llama : adapt to F16 KQ_pos

Thanks Johannes again for the tip

.devops/nix/docker.nix

@@ -1,37 +0,0 @@
-{
-  lib,
-  dockerTools,
-  buildEnv,
-  llama-cpp,
-  interactive ? true,
-  coreutils,
-}:
-
-# A tar that can be fed into `docker load`:
-#
-# $ nix build .#llamaPackages.docker
-# $ docker load < result
-
-# For details and variations cf.
-# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage
-# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922
-# - https://nixery.dev/
-
-# Approximate (compressed) sizes, at the time of writing, are:
-#
-# .#llamaPackages.docker: 125M;
-# .#llamaPackagesCuda.docker: 537M;
-# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M.
-
-dockerTools.buildLayeredImage {
-  name = llama-cpp.pname;
-  tag = "latest";
-
-  contents =
-    [ llama-cpp ]
-    ++ lib.optionals interactive [
-      coreutils
-      dockerTools.binSh
-      dockerTools.caCertificates
-    ];
-}

.devops/nix/package.nix

@@ -1,6 +1,5 @@
 {
   lib,
-  glibc,
   config,
   stdenv,
   mkShell,
@@ -31,11 +30,6 @@
   useRocm ? config.rocmSupport,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
-
-  # It's necessary to consistently use backendStdenv when building with CUDA support,
-  # otherwise we get libstdc++ errors downstream.
-  effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
-  enableStatic ? effectiveStdenv.hostPlatform.isStatic
 }@inputs:
 
 let
@@ -47,7 +41,10 @@ let
     versionOlder
     ;
 
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
   stdenv = throw "Use effectiveStdenv instead";
+  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
 
   suffices =
     lib.optionals useBlas [ "BLAS" ]
@@ -170,9 +167,6 @@ effectiveStdenv.mkDerivation (
         # TODO: Replace with autoAddDriverRunpath
         # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
         cudaPackages.autoAddOpenGLRunpathHook
-      ]
-      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
-        glibc.static
       ];
 
     buildInputs =
@@ -187,7 +181,7 @@ effectiveStdenv.mkDerivation (
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
+        (cmakeBool "BUILD_SHARED_LIBS" true)
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
@@ -196,7 +190,6 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
-        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -262,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,8 +12,5 @@ lib.makeScope newScope (
   self: {
     inherit llamaVersion;
     llama-cpp = self.callPackage ./package.nix { };
-    docker = self.callPackage ./docker.nix { };
-    docker-min = self.callPackage ./docker.nix { interactive = false; };
-    sif = self.callPackage ./sif.nix { };
   }
 )

.devops/nix/sif.nix

@@ -1,27 +0,0 @@
-{
-  lib,
-  singularity-tools,
-  llama-cpp,
-  bashInteractive,
-  interactive ? false,
-}:
-
-let
-  optionalInt = cond: x: if cond then x else 0;
-in
-singularity-tools.buildImage rec {
-  inherit (llama-cpp) name;
-  contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ];
-
-  # These are excessive (but safe) for most variants. Building singularity
-  # images requires superuser privileges, so we build them inside a VM in a
-  # writable image of pre-determined size.
-  #
-  # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846
-  #
-  # Expected image sizes:
-  # - cpu/blas: 150M,
-  # - cuda, all gencodes: 560M,
-  diskSize = 4096 + optionalInt llama-cpp.useRocm 16384;
-  memSize = diskSize;
-}

.github/ISSUE_TEMPLATE/bug.md

@@ -7,5 +7,3 @@ assignees: ''
 ---
 
 Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/workflows/build.yml

@@ -145,28 +145,6 @@ jobs:
           cd build
           ctest -L main --verbose
 
-  ubuntu-22-cmake-vulkan:
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake -DLLAMA_VULKAN=ON ..
-          cmake --build . --config Release -j $(nproc)
-
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -691,7 +669,8 @@ jobs:
         run: |
           cd examples/llama.android
 
-          ./gradlew build --no-daemon
+          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
+          ./gradlew build --no-daemon -Pskip-armeabi-v7a
 
 #  freeBSD-latest:
 #    runs-on: macos-12

.github/workflows/nix-ci-aarch64.yml

@@ -19,6 +19,7 @@ on:
 
 jobs:
   nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -36,8 +37,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -45,7 +46,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Show all output paths
       run: >
           nix run github:nix-community/nix-eval-jobs

.github/workflows/nix-ci.yml

@@ -23,8 +23,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,6 +37,7 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
+    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -50,8 +51,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -59,7 +60,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

.github/workflows/python-check-requirements.yml

@@ -3,14 +3,12 @@ name: Python check requirements.txt
 on:
   push:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
       - 'requirements/*.txt'
   pull_request:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
@@ -28,4 +26,4 @@ jobs:
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh
+        run:  bash scripts/check-requirements.sh nocleanup

.github/workflows/server.yml

@@ -1,136 +0,0 @@
-# Server build and tests
-name: Server
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  schedule:
-    -  cron: '0 0 * * *'
-
-jobs:
-  server:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        build_type: [Debug, Release]
-        include:
-          - build_type: Release
-            sanitizer: ""
-        exclude:
-          - build_type: Release
-            sanitizer: ADDRESS
-          - build_type: Release
-            sanitizer: THREAD
-          - build_type: Release
-            sanitizer: UNDEFINED
-
-    container:
-      image: ubuntu:latest
-      ports:
-        - 8888
-      options: --cpus 4
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
-      - name: Dependencies
-        id: depends
-        run: |
-          apt-get update
-          apt-get -y install \
-            build-essential \
-            git \
-            cmake \
-            python3-pip \
-            wget \
-            language-pack-en
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. \
-              -DLLAMA_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
-          cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
-
-      - name: Tests
-        id: server_integration_tests
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' && matrix.build_type == 'Release' || github.event.inputs.slow_tests == 'true' }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
-
-
-  server-windows:
-    runs-on: windows-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake ..  -DLLAMA_BUILD_SERVER=ON -DCMAKE_BUILD_TYPE=Release ;
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
-
-      - name: Python setup
-        id: setup_python
-        uses: actions/setup-python@v5
-        with:
-          python-version: '3.11'
-
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
-
-      - name: Tests
-        id: server_integration_tests
-        run: |
-          cd examples/server/tests
-          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ github.event.schedule != '' || github.event.inputs.slow_tests == 'true' }}
-        run: |
-          cd examples/server/tests
-          behave.exe --stop --no-skipped --no-capture --tags slow

.gitignore

@@ -45,7 +45,6 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
-/gritlm
 /imatrix
 /infill
 /libllama.so

CMakeLists.txt

@@ -110,7 +110,6 @@ option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
-option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
 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)
@@ -146,6 +145,14 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
+if (LLAMA_FATAL_WARNINGS)
+    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
+        add_compile_options(-Werror)
+    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
+        add_compile_options(/WX)
+    endif()
+endif()
+
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -199,33 +206,9 @@ if (LLAMA_METAL)
     # get full path to the file
     #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
 
-    # copy ggml-common.h and ggml-metal.metal to bin directory
-    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
+    # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_SOURCE "${CMAKE_SOURCE_DIR}/ggml-metal.metal")
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-        set(EMBED_METALLIB_ASSEMBLY "${CMAKE_BINARY_DIR}/autogenerated/ggml-embed-metallib.s")
-
-        add_custom_command(
-            OUTPUT ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".section __DATA,__ggml_metallib" > ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_start" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_start:" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE}\\\"" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo ".globl _ggml_metallib_end" >> ${EMBED_METALLIB_ASSEMBLY}
-            COMMAND echo "_ggml_metallib_end:" >> ${EMBED_METALLIB_ASSEMBLY}
-            DEPENDS ${METALLIB_SOURCE}
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${EMBED_METALLIB_ASSEMBLY})
-    endif()
-
     if (LLAMA_METAL_SHADER_DEBUG)
         # custom command to do the following:
         #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
@@ -764,24 +747,15 @@ function(get_flags CCID CCVER)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
 
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -799,10 +773,6 @@ set(CUDA_CXX_FLAGS "")
 if (LLAMA_CUBLAS)
     set(CUDA_FLAGS -use_fast_math)
 
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
-    endif()
-
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
         set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
         if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
@@ -937,16 +907,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
+            # Raspberry Pi 2
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()

Makefile

@@ -2,7 +2,7 @@
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -97,10 +97,9 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -173,7 +172,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -201,10 +200,6 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
-ifdef LLAMA_SERVER_SSL
-	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
-	MK_LDFLAGS += -lssl -lcrypto
-endif
 
 ifdef LLAMA_CODE_COVERAGE
 	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
@@ -221,7 +216,7 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
 ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
+	MK_CFLAGS += -Werror
 	MK_CXXFLAGS += -Werror
 endif
 
@@ -385,18 +380,10 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	MK_NVCCFLAGS += -use_fast_math
-ifdef LLAMA_FATAL_WARNINGS
-	MK_NVCCFLAGS += -Werror all-warnings
-endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -453,11 +440,11 @@ endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -542,29 +529,11 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal
-	@echo "Embedding Metal library"
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -576,10 +545,9 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
 override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -606,7 +574,7 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
-CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
@@ -630,7 +598,7 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
 OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
@@ -724,17 +692,14 @@ embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gritlm: examples/gritlm/gritlm.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

README-sycl.md

@@ -1,7 +1,6 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
-- [News](#news)
 - [OS](#os)
 - [Intel GPU](#intel-gpu)
 - [Docker](#docker)
@@ -26,21 +25,6 @@ The llama.cpp for SYCL is used to support Intel GPUs.
 
 For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 
-## News
-
-- 2024.3
-  - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
-  - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
-  - Support detecting all GPUs with level-zero and same top **Max compute units**.
-  - Support OPs
-    - hardsigmoid
-    - hardswish
-    - pool2d
-
-- 2024.1
-  - Create SYCL backend for Intel GPU.
-  - Support Windows build
-
 ## OS
 
 |OS|Status|Verified|
@@ -288,7 +272,7 @@ Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact
 
 a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
+Recommend to install to default folder: **/opt/intel/oneapi**.
 
 Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 
@@ -465,7 +449,6 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 |-|-|-|
 |GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
 |GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-|ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
 
 ## Known Issue
 
@@ -475,10 +458,6 @@ Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 
   Solution: add **--no-mmap** or **--mmap 0**.
 
-- Split-mode: [row] is not supported
-
-  It's on developing.
-
 ## Q&A
 
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.

README.md

@@ -8,20 +8,16 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-> [!IMPORTANT]
-> **Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962**
->
-> Vote for which quantization type provides better responses, all other parameters being the same.
-
-### Recent API changes
-
-- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_max_seq()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
-- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
-- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
-
 ### Hot topics
 
-- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
+- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
+- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
+  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
+- Collecting Apple Silicon performance stats:
+  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
+  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
+- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
 
@@ -111,21 +107,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
-- [x] [Gemma](https://ai.google.dev/gemma)
-- [x] [Mamba](https://github.com/state-spaces/mamba)
 
 **Multimodal models:**
 
-- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
+- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
 - [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
 - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
@@ -154,7 +145,6 @@ 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)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
 - [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
@@ -166,9 +156,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Msty](https://msty.app) (proprietary)
-- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 
 ---
 
@@ -794,7 +781,7 @@ And after 4.45 hours, you will have the final perplexity.
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
-In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
+In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
 Here is an example of a few-shot interaction, invoked with the command
 
