0.0f}}};
+
+    // Convert Q to half2 and store in registers:
+    __shared__ float Q_f[ncols][D];
+#pragma unroll
+    for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+        const int j = j0 + threadIdx.y;
+
+#pragma unroll
+        for (int i0 = 0; i0 < D; i0 += 2*WARP_SIZE) {
+            float2 tmp = ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i0/2 + threadIdx.x] : make_float2(0.0f, 0.0f);
+            Q_f[j][i0 + 0*WARP_SIZE + threadIdx.x] = tmp.x * scale;
+            Q_f[j][i0 + 1*WARP_SIZE + threadIdx.x] = tmp.y * scale;
+        }
+    }
+
+    __syncthreads();
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*FATTN_KQ_STRIDE_TILE_F32;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*FATTN_KQ_STRIDE_TILE_F32) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        float kqmax_new[ncols/nwarps];
+#pragma unroll
+        for (int j = 0; j < ncols/nwarps; ++j) {
+            kqmax_new[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D; k_KQ_0 += 2*WARP_SIZE) {
+                const half2 tmp = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ_0/2 + threadIdx.x];
+                KV_tmp[i_KQ][k_KQ_0 + 0*WARP_SIZE + threadIdx.x] =  __low2float(tmp);
+                KV_tmp[i_KQ][k_KQ_0 + 1*WARP_SIZE + threadIdx.x] = __high2float(tmp);
+            }
+        }
+
+        __syncthreads();
+
+        float sum[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE][ncols/nwarps] = {{0.0f}};
+
+#pragma unroll
+        for (int k_KQ = 0; k_KQ < D; ++k_KQ) {
+            float K_k[FATTN_KQ_STRIDE_TILE_F32/WARP_SIZE];
+            float Q_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+                const int i_KQ = i_KQ_0 + threadIdx.x;
+
+                K_k[i_KQ_0/WARP_SIZE] = KV_tmp[i_KQ][k_KQ];
+            }
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                Q_k[j_KQ_0/nwarps] = Q_f[j_KQ][k_KQ];
+            }
+
+#pragma unroll
+            for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+#pragma unroll
+                for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                    sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += K_k[i_KQ_0/WARP_SIZE] * Q_k[j_KQ_0/nwarps];
+                }
+            }
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < FATTN_KQ_STRIDE_TILE_F32; i_KQ_0 += WARP_SIZE) {
+            const int i_KQ = i_KQ_0 + threadIdx.x;
+
+#pragma unroll
+            for (int j_KQ_0 = 0; j_KQ_0 < ncols; j_KQ_0 += nwarps) {
+                const int j_KQ = j_KQ_0 + threadIdx.y;
+
+                sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps] += mask ? slope*__half2float(maskh[j_KQ*ne11 + k_VKQ_0 + i_KQ]) : 0.0f;
+
+                kqmax_new[j_KQ_0/nwarps] = fmaxf(kqmax_new[j_KQ_0/nwarps], sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps]);
+
+                KQ[j_KQ*FATTN_KQ_STRIDE_TILE_F32 + i_KQ] = sum[i_KQ_0/WARP_SIZE][j_KQ_0/nwarps];
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+            const int j = j0 + threadIdx.y;
+
+            kqmax_new[j0/nwarps] = warp_reduce_max(kqmax_new[j0/nwarps]);
+            const float KQ_max_scale = expf(kqmax[j0/nwarps] - kqmax_new[j0/nwarps]);
+            kqmax[j0/nwarps] = kqmax_new[j0/nwarps];
+
+            float kqsum_add = 0.0f;
+#pragma unroll
+            for (int i0 = 0; i0 < FATTN_KQ_STRIDE_TILE_F32; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                const float diff = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] - kqmax[j0/nwarps];
+                const float val = expf(diff);
+                kqsum_add += val;
+                KQ[j*FATTN_KQ_STRIDE_TILE_F32 + i] = val;
+            }
+            kqsum[j0/nwarps] = kqsum[j0/nwarps]*KQ_max_scale + kqsum_add;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                VKQ[j0/nwarps][i0/WARP_SIZE].x *= KQ_max_scale;
+                VKQ[j0/nwarps][i0/WARP_SIZE].y *= KQ_max_scale;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < FATTN_KQ_STRIDE_TILE_F32; k0 += nwarps) {
+            const int k = k0 + threadIdx.y;
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                KV_tmp2[k*(D/2) + i].x =  __low2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+                KV_tmp2[k*(D/2) + i].y = __high2float(V_h2[(k_VKQ_0 + k)*stride_KV2 + i]);