@@ -858,7 +845,7 @@ Sample run:
 ```
 == Running in interactive mode. ==
  - Press Ctrl+C to interject at any time.
- - Press Return to return control to LLaMA.
+ - Press Return to return control to LLaMa.
  - If you want to submit another line, end your input in '\'.
 
  Below is an instruction that describes a task. Write a response that appropriately completes the request.

awq-py/README.md

@@ -0,0 +1,116 @@
+# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
+[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
+
+**Supported models:**
+
+- [X] LLaMA
+- [x] LLaMA 2
+- [X] MPT
+- [X] Mistral AI v0.1
+- [ ] Bloom
+- [ ] Mixtral MoE
+
+**TODO:**
+- [x] Update version work with both MPT and MPT-AWQ model
+- [ ] Add OPT model
+- [ ] Add Bloom model
+- [ ] Add Mixtral MoE
+- [ ] Support w3, w2
+
+
+## Contents
+
+- [Install](##Install)
+- [Convert](##Convert)
+- [Quantize](##Quantize)
+- [Test](##Test)
+- [Benchmark](##Benchmark)
+- [Results](##Results)
+
+## Install
+Install requirements
+```bash
+pip install -r requirements.txt
+```
+Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
+```bash
+git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
+```
+
+## Convert
+Example for llama model
+```bash
+# For llama7b and llama2 models
+python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
+# For mistral and mpt models
+python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/mpt-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
+```
+
+## Quantize
+```bash
+# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
+./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
+```
+
+## Test
+```bash
+# For all models.
+./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
+```
+
+## Benchmark
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+```bash
+# For llama and llama2, and mistral models.
+./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
+```
+
+## Results
+Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
+We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
+
+### Llama 7B (Build with OpenBLAS)
+
+| Model      | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-----------:|--------------|-------:|-------:|-------:|-------:|
+|Llama 7B    | perplexity   | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
+|Llama 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama 7B| perplexity   | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
+|AWQ-LLama 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Llama2 7B (Build with CuBLAS)
+
+| Model       | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|------------:|--------------|-------:|-------:|-------:|-------:|
+|Llama2 7B    | perplexity   | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
+|Llama2 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama2 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama2 7B| perplexity   | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
+|AWQ-LLama2 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama2 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Mistral 7B v0.1 (Build with CuBLAS)
+
+| Model        | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-------------:|--------------|-------:|-------:|-------:|-------:|
+|Mistral 7B    | perplexity   | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
+|Mistral 7B    | file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|Mistral 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-Mistral 7B| perplexity   | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
+|AWQ-Mistral 7B| file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|AWQ-Mistral 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+### MPT 7B (Build with OpenBLAS)
+
+| Model    | Measure      | F16    | Q4_0   | Q4_1   | Q2_K    |
+|---------:|--------------|-------:|-------:|-------:|--------:|
+|MPT 7B    | perplexity   | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
+|MPT 7B    | file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|MPT 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |
+|AWQ-MPT 7B| perplexity   | 8.4944 | 8.7053 |  8.6750 | 10.2873|
+|AWQ-MPT 7B| file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|AWQ-MPT 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |

awq-py/awq/apply_awq.py

@@ -0,0 +1,254 @@
+"""
+Implements the AWQ for llama.cpp use cases.
+Original paper: https://arxiv.org/abs/2306.00978
+
+This code is based on versions of the AWQ implementation found in the following repositories:
+* https://github.com/mit-han-lab/llm-awq
+* https://github.com/casper-hansen/AutoAWQ
+"""
+
+import os
+import torch
+import torch.nn as nn
+
+from transformers import AutoModelForCausalLM, AutoConfig
+from transformers.models.bloom.modeling_bloom import BloomGelu
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
+from transformers.activations import GELUActivation
+
+
+class ScaledActivation(nn.Module):
+    """
+    ScaledActivation module wraps an existing activation function and applies a
+    scale factor to its output.
+
+    Args:
+        module (nn.Module): The activation function to be scaled.
+        scales (torch.Tensor): A tensor of size (num_features,) containing the initial
+            scale factors for each feature.
+
+    Returns:
+        torch.Tensor: The scaled output of the activation function.
+    """
+
+    def __init__(self, module, scales):
+        super().__init__()
+        self.act = module
+        self.scales = nn.Parameter(scales.data)
+
+    def forward(self, x):
+        return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
+
+
+def set_op_by_name(layer, name, new_module):
+    """
+    Set the new module for given module's name.
+
+    Args:
+        layer (nn.Module): The layer in which to replace the submodule.
+        name (str): The path to the submodule to be replaced, using dot notation
+            to access nested modules.
+        new_module (nn.Module): The new module to replace the existing one.
+    """
+    levels = name.split(".")
+    if len(levels) > 1:
+        mod_ = layer
+        for l_idx in range(len(levels) - 1):
+            if levels[l_idx].isdigit():
+                mod_ = mod_[int(levels[l_idx])]
+            else:
+                mod_ = getattr(mod_, levels[l_idx])
+        setattr(mod_, levels[-1], new_module)
+    else:
+        setattr(layer, name, new_module)
+
+
+def get_op_by_name(module, op_name):
+    """
+    Retrieves a submodule within a given layer based on its name.
+
+    Args:
+        module (nn.Module): The layer containing the submodule to find.
+        op_name (str): The name of the submodule.
+
+    Returns:
+        nn.Module: The requested submodule found within the given layer.
+
+    Raises:
+        ValueError: If the specified submodule cannot be found within the layer.
+    """
+    for name, m in module.named_modules():
+        if name == op_name:
+            return m
+    raise ValueError(f"Cannot find op {op_name} in module {module}")
+
+
+@torch.no_grad()
+def scale_ln_fcs(ln, fcs, scales):
+    """
+    Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
+
+    Args:
+        ln (nn.LayerNorm): The LayerNorm module to be scaled.
+        fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+
+    if not isinstance(fcs, list):
+        fcs = [fcs]
+
+    scales = scales.to(ln.weight.device)
+
+    ln.weight.div_(scales)
+    if hasattr(ln, "bias") and ln.bias is not None:
+        ln.bias.div_(scales)
+
+    for fc in fcs:
+        fc.weight.mul_(scales.view(1, -1))
+
+    for p in ln.parameters():
+        assert torch.isnan(p).sum() == 0
+    for fc in fcs:
+        for p in fc.parameters():
+            assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_fc_fc(fc1, fc2, scales):
+    """
+    Scales the weights of two fully-connected layers in a specific pattern.
+
+    Args:
+        fc1 (nn.Linear): The first fully-connected layer to be scaled.
+        fc2 (nn.Linear): The second fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+    assert isinstance(fc1, nn.Linear)
+    assert isinstance(fc2, nn.Linear)
+
+    scales = scales.to(fc1.weight.device)
+
+    fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
+    if fc1.bias is not None:
+        fc1.bias.div_(scales.view(-1))
+
+    fc2.weight.mul_(scales.view(1, -1))
+
+    for p in fc1.parameters():
+        assert torch.isnan(p).sum() == 0
+    for p in fc2.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_gelu_fc(gelu, fc, scales):
+    """
+    Scales the weight of a GELU activation and a fully-connected layer proportionally.
+
+    Args:
+        gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
+        fc (nn.Linear): The fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+
+    Raises:
+        TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
+        TypeError: If the `fc` module is not of type `nn.Linear`.
+    """
+    assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
+    assert isinstance(fc, nn.Linear)
+
+    fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
+
+    for p in fc.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+def apply_scale(module, scales_list, input_feat_dict=None):
+    """
+    Applies different scaling strategies to layers based on their type and hierarchy within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layers to be scaled.
+        scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
+            * prev_op_name (str): The name of the preceding operation or module,
+                relative to which the layers to be scaled are located.
+            * layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
+            * scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
+        input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
+            input features (optional).
+    """
+    for prev_op_name, layer_names, scales in scales_list:
+        prev_op = get_op_by_name(module, prev_op_name)
+        layers = [get_op_by_name(module, name) for name in layer_names]
+
+        prev_op.cuda()
+        for layer in layers:
+            layer.cuda()
+        scales.cuda()
+
+        if isinstance(prev_op, nn.Linear):
+            assert len(layers) == 1
+            scale_fc_fc(prev_op, layers[0], scales)
+        elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
+            scale_ln_fcs(prev_op, layers, scales)
+        elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
+            new_module = ScaledActivation(prev_op, scales)
+            set_op_by_name(module, prev_op_name, new_module)
+            scale_gelu_fc(prev_op, layers[0], scales)
+        else:
+            raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
+
+        # apply the scaling to input feat if given; prepare it for clipping
+        if input_feat_dict is not None:
+            for layer_name in layer_names:
+                inp = input_feat_dict[layer_name]
+                inp.div_(scales.view(1, -1).to(inp.device))
+
+        prev_op.cpu()
+        for layer in layers:
+            layer.cpu()
+        scales.cpu()
+
+
+@torch.no_grad()
+def apply_clip(module, clip_list):
+    """
+    Applies element-wise clipping to the weight of a specific layer within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layer to be clipped.
+        clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
+            * name (str): The name of the layer to be clipped, relative to the root of the module.
+            * max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
+    """
+    for name, max_val in clip_list:
+        layer = get_op_by_name(module, name)
+        layer.cuda()
+        max_val = max_val.to(layer.weight.device)
+        org_shape = layer.weight.shape
+        layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
+        layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
+        layer.weight.data = layer.weight.data.reshape(org_shape)
+        layer.cpu()
+
+
+def add_scale_weights(model_path, scale_path, tmp_path):
+    """
+    Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
+    including scaling factors and clipping bounds.
+
+    Args:
+        model_path (str): Path to the pre-trained model to be equipped with AWQ.
+        scale_path (str): Path to the AWQ scale factors (.pt file).
+        tmp_path (str): Path to the temporary directory where the equipped model will be saved.
+    """
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, config=config, trust_remote_code=True
+    )
+    model.eval()
+    awq_results = torch.load(str(scale_path), map_location="cpu")
+    apply_scale(model, awq_results["scale"])
+    apply_clip(model, awq_results["clip"])
+    model.save_pretrained(str(tmp_path))
+    os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

awq-py/requirements.txt

@@ -0,0 +1,2 @@
+torch>=2.1.1
+transformers>=4.32.0

build.zig

@@ -123,7 +123,6 @@ pub fn build(b: *std.build.Builder) !void {
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.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, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
     _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
@@ -132,7 +131,7 @@ pub fn build(b: *std.build.Builder) !void {
     _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
     _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -45,8 +45,7 @@ fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
     if [ -z ${ONEAPI_ROOT} ]; then
-        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
-        echo "source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 
@@ -273,19 +272,19 @@ function gg_run_open_llama_3b_v2 {
     (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/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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 2 ) 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/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 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
 
@@ -344,17 +343,17 @@ function gg_run_open_llama_3b_v2 {
     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
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 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 2 ) 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
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 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 2 ) 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
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 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

common/CMakeLists.txt

@@ -19,12 +19,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
-    else()
-        message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
-    endif()
+    set(GIT_INDEX "${GIT_DIR}/index")
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
     set(GIT_INDEX "")

common/common.cpp

@@ -295,9 +295,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::string value(argv[i]);
-            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_NONE; }
-            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_LINEAR; }
-            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_YARN; }
+            /**/ if (value == "none")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_NONE; }
+            else if (value == "linear") { params.rope_scaling_type = LLAMA_ROPE_SCALING_LINEAR; }
+            else if (value == "yarn")   { params.rope_scaling_type = LLAMA_ROPE_SCALING_YARN; }
             else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
@@ -335,22 +335,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--pooling") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string value(argv[i]);
-            /**/ if (value == "none") { params.pooling_type = LLAMA_POOLING_TYPE_NONE; }
-            else if (value == "mean") { params.pooling_type = LLAMA_POOLING_TYPE_MEAN; }
-            else if (value == "cls")  { params.pooling_type = LLAMA_POOLING_TYPE_CLS; }
-            else { invalid_param = true; break; }
-        } else if (arg == "--defrag-thold" || arg == "-dt") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.defrag_thold = std::stof(argv[i]);
         } else if (arg == "--samplers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -513,6 +497,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_sequences = std::stoi(argv[i]);
+        } else if (arg == "--p-accept" || arg == "-pa") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.p_accept = std::stof(argv[i]);
         } else if (arg == "--p-split" || arg == "-ps") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -640,15 +630,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             }
             std::string arg_next = argv[i];
             if (arg_next == "none") {
-                params.split_mode = LLAMA_SPLIT_MODE_NONE;
+                params.split_mode = LLAMA_SPLIT_NONE;
             } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
+                params.split_mode = LLAMA_SPLIT_LAYER;
             } else if (arg_next == "row") {
-#ifdef GGML_USE_SYCL
-                fprintf(stderr, "warning: The split mode value:[row] is not supported by llama.cpp with SYCL. It's developing.\nExit!\n");
-                exit(1);
-#endif // GGML_USE_SYCL
-                params.split_mode = LLAMA_SPLIT_MODE_ROW;
+                params.split_mode = LLAMA_SPLIT_ROW;
             } else {
                 invalid_param = true;
                 break;
@@ -851,15 +837,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             sep++;
             if (strncmp(sep, "int:", 4) == 0) {
                 sep += 4;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+                kvo.tag = LLAMA_KV_OVERRIDE_INT;
                 kvo.int_value = std::atol(sep);
             } else if (strncmp(sep, "float:", 6) == 0) {
                 sep += 6;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
                 kvo.float_value = std::atof(sep);
             } else if (strncmp(sep, "bool:", 5) == 0) {
                 sep += 5;
-                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {
@@ -1018,14 +1004,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --yarn-attn-factor N  YaRN: scale sqrt(t) or attention magnitude (default: 1.0)\n");
     printf("  --yarn-beta-slow N    YaRN: high correction dim or alpha (default: %.1f)\n", params.yarn_beta_slow);
     printf("  --yarn-beta-fast N    YaRN: low correction dim or beta (default: %.1f)\n", params.yarn_beta_fast);
-    printf("  --pooling {none,mean,cls}\n");
-    printf("                        pooling type for embeddings, use model default if unspecified\n");
-    printf("  -dt N, --defrag-thold N\n");
-    printf("                        KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     printf("  --no-penalize-nl      do not penalize newline token\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
-    printf("  --all-logits          return logits for all tokens in the batch (default: disabled)\n");
+    printf("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
     printf("  --hellaswag-tasks N   number of tasks to use when computing the HellaSwag score (default: %zu)\n", params.hellaswag_tasks);
     printf("  --winogrande          compute Winogrande score over random tasks from datafile supplied with -f\n");
@@ -1038,6 +1020,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --chunks N            max number of chunks to process (default: %d, -1 = all)\n", params.n_chunks);
     printf("  -np N, --parallel N   number of parallel sequences to decode (default: %d)\n", params.n_parallel);
     printf("  -ns N, --sequences N  number of sequences to decode (default: %d)\n", params.n_sequences);
+    printf("  -pa N, --p-accept N   speculative decoding accept probability (default: %.1f)\n", (double)params.p_accept);
     printf("  -ps N, --p-split N    speculative decoding split probability (default: %.1f)\n", (double)params.p_split);
     printf("  -cb, --cont-batching  enable continuous batching (a.k.a dynamic batching) (default: disabled)\n");
     printf("  --mmproj MMPROJ_FILE  path to a multimodal projector file for LLaVA. see examples/llava/README.md\n");
@@ -1288,12 +1271,12 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
 
     cparams.n_ctx             = params.n_ctx;
     cparams.n_batch           = params.n_batch;
-    cparams.n_parallel        = params.n_parallel;
     cparams.n_threads         = params.n_threads;
     cparams.n_threads_batch   = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+    cparams.mul_mat_q         = params.mul_mat_q;
     cparams.seed              = params.seed;
     cparams.logits_all        = params.logits_all;
-    cparams.embeddings        = params.embedding;
+    cparams.embedding         = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
     cparams.rope_freq_base    = params.rope_freq_base;
     cparams.rope_freq_scale   = params.rope_freq_scale;
@@ -1302,8 +1285,6 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_beta_fast    = params.yarn_beta_fast;
     cparams.yarn_beta_slow    = params.yarn_beta_slow;
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
-    cparams.pooling_type      = params.pooling_type;
-    cparams.defrag_thold      = params.defrag_thold;
     cparams.offload_kqv       = !params.no_kv_offload;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
@@ -1735,6 +1716,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
+    fprintf(stream, "no_mul_mat_q: %s # default: false\n", !params.mul_mat_q ? "true" : "false");
     fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
@@ -1852,18 +1834,3 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
 
     printf("\n=== Done dumping\n");
 }
-
-void llama_embd_normalize(const float * inp, float * out, int n) {
-    double sum = 0.0;
-    for (int i = 0; i < n; i++) {
-        sum += inp[i] * inp[i];
-    }
-    sum = sqrt(sum);
-
-    const float norm = sum > 0.0 ? 1.0f / sum : 0.0f;
-
-    for (int i = 0; i < n; i++) {
-        out[i] = inp[i] * norm;
-    }
-}
-

common/common.h

@@ -43,7 +43,7 @@ extern char const *LLAMA_BUILD_TARGET;
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+    uint32_t seed                 = -1;    // RNG seed
 
     int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;
@@ -53,14 +53,15 @@ struct gpt_params {
     int32_t n_ctx                 = 512;   // context size
     int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
+    int32_t n_draft               = 8;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            = 1;     // number of parallel sequences to decode
     int32_t n_sequences           = 1;     // number of sequences to decode
+    float   p_accept              = 0.5f;  // speculative decoding accept probability
     float   p_split               = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
-    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
@@ -74,12 +75,8 @@ struct gpt_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     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
-
-    ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
-
-    llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
-    llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
+    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -117,6 +114,7 @@ struct gpt_params {
 
     bool   kl_divergence   = false; // compute KL-divergence
 
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     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
@@ -260,10 +258,3 @@ void dump_kv_cache_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);
-
-//
-// Embedding utils
-//
-
-void llama_embd_normalize(const float * inp, float * out, int n);
-

common/grammar-parser.cpp

@@ -278,22 +278,6 @@ namespace grammar_parser {
             while (*pos) {
                 pos = parse_rule(state, pos);
             }
-            // Validate the state to ensure that all rules are defined
-            for (const auto & rule : state.rules) {
-                for (const auto & elem : rule) {
-                    if (elem.type == LLAMA_GRETYPE_RULE_REF) {
-                        // Ensure that the rule at that location exists
-                        if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
-                            // Get the name of the rule that is missing
-                            for (const auto & kv : state.symbol_ids) {
-                                if (kv.second == elem.value) {
-                                    throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
-                                }
-                            }
-                        }
-                    }
-                }
-            }
             return state;
         } catch (const std::exception & err) {
             fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());

common/log.h

@@ -297,7 +297,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // Main TEE macro.
@@ -311,7 +311,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
@@ -319,8 +319,8 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
-    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
-    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
 #endif
 
 // INTERNAL, DO NOT USE

common/sampling.cpp

@@ -266,7 +266,7 @@ static llama_token llama_sampling_sample_impl(
             //    }
             //}
 
-            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
         }
     }
 
@@ -295,77 +295,6 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
-static llama_token_data_array llama_sample_probability_distribution_impl(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    const llama_sampling_params & params = ctx_sampling->params;
-
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
-    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
-    const float   penalty_repeat  = params.penalty_repeat;
-    const float   penalty_freq    = params.penalty_freq;
-    const float   penalty_present = params.penalty_present;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
-
-    // Get a pointer to the logits
-    float * logits = llama_get_logits_ith(ctx_main, idx);
-
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    if (ctx_cfg) {
-        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
-        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
-    }
-
-    cur.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    // apply penalties
-    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
-    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
-    if (penalty_tokens_used_size) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
-                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    // apply grammar checks
-    if (ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
-    llama_sample_softmax(ctx_main, &cur_p);
-    return cur_p;
-}
-
 llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
@@ -375,14 +304,6 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
-llama_token_data_array llama_sampling_probability_distribution(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx);
-}
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -131,13 +131,6 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
-// returns the probability that token of given id will be sampled
-llama_token_data_array llama_sampling_probability_distribution(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = 0);
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/train.cpp

@@ -31,7 +31,7 @@ struct train_state  * init_train_state() {
 
     state->opt = new struct ggml_opt_context;
     state->opt->ctx = NULL;
-    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
@@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
     std::string opt_type;
     GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
     if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
-        opt->params.type = GGML_OPT_TYPE_ADAM;
+        opt->params.type = GGML_OPT_ADAM;
 
         GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
         GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
         copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
     } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
-        opt->params.type = GGML_OPT_TYPE_LBFGS;
+        opt->params.type = GGML_OPT_LBFGS;
 
         GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
         GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
     gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
 
     switch (opt->params.type) {
-        case GGML_OPT_TYPE_ADAM:
+        case GGML_OPT_ADAM:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
                 gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
@@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
                     gguf_add_tensor(fctx, opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_TYPE_LBFGS:
+        case GGML_OPT_LBFGS:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
                 gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);

convert-hf-to-gguf.py

@@ -8,10 +8,9 @@ import json
 import os
 import re
 import sys
-from abc import ABC, abstractmethod
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, Sequence, cast
 
 import numpy as np
 import torch
@@ -37,12 +36,7 @@ class SentencePieceTokenTypes(IntEnum):
     BYTE = 6
 
 
-AnyModel = TypeVar("AnyModel", bound="type[Model]")
-
-
-class Model(ABC):
-    _model_classes: dict[str, type[Model]] = {}
-
+class Model:
     def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
@@ -53,14 +47,10 @@ class Model(ABC):
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
+        self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
-    @property
-    @abstractmethod
-    def model_arch(self) -> gguf.MODEL_ARCH:
-        pass
-
     def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
         key = next((k for k in keys if k in self.hparams), None)
         if key is not None:
@@ -106,11 +96,9 @@ class Model(ABC):
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
 
-        if (rope_theta := self.hparams.get("rope_theta")) is not None:
-            self.gguf_writer.add_rope_freq_base(rope_theta)
         if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
+        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
@@ -186,22 +174,51 @@ class Model(ABC):
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             return json.load(f)
 
-    @classmethod
-    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
-        assert names
-
-        def func(modelcls: type[Model]):
-            for name in names:
-                cls._model_classes[name] = modelcls
-            return modelcls
-        return func
-
-    @classmethod
-    def from_model_architecture(cls, arch):
-        try:
-            return cls._model_classes[arch]
-        except KeyError:
-            raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
+    @staticmethod
+    def from_model_architecture(model_architecture):
+        if model_architecture == "GPTNeoXForCausalLM":
+            return GPTNeoXModel
+        if model_architecture == "BloomForCausalLM":
+            return BloomModel
+        if model_architecture == "MPTForCausalLM":
+            return MPTModel
+        if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return BaichuanModel
+        if model_architecture in ("FalconForCausalLM", "RWForCausalLM"):
+            return FalconModel
+        if model_architecture == "GPTBigCodeForCausalLM":
+            return StarCoderModel
+        if model_architecture == "GPTRefactForCausalLM":
+            return RefactModel
+        if model_architecture == "PersimmonForCausalLM":
+            return PersimmonModel
+        if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return StableLMModel
+        if model_architecture == "QWenLMHeadModel":
+            return QwenModel
+        if model_architecture == "Qwen2ForCausalLM":
+            return Model
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
+        if model_architecture == "GPT2LMHeadModel":
+            return GPT2Model
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
+        if model_architecture == "PlamoForCausalLM":
+            return PlamoModel
+        if model_architecture == "CodeShellForCausalLM":
+            return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
+        if model_architecture == "InternLM2ForCausalLM":
+            return InternLM2Model
+        if model_architecture == "MiniCPMForCausalLM":
+            return MiniCPMModel
+        if model_architecture == "BertModel":
+            return BertModel
+        if model_architecture == "NomicBertModel":
+            return NomicBertModel
+        return Model
 
     def _is_model_safetensors(self) -> bool:
         return Model.count_model_parts(self.dir_model, ".safetensors") > 0
@@ -216,6 +233,53 @@ class Model(ABC):
             return ("pytorch_model.bin",)
         return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
 
+    def _get_model_architecture(self) -> gguf.MODEL_ARCH:
+        arch = self.hparams["architectures"][0]
+        if arch == "GPTNeoXForCausalLM":
+            return gguf.MODEL_ARCH.GPTNEOX
+        if arch == "BloomForCausalLM":
+            return gguf.MODEL_ARCH.BLOOM
+        if arch == "MPTForCausalLM":
+            return gguf.MODEL_ARCH.MPT
+        if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return gguf.MODEL_ARCH.BAICHUAN
+        if arch in ("FalconForCausalLM", "RWForCausalLM"):
+            return gguf.MODEL_ARCH.FALCON
+        if arch == "GPTBigCodeForCausalLM":
+            return gguf.MODEL_ARCH.STARCODER
+        if arch == "GPTRefactForCausalLM":
+            return gguf.MODEL_ARCH.REFACT
+        if arch == "PersimmonForCausalLM":
+            return gguf.MODEL_ARCH.PERSIMMON
+        if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return gguf.MODEL_ARCH.STABLELM
+        if arch == "QWenLMHeadModel":
+            return gguf.MODEL_ARCH.QWEN
+        if arch == "Qwen2ForCausalLM":
+            return gguf.MODEL_ARCH.QWEN2
+        if arch == "MixtralForCausalLM":
+            return gguf.MODEL_ARCH.LLAMA
+        if arch == "GPT2LMHeadModel":
+            return gguf.MODEL_ARCH.GPT2
+        if arch == "PhiForCausalLM":
+            return gguf.MODEL_ARCH.PHI2
+        if arch == "PlamoForCausalLM":
+            return gguf.MODEL_ARCH.PLAMO
+        if arch == "CodeShellForCausalLM":
+            return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
+        if arch == "InternLM2ForCausalLM":
+            return gguf.MODEL_ARCH.INTERNLM2
+        if arch == "MiniCPMForCausalLM":
+            return gguf.MODEL_ARCH.MINICPM
+        if arch == "BertModel":
+            return gguf.MODEL_ARCH.BERT
+        if arch == "NomicBertModel":
+            return gguf.MODEL_ARCH.NOMIC_BERT
+
+        raise NotImplementedError(f'Architecture "{arch}" not supported!')
+
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
@@ -383,10 +447,7 @@ class Model(ABC):
         special_vocab.add_to_gguf(self.gguf_writer)
 
 
-@Model.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(Model):
-    model_arch = gguf.MODEL_ARCH.GPTNEOX
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
 
@@ -403,10 +464,7 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
 
-@Model.register("BloomForCausalLM")
 class BloomModel(Model):
-    model_arch = gguf.MODEL_ARCH.BLOOM
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("Bloom")
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
@@ -498,10 +556,7 @@ class BloomModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("MPTForCausalLM")
 class MPTModel(Model):
-    model_arch = gguf.MODEL_ARCH.MPT
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -563,11 +618,13 @@ class MPTModel(Model):
 
             self.gguf_writer.add_tensor(new_name, data)
 
+            # note: MPT output is tied to (same as) wte in original model;
+            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+            if new_name == "token_embd.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+
 
-@Model.register("OrionForCausalLM")
 class OrionModel(Model):
-    model_arch = gguf.MODEL_ARCH.ORION
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -598,8 +655,6 @@ class OrionModel(Model):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
-        # note: config provides rms norm but it is actually layer norm
-        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -646,10 +701,7 @@ class OrionModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
 class BaichuanModel(Model):
-    model_arch = gguf.MODEL_ARCH.BAICHUAN
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -764,10 +816,7 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
-@Model.register("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
-    model_arch = gguf.MODEL_ARCH.FALCON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_hidden_layers")
         if block_count is None:
@@ -860,10 +909,7 @@ class FalconModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("GPTBigCodeForCausalLM")
 class StarCoderModel(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -878,10 +924,7 @@ class StarCoderModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@Model.register("GPTRefactForCausalLM")
 class RefactModel(Model):
-    model_arch = gguf.MODEL_ARCH.REFACT
-
     def set_gguf_parameters(self):
         hidden_dim = self.hparams["n_embd"]
         inner_dim = 4 * hidden_dim
@@ -965,10 +1008,7 @@ class RefactModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("PersimmonForCausalLM")
 class PersimmonModel(Model):
-    model_arch = gguf.MODEL_ARCH.PERSIMMON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
         head_count = self.hparams["num_attention_heads"]
@@ -991,6 +1031,7 @@ class PersimmonModel(Model):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1016,10 +1057,7 @@ class PersimmonModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
 class StableLMModel(Model):
-    model_arch = gguf.MODEL_ARCH.STABLELM
-
     def set_vocab(self):
         if (self.dir_model / "tokenizer.json").is_file():
             self._set_vocab_gpt2()
@@ -1036,25 +1074,18 @@ class StableLMModel(Model):
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
-        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
-@Model.register("MixtralForCausalLM")
 class MixtralModel(Model):
-    model_arch = gguf.MODEL_ARCH.LLAMA
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
 
-@Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
-    model_arch = gguf.MODEL_ARCH.MINICPM
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
         self.gguf_writer.add_name("MiniCPM")
@@ -1131,10 +1162,7 @@ class MiniCPMModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("QWenLMHeadModel")
 class QwenModel(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN
-
     @staticmethod
     def token_bytes_to_string(b):
         from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
@@ -1214,15 +1242,7 @@ class QwenModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("Qwen2ForCausalLM")
-class Qwen2Model(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN2
-
-
-@Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
-    model_arch = gguf.MODEL_ARCH.GPT2
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_block_count(self.hparams["n_layer"])
@@ -1284,10 +1304,7 @@ class GPT2Model(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
-@Model.register("PhiForCausalLM")
 class Phi2Model(Model):
-    model_arch = gguf.MODEL_ARCH.PHI2
-
     def set_gguf_parameters(self):
         block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
 
@@ -1309,10 +1326,7 @@ class Phi2Model(Model):
         self.gguf_writer.add_add_bos_token(False)
 
 
-@Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
-    model_arch = gguf.MODEL_ARCH.PLAMO
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -1391,10 +1405,7 @@ class PlamoModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("CodeShellForCausalLM")
 class CodeShellModel(Model):
-    model_arch = gguf.MODEL_ARCH.CODESHELL
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -1459,10 +1470,7 @@ class CodeShellModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("InternLM2ForCausalLM")
 class InternLM2Model(Model):
-    model_arch = gguf.MODEL_ARCH.INTERNLM2
-
     def set_vocab(self):
         # (TODO): Is there a better way?
         # Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
@@ -1634,10 +1642,7 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
-@Model.register("BertModel")
 class BertModel(Model):
-    model_arch = gguf.MODEL_ARCH.BERT
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
         self.vocab_size = None
@@ -1647,17 +1652,16 @@ class BertModel(Model):
         self.gguf_writer.add_causal_attention(False)
 
         # get pooling path
+        with open(self.dir_model / "modules.json", encoding="utf-8") as f:
+            modules = json.load(f)
         pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
+        for mod in modules:
+            if mod["type"] == "sentence_transformers.models.Pooling":
+                pooling_path = mod["path"]
+                break
 
         # get pooling type
+        pooling_type = gguf.PoolingType.NONE
         if pooling_path is not None:
             with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
                 pooling = json.load(f)
@@ -1667,7 +1671,8 @@ class BertModel(Model):
                 pooling_type = gguf.PoolingType.CLS
             else:
                 raise NotImplementedError("Only MEAN and CLS pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
+
+        self.gguf_writer.add_pooling_type(pooling_type.value)
 
     def set_vocab(self):
         path = self.dir_model
@@ -1743,10 +1748,7 @@ class BertModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("NomicBertModel")
 class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
 
@@ -1783,188 +1785,6 @@ class NomicBertModel(BertModel):
             yield name, data
 
 
-@Model.register("GemmaForCausalLM")
-class GemmaModel(Model):
-    model_arch = gguf.MODEL_ARCH.GEMMA
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
-        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
-        self.gguf_writer.add_key_length(hparams["head_dim"])
-        self.gguf_writer.add_value_length(hparams["head_dim"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("Starcoder2ForCausalLM")
-class StarCoder2Model(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER2
-
-
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
-class MambaModel(Model):
-    model_arch = gguf.MODEL_ARCH.MAMBA
-
-    def set_vocab(self):
-        vocab_size = self.hparams["vocab_size"]
-        # Round vocab size to next multiple of 8
-        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
-        # pad using ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
-        self.hparams["vocab_size"] = vocab_size
-
-        if (self.dir_model / "tokenizer.json").is_file():
-            self._set_vocab_gpt2()
-        else:
-            # Use the GPT-NeoX tokenizer when no tokenizer files are present
-            tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
-            print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
-            neox_reader = gguf.GGUFReader(tokenizer_path, "r")
-
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
-            self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
-            self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
-            self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
-            self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
-            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
-            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
-            self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
-
-    def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model"])
-        d_conv  = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
-        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
-        # ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
-        dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
-        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
-        # Fail early for models which don't have a block expansion factor of 2
-        assert d_inner == 2 * d_model
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
-        self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
-        self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
-        self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams["n_layer"]
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        tok_embd = None
-        tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
-        output_name   = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT]     + ".weight"
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            if name.endswith(".A_log"):
-                print("A_log --> A ==> " + new_name)
-                data_torch = -torch.exp(data_torch)
-
-            # assuming token_embd.weight is seen before output.weight
-            if tok_embd is not None and new_name == output_name:
-                if torch.equal(tok_embd, data_torch):
-                    print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
-                    continue
-            if new_name == tok_embd_name:
-                tok_embd = data_torch
-
-            data = data_torch.squeeze().numpy()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert big float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
 ###### CONVERSION LOGIC ######
 
 

convert-llama-ggml-to-gguf.py

@@ -373,7 +373,7 @@ def handle_metadata(cfg, hp):
         raise ValueError('Unable to load metadata')
     vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
     vocab_factory = convert.VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), cfg.model_metadata_dir)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
     return params, vocab, special_vocab
 
@@ -398,8 +398,8 @@ def handle_args():
                         help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
     parser.add_argument("--vocab-dir", type=Path,
                         help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
-    parser.add_argument("--vocabtype", default="spm,hfft",
-                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
 

convert.py

@@ -1282,32 +1282,35 @@ def load_some_model(path: Path) -> ModelPlus:
 
 
 class VocabFactory:
-    _FILES = {"spm": "tokenizer.model", "bpe": "vocab.json", "hfft": "tokenizer.json"}
-
     def __init__(self, path: Path):
         self.path = path
-        self.file_paths = self._detect_files()
-        print(f"Found vocab files: {self.file_paths}")
+        self.files: dict[str, Path | None] = {
+            "tokenizer.model": None,
+            "vocab.json": None,
+            "tokenizer.json": None,
+        }
+        self._detect_files()
 
-    def _detect_files(self) -> dict[str, Path | None]:
-        def locate(file: str) -> Path | None:
-            if (path := self.path / file).exists():
-                return path
-            if (path := self.path.parent / file).exists():
-                return path
-            return None
+    def _detect_files(self):
+        for file in self.files.keys():
+            file_path = self.path / file
+            parent_file_path = self.path.parent / file
+            if file_path.exists():
+                self.files[file] = file_path
+            elif parent_file_path.exists():
+                self.files[file] = parent_file_path
+        print(f"Found vocab files: {self.files}")
 
-        return {vt: locate(f) for vt, f in self._FILES.items()}
-
-    def _select_file(self, vocab_types: list[str]) -> tuple[str, Path]:
-        for vtype in vocab_types:
-            try:
-                path = self.file_paths[vtype]
-            except KeyError:
-                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
-            if path is not None:
-                return vtype, path
-        raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
+    def _select_file(self, vocabtype: str | None) -> Path:
+        if vocabtype in ["spm", "bpe"]:
+            for file_key in self.files.keys():
+                if (file := self.files[file_key]) is not None:
+                    return file
+            raise FileNotFoundError(f"{vocabtype} vocab not found.")
+        if vocabtype == "hfft":
+            # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
+            return self.path
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
     def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
         load_merges = vocabtype == "bpe"
@@ -1319,30 +1322,30 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
-        vocab_type, path = self._select_file(vocab_types)
-        print(f"Loading vocab file {path!r}, type {vocab_type!r}")
+    def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        path = self._select_file(vocabtype)
+        print(f"Loading vocab file '{path}', type '{vocabtype}'")
 
         added_tokens_path = path.parent / "added_tokens.json"
         vocab: Vocab
-        if vocab_type == "bpe":
+        if vocabtype == "bpe":
             vocab = BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "spm":
+        elif vocabtype == "spm":
             vocab = SentencePieceVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocab_type == "hfft":
+        elif vocabtype == "hfft":
             vocab = HfVocab(
-                path.parent, added_tokens_path if added_tokens_path.exists() else None
+                path, added_tokens_path if added_tokens_path.exists() else None
             )
         else:
-            raise ValueError(vocab_type)
+            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocab_type,
+            vocabtype,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1376,13 +1379,15 @@ def main(args_in: list[str] | None = None) -> None:
     if np.uint32(1) == np.uint32(1).newbyteorder("<"):
         # We currently only support Q8_0 output on little endian systems.
         output_choices.append("q8_0")
-    parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
+    vocab_types = ["spm", "bpe", "hfft"]
+    parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
+    parser.add_argument("--awq-path",     type=Path,              help="Path to scale awq cache file", default=None)
     parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
     parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
     parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
     parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
     parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
+    parser.add_argument("--vocab-type",   choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
     parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
     parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
     parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
@@ -1392,6 +1397,18 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
+    if args.awq_path:
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
+        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
+        tmp_model_path = args.model / "weighted_model"
+        if tmp_model_path.is_dir():
+            print(f"{tmp_model_path} exists as a weighted model.")
+        else:
+            tmp_model_path.mkdir(parents=True, exist_ok=True)
+            print("Saving new weighted model ...")
+            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
+            print(f"Saved weighted model at {tmp_model_path}.")
+        args.model = tmp_model_path
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1431,7 +1448,7 @@ def main(args_in: list[str] | None = None) -> None:
     model_parent_path = model_plus.paths[0].parent
     vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
+    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type, model_parent_path)
 
     if args.vocab_only:
         if not args.outfile:

examples/CMakeLists.txt

@@ -20,7 +20,6 @@ else()
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(finetune)
-    add_subdirectory(gritlm)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)

examples/baby-llama/baby-llama.cpp

@@ -1547,7 +1547,7 @@ int main(int argc, char ** argv) {
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;

examples/batched-bench/batched-bench.cpp

@@ -32,15 +32,16 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
         printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
     }
 
     int n_kv_max     = 2048;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
+    int mmq          = 0;
 
     std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
     std::vector<int> n_tg = { 128, 256, };
@@ -64,15 +65,19 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        n_pp = parse_list(argv[5]);
+        mmq = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_tg = parse_list(argv[6]);
+        n_pp = parse_list(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pl = parse_list(argv[7]);
+        n_tg = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_pl = parse_list(argv[8]);
     }
 
     // init LLM
@@ -100,14 +105,12 @@ int main(int argc, char ** argv) {
 
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
+    ctx_params.n_batch   = 2048;
+    ctx_params.mul_mat_q = mmq;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
-    // ensure enough sequences are available
-    ctx_params.n_parallel = *std::max_element(n_pl.begin(), n_pl.end());
-
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     if (ctx == NULL) {
@@ -156,7 +159,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
@@ -177,10 +180,10 @@ int main(int argc, char ** argv) {
 
                 llama_batch_clear(batch);
 
-                for (int i = 0; i < pp; ++i) {
-                    for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        llama_batch_add(batch, 0, i, { j }, false);
-                    }
+                const int n_tokens = is_pp_shared ? pp : pl*pp;
+
+                for (int i = 0; i < n_tokens; ++i) {
+                    llama_batch_add(batch, 0, i, { 0 }, false);
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
@@ -195,7 +198,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
                     }
                 }
 

examples/batched/batched.cpp

@@ -80,7 +80,6 @@ int main(int argc, char ** argv) {
     ctx_params.seed  = 1234;
     ctx_params.n_ctx = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_parallel      = n_parallel;
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
@@ -133,7 +132,7 @@ int main(int argc, char ** argv) {
     // assign the system KV cache to all parallel sequences
     // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
     for (int32_t i = 1; i < n_parallel; ++i) {
-        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
     }
 
     if (n_parallel > 1) {

examples/benchmark/benchmark-matmult.cpp

@@ -189,10 +189,12 @@ int main(int argc, char ** argv)  {
 
     int32_t nelements = sizex*sizey;
 
+    std::vector<int64_t> hist_cur(1 << 4, 0);
+
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
     struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -205,7 +207,7 @@ int main(int argc, char ** argv)  {
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
     struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

examples/embedding/embedding.cpp

@@ -19,7 +19,18 @@ static std::vector<std::string> split_lines(const std::string & s) {
 
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
     for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
     }
 }
 
@@ -33,23 +44,11 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
-    for (int i = 0; i < batch.n_tokens; i++) {
-        if (!batch.logits[i]) {
-            continue;
-        }
-
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        if (embd == NULL) {
-            embd = llama_get_embeddings_ith(ctx, i);
-            if (embd == NULL) {
-                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
-                continue;
-            }
-        }
-
-        float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd);
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
     }
 }
 
@@ -133,7 +132,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
     // allocate output
     const int n_embd = llama_n_embd(model);
@@ -146,7 +145,6 @@ int main(int argc, char ** argv) {
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
         auto & inp = inputs[k];
-
         const uint64_t n_toks = inp.size();
 
         // encode if at capacity

examples/finetune/finetune.cpp

@@ -1531,7 +1531,7 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;

examples/gritlm/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET gritlm)
-add_executable(${TARGET} gritlm.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gritlm/gritlm.cpp

@@ -1,229 +0,0 @@
-#include "common.h"
-#include "llama.h"
-
-#include <string>
-#include <vector>
-
-// #define GRIT_DEBUG
-
-static float dot_product(const std::vector<float> & v1, const std::vector<float> & v2) {
-    float dot = 0.0f;
-    for (uint64_t i = 0; i < v1.size(); ++i) {
-        dot += v1[i] * v2[i];
-    }
-    return dot;
-}
-
-static float norm(const std::vector<float> & v) {
-    return std::sqrt(dot_product(v, v));
-}
-
-static float cosine_similarity(const std::vector<float> & v1, const std::vector<float> & v2) {
-    return dot_product(v1, v2) / (norm(v1) * norm(v2));
-}
-
-static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
-    std::vector<std::vector<float>> result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-
-    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    for (uint64_t i = 0; i < sentences.size(); i++) {
-        llama_batch_clear(batch);
-
-        const std::string input_string = instruction + sentences[i];
-
-        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
-
-        const int32_t n_toks = inputs.size();
-
-        // GritLM seems to have EOS = ""
-        // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_token_eos(mdl));
-
-        // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
-
-#ifdef GRIT_DEBUG
-        // debug tokens - should be matching as referenced in the GritLM sample
-        std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
-            std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
-        });
-        std::printf("\n");
-#endif
-
-        // add input to batch (this increments n_tokens)
-        for (int32_t j = 0; j < n_toks; j++) {
-            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
-        }
-
-        // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
-        llama_set_causal_attn(ctx, false);
-
-        // run model
-        llama_decode(ctx, batch);
-
-        // get embedding dimensions
-        uint64_t n_embd = llama_n_embd(mdl);
-
-        // allocate embedding output
-        std::vector<float> emb_unorm(n_embd, 0.0f);
-
-        // sum up all token embeddings
-        for (int32_t k = n_inst; k < n_toks; k++) {
-            float * emb = llama_get_embeddings_ith(ctx, k);
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] += emb[j];
-            }
-        }
-
-        // divide by number of tokens (mean pooling)
-        {
-            const uint64_t n_sent = n_toks - n_inst;
-
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] /= n_sent;
-            }
-        }
-
-        std::vector<float> emb_norm(emb_unorm.size());
-        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
-        result.push_back(emb_norm);
-
-#ifdef GRIT_DEBUG
-        // print out emb_norm
-        std::printf("embedding %ld: ", i);
-        for (uint64_t j = 0; j < n_embd; j++) {
-            std::printf("%.5f ", emb_norm[j]);
-        }
-        std::printf("\n\n");
-#endif
-    }
-
-    llama_batch_free(batch);
-
-    return result;
-}
-
-static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
-    std::string result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(mdl);
-
-    llama_kv_cache_clear(ctx);
-    llama_set_causal_attn(ctx, true);
-    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
-    int32_t i_current_token = 0;
-
-    while (true) {
-        llama_batch_clear(bat);
-        auto n_inputs = (int32_t)inputs.size();
-        for (int32_t i = 0; i < n_inputs; i++) {
-            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
-        }
-        inputs.clear();
-
-        llama_decode(ctx, bat);
-        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
-
-        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
-        auto n_candidates = (int32_t)candidates.size();
-        for (int32_t token = 0; token < n_candidates; token++) {
-            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
-        }
-        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
-
-        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
-        if (token == eos_token) {
-            break;
-        }
-
-        std::string piece = llama_token_to_piece(ctx, token);
-        if (stream) {
-            std::printf("%s", piece.c_str());
-            std::fflush(stdout);
-        }
-
-        inputs.push_back(token);
-
-        result += piece;
-    }
-
-    if (stream) {
-        std::printf("\n");
-    }
-
-    llama_batch_free(bat);
-
-    return result;
-}
-
-static std::string gritlm_instruction(const std::string & instruction) {
-    return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
-}
-
-int main(int argc, char * argv[]) {
-    gpt_params params;
-    if (!gpt_params_parse(argc, argv, params)) {
-        return 1;
-    }
-
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
-    llama_backend_init();
-
-    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
-
-    // create new context - set to embedding mode
-    cparams.embeddings = true;
-    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
-
-    // ### Embedding/Representation ###
-    // samples taken from: https://github.com/ContextualAI/gritlm#basic
-    {
-        const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
-
-        const std::vector<std::string> queries = {
-            "Bitcoin: A Peer-to-Peer Electronic Cash System",
-            "Generative Representational Instruction Tuning",
-        };
-
-        const std::vector<std::string> documents = {
-            "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
-            "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
-        };
-
-        // No need to add instruction for retrieval documents
-        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
-        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
-
-        const float cosine_sim_q0_d0 = cosine_similarity(q_rep[0], d_rep[0]);
-        const float cosine_sim_q0_d1 = cosine_similarity(q_rep[0], d_rep[1]);
-        const float cosine_sim_q1_d0 = cosine_similarity(q_rep[1], d_rep[0]);
-        const float cosine_sim_q1_d1 = cosine_similarity(q_rep[1], d_rep[1]);
-
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
-    }
-
-    // ### Generation ###
-    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
-    {
-        const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, prompt, true);
-    }
-
-    llama_free(ctx);
-    llama_free_model(mdl);
-    llama_backend_free();
-
-    return 0;
-}

examples/infill/infill.cpp

@@ -378,10 +378,10 @@ int main(int argc, char ** argv) {
     if (params.interactive) {
         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 LLaMa, 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 LLaMa.\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";
         }
@@ -447,8 +447,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

@@ -87,21 +87,7 @@ class SchemaConverter:
         elif schema_type == 'array' and 'items' in schema:
             # TODO `prefixItems` keyword
             item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
             return self._add_rule(rule_name, rule)
 
         else:

examples/llama-bench/README.md

@@ -35,6 +35,7 @@ options:
   -mg, --main-gpu <i>                 (default: 0)
   -nkvo, --no-kv-offload <0|1>        (default: 0)
   -mmp, --mmap <0|1>                  (default: 1)
+  -mmq, --mul-mat-q <0|1>             (default: 1)
   -ts, --tensor_split <ts0/ts1/..>    (default: 0)
   -r, --repetitions <n>               (default: 5)
   -o, --output <csv|json|md|sql>      (default: md)

examples/llama-bench/llama-bench.cpp

@@ -123,15 +123,20 @@ static std::string get_gpu_info() {
     }
 #endif
 #ifdef GGML_USE_SYCL
-    int count = ggml_backend_sycl_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_sycl_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
+    int device_list[GGML_SYCL_MAX_DEVICES];
+    ggml_sycl_get_gpu_list(device_list, GGML_SYCL_MAX_DEVICES);
+
+    for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
+        if (device_list[i] >0 ){
+            char buf[128];
+            ggml_sycl_get_device_description(i, buf, sizeof(buf));
+            id += buf;
             id += "/";
         }
     }
+    if (id.length() >2 ) {
+        id.pop_back();
+    }
 #endif
     // TODO: other backends
     return id;
@@ -152,9 +157,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_MODE_NONE:  return "none";
-        case LLAMA_SPLIT_MODE_LAYER: return "layer";
-        case LLAMA_SPLIT_MODE_ROW:   return "row";
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -171,9 +176,9 @@ struct cmd_params {
     std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
+    std::vector<bool> mul_mat_q;
     std::vector<std::vector<float>> tensor_split;
     std::vector<bool> use_mmap;
-    std::vector<bool> embeddings;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -188,12 +193,12 @@ static const cmd_params cmd_params_defaults = {
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
+    /* mul_mat_q     */ {true},
     /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* use_mmap      */ {true},
-    /* embeddings    */ {false},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -216,7 +221,7 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
     printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
@@ -353,11 +358,11 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             for (const auto & m : p) {
                 llama_split_mode mode;
                 if (m == "none") {
-                    mode = LLAMA_SPLIT_MODE_NONE;
+                    mode = LLAMA_SPLIT_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_MODE_LAYER;
+                    mode = LLAMA_SPLIT_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_MODE_ROW;
+                    mode = LLAMA_SPLIT_ROW;
                 } else {
                     invalid_param = true;
                     break;
@@ -378,6 +383,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+        } else if (arg == "-mmq" || arg == "--mul-mat-q") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -385,13 +397,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
-        } else if (arg == "-embd" || arg == "--embeddings") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -461,9 +466,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
+    if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
-    if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
     return params;
@@ -481,9 +486,9 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -513,8 +518,8 @@ struct cmd_params_instance {
         cparams.n_batch = n_batch;
         cparams.type_k = type_k;
         cparams.type_v = type_v;
+        cparams.mul_mat_q = mul_mat_q;
         cparams.offload_kqv = !no_kv_offload;
-        cparams.embeddings = embeddings;
 
         return cparams;
     }
@@ -530,10 +535,10 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & mmp : params.use_mmap)
-    for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
+    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -552,9 +557,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -575,9 +580,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -611,9 +616,9 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -634,9 +639,9 @@ struct test {
         split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
         no_kv_offload = inst.no_kv_offload;
+        mul_mat_q = inst.mul_mat_q;
         tensor_split = inst.tensor_split;
         use_mmap = inst.use_mmap;
-        embeddings = inst.embeddings;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
         // RFC 3339 date-time format
@@ -708,7 +713,7 @@ struct test {
             "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
-            "tensor_split", "use_mmap", "embeddings",
+            "mul_mat_q", "tensor_split", "use_mmap",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -728,7 +733,7 @@ struct test {
         }
         if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "use_mmap" || field == "embeddings") {
+            field == "mul_mat_q" || field == "use_mmap") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -762,7 +767,7 @@ struct test {
             std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload),
-            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
+            std::to_string(mul_mat_q), tensor_split_str, std::to_string(use_mmap),
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -926,15 +931,15 @@ struct markdown_printer : public printer {
         if (field == "n_threads") {
             return "threads";
         }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
         if (field == "no_kv_offload") {
             return "nkvo";
         }
         if (field == "use_mmap") {
             return "mmap";
         }
-        if (field == "embeddings") {
-            return "embd";
-        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -969,6 +974,9 @@ struct markdown_printer : public printer {
         if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
             fields.emplace_back("split_mode");
         }
+        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
+            fields.emplace_back("mul_mat_q");
+        }
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
             fields.emplace_back("no_kv_offload");
         }
@@ -978,9 +986,6 @@ struct markdown_printer : public printer {
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
-        if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
-            fields.emplace_back("embeddings");
-        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 

examples/llama.android/app/build.gradle.kts

@@ -21,8 +21,12 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
+            // Workaround for https://github.com/llvm/llvm-project/issues/65820
+            // affecting armeabi-v7a. Skip armeabi-v7a when invoked with
+            // -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
+            if (project.hasProperty("skip-armeabi-v7a")) {
+                abiFilters += listOf("arm64-v8a", "x86_64", "x86")
+            }
         }
         externalNativeBuild {
             cmake {

examples/llava/README.md

@@ -59,39 +59,14 @@ python ./convert.py ../llava-v1.5-7b --skip-unknown
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
 ## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-2) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
+
+1) Backup your pth/safetensor model files as llava-surgery modifies them
+2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
 - you will find a llava.projector and a llava.clip file in your model directory
-3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-4) Create the visual gguf model:
-```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
+3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
+4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
 - This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-5) Then convert the model to gguf format:
-```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-6) And finally we can run the llava-cli using the 1.6 model version:
-```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
+5) Everything else as usual: convert.py the hf model, quantize as needed
 **note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
 **note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
 

examples/llava/clip.cpp

@@ -616,9 +616,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             KQ = ggml_soft_max_inplace(ctx0, KQ);
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
             KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_positions, n_head, batch_size);
-            KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+            KQV = ggml_cont(ctx0, ggml_permute(ctx0, KQV, 0, 2, 1, 3));
 
-            cur = ggml_cont_3d(ctx0, KQV, hidden_size, num_positions, batch_size);
+            cur = ggml_cpy(ctx0, KQV, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size));
         }
 
         // attention output
@@ -1862,6 +1862,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     std::vector<uint8_t> work(512);
     std::vector<float> conv_buf(512);
+    std::vector<int64_t> hist_all(1 << 4, 0);
     size_t total_size_org = 0;
     size_t total_size_new = 0;
 
@@ -1916,7 +1917,48 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             }
             new_data = work.data();
 
-            new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, n_elms/cur->ne[0], cur->ne[0], nullptr);
+            std::vector<int64_t> hist_cur(1 << 4, 0);
+
+            switch (new_type) {
+                case GGML_TYPE_Q4_0: {
+                    new_size = ggml_quantize_q4_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_1: {
+                    new_size = ggml_quantize_q4_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_0: {
+                    new_size = ggml_quantize_q5_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_1: {
+                    new_size = ggml_quantize_q5_1(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q8_0: {
+                    new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q2_K: {
+                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q3_K: {
+                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_K: {
+                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_K: {
+                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q6_K: {
+                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                default: {
+                    fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
+                    return false;
+                }
+            }
+
+            for (size_t j = 0; j < hist_cur.size(); ++j) {
+                hist_all[j] += hist_cur[j];
+            }
         } else {
             new_type = cur->type;
             new_data = cur->data;
@@ -1951,6 +1993,17 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     {
         printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
         printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+
+        int64_t sum_all = 0;
+        for (size_t i = 0; i < hist_all.size(); ++i) {
+            sum_all += hist_all[i];
+        }
+
+        printf("%s: hist: ", __func__);
+        for (size_t i = 0; i < hist_all.size(); ++i) {
+            printf("%5.3f ", hist_all[i] / (float)sum_all);
+        }
+        printf("\n");
     }
 
     return true;

examples/llava/llava-surgery-v2.py

@@ -65,7 +65,9 @@ def clean_vision_tower_from_checkpoint(checkpoint_path):
         for name in clip_tensors:
             del checkpoint[name]
 
+        # Save the updated checkpoint
         checkpoint_path = checkpoint_path
+        save_model(checkpoint, checkpoint_path, file_type)
         return True
     return False
 
@@ -150,6 +152,16 @@ for name in first_mm_tensors:
 if len(projector) > 0:
     save_model(projector, f"{args.model}/llava.projector", 'pytorch')
 
+for name in mm_tensors:
+    del last_checkpoint[name]
+for name in first_mm_tensors:
+    del first_checkpoint[name]
+
+if len(mm_tensors) > 0:
+    save_model(last_checkpoint, projector_checkpoint_path, file_type)
+if len(first_mm_tensors) > 0:
+    save_model(first_checkpoint, newline_checkpoint_path, file_type)
+
 print("Done!")
 print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
 print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava.cpp

@@ -152,7 +152,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
 
     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->backend != GGML_BACKEND_CPU) {
         if (newline_tmp->buffer == NULL) {
             printf("newline_tmp tensor buffer is NULL\n");
         }
@@ -311,7 +311,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
+static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
         fprintf(stderr, "Unable to allocate memory for image embeddings\n");

examples/llava/llava.h

@@ -31,8 +31,6 @@ struct llava_image_embed {
 /** sanity check for clip <-> llava embed size match */
 LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
 
-LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
-
 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
 /** build an image embed from a path to an image filename */

examples/main/main.cpp

@@ -334,8 +334,6 @@ int main(int argc, char ** argv) {
     // number of tokens to keep when resetting context
     if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct || params.chatml) {
         params.n_keep = (int)embd_inp.size();
-    } else {
-        params.n_keep += add_bos; // always keep the BOS token
     }
 
     // prefix & suffix for instruct mode
@@ -385,8 +383,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > add_bos) {
-            LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > 0) {
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
                 LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
@@ -511,14 +509,6 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> embd;
     std::vector<llama_token> embd_guidance;
 
-    // tokenized antiprompts
-    std::vector<std::vector<llama_token>> antiprompt_ids;
-
-    antiprompt_ids.reserve(params.antiprompt.size());
-    for (const std::string & antiprompt : params.antiprompt) {
-        antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
-    }
-
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
 
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
@@ -550,14 +540,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep;
+                    const int n_left    = n_past - params.n_keep - 1;
                     const int n_discard = n_left/2;
 
                     LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                             n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                     n_past -= n_discard;
 
@@ -584,9 +574,9 @@ int main(int argc, char ** argv) {
                     LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
                     LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                    llama_kv_cache_seq_add(ctx, 0, ga_i,                n_past,              ib*bd);
-                    llama_kv_cache_seq_div(ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
-                    llama_kv_cache_seq_add(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
                     n_past -= bd;
 
@@ -777,18 +767,6 @@ int main(int argc, char ** argv) {
                     }
                 }
 
-                // check for reverse prompt using special tokens
-                llama_token last_token = llama_sampling_last(ctx_sampling);
-                for (std::vector<llama_token> ids : antiprompt_ids) {
-                    if (ids.size() == 1 && last_token == ids[0]) {
-                        if (params.interactive) {
-                            is_interacting = true;
-                        }
-                        is_antiprompt = true;
-                        break;
-                    }
-                }
-
                 if (is_antiprompt) {
                     LOG("found antiprompt: %s\n", last_output.c_str());
                 }

examples/parallel/parallel.cpp

@@ -107,9 +107,6 @@ int main(int argc, char ** argv) {
     // number of simultaneous "clients" to simulate
     const int32_t n_clients = params.n_parallel;
 
-    // dedicate one sequence to the system prompt
-    params.n_parallel += 1;
-
     // requests to simulate
     const int32_t n_seq = params.n_sequences;
 
@@ -199,8 +196,8 @@ int main(int argc, char ** argv) {
         }
 
         // assign the system KV cache to all parallel sequences
-        for (int32_t i = 1; i <= n_clients; ++i) {
-            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+        for (int32_t i = 1; i < n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, 0, n_tokens_system);
         }
 
         LOG_TEE("\n");
@@ -224,17 +221,15 @@ int main(int argc, char ** argv) {
 
             client.i_batch = batch.n_tokens;
 
-            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
+            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id }, true);
 
             client.n_decoded += 1;
         }
 
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
-            for (int i = 1; i <= n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, -1, -1);
-                // but keep the system prompt
-                llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+            for (int i = 0; i < n_clients; ++i) {
+                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
             }
 
             LOG_TEE("%s: clearing the KV cache\n", __func__);
@@ -260,7 +255,7 @@ int main(int argc, char ** argv) {
                     tokens_prompt = ::llama_tokenize(ctx, client.prompt, false);
 
                     for (size_t i = 0; i < tokens_prompt.size(); ++i) {
-                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
+                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id }, false);
                     }
 
                     // extract the logits only for the last token
@@ -371,8 +366,7 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
-                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -126,7 +126,7 @@ int main(int argc, char ** argv) {
     const int n_batch     = ctx_params.n_batch;
     const int n_batch_grp = ctx_params.n_batch/n_grp;
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d, n_junk = %d, i_pos = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch, n_junk, i_pos);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
 
     // print the prompt token-by-token
 
@@ -146,11 +146,10 @@ int main(int argc, char ** argv) {
             const int ib = i/n_batch - 1;
             const int bd = n_batch_grp*(n_grp - 1);
 
-            llama_kv_cache_seq_add (ctx, 0, n_past - n_batch,         n_past,         ib*bd);
-            llama_kv_cache_seq_div (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
-            llama_kv_cache_update  (ctx);
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= bd;
         }
 
         llama_batch_clear(batch);
@@ -180,12 +179,10 @@ int main(int argc, char ** argv) {
 
         LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
 
-        llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-        llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-      //llama_kv_cache_defrag (ctx);
-        llama_kv_cache_update (ctx);
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+        n_past -= n_discard;
 
         llama_batch_clear(batch);
 
@@ -211,12 +208,10 @@ int main(int argc, char ** argv) {
         if (n_discard > 0) {
             LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
 
-            llama_kv_cache_seq_rm (ctx, 0, n_keep            , n_keep + n_discard);
-            llama_kv_cache_seq_add(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
-          //llama_kv_cache_defrag (ctx);
-            llama_kv_cache_update (ctx);
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
 
-            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
+            n_past -= n_discard;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -442,7 +442,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     return {tokens, std::exp(nll / count), logit_history, prob_history};
 }
 
-static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
@@ -453,6 +453,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -498,19 +499,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     double nll2 = 0.0;
 
     const int num_batches = (n_ctx + n_batch - 1) / n_batch;
-    const int n_seq = std::max(1, n_batch / n_ctx);
-
-    GGML_ASSERT(n_batch < n_ctx || n_batch % n_ctx == 0);
-    GGML_ASSERT(params.n_ctx == n_seq * n_ctx);
-
-    llama_batch batch = llama_batch_init(std::min(n_batch, n_ctx*n_seq), 0, 1);
 
     std::vector<float> logits;
     if (num_batches > 1) {
         logits.reserve((size_t)n_ctx * n_vocab);
     }
 
-    fprintf(stderr, "%s: calculating perplexity over %d chunks, n_ctx=%d, batch_size=%d, n_seq=%d\n", __func__, n_chunk, n_ctx, n_batch, n_seq);
+    fprintf(stderr, "%s: calculating perplexity over %d chunks, batch_size=%d\n", __func__, n_chunk, n_batch);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -523,26 +518,10 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         log_probs.resize(n_ctx * nv);
     }
 
-    // We get the logits for all the tokens in the context window (params.n_ctx)
-    // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
-    // calculate the perplexity over the last half of the window (so the model always has
-    // some context to predict the token).
-    //
-    // We rely on the fact that attention in the forward pass only looks at previous
-    // tokens here, so the logits returned for each token are an accurate representation
-    // of what the model would have predicted at that point.
-    //
-    // Example, we have a context window of 512, we will compute perplexity for each of the
-    // last 256 tokens.  Then, we split the input up into context window size chunks to
-    // process the entire prompt.
-    const int first = n_ctx/2;
-
-    for (int i = 0; i < n_chunk; i += n_seq) {
+    for (int i = 0; i < n_chunk; ++i) {
         const int start =     i * n_ctx;
         const int end   = start + n_ctx;
 
-        const int n_seq_batch = std::min(n_seq, n_chunk - i);
-
         const auto t_start = std::chrono::high_resolution_clock::now();
 
         // clear the KV cache
@@ -552,37 +531,22 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
-            batch.n_tokens = 0;
-            for (int seq = 0; seq < n_seq_batch; seq++) {
-                int seq_start = batch_start + seq*n_ctx;
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
 
-                // save original token and restore it after eval
-                const auto token_org = tokens[seq_start];
-
-                // add BOS token for the first batch of each chunk
-                if (add_bos && j == 0) {
-                    tokens[seq_start] = llama_token_bos(llama_get_model(ctx));
-                }
-
-                for (int k = 0; k < batch_size; ++k) {
-                    const int idx = seq*n_ctx + k;
-                    batch.token[idx] = tokens[seq_start + k];
-                    batch.pos[idx] = j*n_batch + k;
-                    batch.n_seq_id[idx] = 1;
-                    batch.seq_id[idx][0] = seq;
-                    batch.logits[idx] = batch.pos[idx] >= first ? 1 : 0;
-                }
-                batch.n_tokens += batch_size;
-
-                // restore the original token in case it was set to BOS
-                tokens[seq_start] = token_org;
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            if (llama_decode(ctx, batch)) {
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
                 fprintf(stderr, "%s : failed to eval\n", __func__);
                 return {tokens, -1, logit_history, prob_history};
             }
 
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
             if (num_batches > 1) {
                 const auto * batch_logits = llama_get_logits(ctx);
                 logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
@@ -594,7 +558,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         if (i == 0) {
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
             fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
-            int total_seconds = (int)(t_total*n_chunk/n_seq);
+            int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
                 fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
@@ -602,31 +566,37 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        for (int seq = 0; seq < n_seq_batch; seq++) {
-            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx);
-            llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
-            if (!params.logits_file.empty()) {
-                process_logits(logits_stream, n_vocab, all_logits + first*n_vocab,
-                        tokens_data, n_ctx - 1 - first,
-                        workers, log_probs, nll, nll2);
-            } else {
-                process_logits(n_vocab, all_logits + first*n_vocab,
-                        tokens_data, n_ctx - 1 - first,
-                        workers, nll, nll2,
-                        logit_history.data() + start + seq*n_ctx + first,
-                        prob_history.data()  + start + seq*n_ctx + first);
-            }
-            count += n_ctx - first - 1;
+        // We get the logits for all the tokens in the context window (params.n_ctx)
+        // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,
+        // calculate the perplexity over the last half of the window (so the model always has
+        // some context to predict the token).
+        //
+        // We rely on the fact that attention in the forward pass only looks at previous
+        // tokens here, so the logits returned for each token are an accurate representation
+        // of what the model would have predicted at that point.
+        //
+        // Example, we have a context window of 512, we will compute perplexity for each of the
+        // last 256 tokens.  Then, we split the input up into context window size chunks to
+        // process the entire prompt.
+        const int first = n_ctx/2;
+        const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+        if (!params.logits_file.empty()) {
+            process_logits(logits_stream, n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                    workers, log_probs, nll, nll2);
+        } else {
+            process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                    workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        }
+        count += n_ctx - first - 1;
 
-            // perplexity is e^(average negative log-likelihood)
-            if (params.ppl_output_type == 0) {
-                printf("[%d]%.4lf,", i + seq + 1, std::exp(nll / count));
-            } else {
-                double av = nll/count;
-                double av2 = nll2/count - av*av;
-                if (av2 > 0) av2 = sqrt(av2/(count-1));
-                printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
-            }
+        // perplexity is e^(average negative log-likelihood)
+        if (params.ppl_output_type == 0) {
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        } else {
+            double av = nll/count;
+            double av2 = nll2/count - av*av;
+            if (av2 > 0) av2 = sqrt(av2/(count-1));
+            printf("%8d  %.4lf  %4lf  %4lf\n", i*n_ctx, std::exp(nll / count), av, av2);
         }
         fflush(stdout);
 
@@ -645,8 +615,6 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         printf("Unexpected negative standard deviation of log(prob)\n");
     }
 
-    llama_batch_free(batch);
-
     return {tokens, ppl, logit_history, prob_history};
 }
 
@@ -841,7 +809,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = 4*max_tasks_per_batch;
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1118,7 +1086,7 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 128;
-    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = 2*max_tasks_per_batch;
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1470,7 +1438,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     const int n_batch = params.n_batch;
 
     const int max_tasks_per_batch = 32;
-    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
+    const int max_seq = 4*max_tasks_per_batch;
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1814,24 +1782,13 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
+    params.n_batch = 512;
     if (!gpt_params_parse(argc, argv, params)) {
         return 1;
     }
 
     params.logits_all = true;
-
-    const int32_t n_ctx = params.n_ctx;
-
-    const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence;
-    if (ppl) {
-        int n_seq = std::max(1, params.n_batch / n_ctx);
-        int32_t n_kv = n_seq * n_ctx;
-        params.n_parallel = n_seq;
-        params.n_ctx = n_kv;
-        params.n_batch = std::min(params.n_batch, n_kv);
-    } else {
-        params.n_batch = std::min(params.n_batch, params.n_ctx);
-    }
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     if (params.ppl_stride > 0) {
         fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
@@ -1858,9 +1815,6 @@ int main(int argc, char ** argv) {
     llama_model * model;
     llama_context * ctx;
 
-    // ensure there's at least enough seq_ids for HellaSwag
-    params.n_parallel = std::max(4, params.n_parallel);
-
     // load the model and apply lora adapter, if any
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == NULL) {
@@ -1890,7 +1844,7 @@ int main(int argc, char ** argv) {
     } else if (params.kl_divergence) {
         kl_divergence(ctx, params);
     } else {
-        results = perplexity(ctx, params, n_ctx);
+        results = perplexity(ctx, params);
     }
 
     llama_print_timings(ctx);

examples/quantize/quantize.cpp

@@ -23,21 +23,15 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "Q5_1",   LLAMA_FTYPE_MOSTLY_Q5_1,   " 4.70G, +0.0349 ppl @ LLaMA-v1-7B", },
     { "IQ2_XXS",LLAMA_FTYPE_MOSTLY_IQ2_XXS," 2.06 bpw quantization",            },
     { "IQ2_XS", LLAMA_FTYPE_MOSTLY_IQ2_XS, " 2.31 bpw quantization",            },
-    { "IQ2_S",  LLAMA_FTYPE_MOSTLY_IQ2_S,  " 2.5  bpw quantization",            },
-    { "IQ2_M",  LLAMA_FTYPE_MOSTLY_IQ2_M,  " 2.7  bpw quantization",            },
     { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
-    { "IQ3_S",  LLAMA_FTYPE_MOSTLY_IQ3_S,  " 3.44 bpw quantization",            },
-    { "IQ3_M",  LLAMA_FTYPE_MOSTLY_IQ3_M,  " 3.66 bpw quantization mix",        },
     { "Q3_K",   LLAMA_FTYPE_MOSTLY_Q3_K_M, "alias for Q3_K_M" },
-    { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw quantization"   ,          },
+    { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
     { "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
-    { "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.50 bpw non-linear quantization", },
-    { "IQ4_XS", LLAMA_FTYPE_MOSTLY_IQ4_XS, " 4.25 bpw non-linear quantization", },
     { "Q4_K",   LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
     { "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
     { "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
@@ -295,7 +289,6 @@ int main(int argc, char ** argv) {
     }
 
     if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
          params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
         fprintf(stderr, "\n===============================================================================================\n");
         fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");

examples/server-embd.py

@@ -1,34 +0,0 @@
-import asyncio
-import requests
-import numpy as np
-
-n = 8
-
-result = []
-
-async def requests_post_async(*args, **kwargs):
-    return await asyncio.to_thread(requests.post, *args, **kwargs)
-
-async def main():
-    model_url = "http://127.0.0.1:6900"
-    responses: list[requests.Response] = await asyncio.gather(*[requests_post_async(
-        url= f"{model_url}/embedding",
-        json= {"content": str(0)*1024}
-    ) for i in range(n)])
-
-    for response in responses:
-        embedding = response.json()["embedding"]
-        print(embedding[-8:])
-        result.append(embedding)
-
-asyncio.run(main())
-
-# compute cosine similarity
-
-for i in range(n-1):
-    for j in range(i+1, n):
-        embedding1 = np.array(result[i])
-        embedding2 = np.array(result[j])
-        similarity = np.dot(embedding1, embedding2) / (np.linalg.norm(embedding1) * np.linalg.norm(embedding2))
-        print(f"Similarity between {i} and {j}: {similarity:.2f}")
-

examples/server/CMakeLists.txt

@@ -1,18 +1,12 @@
 set(TARGET server)
 option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
-option(LLAMA_SERVER_SSL "Build SSL support for the server" OFF)
 include_directories(${CMAKE_CURRENT_SOURCE_DIR})
-add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp oai.hpp utils.hpp json.hpp httplib.h)
 install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common ${CMAKE_THREAD_LIBS_INIT})
-if (LLAMA_SERVER_SSL)
-    find_package(OpenSSL REQUIRED)
-    target_link_libraries(${TARGET} PRIVATE OpenSSL::SSL OpenSSL::Crypto)
-    target_compile_definitions(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT)
-endif()
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -1,24 +1,11 @@
-# LLaMA.cpp HTTP Server
+# llama.cpp/example/server
 
-Fast, lightweight, pure C/C++ HTTP server based on [httplib](https://github.com/yhirose/cpp-httplib), [nlohmann::json](https://github.com/nlohmann/json) and **llama.cpp**.
+This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
 
-Set of LLM REST APIs and a simple web front end to interact with llama.cpp.
-
-**Features:**
- * LLM inference of F16 and quantum models on GPU and CPU
- * [OpenAI API](https://github.com/openai/openai-openapi) compatible chat completions and embeddings routes
- * Parallel decoding with multi-user support
- * Continuous batching
- * Multimodal (wip)
- * Monitoring endpoints
-
-The project is under active development, and we are [looking for feedback and contributors](https://github.com/ggerganov/llama.cpp/issues/4216).
-
-**Command line options:**
+Command line options:
 
 - `--threads N`, `-t N`: Set the number of threads to use during generation.
 - `-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.
-- `--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`).
 - `-a ALIAS`, `--alias ALIAS`: Set an alias for the model. The alias will be returned in API responses.
 - `-c N`, `--ctx-size N`: Set the size of the prompt context. The default is 512, but LLaMA models were built with a context of 2048, which will provide better results for longer input/inference. The size may differ in other models, for example, baichuan models were build with a context of 4096.
@@ -42,7 +29,7 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 - `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 - `--port`: Set the port to listen. Default: `8080`.
-- `--path`: path from which to serve static files (default: disabled)
+- `--path`: path from which to serve static files (default examples/server/public)
 - `--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.
@@ -52,16 +39,8 @@ see https://github.com/ggerganov/llama.cpp/issues/1437
 - `--mmproj MMPROJ_FILE`: Path to a multimodal projector file for LLaVA.
 - `--grp-attn-n`: Set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`
 - `--grp-attn-w`: Set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`
-- `-n N, --n-predict N`: Set the maximum tokens to predict (default: -1)
+- `-n, --n-predict`: Set the maximum tokens to predict (default: -1)
 - `--slots-endpoint-disable`: To disable slots state monitoring endpoint. Slots state may contain user data, prompts included.
-- `--metrics`: enable prometheus `/metrics` compatible endpoint (default: disabled)
-- `--chat-template JINJA_TEMPLATE`: Set custom jinja chat template. This parameter accepts a string, not a file name (default: template taken from model's metadata). We only support [some pre-defined templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
-- `--log-disable`: Output logs to stdout only, default: enabled.
-- `--log-format FORMAT`: Define the log output to FORMAT: json or text (default: json)
-
-**If compiled with `LLAMA_SERVER_SSL=ON`**
-- `--ssl-key-file FNAME`: path to file a PEM-encoded SSL private key
-- `--ssl-cert-file FNAME`: path to file a PEM-encoded SSL certificate
 
 ## Build
 
@@ -79,28 +58,6 @@ server is build alongside everything else from the root of the project
   cmake --build . --config Release
   ```
 