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k = 0; k < FATTN_KQ_STRIDE_TILE_F32; ++k) {
+            float2 V_k[(D/2)/WARP_SIZE];
+            float  KQ_k[ncols/nwarps];
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+                const int i = i0 + threadIdx.x;
+
+                V_k[i0/WARP_SIZE] = KV_tmp2[k*(D/2) + i];
+            }
+#pragma unroll
+            for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                const int j = j0 + threadIdx.y;
+
+                KQ_k[j0/nwarps] = KQ[j*FATTN_KQ_STRIDE_TILE_F32 + k];
+            }
+
+#pragma unroll
+            for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+#pragma unroll
+                for (int j0 = 0; j0 < ncols; j0 += nwarps) {
+                    VKQ[j0/nwarps][i0/WARP_SIZE].x += V_k[i0/WARP_SIZE].x*KQ_k[j0/nwarps];
+                    VKQ[j0/nwarps][i0/WARP_SIZE].y += V_k[i0/WARP_SIZE].y*KQ_k[j0/nwarps];
+                }
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j_VKQ_0 = 0; j_VKQ_0 < ncols; j_VKQ_0 += nwarps) {
+        const int j_VKQ = j_VKQ_0 + threadIdx.y;
+
+        if (ic0 + j_VKQ >= ne01) {
+            return;
+        }
+
+        float kqsum_j = kqsum[j_VKQ_0/nwarps];
+        kqsum_j = warp_reduce_sum(kqsum_j);
+
+#pragma unroll
+        for (int i00 = 0; i00 < D; i00 += 2*WARP_SIZE) {
+            const int i0 = i00 + 2*threadIdx.x;
+
+            float2 dst_val = VKQ[j_VKQ_0/nwarps][i0/(2*WARP_SIZE)];
+            if (parallel_blocks == 1) {
+                dst_val.x /= kqsum_j;
+                dst_val.y /= kqsum_j;
+            }
+            const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 0] = dst_val.x;
+            dst[j_dst*D*gridDim.y + D*blockIdx.y + i0 + 1] = dst_val.y;
+        }
+
+        if (parallel_blocks != 1 && threadIdx.x == 0) {
+            dst_meta[(ic0 + j_VKQ)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[j_VKQ_0/nwarps], kqsum_j);
+        }
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_tile_f32_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = 8;
+            fattn_kernel_t fattn_kernel = flash_attn_tile_ext_f32<D, cols_per_block, nwarps, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+
+    if (Q->ne[1] <= 16) {
+        constexpr int cols_per_block = 16;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 32) {
+        constexpr int cols_per_block = 32;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block = 32;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_tile_f32_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}
--- a/ggml-cuda/fattn-tile-f32.cuh
+++ b/ggml-cuda/fattn-tile-f32.cuh
@@ -0,0 +1,3 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_tile_f32(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/ggml-cuda/fattn-vec-f16.cu
+++ b/ggml-cuda/fattn-vec-f16.cu
@@ -0,0 +1,330 @@
+#include "common.cuh"
+#include "fattn-common.cuh"
+#include "fattn-vec-f16.cuh"
+
+template<int D, int ncols, int parallel_blocks> // D == head size
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+__launch_bounds__(D, 1)
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+static __global__ void flash_attn_vec_ext_f16(
+        const char * __restrict__ Q,
+        const char * __restrict__ K,
+        const char * __restrict__ V,
+        const char * __restrict__ mask,
+        float      * __restrict__ dst,
+        float2     * __restrict__ dst_meta,
+        const float scale,
+        const float max_bias,
+        const float m0,
+        const float m1,
+        const uint32_t n_head_log2,
+        const int ne00,
+        const int ne01,
+        const int ne02,
+        const int ne03,
+        const int ne10,
+        const int ne11,
+        const int ne12,
+        const int ne13,
+        const int ne31,
+        const int nb31,
+        const int nb01,
+        const int nb02,
+        const int nb03,
+        const int nb11,
+        const int nb12,
+        const int nb13,
+        const int ne0,
+        const int ne1,
+        const int ne2,
+        const int ne3) {
+#if FP16_AVAILABLE
+    //In this kernel Q, K, V are matrices while i, j, k are matrix indices.