-## Build with SSL
-
-server can also be built with SSL support using OpenSSL 3
-
-- Using `make`:
-
-  ```bash
-  # NOTE: For non-system openssl, use the following:
-  #   CXXFLAGS="-I /path/to/openssl/include"
-  #   LDFLAGS="-L /path/to/openssl/lib"
-  make LLAMA_SERVER_SSL=true server
-  ```
-
-- Using `CMake`:
-
-  ```bash
-  mkdir build
-  cd build
-  cmake .. -DLLAMA_SERVER_SSL=ON
-  make server
-  ```
-
 ## Quick Start
 
 To get started right away, run the following command, making sure to use the correct path for the model you have:
@@ -140,12 +97,6 @@ curl --request POST \
     --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
 ```
 
-## Advanced testing
-
-We implemented a [server test framework](./tests/README.md) using human-readable scenario.
-
-*Before submitting an issue, please try to reproduce it with this format.*
-
 ## Node JS Test
 
 You need to have [Node.js](https://nodejs.org/en) installed.
@@ -183,27 +134,20 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slot are currently available.
-
-  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+  - `{"status": "loading model"}` if the model is still being loaded.
+  - `{"status": "error"}` if the model failed to load.
+  - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
+  - `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slot are currently available
 
 - **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
     *Options:*
 
-    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, if `cache_prompt` is `true`, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. A `BOS` token is inserted at the start, if all of the following conditions are true:
-
-      - The prompt is a string or an array with the first element given as a string
-      - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
-      - The system prompt is empty
+    `prompt`: Provide the prompt for this completion as a string or as an array of strings or numbers representing tokens. Internally, the prompt is compared to the previous completion and only the "unseen" suffix is evaluated. If the prompt is a string or an array with the first element given as a string, a `bos` token is inserted in the front like `main` does.
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range. The final temperature will be in the range of `[temperature - dynatemp_range; temperature + dynatemp_range]` (default: 0.0, 0.0 = disabled).
+    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
 
     `dynatemp_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -261,7 +205,7 @@ node index.js
 
     `slot_id`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot (default: -1)
 
-    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (default: false)
+    `cache_prompt`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
@@ -294,7 +238,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
 - `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 - `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
-- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`. These options may differ from the original ones in some way (e.g. bad values filtered out, strings converted to tokens, etc.).
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
 - `model`: The path to the model loaded with `-m`
 - `prompt`: The provided `prompt`
 - `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
@@ -312,7 +256,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `content`: Set the text to tokenize.
 
-    Note that a special `BOS` token is never inserted.
+    Note that the special `BOS` token is not added in front of the text and also a space character is not inserted automatically as it is for `/completion`.
 
 - **POST** `/detokenize`: Convert tokens to text.
 
@@ -356,7 +300,7 @@ Notice that each `probs` is an array of length `n_probs`.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 
-- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint.
+- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only ChatML-tuned models, such as Dolphin, OpenOrca, OpenHermes, OpenChat-3.5, etc can be used with this endpoint. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
 
     *Options:*
 
@@ -466,7 +410,7 @@ Notice that each `probs` is an array of length `n_probs`.
         "next_token": {
             "has_next_token": true,
             "n_remain": -1,
-            "n_decoded": 0,
+            "num_tokens_predicted": 0,
             "stopped_eos": false,
             "stopped_limit": false,
             "stopped_word": false,
@@ -503,18 +447,6 @@ Notice that each `probs` is an array of length `n_probs`.
 ]
 ```
 
-- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
-
-Available metrics:
-- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
-- `llamacpp:tokens_predicted_total`: Number of generation tokens processed.
-- `llamacpp:prompt_tokens_seconds`: Average prompt throughput in tokens/s.
-- `llamacpp:predicted_tokens_seconds`: Average generation throughput in tokens/s.
-- `llamacpp:kv_cache_usage_ratio`: KV-cache usage. 1 means 100 percent usage.
-- `llamacpp:kv_cache_tokens`: KV-cache tokens.
-- `llamacpp:requests_processing`: Number of request processing.
-- `llamacpp:requests_deferred`: Number of request deferred.
-
 ## More examples
 
 ### Change system prompt on runtime
@@ -558,11 +490,24 @@ bash chat.sh
 
 ### API like OAI
 
-The HTTP server supports OAI-like API
+API example using Python Flask: [api_like_OAI.py](api_like_OAI.py)
+This example must be used with server.cpp
+
+```sh
+python api_like_OAI.py
+```
+
+After running the API server, you can use it in Python by setting the API base URL.
+
+```python
+openai.api_base = "http://<Your api-server IP>:port"
+```
+
+Then you can utilize llama.cpp as an OpenAI's **chat.completion** or **text_completion** API
 
 ### Extending or building alternative Web Front End
 
-You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.
+The default location for the static files is `examples/server/public`. You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.
 
 Read the documentation in `/completion.js` to see convenient ways to access llama.
 

examples/server/api_like_OAI.py

@@ -0,0 +1,228 @@
+#!/usr/bin/env python3
+import argparse
+from flask import Flask, jsonify, request, Response
+import urllib.parse
+import requests
+import time
+import json
+
+
+app = Flask(__name__)
+slot_id = -1
+
+parser = argparse.ArgumentParser(description="An example of using server.cpp with a similar API to OAI. It must be used together with server.cpp.")
+parser.add_argument("--chat-prompt", type=str, help="the top prompt in chat completions(default: 'A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')", default='A chat between a curious user and an artificial intelligence assistant. The assistant follows the given rules no matter what.')
+parser.add_argument("--user-name", type=str, help="USER name in chat completions(default: 'USER: ')", default="USER: ")
+parser.add_argument("--ai-name", type=str, help="ASSISTANT name in chat completions(default: 'ASSISTANT: ')", default="ASSISTANT: ")
+parser.add_argument("--system-name", type=str, help="SYSTEM name in chat completions(default: 'ASSISTANT's RULE: ')", default="ASSISTANT's RULE: ")
+parser.add_argument("--stop", type=str, help="the end of response in chat completions(default: '</s>')", default="</s>")
+parser.add_argument("--llama-api", type=str, help="Set the address of server.cpp in llama.cpp(default: http://127.0.0.1:8080)", default='http://127.0.0.1:8080')
+parser.add_argument("--api-key", type=str, help="Set the api key to allow only few user(default: NULL)", default="")
+parser.add_argument("--host", type=str, help="Set the ip address to listen.(default: 127.0.0.1)", default='127.0.0.1')
+parser.add_argument("--port", type=int, help="Set the port to listen.(default: 8081)", default=8081)
+
+args = parser.parse_args()
+
+def is_present(json, key):
+    try:
+        buf = json[key]
+    except KeyError:
+        return False
+    if json[key] == None:
+        return False
+    return True
+
+#convert chat to prompt
+def convert_chat(messages):
+
+    system_n = args.system_name
+    user_n = args.user_name
+    ai_n = args.ai_name
+    stop = args.stop
+
+    prompt = "" + args.chat_prompt + stop
+
+    for line in messages:
+        if (line["role"] == "system"):
+            prompt += f"{system_n}{line['content']}{stop}"
+        if (line["role"] == "user"):
+            prompt += f"{user_n}{line['content']}{stop}"
+        if (line["role"] == "assistant"):
+            prompt += f"{ai_n}{line['content']}{stop}"
+    prompt += ai_n.rstrip()
+
+    return prompt
+
+def make_postData(body, chat=False, stream=False):
+    postData = {}
+    if (chat):
+        postData["prompt"] = convert_chat(body["messages"])
+    else:
+        postData["prompt"] = body["prompt"]
+    if(is_present(body, "temperature")): postData["temperature"] = body["temperature"]
+    if(is_present(body, "top_k")): postData["top_k"] = body["top_k"]
+    if(is_present(body, "top_p")): postData["top_p"] = body["top_p"]
+    if(is_present(body, "max_tokens")): postData["n_predict"] = body["max_tokens"]
+    if(is_present(body, "presence_penalty")): postData["presence_penalty"] = body["presence_penalty"]
+    if(is_present(body, "frequency_penalty")): postData["frequency_penalty"] = body["frequency_penalty"]
+    if(is_present(body, "repeat_penalty")): postData["repeat_penalty"] = body["repeat_penalty"]
+    if(is_present(body, "mirostat")): postData["mirostat"] = body["mirostat"]
+    if(is_present(body, "mirostat_tau")): postData["mirostat_tau"] = body["mirostat_tau"]
+    if(is_present(body, "mirostat_eta")): postData["mirostat_eta"] = body["mirostat_eta"]
+    if(is_present(body, "seed")): postData["seed"] = body["seed"]
+    if(is_present(body, "grammar")): postData["grammar"] = body["grammar"]
+    if(is_present(body, "logit_bias")): postData["logit_bias"] = [[int(token), body["logit_bias"][token]] for token in body["logit_bias"].keys()]
+    if (args.stop != ""):
+        postData["stop"] = [args.stop]
+    else:
+        postData["stop"] = []
+    if(is_present(body, "stop")): postData["stop"] += body["stop"]
+    postData["n_keep"] = -1
+    postData["stream"] = stream
+    postData["cache_prompt"] = True
+    postData["slot_id"] = slot_id
+    return postData
+
+def make_resData(data, chat=False, promptToken=[]):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion" if (chat) else "text_completion",
+        "created": int(time.time()),
+        "truncated": data["truncated"],
+        "model": "LLaMA_CPP",
+        "usage": {
+            "prompt_tokens": data["tokens_evaluated"],
+            "completion_tokens": data["tokens_predicted"],
+            "total_tokens": data["tokens_evaluated"] + data["tokens_predicted"]
+        }
+    }
+    if (len(promptToken) != 0):
+        resData["promptToken"] = promptToken
+    if (chat):
+        #only one choice is supported
+        resData["choices"] = [{
+            "index": 0,
+            "message": {
+                "role": "assistant",
+                "content": data["content"],
+            },
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    else:
+        #only one choice is supported
+        resData["choices"] = [{
+            "text": data["content"],
+            "index": 0,
+            "logprobs": None,
+            "finish_reason": "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+        }]
+    return resData
+
+def make_resData_stream(data, chat=False, time_now = 0, start=False):
+    resData = {
+        "id": "chatcmpl" if (chat) else "cmpl",
+        "object": "chat.completion.chunk" if (chat) else "text_completion.chunk",
+        "created": time_now,
+        "model": "LLaMA_CPP",
+        "choices": [
+            {
+                "finish_reason": None,
+                "index": 0
+            }
+        ]
+    }
+    slot_id = data.get("slot_id")
+    if (chat):
+        if (start):
+            resData["choices"][0]["delta"] =  {
+                "role": "assistant"
+            }
+        else:
+            resData["choices"][0]["delta"] =  {
+                "content": data["content"]
+            }
+            if (data["stop"]):
+                resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+    else:
+        resData["choices"][0]["text"] = data["content"]
+        if (data["stop"]):
+            resData["choices"][0]["finish_reason"] = "stop" if (data["stopped_eos"] or data["stopped_word"]) else "length"
+
+    return resData
+
+
+@app.route('/chat/completions', methods=['POST', 'OPTIONS'])
+@app.route('/v1/chat/completions', methods=['POST', 'OPTIONS'])
+def chat_completions():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    if request.method == 'OPTIONS':
+        return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=True, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=True, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            resData = make_resData_stream({}, chat=True, time_now=time_now, start=True)
+            yield 'data: {}\n\n'.format(json.dumps(resData))
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=True, time_now=time_now)
+                    yield 'data: {}\n\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+
+
+@app.route('/completions', methods=['POST', 'OPTIONS'])
+@app.route('/v1/completions', methods=['POST', 'OPTIONS'])
+def completion():
+    if (args.api_key != "" and request.headers["Authorization"].split()[1] != args.api_key):
+        return Response(status=403)
+    if request.method == 'OPTIONS':
+        return Response(headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+    body = request.get_json()
+    stream = False
+    tokenize = False
+    if(is_present(body, "stream")): stream = body["stream"]
+    if(is_present(body, "tokenize")): tokenize = body["tokenize"]
+    postData = make_postData(body, chat=False, stream=stream)
+
+    promptToken = []
+    if (tokenize):
+        tokenData = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/tokenize"), data=json.dumps({"content": postData["prompt"]})).json()
+        promptToken = tokenData["tokens"]
+
+    if (not stream):
+        data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData))
+        print(data.json())
+        resData = make_resData(data.json(), chat=False, promptToken=promptToken)
+        return jsonify(resData)
+    else:
+        def generate():
+            data = requests.request("POST", urllib.parse.urljoin(args.llama_api, "/completion"), data=json.dumps(postData), stream=True)
+            time_now = int(time.time())
+            for line in data.iter_lines():
+                if line:
+                    decoded_line = line.decode('utf-8')
+                    resData = make_resData_stream(json.loads(decoded_line[6:]), chat=False, time_now=time_now)
+                    yield 'data: {}\n\n'.format(json.dumps(resData))
+        return Response(generate(), mimetype='text/event-stream', headers={"Access-Control-Allow-Origin": "*", "Access-Control-Allow-Headers": "*"})
+
+if __name__ == '__main__':
+    app.run(args.host, port=args.port)

examples/server/bench/README.md

@@ -1,88 +0,0 @@
-### Server benchmark tools
-
-Benchmark is using [k6](https://k6.io/).
-
-##### Install k6
-
-Follow instruction from: https://k6.io/docs/get-started/installation/
-
-Example for ubuntu:
-```shell
-snap install k6
-```
-
-#### Download a dataset
-
-This dataset was originally proposed in [vLLM benchmarks](https://github.com/vllm-project/vllm/blob/main/benchmarks/README.md).
-
-```shell
-wget https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
-```
-
-#### Download a model
-Example for PHI-2
-
-```shell
-../../../scripts/hf.sh --repo ggml-org/models --file phi-2/ggml-model-q4_0.gguf
-```
-
-#### Start the server
-The server must answer OAI Chat completion requests on `http://localhost:8080/v1` or according to the environment variable `SERVER_BENCH_URL`.
-
-Example:
-```shell
-server --host localhost --port 8080 \
-  --model ggml-model-q4_0.gguf \
-  --cont-batching \
-  --metrics \
-  --parallel 8 \
-  --batch-size 512 \
-  --ctx-size 4096 \
-  --log-format text \
-  -ngl 33
-```
-
-#### Run the benchmark
-
-For 500 chat completions request with 8 concurrent users during maximum 10 minutes, run:
-```shell
-k6 run script.js --duration 10m --iterations 500 --vus 8
-```
-
-The benchmark values can be overridden with:
-- `SERVER_BENCH_URL` server url prefix for chat completions, default `http://localhost:8080/v1`
-- `SERVER_BENCH_N_PROMPTS` total prompts to randomly select in the benchmark, default `480`
-- `SERVER_BENCH_MODEL_ALIAS` model alias to pass in the completion request, default `my-model`
-- `SERVER_BENCH_MAX_TOKENS` max tokens to predict, default: `512`
-- `SERVER_BENCH_DATASET` path to the benchmark dataset file
-- `SERVER_BENCH_MAX_PROMPT_TOKENS` maximum prompt tokens to filter out in the dataset: default `1024`
-- `SERVER_BENCH_MAX_CONTEXT` maximum context size of the completions request to filter out in the dataset: prompt + predicted tokens, default `2048`
-
-Note: the local tokenizer is just a string space split, real number of tokens will differ.
-
-Or with [k6 options](https://k6.io/docs/using-k6/k6-options/reference/):
-
-```shell
-SERVER_BENCH_N_PROMPTS=500 k6 run script.js --duration 10m --iterations 500 --vus 8
-```
-
-To [debug http request](https://k6.io/docs/using-k6/http-debugging/) use `--http-debug="full"`.
-
-#### Metrics
-
-Following metrics are available computed from the OAI chat completions response `usage`:
-- `llamacpp_tokens_second` Trend of `usage.total_tokens / request duration`
-- `llamacpp_prompt_tokens` Trend of `usage.prompt_tokens`
-- `llamacpp_prompt_tokens_total_counter` Counter of `usage.prompt_tokens`
-- `llamacpp_completion_tokens` Trend of `usage.completion_tokens`
-- `llamacpp_completion_tokens_total_counter` Counter of `usage.completion_tokens`
-- `llamacpp_completions_truncated_rate` Rate of completions truncated, i.e. if `finish_reason === 'length'`
-- `llamacpp_completions_stop_rate` Rate of completions stopped by the model, i.e. if `finish_reason === 'stop'`
-
-The script will fail if too many completions are truncated, see `llamacpp_completions_truncated_rate`.
-
-K6 metrics might be compared against [server metrics](../README.md), with:
-
-```shell
-curl http://localhost:8080/metrics
-```

examples/server/bench/script.js

@@ -1,120 +0,0 @@
-import http from 'k6/http'
-import {check, sleep} from 'k6'
-import {SharedArray} from 'k6/data'
-import {Counter, Rate, Trend} from 'k6/metrics'
-import exec from 'k6/execution';
-
-// Server chat completions prefix
-const server_url = __ENV.SERVER_BENCH_URL ? __ENV.SERVER_BENCH_URL : 'http://localhost:8080/v1'
-
-// Number of total prompts in the dataset - default 10m / 10 seconds/request * number of users
-const n_prompt = __ENV.SERVER_BENCH_N_PROMPTS ? parseInt(__ENV.SERVER_BENCH_N_PROMPTS) : 600 / 10 * 8
-
-// Model name to request
-const model = __ENV.SERVER_BENCH_MODEL_ALIAS ? __ENV.SERVER_BENCH_MODEL_ALIAS : 'my-model'
-
-// Dataset path
-const dataset_path = __ENV.SERVER_BENCH_DATASET ? __ENV.SERVER_BENCH_DATASET : './ShareGPT_V3_unfiltered_cleaned_split.json'
-
-// Max tokens to predict
-const max_tokens = __ENV.SERVER_BENCH_MAX_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_TOKENS) : 512
-
-// Max prompt tokens
-const n_prompt_tokens = __ENV.SERVER_BENCH_MAX_PROMPT_TOKENS ? parseInt(__ENV.SERVER_BENCH_MAX_PROMPT_TOKENS) : 1024
-
-// Max slot context
-const n_ctx_slot = __ENV.SERVER_BENCH_MAX_CONTEXT ? parseInt(__ENV.SERVER_BENCH_MAX_CONTEXT) : 2048
-
-export function setup() {
-    console.info(`Benchmark config: server_url=${server_url} n_prompt=${n_prompt} model=${model} dataset_path=${dataset_path} max_tokens=${max_tokens}`)
-}
-
-const data = new SharedArray('conversations', function () {
-    const tokenizer = (message) => message.split(/[\s,'".?]/)
-
-    return JSON.parse(open(dataset_path))
-        // Filter out the conversations with less than 2 turns.
-        .filter(data => data["conversations"].length >= 2)
-        .filter(data => data["conversations"][0]["from"] === "human")
-        .map(data => {
-            return {
-                prompt: data["conversations"][0]["value"],
-                n_prompt_tokens: tokenizer(data["conversations"][0]["value"]).length,
-                n_completion_tokens: tokenizer(data["conversations"][1]["value"]).length,
-            }
-        })
-        // Filter out too short sequences
-        .filter(conv => conv.n_prompt_tokens >= 4 && conv.n_completion_tokens >= 4)
-        // Filter out too long sequences.
-        .filter(conv => conv.n_prompt_tokens <= n_prompt_tokens && conv.n_prompt_tokens + conv.n_completion_tokens <= n_ctx_slot)
-        // Keep only first n prompts
-        .slice(0, n_prompt)
-})
-
-const llamacpp_prompt_tokens = new Trend('llamacpp_prompt_tokens')
-const llamacpp_completion_tokens = new Trend('llamacpp_completion_tokens')
-const llamacpp_tokens_second = new Trend('llamacpp_tokens_second')
-
-const llamacpp_prompt_tokens_total_counter = new Counter('llamacpp_prompt_tokens_total_counter')
-const llamacpp_completion_tokens_total_counter = new Counter('llamacpp_completion_tokens_total_counter')
-
-const llamacpp_completions_truncated_rate = new Rate('llamacpp_completions_truncated_rate')
-const llamacpp_completions_stop_rate = new Rate('llamacpp_completions_stop_rate')
-
-export const options = {
-    thresholds: {
-        llamacpp_completions_truncated_rate: [
-            // more than 80% of truncated input will abort the test
-            {threshold: 'rate < 0.8', abortOnFail: true, delayAbortEval: '1m'},
-        ],
-    },
-    duration: '10m',
-    vus: 8,
-}
-
-export default function () {
-    const conversation = data[exec.scenario.iterationInInstance % data.length]
-    const payload = {
-        "messages": [
-            {
-                "role": "system",
-                "content": "You are ChatGPT, an AI assistant.",
-            },
-            {
-                "role": "user",
-                "content": conversation.prompt,
-            }
-        ],
-        "model": model,
-        "stream": false,
-        "max_tokens": max_tokens
-    }
-
-    const body = JSON.stringify(payload)
-
-    let res = http.post(`${server_url}/chat/completions`, body, {
-        headers: {'Content-Type': 'application/json'},
-        timeout: '300s'
-    })
-
-    check(res, {'success completion': (r) => r.status === 200})
-
-    if (res.status === 200) {
-        const completions = res.json()
-
-        llamacpp_prompt_tokens.add(completions.usage.prompt_tokens)
-        llamacpp_prompt_tokens_total_counter.add(completions.usage.prompt_tokens)
-
-        llamacpp_completion_tokens.add(completions.usage.completion_tokens)
-        llamacpp_completion_tokens_total_counter.add(completions.usage.completion_tokens)
-
-        llamacpp_completions_truncated_rate.add(completions.choices[0].finish_reason === 'length')
-        llamacpp_completions_stop_rate.add(completions.choices[0].finish_reason === 'stop')
-
-        llamacpp_tokens_second.add(completions.usage.total_tokens / res.timings.duration * 1.e3)
-    } else {
-        console.error(`response: ${res.body} request=${payload}`)
-    }
-
-    sleep(0.3)
-}