+
+    const int ic0 = (blockIdx.x / parallel_blocks) * ncols; // Index of the Q/QKV column to work on.
+    const int ip  =  blockIdx.x % parallel_blocks; // Index in group of blocks running for the same column in parallel.
+
+    const int gqa_ratio = ne02 / ne12; // With grouped query attention there are > 1 Q matrices per K, V matrix.
+    const float2 * Q_f2  = (const float2 *) (Q    + nb02* blockIdx.y              + nb01*ic0);
+    const half2  * K_h2  = (const half2  *) (K    + nb12*(blockIdx.y / gqa_ratio));
+    const half   * V_h   = (const half   *) (V    + nb12*(blockIdx.y / gqa_ratio)); // K and V have same shape
+    const half   * maskh = (const half   *)  mask + ne11*ic0;
+
+    const int stride_KV  = nb11 / sizeof(half);
+    const int stride_KV2 = nb11 / sizeof(half2);
+
+    const float slopef = get_alibi_slope(max_bias, blockIdx.y, n_head_log2, m0, m1);
+    const half  slopeh = __float2half(slopef);
+
+    static_assert(D % (2*WARP_SIZE) == 0, "D not divisible by 2*WARP_SIZE == 64.");
+    constexpr int nwarps = D / WARP_SIZE;
+    const int tid = WARP_SIZE*threadIdx.y + threadIdx.x;
+    __builtin_assume(tid < D);
+
+    __shared__ half KQ[ncols*D];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        KQ[j*D + tid] = -HALF_MAX_HALF;
+    }
+    half2 * KQ2 = (half2 *) KQ;
+
+    half kqmax[ncols];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqmax[j] = -HALF_MAX_HALF;
+    }
+    half kqsum[ncols] = {0.0f};
+
+    __shared__ half kqmax_shared[ncols][WARP_SIZE];
+    __shared__ half kqsum_shared[ncols][WARP_SIZE];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        if (threadIdx.y == 0) {
+            kqmax_shared[j][threadIdx.x] = -HALF_MAX_HALF;
+            kqsum_shared[j][threadIdx.x] = 0.0f;
+        }
+    }
+    __syncthreads();
+
+    // Convert Q to half2 and store in registers:
+    half2 Q_h2[ncols][D/(2*WARP_SIZE)];
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+#pragma unroll
+        for (int i0 = 0; i0 < D/2; i0 += WARP_SIZE) {
+            const int i = i0 + threadIdx.x;
+
+            const float2 tmp = ncols <= 2 || ic0 + j < ne01 ? Q_f2[j*(nb01/sizeof(float2)) + i] : make_float2(0.0f, 0.0f);
+            Q_h2[j][i0/WARP_SIZE] = make_half2(scale, scale) * make_half2(tmp.x, tmp.y);
+        }
+    }
+
+    half2 VKQ[ncols] = {{0.0f, 0.0f}};
+
+    const int k_start = parallel_blocks == 1 ? 0 : ip*D;
+    for (int k_VKQ_0 = k_start; k_VKQ_0 < ne11; k_VKQ_0 += parallel_blocks*D) {
+        // Calculate KQ tile and keep track of new maximum KQ values:
+
+        // For unknown reasons using a half array of size 1 for kqmax_new causes a performance regression,
+        // see https://github.com/ggerganov/llama.cpp/pull/7061 .
+        // Therefore this variable is defined twice but only used once (so that the compiler can optimize out the unused variable).
+        half kqmax_new = kqmax[0];
+        half kqmax_new_arr[ncols];
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            kqmax_new_arr[j] = kqmax[j];
+        }
+
+#pragma unroll
+        for (int i_KQ_0 = 0; i_KQ_0 < D; i_KQ_0 += nwarps) {
+            const int i_KQ = i_KQ_0 + threadIdx.y;
+
+            if ((i_KQ_0 + nwarps > D && i_KQ >= D) || (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + i_KQ >= ne11)) {
+                break;
+            }
+
+            half2 sum2[ncols] = {{0.0f, 0.0f}};
+#pragma unroll
+            for (int k_KQ_0 = 0; k_KQ_0 < D/2; k_KQ_0 += WARP_SIZE) {
+                const int k_KQ = k_KQ_0 + threadIdx.x;
+
+                const half2 K_ik = K_h2[(k_VKQ_0 + i_KQ)*stride_KV2 + k_KQ];
+#pragma unroll
+                for (int j = 0; j < ncols; ++j) {
+                    sum2[j] += K_ik * Q_h2[j][k_KQ_0/WARP_SIZE];
+                }
+            }
+
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                sum2[j] = warp_reduce_sum(sum2[j]);
+                half sum = __low2half(sum2[j]) + __high2half(sum2[j]);
+                sum += mask ? slopeh*maskh[j*ne11 + k_VKQ_0 + i_KQ] : __float2half(0.0f);
+
+                if (ncols == 1) {
+                    kqmax_new        = ggml_cuda_hmax(kqmax_new,        sum);
+                } else {
+                    kqmax_new_arr[j] = ggml_cuda_hmax(kqmax_new_arr[j], sum);
+                }
+
+                if (threadIdx.x == 0) {
+                    KQ[j*D + i_KQ] = sum;
+                }
+            }
+        }
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = ncols == 1 ? kqmax_new : kqmax_new_arr[j];
+
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+            if (threadIdx.x == 0) {
+                kqmax_shared[j][threadIdx.y] = kqmax_new_j;
+            }
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int j = 0; j < ncols; ++j) {
+            half kqmax_new_j = kqmax_shared[j][threadIdx.x];
+            kqmax_new_j = warp_reduce_max(kqmax_new_j);
+
+            const half KQ_max_scale = hexp(kqmax[j] - kqmax_new_j);
+            kqmax[j] = kqmax_new_j;
+
+            const half val = hexp(KQ[j*D + tid] - kqmax[j]);
+            kqsum[j] = kqsum[j]*KQ_max_scale + val;
+            KQ[j*D + tid] = val;
+
+            VKQ[j] *= __half2half2(KQ_max_scale);
+        }
+
+        __syncthreads();
+
+#pragma unroll
+        for (int k0 = 0; k0 < D; k0 += 2) {
+            if (FATTN_KQ_STRIDE % D != 0 && k_VKQ_0 + k0 >= ne11) {
+                break;
+            }
+
+            half2 V_k;
+            reinterpret_cast<half&>(V_k.x) = V_h[(k_VKQ_0 + k0 + 0)*stride_KV + tid];
+            reinterpret_cast<half&>(V_k.y) = V_h[(k_VKQ_0 + k0 + 1)*stride_KV + tid];
+#pragma unroll
+            for (int j = 0; j < ncols; ++j) {
+                VKQ[j] += V_k*KQ2[j*(D/2) + k0/2];
+            }
+        }
+
+        __syncthreads();
+    }
+
+#pragma unroll
+    for (int j = 0; j < ncols; ++j) {
+        kqsum[j] = warp_reduce_sum(kqsum[j]);
+        if (threadIdx.x == 0) {
+            kqsum_shared[j][threadIdx.y] = kqsum[j];
+        }
+    }
+
+    __syncthreads();
+
+#pragma unroll