examples/server/oai.hpp

@@ -0,0 +1,227 @@
+#pragma once
+
+#include <string>
+#include <vector>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
+
+#include "json.hpp"
+#include "utils.hpp"
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+
+using json = nlohmann::json;
+
+inline static json oaicompat_completion_params_parse(
+    const json &body, /* openai api json semantics */
+    const std::string &chat_template)
+{
+    json llama_params;
+    std::string formatted_prompt = chat_template == "chatml"
+        ? format_chatml(body["messages"])  // OpenAI 'messages' to chatml (with <|im_start|>,...)
+        : format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
+
+    llama_params["__oaicompat"] = true;
+
+    // Map OpenAI parameters to llama.cpp parameters
+    //
+    // For parameters that are defined by the OpenAI documentation (e.g.
+    // temperature), we explicitly specify OpenAI's intended default; we
+    // need to do that because sometimes OpenAI disagrees with llama.cpp
+    //
+    // https://platform.openai.com/docs/api-reference/chat/create
+    llama_sampling_params default_sparams;
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
+    llama_params["prompt"]            = formatted_prompt;
+    llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
+    llama_params["temperature"]       = json_value(body, "temperature", 0.0);
+    llama_params["top_k"]             = json_value(body, "top_k", default_sparams.top_k);
+    llama_params["top_p"]             = json_value(body, "top_p", 1.0);
+    llama_params["n_predict"]         = json_value(body, "max_tokens", -1);
+    llama_params["logit_bias"]        = json_value(body, "logit_bias",json::object());
+    llama_params["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
+    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["mirostat"]          = json_value(body, "mirostat", default_sparams.mirostat);
+    llama_params["mirostat_tau"]      = json_value(body, "mirostat_tau", default_sparams.mirostat_tau);
+    llama_params["mirostat_eta"]      = json_value(body, "mirostat_eta", default_sparams.mirostat_eta);
+    llama_params["penalize_nl"]       = json_value(body, "penalize_nl", default_sparams.penalize_nl);
+    llama_params["typical_p"]         = json_value(body, "typical_p", default_sparams.typical_p);
+    llama_params["repeat_last_n"]     = json_value(body, "repeat_last_n", default_sparams.penalty_last_n);
+    llama_params["ignore_eos"]        = json_value(body, "ignore_eos", false);
+    llama_params["tfs_z"]             = json_value(body, "tfs_z", default_sparams.tfs_z);
+
+    if (body.count("grammar") != 0) {
+        llama_params["grammar"] = json_value(body, "grammar", json::object());
+    }
+
+    // Handle 'stop' field
+    if (body.contains("stop") && body["stop"].is_string()) {
+        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
+    } else {
+        llama_params["stop"] = json_value(body, "stop", json::array());
+    }
+
+    // Ensure there is ChatML-specific end sequence among stop words
+    llama_params["stop"].push_back("<|im_end|>");
+
+    return llama_params;
+}
+
+inline static json format_final_response_oaicompat(const json &request, const task_result &response, bool streaming = false)
+{
+    json result = response.result_json;
+
+    bool stopped_word        = result.count("stopped_word") != 0;
+    bool stopped_eos         = json_value(result, "stopped_eos", false);
+    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
+    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
+    std::string content      = json_value(result, "content", std::string(""));
+
+    std::string finish_reason = "length";
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+
+    json choices =
+        streaming ? json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"delta", json::object()}}})
+                  : json::array({json{{"finish_reason", finish_reason},
+                                        {"index", 0},
+                                        {"message", json{{"content", content},
+                                                         {"role", "assistant"}}}}});
+
+    std::time_t t = std::time(0);
+
+    json res =
+        json{{"choices", choices},
+            {"created", t},
+            {"model",
+                json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
+            {"usage",
+                json{{"completion_tokens", num_tokens_predicted},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
+            {"id", gen_chatcmplid()}};
+
+    if (server_verbose) {
+        res["__verbose"] = result;
+    }
+
+    if (result.contains("completion_probabilities")) {
+        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
+    }
+
+    return res;
+}
+
+// return value is vector as there is one case where we might need to generate two responses
+inline static std::vector<json> format_partial_response_oaicompat(const task_result &response) {
+    json result = response.result_json;
+
+    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
+        return std::vector<json>({response.result_json});
+    }
+
+    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
+    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
+
+    bool stopped_word   = json_value(result, "stopped_word", false);
+    bool stopped_eos    = json_value(result, "stopped_eos", false);
+    bool stopped_limit  = json_value(result, "stopped_limit", false);
+    std::string content = json_value(result, "content", std::string(""));
+
+    std::string finish_reason;
+    if (stopped_word || stopped_eos) {
+        finish_reason = "stop";
+    }
+    if (stopped_limit) {
+        finish_reason = "length";
+    }
+
+    std::time_t t = std::time(0);
+
+    json choices;
+
+    if (!finish_reason.empty()) {
+        choices = json::array({json{{"finish_reason", finish_reason},
+                                    {"index", 0},
+                                    {"delta", json::object()}}});
+    } else {
+        if (first) {
+            if (content.empty()) {
+                choices = json::array({json{{"finish_reason", nullptr},
+                                            {"index", 0},
+                                            {"delta", json{{"role", "assistant"}}}}});
+            } else {
+                // We have to send this as two updates to conform to openai behavior
+                json initial_ret = json{{"choices", json::array({json{
+                                        {"finish_reason", nullptr},
+                                        {"index", 0},
+                                        {"delta", json{
+                                            {"role", "assistant"}
+                                        }}}})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                json second_ret = json{
+                            {"choices", json::array({json{{"finish_reason", nullptr},
+                                                            {"index", 0},
+                                                            {"delta", json{
+                                                            {"content", content}}}
+                                                            }})},
+                            {"created", t},
+                            {"id", gen_chatcmplid()},
+                            {"model", modelname},
+                            {"object", "chat.completion.chunk"}};
+
+                return std::vector<json>({initial_ret, second_ret});
+            }
+        } else {
+            // Some idiosyncrasy in task processing logic makes several trailing calls
+            // with empty content, we ignore these at the calee site.
+            if (content.empty()) {
+                return std::vector<json>({json::object()});
+            }
+
+            choices = json::array({json{
+                {"finish_reason", nullptr},
+                {"index", 0},
+                {"delta",
+                json{
+                    {"content", content},
+                }},
+            }});
+        }
+    }
+
+    json ret = json{{"choices", choices},
+                    {"created", t},
+                    {"id", gen_chatcmplid()},
+                    {"model", modelname},
+                    {"object", "chat.completion.chunk"}};
+
+    return std::vector<json>({ret});
+}
+
+inline static json format_embeddings_response_oaicompat(const json &request, const json &embeddings)
+{
+    json res =
+        json{
+            {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
+            {"object", "list"},
+            {"usage",
+                json{{"prompt_tokens", 0},
+                     {"total_tokens", 0}}},
+            {"data", embeddings}
+        };
+    return res;
+}
+

examples/server/tests/README.md

@@ -1,67 +0,0 @@
-# Server tests
-
-Python based server tests scenario using [BDD](https://en.wikipedia.org/wiki/Behavior-driven_development)
-and [behave](https://behave.readthedocs.io/en/latest/):
-
-* [issues.feature](./features/issues.feature) Pending issues scenario
-* [parallel.feature](./features/parallel.feature) Scenario involving multi slots and concurrent requests
-* [security.feature](./features/security.feature) Security, CORS and API Key
-* [server.feature](./features/server.feature) Server base scenario: completion, embedding, tokenization, etc...
-
-Tests target GitHub workflows job runners with 4 vCPU.
-
-Requests are
-using [aiohttp](https://docs.aiohttp.org/en/stable/client_reference.html), [asyncio](https://docs.python.org/fr/3/library/asyncio.html)
-based http client.
-
-Note: If the host architecture inference speed is faster than GitHub runners one, parallel scenario may randomly fail.
-To mitigate it, you can increase values in `n_predict`, `kv_size`.
-
-### Install dependencies
-
-`pip install -r requirements.txt`
-
-### Run tests
-
-1. Build the server
-
-```shell
-cd ../../..
-mkdir build
-cd build
-cmake ../
-cmake --build . --target server
-```
-
-2. Start the test: `./tests.sh`
-
-It's possible to override some scenario steps values with environment variables:
-
-| variable                 | description                                                                                    |
-|--------------------------|------------------------------------------------------------------------------------------------|
-| `PORT`                   | `context.server_port` to set the listening port of the server during scenario, default: `8080` |
-| `LLAMA_SERVER_BIN_PATH`  | to change the server binary path, default: `../../../build/bin/server`                         |
-| `DEBUG`                  | "ON" to enable steps and server verbose mode `--verbose`                                       |
-| `SERVER_LOG_FORMAT_JSON` | if set switch server logs to json format                                                       |
-| `N_GPU_LAYERS`           | number of model layers to offload to VRAM `-ngl --n-gpu-layers`                                |
-
-### Run @bug, @wip or @wrong_usage annotated scenario
-
-Feature or Scenario must be annotated with `@llama.cpp` to be included in the default scope.
-
-- `@bug` annotation aims to link a scenario with a GitHub issue.
-- `@wrong_usage` are meant to show user issue that are actually an expected behavior
-- `@wip` to focus on a scenario working in progress
-- `@slow` heavy test, disabled by default
-
-To run a scenario annotated with `@bug`, start:
-
-```shell
-DEBUG=ON ./tests.sh --no-skipped --tags bug
-```
-
-After changing logic in `steps.py`, ensure that `@bug` and `@wrong_usage` scenario are updated.
-
-```shell
-./tests.sh --no-skipped --tags bug,wrong_usage || echo "should failed but compile"
-```

examples/server/tests/features/embeddings.feature

@@ -1,94 +0,0 @@
-@llama.cpp
-@embeddings
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file bert-bge-small/ggml-model-f16.gguf from HF repo ggml-org/models
-    And   a model alias bert-bge-small
-    And   42 as server seed
-    And   2 slots
-    And   1024 as batch size
-    And   2048 KV cache size
-    And   embeddings extraction
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Embedding
-    When embeddings are computed for:
-    """
-    What is the capital of Bulgaria ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility
-    Given a model bert-bge-small
-    When an OAI compatible embeddings computation request for:
-    """
-    What is the capital of Spain ?
-    """
-    Then embeddings are generated
-
-  Scenario: OAI Embeddings compatibility with multiple inputs
-    Given a model bert-bge-small
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    When an OAI compatible embeddings computation request for multiple inputs
-    Then embeddings are generated
-
-  Scenario: Multi users embeddings
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    Given concurrent embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: Multi users OAI compatibility embeddings
-    Given a prompt:
-      """
-      In which country Paris is located ?
-      """
-    And a prompt:
-      """
-      Is Madrid the capital of Spain ?
-      """
-    And a prompt:
-      """
-      What is the biggest US city ?
-      """
-    And a prompt:
-      """
-      What is the capital of Bulgaria ?
-      """
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then the server is busy
-    Then the server is idle
-    Then all embeddings are generated
-
-  Scenario: All embeddings should be the same
-    Given 10 fixed prompts
-    And   a model bert-bge-small
-    Given concurrent OAI embedding requests
-    Then all embeddings are the same

examples/server/tests/features/environment.py

@@ -1,100 +0,0 @@
-import errno
-import os
-import socket
-import subprocess
-import time
-from contextlib import closing
-import signal
-
-
-def before_scenario(context, scenario):
-    context.debug = 'DEBUG' in os.environ and os.environ['DEBUG'] == 'ON'
-    if context.debug:
-        print("DEBUG=ON\n")
-    print(f"\x1b[33;42mStarting new scenario: {scenario.name}!\x1b[0m\n")
-    port = 8080
-    if 'PORT' in os.environ:
-        port = int(os.environ['PORT'])
-    if is_server_listening("localhost", port):
-        assert False, "Server already started"
-
-
-def after_scenario(context, scenario):
-    if context.server_process is None:
-        return
-    if scenario.status == "failed":
-        if 'GITHUB_ACTIONS' in os.environ:
-            print(f"\x1b[33;101mSCENARIO FAILED: {scenario.name} server logs:\x1b[0m\n\n")
-            if os.path.isfile('llama.log'):
-                with closing(open('llama.log', 'r')) as f:
-                    for line in f:
-                        print(line)
-        if not is_server_listening(context.server_fqdn, context.server_port):
-            print("\x1b[33;101mERROR: Server stopped listening\x1b[0m\n")
-
-    if not pid_exists(context.server_process.pid):
-        assert False, f"Server not running pid={context.server_process.pid} ..."
-
-    server_graceful_shutdown(context)
-
-    # Wait few for socket to free up
-    time.sleep(0.05)
-
-    attempts = 0
-    while pid_exists(context.server_process.pid) or is_server_listening(context.server_fqdn, context.server_port):
-        server_kill(context)
-        time.sleep(0.1)
-        attempts += 1
-        if attempts > 5:
-            server_kill_hard(context)
-
-
-def server_graceful_shutdown(context):
-    print(f"shutting down server pid={context.server_process.pid} ...\n")
-    if os.name == 'nt':
-        os.kill(context.server_process.pid, signal.CTRL_C_EVENT)
-    else:
-        os.kill(context.server_process.pid, signal.SIGINT)
-
-
-def server_kill(context):
-    print(f"killing server pid={context.server_process.pid} ...\n")
-    context.server_process.kill()
-
-
-def server_kill_hard(context):
-    pid = context.server_process.pid
-    path = context.server_path
-
-    print(f"Server dangling exits, hard killing force {pid}={path}...\n")
-    if os.name == 'nt':
-        process = subprocess.check_output(['taskkill', '/F', '/pid', str(pid)]).decode()
-        print(process)
-    else:
-        os.kill(-pid, signal.SIGKILL)
-
-
-def is_server_listening(server_fqdn, server_port):
-    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-        result = sock.connect_ex((server_fqdn, server_port))
-        _is_server_listening = result == 0
-        if _is_server_listening:
-            print(f"server is listening on {server_fqdn}:{server_port}...\n")
-        return _is_server_listening
-
-
-def pid_exists(pid):
-    """Check whether pid exists in the current process table."""
-    if pid < 0:
-        return False
-    if os.name == 'nt':
-        output = subprocess.check_output(['TASKLIST', '/FI', f'pid eq {pid}']).decode()
-        print(output)
-        return "No tasks are running" not in output
-    else:
-        try:
-            os.kill(pid, 0)
-        except OSError as e:
-            return e.errno == errno.EPERM
-        else:
-            return True

examples/server/tests/features/issues.feature

@@ -1,5 +0,0 @@
-# List of ongoing issues
-# run with: DEBUG=ON ./tests.sh --no-skipped --tags bug
-@bug
-Feature: Issues
-  # No confirmed issue at the moment

examples/server/tests/features/parallel.feature

@@ -1,101 +0,0 @@
-@llama.cpp
-@parallel
-Feature: Parallel
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   42 as server seed
-    And   128 as batch size
-    And   256 KV cache size
-    And   2 slots
-    And   continuous batching
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Multi users completion
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And <n_predict> max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    And  all slots are idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | n_predict |
-      | 128       |
-
-  Scenario Outline: Multi users OAI completions compatibility
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-  Scenario Outline: Multi users OAI completions compatibility no v1
-    Given a system prompt You are a writer.
-    And   a model tinyllama-2
-    Given a prompt:
-      """
-      Write a very long book.
-      """
-    And a prompt:
-      """
-      Write another a poem.
-      """
-    And <n_predict> max tokens to predict
-    And streaming is <streaming>
-    Given concurrent OAI completions requests no v1
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted with <n_predict> tokens
-    Examples:
-      | streaming | n_predict |
-      | disabled  | 128       |
-      | enabled   | 64        |
-
-
-  Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
-    Given a prompt:
-      """
-      Write a very long story about AI.
-      """
-    And a prompt:
-      """
-      Write another very long music lyrics.
-      """
-    And a prompt:
-      """
-      Write a very long poem.
-      """
-    And a prompt:
-      """
-      Write a very long joke.
-      """
-    And 128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is busy
-    Then the server is idle
-    Then all prompts are predicted

examples/server/tests/features/passkey.feature

@@ -1,55 +0,0 @@
-# run with: ./tests.sh --no-skipped --tags passkey
-@passkey
-@slow
-Feature: Passkey / Self-extend with context shift
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-
-  # Generates a long text of junk and inserts a secret passkey number inside it.
-  # Then we query the LLM for the secret passkey.
-  # see #3856 and #4810
-  Scenario Outline: Passkey
-    Given a model file <hf_file> from HF repo <hf_repo>
-    And   <n_batch> as batch size
-    And   <n_junk> as number of junk
-    And   <n_predicted> server max tokens to predict
-    And   42 as seed
-    And   <n_ctx> KV cache size
-    And   1 slots
-    And   <n_ga> group attention factor to extend context size through self-extend
-    And   <n_ga_w> group attention width to extend context size through self-extend
-    # Can be override with N_GPU_LAYERS
-    And   <ngl> GPU offloaded layers
-    Then  the server is starting
-    Then  the server is healthy
-    Given available models
-    Then  model 0 is trained on <n_ctx_train> tokens context
-    Given a prefix prompt:
-    """
-    here is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.
-    """
-    And a passkey prompt template:
-    """
-    The pass key is <passkey> Remember it. <passkey> is the pass key.
-    """
-    And a junk suffix prompt:
-    """
-    The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.
-    """
-    And a suffix prompt:
-    """
-    What is the pass key? The pass key is
-    """
-    Given a "<passkey>" passkey challenge prompt with the passkey inserted every <i_pos> junk
-    And  a completion request with no api error
-    Then <n_predicted> tokens are predicted matching <re_content>
-
-    Examples:
-      | hf_repo                         | hf_file                     | n_ctx_train | ngl | n_ctx | n_batch | n_ga | n_ga_w | n_junk | i_pos | passkey | n_predicted | re_content     |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 4    | 512    | 250    | 50    | 42      | 1           | 42             |
-      | TheBloke/phi-2-GGUF             | phi-2.Q4_K_M.gguf           | 2048        | 5   | 8192  | 512     | 2    | 512    | 250    | 50    | 42      | 1           | \b((?!42)\w)+\b  |
-      #| TheBloke/Llama-2-7B-GGUF        | llama-2-7b.Q2_K.gguf        | 4096        | 3   | 16384 | 512     | 4    | 512    | 500    | 300   | 1234    | 5           | 1234           |
-      #| TheBloke/Mixtral-8x7B-v0.1-GGUF | mixtral-8x7b-v0.1.Q2_K.gguf | 32768       | 2   | 16384 | 512     | 4    | 512    | 500    | 100   | 0987    | 5           | 0
-      # 987           |
-

examples/server/tests/features/security.feature

@@ -1,52 +0,0 @@
-@llama.cpp
-@security
-Feature: Security
-
-  Background: Server startup with an api key defined
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a server api key llama.cpp
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario Outline: Completion with some user api key
-    Given a prompt test
-    And   a user api key <api_key>
-    And   4 max tokens to predict
-    And   a completion request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackeme   | raised    |
-      |           | raised    |
-
-  Scenario Outline: OAI Compatibility
-    Given a system prompt test
-    And   a user prompt test
-    And   a model test
-    And   2 max tokens to predict
-    And   streaming is disabled
-    And   a user api key <api_key>
-    Given an OAI compatible chat completions request with <api_error> api error
-
-    Examples: Prompts
-      | api_key   | api_error |
-      | llama.cpp | no        |
-      | llama.cpp | no        |
-      | hackme    | raised    |
-
-
-  Scenario Outline: CORS Options
-    Given a user api key llama.cpp
-    When  an OPTIONS request is sent from <origin>
-    Then  CORS header <cors_header> is set to <cors_header_value>
-
-    Examples: Headers
-      | origin          | cors_header                      | cors_header_value |
-      | localhost       | Access-Control-Allow-Origin      | localhost         |
-      | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
-      | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
-      | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |

examples/server/tests/features/server.feature

@@ -1,83 +0,0 @@
-@llama.cpp
-@server
-Feature: llama.cpp server
-
-  Background: Server startup
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    And   a model alias tinyllama-2
-    And   42 as server seed
-      # KV Cache corresponds to the total amount of tokens
-      # that can be stored across all independent sequences: #4130
-      # see --ctx-size and #5568
-    And   256 KV cache size
-    And   32 as batch size
-    And   2 slots
-    And   64 server max tokens to predict
-    And   prometheus compatible metrics exposed
-    Then  the server is starting
-    Then  the server is healthy
-
-  Scenario: Health
-    Then the server is ready
-    And  all slots are idle
-
-
-  Scenario Outline: Completion
-    Given a prompt <prompt>
-    And   <n_predict> max tokens to predict
-    And   a completion request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-    And   the completion is <truncated> truncated
-    And   <n_prompt> prompt tokens are processed
-    And   prometheus metrics are exposed
-    And   metric llamacpp:tokens_predicted is <n_predicted>
-
-    Examples: Prompts
-      | prompt                                                                    | n_predict | re_content                    | n_prompt | n_predicted | truncated |
-      | I believe the meaning of life is                                          | 8         | (read\|going)+                | 18       | 8           | not       |
-      | Write a joke about AI from a very long prompt which will not be truncated | 256       | (princesses\|everyone\|kids)+ | 46       | 64          | not       |
-
-  Scenario: Completion prompt truncated
-    Given a prompt:
-    """
-    Lorem ipsum dolor sit amet, consectetur adipiscing elit, sed do eiusmod tempor incididunt ut labore et dolore magna aliqua.
-    Ut enim ad minim veniam, quis nostrud exercitation ullamco laboris nisi ut aliquip ex ea commodo consequat.
-    Duis aute irure dolor in reprehenderit in voluptate velit esse cillum dolore eu fugiat nulla pariatur.
-    Excepteur sint occaecat cupidatat non proident, sunt in culpa qui officia deserunt mollit anim id est laborum.
-    """
-    And   a completion request with no api error
-    Then  64 tokens are predicted matching fun|Annaks|popcorns|pictry
-    And   the completion is  truncated
-    And   109 prompt tokens are processed
-
-
-  Scenario Outline: OAI Compatibility
-    Given a model <model>
-    And   a system prompt <system_prompt>
-    And   a user prompt <user_prompt>
-    And   <max_tokens> max tokens to predict
-    And   streaming is <enable_streaming>
-    Given an OAI compatible chat completions request with no api error
-    Then  <n_predicted> tokens are predicted matching <re_content>
-    And   <n_prompt> prompt tokens are processed
-    And   the completion is <truncated> truncated
-
-    Examples: Prompts
-      | model        | system_prompt               | user_prompt                          | max_tokens | re_content             | n_prompt | n_predicted | enable_streaming | truncated |
-      | llama-2      | Book                        | What is the best book                | 8          | (Here\|what)+          | 77       | 8           | disabled         | not       |
-      | codellama70b | You are a coding assistant. | Write the fibonacci function in c++. | 128        | (thanks\|happy\|bird)+ | -1       | 64          | enabled          |           |
-
-
-  Scenario: Tokenize / Detokenize
-    When tokenizing:
-    """
-    What is the capital of France ?
-    """
-    Then tokens can be detokenize
-
-  Scenario: Models available
-    Given available models
-    Then  1 models are supported
-    Then  model 0 is identified by tinyllama-2
-    Then  model 0 is trained on 128 tokens context

examples/server/tests/features/wrong_usages.feature

@@ -1,22 +0,0 @@
-# run with: ./tests.sh --no-skipped --tags wrong_usage
-@wrong_usage
-Feature: Wrong usage of llama.cpp server
-
-  #3969 The user must always set --n-predict option
-  # to cap the number of tokens any completion request can generate
-  # or pass n_predict/max_tokens in the request.
-  Scenario: Infinite loop
-    Given a server listening on localhost:8080
-    And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
-    # Uncomment below to fix the issue
-    #And   64 server max tokens to predict
-    Then  the server is starting
-    Given a prompt:
-      """
-      Go to: infinite loop
-      """
-    # Uncomment below to fix the issue
-    #And   128 max tokens to predict
-    Given concurrent completion requests
-    Then the server is idle
-    Then all prompts are predicted

examples/server/tests/requirements.txt

@@ -1,6 +0,0 @@
-aiohttp~=3.9.3
-behave~=1.2.6
-huggingface_hub~=0.20.3
-numpy~=1.24.4
-openai~=0.25.0
-prometheus-client~=0.20.0

examples/server/tests/tests.sh

@@ -1,12 +0,0 @@
-#!/bin/bash
-
-set -eu
-
-if [ $# -lt 1 ]
-then
-  # Start @llama.cpp scenario
-  behave --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
-else
-  behave "$@"
-fi
-

examples/server/utils.hpp

@@ -1,21 +1,19 @@
 #pragma once
 
-#include "llama.h"
-#include "common.h"
-
-#include "json.hpp"
-
 #include <string>
 #include <vector>
-#include <sstream>
-#include <random>
+#include <set>
+#include <mutex>
+#include <condition_variable>
+#include <unordered_map>
 
-#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
+#include "json.hpp"
+
+#include "../llava/clip.h"
 
 using json = nlohmann::json;
 
 extern bool server_verbose;
-extern bool server_log_json;
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -29,110 +27,428 @@ extern bool server_log_json;
     {                                                                    \
         if (server_verbose)                                              \
         {                                                                \
-            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
+            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
         }                                                                \
     } while (0)
 #endif
 
-#define LOG_ERROR(  MSG, ...) server_log("ERR",  __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_WARNING(MSG, ...) server_log("WARN", __func__, __LINE__, MSG, __VA_ARGS__)
-#define LOG_INFO(   MSG, ...) server_log("INFO", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_ERROR(  MSG, ...) server_log("ERROR",   __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_WARNING(MSG, ...) server_log("WARNING", __func__, __LINE__, MSG, __VA_ARGS__)
+#define LOG_INFO(   MSG, ...) server_log("INFO",    __func__, __LINE__, MSG, __VA_ARGS__)
+
+//
+// parallel
+//
+
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
+enum task_type {
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
+    TASK_TYPE_NEXT_RESPONSE
+};
+
+struct task_server {
+    int id = -1; // to be filled by llama_server_queue
+    int target_id;
+    task_type type;
+    json data;
+    bool infill_mode = false;
+    bool embedding_mode = false;
+    int multitask_id = -1;
+};
+
+struct task_result {
+    int id;
+    int multitask_id = -1;
+    bool stop;
+    bool error;
+    json result_json;
+};
+
+struct task_multi {
+    int id;
+    std::set<int> subtasks_remaining{};
+    std::vector<task_result> results{};
+};
+
+// TODO: can become bool if we can't find use of more states
+enum slot_state
+{
+    IDLE,
+    PROCESSING,
+};
+
+enum slot_command
+{
+    NONE,
+    LOAD_PROMPT,
+    RELEASE,
+};
+
+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_predict = -1; // new tokens to predict
+
+    std::vector<std::string> antiprompt;
+
+    json input_prefix;
+    json input_suffix;
+};
+
+struct slot_image
+{
+    int32_t id;
+
+    bool request_encode_image = false;
+    float * image_embedding = nullptr;
+    int32_t image_tokens = 0;
+
+    clip_image_u8 * img_data;
+
+    std::string prefix_prompt; // before of this image
+};
+
+// completion token output with probabilities
+struct completion_token_output
+{
+    struct token_prob
+    {
+        llama_token tok;
+        float prob;
+    };
+
+    std::vector<token_prob> probs;
+    llama_token tok;
+    std::string text_to_send;
+};
+
+static inline void server_log(const char *level, const char *function, int line,
+                       const char *message, const nlohmann::ordered_json &extra)
+{
+    nlohmann::ordered_json log
+    {
+        {"timestamp", time(nullptr)},
+        {"level",     level},
+        {"function",  function},
+        {"line",      line},
+        {"message",   message},
+    };
+
+    if (!extra.empty())
+    {
+        log.merge_patch(extra);
+    }
+
+    const std::string str = log.dump(-1, ' ', false, json::error_handler_t::replace);
+    printf("%.*s\n", (int)str.size(), str.data());
+    fflush(stdout);
+}
+
+//
+// server utils
+//
 
 template <typename T>
-static T json_value(const json &body, const std::string &key, const T &default_value) {
+static T json_value(const json &body, const std::string &key, const T &default_value)
+{
     // Fallback null to default value
     return body.contains(key) && !body.at(key).is_null()
         ? body.value(key, default_value)
         : default_value;
 }
 
-static inline void server_log(const char *level, const char *function, int line, const char *message, const nlohmann::ordered_json &extra) {
-    std::stringstream ss_tid;
-    ss_tid << std::this_thread::get_id();
-    json log = nlohmann::ordered_json{
-        {"tid",       ss_tid.str()},
-        {"timestamp", time(nullptr)},
-    };
+inline std::string format_llama2(std::vector<json> messages)
+{
+    std::ostringstream output;
+    bool is_inside_turn = false;
 
-    if (server_log_json) {
-        log.merge_patch( {
-            {"level",    level},
-            {"function", function},
-            {"line",     line},
-            {"msg",      message},
-        });
-
-        if (!extra.empty()) {
-            log.merge_patch(extra);
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        if (!is_inside_turn) {
+            output << "[INST] ";
         }
-
-        printf("%s\n", log.dump(-1, ' ', false, json::error_handler_t::replace).c_str());
-    } else {
-        char buf[1024];
-        snprintf(buf, 1024, "%4s [%24s] %s", level, function, message);
-
-        if (!extra.empty()) {
-            log.merge_patch(extra);
+        std::string role    = json_value(*it, "role", std::string("user"));
+        std::string content = json_value(*it, "content", std::string(""));
+        if (role == "system") {
+            output << "<<SYS>>\n" << content << "\n<<SYS>>\n\n";
+            is_inside_turn = true;
+        } else if (role == "user") {
+            output << content << " [/INST]";
+            is_inside_turn = true;
+        } else {
+            output << " " << content << " </s>";
+            is_inside_turn = false;
         }
-        std::stringstream ss;
-        ss << buf << " |";
-        for (const auto& el : log.items())
+    }
+
+    LOG_VERBOSE("format_llama2", {{"text", output.str()}});
+
+    return output.str();
+}
+
+inline std::string format_chatml(std::vector<json> messages)
+{
+    std::ostringstream chatml_msgs;
+
+    for (auto it = messages.begin(); it != messages.end(); ++it) {
+        chatml_msgs << "<|im_start|>"
+                    << json_value(*it, "role",    std::string("user")) << '\n';
+        chatml_msgs << json_value(*it, "content", std::string(""))
+                    << "<|im_end|>\n";
+    }
+
+    chatml_msgs << "<|im_start|>assistant" << '\n';
+
+    LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
+
+    return chatml_msgs.str();
+}
+
+//
+// work queue utils
+//
+
+struct llama_server_queue {
+    int id = 0;
+    std::mutex mutex_tasks;
+    bool running;
+    // queues
+    std::vector<task_server> queue_tasks;
+    std::vector<task_server> queue_tasks_deferred;
+    std::vector<task_multi> queue_multitasks;
+    std::condition_variable condition_tasks;
+    // callback functions
+    std::function<void(task_server&)> callback_new_task;
+    std::function<void(task_multi&)> callback_finish_multitask;
+    std::function<void(void)> callback_all_task_finished;
+
+    // Add a new task to the end of the queue
+    int post(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        if (task.id == -1) {
+            task.id = id++;
+        }
+        queue_tasks.push_back(std::move(task));
+        condition_tasks.notify_one();
+        return task.id;
+    }
+
+    // Add a new task, but defer until one slot is available
+    void defer(task_server task) {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        queue_tasks_deferred.push_back(std::move(task));
+    }
+
+    // Get the next id for creating anew task
+    int get_new_id() {
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        return id++;
+    }
+
+    // Register function to process a new task
+    void on_new_task(std::function<void(task_server&)> callback) {
+        callback_new_task = callback;
+    }
+
+    // Register function to process a multitask
+    void on_finish_multitask(std::function<void(task_multi&)> callback) {
+        callback_finish_multitask = callback;
+    }
+
+    // Register the function to be called when the batch of tasks is finished
+    void on_all_tasks_finished(std::function<void(void)> callback) {
+        callback_all_task_finished = callback;
+    }
+
+    // Call when the state of one slot is changed
+    void notify_slot_changed() {
+        // move deferred tasks back to main loop
+        std::unique_lock<std::mutex> lock(mutex_tasks);
+        for (auto & task : queue_tasks_deferred) {
+            queue_tasks.push_back(std::move(task));
+        }
+        queue_tasks_deferred.clear();
+    }
+
+    // end the start_loop routine
+    void terminate() {
         {
-            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
-            ss << " " << el.key() << "=" << value;
+            std::unique_lock<std::mutex> lock(mutex_tasks);
+            running = false;
+        }
+        condition_tasks.notify_all();
+    }
+
+    // Start the main loop.
+    void start_loop() {
+        running = true;
+        while (true) {
+            // new task arrived
+            LOG_VERBOSE("have new task", {});
+            {
+                while (true)
+                {
+                    std::unique_lock<std::mutex> lock(mutex_tasks);
+                    if (queue_tasks.empty()) {
+                        lock.unlock();
+                        break;
+                    }
+                    task_server task = queue_tasks.front();
+                    queue_tasks.erase(queue_tasks.begin());
+                    lock.unlock();
+                    LOG_VERBOSE("callback_new_task", {});
+                    callback_new_task(task);
+                }
+                LOG_VERBOSE("callback_all_task_finished", {});
+                // process and update all the multitasks
+                auto queue_iterator = queue_multitasks.begin();
+                while (queue_iterator != queue_multitasks.end())
+                {
+                    if (queue_iterator->subtasks_remaining.empty())
+                    {
+                        // all subtasks done == multitask is done
+                        task_multi current_multitask = *queue_iterator;
+                        callback_finish_multitask(current_multitask);
+                        // remove this multitask
+                        queue_iterator = queue_multitasks.erase(queue_iterator);
+                    }
+                    else
+                    {
+                        ++queue_iterator;
+                    }
+                }
+                // all tasks in the current loop is finished
+                callback_all_task_finished();
+            }
+            LOG_VERBOSE("wait for new task", {});
+            // wait for new task
+            {
+                std::unique_lock<std::mutex> lock(mutex_tasks);
+                if (queue_tasks.empty()) {
+                    if (!running) {
+                        LOG_VERBOSE("ending start_loop", {});
+                        return;
+                    }
+                    condition_tasks.wait(lock, [&]{
+                        return (!queue_tasks.empty() || !running);
+                    });
+                }
+            }
+        }
+    }
+
+    //
+    // functions to manage multitasks
+    //
+
+    // add a multitask by specifying the id of all subtask (subtask is a task_server)
+    void add_multitask(int multitask_id, std::vector<int>& sub_ids)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        task_multi multi;
+        multi.id = multitask_id;
+        std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
+        queue_multitasks.push_back(multi);
+    }
+
+    // updatethe remaining subtasks, while appending results to multitask
+    void update_multitask(int multitask_id, int subtask_id, task_result& result)
+    {
+        std::lock_guard<std::mutex> lock(mutex_tasks);
+        for (auto& multitask : queue_multitasks)
+        {
+            if (multitask.id == multitask_id)
+            {
+                multitask.subtasks_remaining.erase(subtask_id);
+                multitask.results.push_back(result);
+            }
+        }
+    }
+};
+
+struct llama_server_response {
+    typedef std::function<void(int, int, task_result&)> callback_multitask_t;
+    callback_multitask_t callback_update_multitask;
+    // for keeping track of all tasks waiting for the result
+    std::set<int> waiting_task_ids;
+    // the main result queue
+    std::vector<task_result> queue_results;
+    std::mutex mutex_results;
+    std::condition_variable condition_results;
+
+    void add_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.insert(task_id);
+    }
+
+    void remove_waiting_task_id(int task_id) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        waiting_task_ids.erase(task_id);
+    }
+
+    // This function blocks the thread until there is a response for this task_id
+    task_result recv(int task_id) {
+        while (true)
+        {
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
+            LOG_VERBOSE("condition_results unblock", {});
+
+            for (int i = 0; i < (int) queue_results.size(); i++)
+            {
+                if (queue_results[i].id == task_id)
+                {
+                    assert(queue_results[i].multitask_id == -1);
+                    task_result res = queue_results[i];
+                    queue_results.erase(queue_results.begin() + i);
+                    return res;
+                }
+            }
         }
 
-        const std::string str = ss.str();
-        printf("%.*s\n", (int)str.size(), str.data());
-        fflush(stdout);
-    }
-}
-
-//
-// chat template utils
-//
-
-// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-inline bool verify_custom_template(const std::string & tmpl) {
-    llama_chat_message chat[] = {{"user", "test"}};
-    int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, nullptr, 0);
-    return res >= 0;
-}
-
-// Format given chat. If tmpl is empty, we take the template from model metadata
-inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
-    size_t alloc_size = 0;
-    // vector holding all allocated string to be passed to llama_chat_apply_template
-    std::vector<std::string> str(messages.size() * 2);
-    std::vector<llama_chat_message> chat(messages.size());
-
-    for (size_t i = 0; i < messages.size(); ++i) {
-        const auto & curr_msg = messages[i];
-        str[i*2 + 0]    = json_value(curr_msg, "role",    std::string(""));
-        str[i*2 + 1]    = json_value(curr_msg, "content", std::string(""));
-        alloc_size     += str[i*2 + 1].length();
-        chat[i].role    = str[i*2 + 0].c_str();
-        chat[i].content = str[i*2 + 1].c_str();
+        // should never reach here
     }
 
-    const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
-    std::vector<char> buf(alloc_size * 2);
-
-    // run the first time to get the total output length
-    int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
-
-    // if it turns out that our buffer is too small, we resize it
-    if ((size_t) res > buf.size()) {
-        buf.resize(res);
-        res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
+    // Register the function to update multitask
+    void on_multitask_update(callback_multitask_t callback) {
+        callback_update_multitask = callback;
     }
 
-    const std::string formatted_chat(buf.data(), res);
+    // Send a new result to a waiting task_id
+    void send(task_result result) {
+        std::unique_lock<std::mutex> lock(mutex_results);
+        LOG_VERBOSE("send new result", {});
+        for (auto& task_id : waiting_task_ids) {
+            // LOG_TEE("waiting task id %i \n", task_id);
+            // for now, tasks that have associated parent multitasks just get erased once multitask picks up the result
+            if (result.multitask_id == task_id)
+            {
+                LOG_VERBOSE("callback_update_multitask", {});
+                callback_update_multitask(task_id, result.id, result);
+                continue;
+            }
 
-    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
-
-    return formatted_chat;
-}
+            if (result.id == task_id)
+            {
+                LOG_VERBOSE("queue_results.push_back", {});
+                queue_results.push_back(result);
+                condition_results.notify_one();
+                return;
+            }
+        }
+    }
+};
 
 //
 // base64 utils (TODO: move to common in the future)
@@ -143,11 +459,13 @@ static const std::string base64_chars =
              "abcdefghijklmnopqrstuvwxyz"
              "0123456789+/";
 
-static inline bool is_base64(uint8_t c) {
+static inline bool is_base64(uint8_t c)
+{
     return (isalnum(c) || (c == '+') || (c == '/'));
 }
 
-static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string) {
+static inline std::vector<uint8_t> base64_decode(const std::string & encoded_string)
+{
     int i = 0;
     int j = 0;
     int in_ = 0;
@@ -159,10 +477,13 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
 
     std::vector<uint8_t> ret;
 
-    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_])) {
+    while (in_len-- && (encoded_string[in_] != '=') && is_base64(encoded_string[in_]))
+    {
         char_array_4[i++] = encoded_string[in_]; in_++;
-        if (i == 4) {
-            for (i = 0; i < 4; i++) {
+        if (i == 4)
+        {
+            for (i = 0; i <4; i++)
+            {
                 char_array_4[i] = base64_chars.find(char_array_4[i]);
             }
 
@@ -170,20 +491,23 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
             char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
             char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
 
-            for (i = 0; (i < 3); i++) {
+            for (i = 0; (i < 3); i++)
+            {
                 ret.push_back(char_array_3[i]);
             }
-
             i = 0;
         }
     }
 
-    if (i) {
-        for (j = i; j < 4; j++) {
+    if (i)
+    {
+        for (j = i; j <4; j++)
+        {
             char_array_4[j] = 0;
         }
 
-        for (j = 0; j < 4; j++) {
+        for (j = 0; j <4; j++)
+        {
             char_array_4[j] = base64_chars.find(char_array_4[j]);
         }
 
@@ -191,7 +515,8 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
         char_array_3[1] = ((char_array_4[1] & 0xf) << 4) + ((char_array_4[2] & 0x3c) >> 2);
         char_array_3[2] = ((char_array_4[2] & 0x3) << 6) +   char_array_4[3];
 
-        for (j = 0; j < i - 1; j++) {
+        for (j = 0; (j < i - 1); j++)
+        {
             ret.push_back(char_array_3[j]);
         }
     }
@@ -203,7 +528,8 @@ static inline std::vector<uint8_t> base64_decode(const std::string & encoded_str
 // random string / id
 //
 
-static std::string random_string() {
+static std::string random_string()
+{
     static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
 
     std::random_device rd;
@@ -218,327 +544,9 @@ static std::string random_string() {
     return result;
 }
 
-static std::string gen_chatcmplid() {
+static std::string gen_chatcmplid()
+{
     std::stringstream chatcmplid;
     chatcmplid << "chatcmpl-" << random_string();
-
     return chatcmplid.str();
 }
-
-//
-// other common utils
-//
-
-static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
-
-    return i;
-}
-
-static bool ends_with(const std::string & str, const std::string & suffix) {
-    return str.size() >= suffix.size() && 0 == str.compare(str.size() - suffix.size(), suffix.size(), suffix);
-}
-
-static size_t find_partial_stop_string(const std::string &stop, const std::string &text) {
-    if (!text.empty() && !stop.empty()) {
-        const char text_last_char = text.back();
-        for (int64_t char_index = stop.size() - 1; char_index >= 0; char_index--) {
-            if (stop[char_index] == text_last_char) {
-                const std::string current_partial = stop.substr(0, char_index + 1);
-                if (ends_with(text, current_partial)) {
-                    return text.size() - char_index - 1;
-                }
-            }
-        }
-    }
-
-    return std::string::npos;
-}
-
-// TODO: reuse llama_detokenize
-template <class Iter>
-static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
-    std::string ret;
-    for (; begin != end; ++begin) {
-        ret += llama_token_to_piece(ctx, *begin);
-    }
-
-    return ret;
-}
-
-// format incomplete utf-8 multibyte character for output
-static std::string tokens_to_output_formatted_string(const llama_context * ctx, const llama_token token) {
-    std::string out = token == -1 ? "" : llama_token_to_piece(ctx, token);
-
-    // if the size is 1 and first bit is 1, meaning it's a partial character
-    //   (size > 1 meaning it's already a known token)
-    if (out.size() == 1 && (out[0] & 0x80) == 0x80) {
-        std::stringstream ss;
-        ss << std::hex << (out[0] & 0xff);
-        std::string res(ss.str());
-        out = "byte: \\x" + res;
-    }
-
-    return out;
-}
-
-struct completion_token_output {
-    llama_token tok;
-    std::string text_to_send;
-
-    struct token_prob {
-        llama_token tok;
-        float prob;
-    };
-
-    std::vector<token_prob> probs;
-};
-
-// convert a vector of completion_token_output to json
-static json probs_vector_to_json(const llama_context * ctx, const std::vector<completion_token_output> & probs) {
-    json out = json::array();
-
-    for (const auto & prob : probs) {
-        json probs_for_token = json::array();
-
-        for (const auto & p : prob.probs) {
-            const std::string tok_str = tokens_to_output_formatted_string(ctx, p.tok);
-            probs_for_token.push_back(json {
-                {"tok_str", tok_str},
-                {"prob",    p.prob},
-            });
-        }
-
-        const std::string tok_str = tokens_to_output_formatted_string(ctx, prob.tok);
-        out.push_back(json {
-            {"content", tok_str},
-            {"probs",   probs_for_token},
-        });
-    }
-
-    return out;
-}
-
-//
-// OAI utils
-//
-
-static json oaicompat_completion_params_parse(
-    const struct llama_model * model,
-    const json & body, /* openai api json semantics */
-    const std::string & chat_template) {
-    json llama_params;
-
-    llama_params["__oaicompat"] = true;
-
-    // Map OpenAI parameters to llama.cpp parameters
-    //
-    // For parameters that are defined by the OpenAI documentation (e.g.
-    // temperature), we explicitly specify OpenAI's intended default; we
-    // need to do that because sometimes OpenAI disagrees with llama.cpp
-    //
-    // https://platform.openai.com/docs/api-reference/chat/create
-    llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body,   "model",             std::string("unknown"));
-    llama_params["prompt"]            = format_chat(model, chat_template,       body["messages"]);
-    llama_params["cache_prompt"]      = json_value(body,   "cache_prompt",      false);
-    llama_params["temperature"]       = json_value(body,   "temperature",       0.0);
-    llama_params["top_k"]             = json_value(body,   "top_k",             default_sparams.top_k);
-    llama_params["top_p"]             = json_value(body,   "top_p",             1.0);
-    llama_params["n_predict"]         = json_value(body,   "max_tokens",        -1);
-    llama_params["logit_bias"]        = json_value(body,   "logit_bias",        json::object());
-    llama_params["frequency_penalty"] = json_value(body,   "frequency_penalty", 0.0);
-    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["mirostat"]          = json_value(body,   "mirostat",          default_sparams.mirostat);
-    llama_params["mirostat_tau"]      = json_value(body,   "mirostat_tau",      default_sparams.mirostat_tau);
-    llama_params["mirostat_eta"]      = json_value(body,   "mirostat_eta",      default_sparams.mirostat_eta);
-    llama_params["penalize_nl"]       = json_value(body,   "penalize_nl",       default_sparams.penalize_nl);
-    llama_params["typical_p"]         = json_value(body,   "typical_p",         default_sparams.typical_p);
-    llama_params["repeat_last_n"]     = json_value(body,   "repeat_last_n",     default_sparams.penalty_last_n);
-    llama_params["ignore_eos"]        = json_value(body,   "ignore_eos",        false);
-    llama_params["tfs_z"]             = json_value(body,   "tfs_z",             default_sparams.tfs_z);
-
-    if (body.count("grammar") != 0) {
-        llama_params["grammar"] = json_value(body, "grammar", json::object());
-    }
-
-    // Handle 'stop' field
-    if (body.contains("stop") && body["stop"].is_string()) {
-        llama_params["stop"] = json::array({body["stop"].get<std::string>()});
-    } else {
-        llama_params["stop"] = json_value(body, "stop", json::array());
-    }
-
-    // Ensure there is ChatML-specific end sequence among stop words
-    llama_params["stop"].push_back("<|im_end|>");
-
-    return llama_params;
-}
-
-static json format_final_response_oaicompat(const json & request, json result, bool streaming = false) {
-    bool stopped_word        = result.count("stopped_word") != 0;
-    bool stopped_eos         = json_value(result, "stopped_eos", false);
-    int num_tokens_predicted = json_value(result, "tokens_predicted", 0);
-    int num_prompt_tokens    = json_value(result, "tokens_evaluated", 0);
-    std::string content      = json_value(result, "content", std::string(""));
-
-    std::string finish_reason = "length";
-    if (stopped_word || stopped_eos) {
-        finish_reason = "stop";
-    }
-
-    json choices =
-        streaming ? json::array({json{{"finish_reason", finish_reason},
-                                        {"index", 0},
-                                        {"delta", json::object()}}})
-                  : json::array({json{{"finish_reason", finish_reason},
-                                        {"index", 0},
-                                        {"message", json{{"content", content},
-                                                         {"role", "assistant"}}}}});
-
-    std::time_t t = std::time(0);
-
-    json res = json {
-        {"choices", choices},
-        {"created", t},
-        {"model",
-            json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
-        {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
-        {"usage", json {
-            {"completion_tokens", num_tokens_predicted},
-            {"prompt_tokens",     num_prompt_tokens},
-            {"total_tokens",      num_tokens_predicted + num_prompt_tokens}
-        }},
-        {"id", gen_chatcmplid()}
-    };
-
-    if (server_verbose) {
-        res["__verbose"] = result;
-    }
-
-    if (result.contains("completion_probabilities")) {
-        res["completion_probabilities"] = json_value(result, "completion_probabilities", json::array());
-    }
-
-    return res;
-}
-
-// return value is vector as there is one case where we might need to generate two responses
-static std::vector<json> format_partial_response_oaicompat(json result) {
-    if (!result.contains("model") || !result.contains("oaicompat_token_ctr")) {
-        return std::vector<json>({result});
-    }
-
-    bool first = json_value(result, "oaicompat_token_ctr", 0) == 0;
-    std::string modelname = json_value(result, "model", std::string(DEFAULT_OAICOMPAT_MODEL));
-
-    bool stopped_word   = json_value(result, "stopped_word",  false);
-    bool stopped_eos    = json_value(result, "stopped_eos",   false);
-    bool stopped_limit  = json_value(result, "stopped_limit", false);
-    std::string content = json_value(result, "content",       std::string(""));
-
-    std::string finish_reason;
-    if (stopped_word || stopped_eos) {
-        finish_reason = "stop";
-    }
-    if (stopped_limit) {
-        finish_reason = "length";
-    }
-
-    std::time_t t = std::time(0);
-
-    json choices;
-
-    if (!finish_reason.empty()) {
-        choices = json::array({json{{"finish_reason", finish_reason},
-                                    {"index", 0},
-                                    {"delta", json::object()}}});
-    } else {
-        if (first) {
-            if (content.empty()) {
-                choices = json::array({json{{"finish_reason", nullptr},
-                                            {"index", 0},
-                                            {"delta", json{{"role", "assistant"}}}}});
-            } else {
-                // We have to send this as two updates to conform to openai behavior
-                json initial_ret = json{{"choices", json::array({json{
-                                        {"finish_reason", nullptr},
-                                        {"index", 0},
-                                        {"delta", json{
-                                            {"role", "assistant"}
-                                        }}}})},
-                            {"created", t},
-                            {"id", gen_chatcmplid()},
-                            {"model", modelname},
-                            {"object", "chat.completion.chunk"}};
-
-                json second_ret = json{
-                            {"choices", json::array({json{{"finish_reason", nullptr},
-                                                            {"index", 0},
-                                                            {"delta", json{
-                                                            {"content", content}}}
-                                                            }})},
-                            {"created", t},
-                            {"id", gen_chatcmplid()},
-                            {"model", modelname},
-                            {"object", "chat.completion.chunk"}};
-
-                return std::vector<json>({initial_ret, second_ret});
-            }
-        } else {
-            // Some idiosyncrasy in task processing logic makes several trailing calls
-            // with empty content, we ignore these at the calee site.
-            if (content.empty()) {
-                return std::vector<json>({json::object()});
-            }
-
-            choices = json::array({json{
-                {"finish_reason", nullptr},
-                {"index", 0},
-                {"delta",
-                json{
-                    {"content", content},
-                }},
-            }});
-        }
-    }
-
-    json ret = json {
-        {"choices", choices},
-        {"created", t},
-        {"id",      gen_chatcmplid()},
-        {"model",   modelname},
-        {"object",  "chat.completion.chunk"}
-    };
-
-    return std::vector<json>({ret});
-}
-
-static json format_embeddings_response_oaicompat(const json & request, const json & embeddings) {
-    json res = json {
-        {"model", json_value(request, "model", std::string(DEFAULT_OAICOMPAT_MODEL))},
-        {"object", "list"},
-        {"usage", json {
-            {"prompt_tokens", 0},
-            {"total_tokens", 0}
-        }},
-        {"data", embeddings}
-    };
-
-    return res;
-}
-
-static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
-    return json {
-        {"tokens", tokens}
-    };
-}
-
-static json format_detokenized_response(const std::string & content) {
-    return json {
-        {"content", content}
-    };
-}