+    for (int j_VKQ = 0; j_VKQ < ncols; ++j_VKQ) {
+        if (ncols > 2 && ic0 + j_VKQ >= ne01) {
+            break;
+        }
+
+        kqsum[j_VKQ] = kqsum_shared[j_VKQ][threadIdx.x];
+        kqsum[j_VKQ] = warp_reduce_sum(kqsum[j_VKQ]);
+
+        half dst_val = (__low2half(VKQ[j_VKQ]) + __high2half(VKQ[j_VKQ]));
+        if (parallel_blocks == 1) {
+            dst_val /= kqsum[j_VKQ];
+        }
+        const int j_dst = (ic0 + j_VKQ)*parallel_blocks + ip;
+        dst[j_dst*D*gridDim.y + D*blockIdx.y + tid] = dst_val;
+    }
+
+    if (parallel_blocks != 1 && tid < ncols && (ncols <= 2 || ic0 + tid < ne01)) {
+        dst_meta[(ic0 + tid)*gridDim.y*parallel_blocks + blockIdx.y*parallel_blocks + ip] = make_float2(kqmax[tid], kqsum[tid]);
+    }
+#else
+   NO_DEVICE_CODE;
+#endif // FP16_AVAILABLE
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    ggml_tensor * KQV = dst;
+    ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    constexpr int cols_per_block  = 1;
+    constexpr int parallel_blocks = 4;
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 256: {
+            constexpr int      D = 256;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default:
+            GGML_ASSERT(false);
+            break;
+    }
+}
+
+template <int cols_per_block, int parallel_blocks>
+void launch_fattn_vec_f16_64_128(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * Q = dst->src[0];
+    switch (Q->ne[0]) {
+        case  64: {
+            constexpr int      D = 64;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        case 128: {
+            constexpr int      D = 128;
+            constexpr int nwarps = D/WARP_SIZE;
+            fattn_kernel_t fattn_kernel = flash_attn_vec_ext_f16<D, cols_per_block, parallel_blocks>;
+            launch_fattn<D, parallel_blocks>(ctx, dst, fattn_kernel, nwarps, cols_per_block);
+        } break;
+        default: {
+            GGML_ASSERT(false && "FlashAttention without tensor cores only supports head sizes 64 and 128.");
+        } break;
+    }
+}
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+    const ggml_tensor * KQV = dst;
+    const ggml_tensor * Q   = dst->src[0];
+
+    const int32_t precision = KQV->op_params[2];
+    GGML_ASSERT(precision == GGML_PREC_DEFAULT);
+
+    if (Q->ne[1] == 1) {
+        ggml_cuda_flash_attn_ext_vec_f16(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] == 2) {
+        constexpr int cols_per_block  = 2;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 4) {
+        constexpr int cols_per_block  = 4;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    if (Q->ne[1] <= 8) {
+        constexpr int cols_per_block  = 8;
+        constexpr int parallel_blocks = 4;
+        launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+        return;
+    }
+
+    constexpr int cols_per_block  = 8;
+    constexpr int parallel_blocks = 1;
+    launch_fattn_vec_f16_64_128<cols_per_block, parallel_blocks>(ctx, dst);
+}
--- a/ggml-cuda/fattn-vec-f16.cuh
+++ b/ggml-cuda/fattn-vec-f16.cuh
@@ -0,0 +1,5 @@
+#include "common.cuh"
+
+void ggml_cuda_flash_attn_ext_vec_f16(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
+void ggml_cuda_flash_attn_ext_vec_f16_no_mma(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
--- a/Show More
+++ b/Show More