examples/speculative/README.md

@@ -6,4 +6,3 @@ More info:
 
 - https://github.com/ggerganov/llama.cpp/pull/2926
 - https://github.com/ggerganov/llama.cpp/pull/3624
-- https://github.com/ggerganov/llama.cpp/pull/5625

examples/speculative/speculative.cpp

@@ -5,7 +5,6 @@
 #include <cstdio>
 #include <string>
 #include <vector>
-#include <set>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -19,7 +18,6 @@ struct seq_draft {
     std::vector<int> i_batch_tgt;
 
     std::vector<llama_token> tokens;
-    std::vector<std::vector<llama_token_data>> dists;
 
     struct llama_sampling_context * ctx_sampling;
 };
@@ -39,15 +37,12 @@ int main(int argc, char ** argv) {
     // max number of parallel drafting sequences (i.e. tree branches)
     const int n_seq_dft = params.n_parallel;
 
+    // probability threshold for accepting a token from the draft model
+    const float p_accept = params.p_accept;
+
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
     const float p_split  = params.p_split;
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-    std::default_random_engine rng(params.seed);
-    std::uniform_real_distribution<> u_dist;
-
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("speculative", "log"));
     LOG_TEE("Log start\n");
@@ -171,9 +166,7 @@ int main(int argc, char ** argv) {
     std::vector<seq_draft> drafts(n_seq_dft);
 
     params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    if (params.sparams.temp == 0) {
-        params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model
-    }
+    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
 
     for (int s = 0; s < n_seq_dft; ++s) {
         drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
@@ -189,15 +182,12 @@ int main(int argc, char ** argv) {
     drafts[0].i_batch_tgt[0] = 0;
 
     while (true) {
-        std::set<int> active_seqs = {};
-
         // print current draft sequences
         for (int s = 0; s < n_seq_dft; ++s) {
             if (!drafts[s].active) {
                 continue;
             }
 
-            active_seqs.insert(s);
             const auto & tokens = drafts[s].tokens;
 
             LOG("draft %d: %s\n", s, LOG_TOKENS_TOSTR_PRETTY(ctx_dft, tokens).c_str());
@@ -206,156 +196,48 @@ int main(int argc, char ** argv) {
         int i_dft  = 0;
         int s_keep = 0;
 
-        llama_token token_id;
-        std::string token_str;
-
-        // loop until we fail to accept a drafted token or we run out of drafted tokens
         while (true) {
+            LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
+
+            // sample from the target model
+            llama_token id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+
+            llama_sampling_accept(ctx_sampling, ctx_tgt, id, true);
+
+            //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
+
+            const std::string token_str = llama_token_to_piece(ctx_tgt, id);
+
+            if (!params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+
+            if (id == llama_token_eos(model_tgt)) {
+                has_eos = true;
+            }
+
+            ++n_predict;
 
             // check if the target token matches any of the drafts
-            // for stochastic sampling, attempt to match the token with the drafted tokens
             {
-                bool accept = false;
-                if (params.sparams.temp > 0) {
-                    // stochastic verification
+                bool matches = false;
 
-                    llama_token_data_array dist_tgt = llama_sampling_probability_distribution(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
-                    float p_tgt = 0, p_dft = 0;
-
-                    // GGML_ASSERT(dist_tgt.size() == dist_dft.size());
-
-                    while (active_seqs.size() > 0) {
-                        // randomly select a sequence to verify from active sequences
-                        std::uniform_int_distribution<unsigned int> u_int_dist(0, active_seqs.size() - 1);
-                        int s = *std::next(active_seqs.begin(), u_int_dist(rng));
-                        if (i_dft >= (int) drafts[s].tokens.size()) {
-                            drafts[s].active = false;
-                            active_seqs.erase(s);
-                            continue;
-                        }
-                        if (accept) {
-                            // if we already accepted a token, we can skip the rest
-                            if (drafts[s].tokens[i_dft] != drafts[s_keep].tokens[i_dft]) {
-                                drafts[s].active = false;
-                                active_seqs.erase(s);
-                            }
-                            continue;
-                        }
-                        LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
-                        float r = u_dist(rng);
-                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), true };
-                        // acquire the token probabilities assigned by the draft and target models
-                        for (size_t i = 0; i < dist_tgt.size; i++) {
-                            if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
-                                p_tgt = dist_tgt.data[i].p;
-                            }
-                            if (dist_dft.data[i].id == drafts[s].tokens[i_dft]) {
-                                p_dft = dist_dft.data[i].p;
-                            }
-                            if (p_tgt && p_dft) {
-                                break;
-                            }
-                        }
-                        LOG("r = %f, p_dft = %f, p_tgt = %f\n", r, p_dft, p_tgt);
-                        if (r <= p_tgt / p_dft) {
-                            s_keep = s;
-                            accept = true;
-                            token_id = drafts[s].tokens[i_dft];
-                            token_str = llama_token_to_piece(ctx_tgt, token_id);
-                            llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
-
-                            LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
-                            break;
-                        } else {
-                            LOG("draft token %d of sequence %d (%d, '%s') rejected\n", i_dft, s, drafts[s].tokens[i_dft], llama_token_to_piece(ctx_tgt, drafts[s].tokens[i_dft]).c_str());
-                            drafts[s].active = false;
-
-                            // calculate residual probability
-                            GGML_ASSERT(dist_tgt.sorted);
-                            GGML_ASSERT(dist_dft.sorted);
-                            float sum_probs = 0.0f;
-
-                            // sort dist by id
-                            std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
-                                return a.id < b.id;
-                            });
-                            std::sort(dist_dft.data, dist_dft.data + dist_dft.size, [](const llama_token_data &a, const llama_token_data &b) {
-                                return a.id < b.id;
-                            });
-
-                            for (size_t i = 0; i < dist_tgt.size; i++) {
-                                dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
-                                sum_probs += dist_tgt.data[i].p;
-                            }
-                            for (size_t i = 0; i < dist_tgt.size; i++) {
-                                dist_tgt.data[i].p /= sum_probs;
-                            }
-
-                            // sort dist_tgt by p desc
-                            std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
-                                return a.p > b.p;
-                            });
-                        }
-
-                        active_seqs.erase(s);
-                        for(int i = 0; i < n_seq_dft; i++) {
-                            if (i == s) {
-                                continue;
-                            }
-                            if (drafts[i].tokens[i_dft] == drafts[s].tokens[i_dft]) {
-                                // synchronize active status for sequences with the same drafted token
-                                drafts[i].active = drafts[i].active && accept;
-                                if (!drafts[i].active) {
-                                    active_seqs.erase(s);
-                                }
-                            }
-                        }
+                for (int s = 0; s < n_seq_dft; ++s) {
+                    if (!drafts[s].active) {
+                        continue;
                     }
 
-                    if (!accept) {
-                        // all drafted tokens were rejected
-                        // sample from the target model
-                        LOG("all drafted tokens were rejected, sampling from residual distribution\n");
-                        token_id = llama_sample_token(ctx_tgt, &dist_tgt);
-                        llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
-                        token_str = llama_token_to_piece(ctx_tgt, token_id);
-                    }
+                    if (i_dft < (int) drafts[s].tokens.size() && id == drafts[s].tokens[i_dft]) {
+                        LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, id, token_str.c_str());
 
-                } else {
-                    // greedy verification
-
-                    // sample from the target model
-                    LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
-                    token_id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
-
-                    llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
-
-                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
-
-                    token_str = llama_token_to_piece(ctx_tgt, token_id);
-
-                    for (int s = 0; s < n_seq_dft; ++s) {
-                        if (!drafts[s].active) {
-                            continue;
-                        }
-
-                        if (i_dft < (int) drafts[s].tokens.size() && token_id == drafts[s].tokens[i_dft]) {
-                            LOG("the sampled target token matches the %dth drafted token of sequence %d (%d, '%s') - accepted\n", i_dft, s, token_id, token_str.c_str());
-
-                            s_keep = s;
-                            accept = true;
-                        } else {
-                            drafts[s].active = false;
-                        }
+                        s_keep = s;
+                        matches = true;
+                    } else {
+                        drafts[s].active = false;
                     }
                 }
 
-                if (token_id == llama_token_eos(model_tgt)) {
-                    has_eos = true;
-                }
-                ++n_predict;
-
-                if (accept) {
+                if (matches) {
                     ++n_accept;
                     ++n_past_tgt;
                     ++n_past_dft;
@@ -363,21 +245,17 @@ int main(int argc, char ** argv) {
                     if (params.use_color) {
                         // Color token according to its origin sequence
                         printf("\u001b[%dm%s\u001b[37m", (36 - s_keep % 6), token_str.c_str());
-                    } else {
-                        printf("%s", token_str.c_str());
+                        fflush(stdout);
                     }
-                    fflush(stdout);
                     continue;
-                } else {
-                    printf("%s", token_str.c_str());
-                    fflush(stdout);
-                    break;
                 }
             }
-        }
+            if (params.use_color) {
+                printf("%s", token_str.c_str());
+            }
+            fflush(stdout);
 
-        {
-            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", id, token_str.c_str());
 
             // TODO: simplify
             {
@@ -397,21 +275,21 @@ int main(int argc, char ** argv) {
                 drafts[s].active = false;
                 drafts[s].tokens.clear();
                 drafts[s].i_batch_tgt.clear();
-                drafts[s].dists.clear();
             }
             // note: will be erased after the speculation phase
-            drafts[0].tokens.push_back(token_id);
-            drafts[0].dists.push_back(std::vector<llama_token_data>());
+            drafts[0].tokens.push_back(id);
             drafts[0].i_batch_tgt.push_back(0);
 
             llama_batch_clear(batch_dft);
-            llama_batch_add  (batch_dft, token_id, n_past_dft, { 0 }, true);
+            llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
 
             llama_kv_cache_seq_rm(ctx_dft, 0, n_past_dft, -1);
             // LOG("dft batch: %s\n", LOG_BATCH_TOSTR_PRETTY(ctx_dft, batch_dft).c_str());
-            llama_decode(ctx_dft, batch_dft);
+            llama_decode         (ctx_dft, batch_dft);
 
             ++n_past_dft;
+
+            break;
         }
 
         if (n_predict > params.n_predict || has_eos) {
@@ -456,6 +334,12 @@ int main(int argc, char ** argv) {
                             k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
                 }
 
+                if (cur_p[0].p < p_accept) {
+                    LOG("stopping drafting for seq %3d, probability too low: %.3f < %.3f\n", s, cur_p[0].p, p_accept);
+                    drafts[s].drafting = false;
+                    continue;
+                }
+
                 std::vector<int> sa(1, s);
 
                 // attempt to split the branch if the probability is high enough
@@ -483,7 +367,6 @@ int main(int argc, char ** argv) {
                         drafts[n_seq_cur].skip     = true;
 
                         drafts[n_seq_cur].tokens      = drafts[s].tokens;
-                        drafts[n_seq_cur].dists       = drafts[s].dists;
                         drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
                         drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
 
@@ -506,8 +389,6 @@ int main(int argc, char ** argv) {
                     llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
 
                     drafts[s].tokens.push_back(id);
-                    // save cur_p.data into drafts[s].dists
-                    drafts[s].dists.push_back(cur_p);
 
                     // add unique drafted tokens to the target batch
                     drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
@@ -559,7 +440,6 @@ int main(int argc, char ** argv) {
             }
 
             drafts[s].tokens.erase(drafts[s].tokens.begin());
-            drafts[s].dists.erase(drafts[s].dists.begin());
         }
     }
 

examples/sycl/ls-sycl-device.cpp

@@ -7,7 +7,7 @@
 
 #include "ggml-sycl.h"
 
-int main() {
+int main(int argc, char ** argv) {
     ggml_backend_sycl_print_sycl_devices();
     return 0;
 }

examples/sycl/run-llama2.sh

@@ -8,19 +8,12 @@ INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 source /opt/intel/oneapi/setvars.sh
 
 if [ $# -gt 0 ]; then
-    GGML_SYCL_DEVICE=$1
+    export GGML_SYCL_DEVICE=$1
 else
-    GGML_SYCL_DEVICE=0
+    export GGML_SYCL_DEVICE=0
 fi
-echo "use $GGML_SYCL_DEVICE as main GPU"
+echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
 #export GGML_SYCL_DEBUG=1
-
-
-#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
-
-#use all GPUs with same max compute units
-ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
-
-#use main GPU only
-#ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+#./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 5 -e -ngl 33 -t 1 -s 0
 

examples/train-text-from-scratch/train-text-from-scratch.cpp

@@ -960,7 +960,7 @@ int main(int argc, char ** argv) {
     struct ggml_opt_context * opt   = train->opt;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1709336216,
-        "narHash": "sha256-Dt/wOWeW6Sqm11Yh+2+t0dfEWxoMxGBvv3JpIocFl9E=",
+        "lastModified": 1706830856,
+        "narHash": "sha256-a0NYyp+h9hlb7ddVz4LUn1vT/PLwqfrWYcHMvFB1xYg=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "f7b3c975cf067e56e7cda6cb098ebe3fb4d74ca2",
+        "rev": "b253292d9c0a5ead9bc98c4e9a26c6312e27d69f",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1709703039,
-        "narHash": "sha256-6hqgQ8OK6gsMu1VtcGKBxKQInRLHtzulDo9Z5jxHEFY=",
+        "lastModified": 1708118438,
+        "narHash": "sha256-kk9/0nuVgA220FcqH/D2xaN6uGyHp/zoxPNUmPCMmEE=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "9df3e30ce24fd28c7b3e2de0d986769db5d6225d",
+        "rev": "5863c27340ba4de8f83e7e3c023b9599c3cb3c80",
         "type": "github"
       },
       "original": {
@@ -37,11 +37,11 @@
     "nixpkgs-lib": {
       "locked": {
         "dir": "lib",
-        "lastModified": 1709237383,
-        "narHash": "sha256-cy6ArO4k5qTx+l5o+0mL9f5fa86tYUX3ozE1S+Txlds=",
+        "lastModified": 1706550542,
+        "narHash": "sha256-UcsnCG6wx++23yeER4Hg18CXWbgNpqNXcHIo5/1Y+hc=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "1536926ef5621b09bba54035ae2bb6d806d72ac8",
+        "rev": "97b17f32362e475016f942bbdfda4a4a72a8a652",
         "type": "github"
       },
       "original": {

flake.nix

@@ -107,12 +107,11 @@
         # ```
         #
         # Cf. https://nixos.org/manual/nix/unstable/command-ref/new-cli/nix3-flake.html?highlight=flake#flake-format
-        flake.overlays.default = (
-          final: prev: {
+        flake.overlays.default =
+          (final: prev: {
             llamaPackages = final.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
             inherit (final.llamaPackages) llama-cpp;
-          }
-        );
+          });
 
         systems = [
           "aarch64-darwin"
@@ -132,9 +131,6 @@
             ...
           }:
           {
-            # For standardised reproducible formatting with `nix fmt`
-            formatter = pkgs.nixfmt-rfc-style;
-
             # Unlike `.#packages`, legacyPackages may contain values of
             # arbitrary types (including nested attrsets) and may even throw
             # exceptions. This attribute isn't recursed into by `nix flake
@@ -154,7 +150,6 @@
             packages =
               {
                 default = config.legacyPackages.llamaPackages.llama-cpp;
-                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs pkgs.stdenv.isLinux {
                 opencl = config.packages.default.override { useOpenCL = true; };
@@ -162,6 +157,7 @@
 
                 mpi-cpu = config.packages.default.override { useMpi = true; };
                 mpi-cuda = config.packages.default.override { useMpi = true; };
+                vulkan = config.packages.default.override { useVulkan = true; };
               }
               // lib.optionalAttrs (system == "x86_64-linux") {
                 rocm = config.legacyPackages.llamaPackagesRocm.llama-cpp;

ggml-alloc.c

@@ -377,9 +377,6 @@ struct ggml_gallocr {
 
     struct node_alloc * node_allocs; // [n_nodes]
     int n_nodes;
-
-    struct tensor_alloc * leaf_allocs; // [n_leafs]
-    int n_leafs;
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
@@ -430,7 +427,6 @@ void ggml_gallocr_free(ggml_gallocr_t galloc) {
     free(galloc->buffers);
     free(galloc->buf_tallocs);
     free(galloc->node_allocs);
-    free(galloc->leaf_allocs);
     free(galloc);
 }
 
@@ -468,7 +464,7 @@ static void ggml_gallocr_allocate_node(ggml_gallocr_t galloc, struct ggml_tensor
             for (int i = 0; i < GGML_MAX_SRC; i++) {
                 struct ggml_tensor * parent = node->src[i];
                 if (parent == NULL) {
-                    continue;
+                    break;
                 }
 
                 // if the node's data is external, then we cannot re-use it
@@ -548,8 +544,22 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
     memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
 
+    // allocate all graph inputs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        if (graph->nodes[i]->flags & GGML_TENSOR_FLAG_INPUT) {
+            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+        for (int j = 0; j < GGML_MAX_SRC; j++) {
+            if (graph->nodes[i]->src[j] == NULL) {
+                continue;
+            }
+            if (graph->nodes[i]->src[j]->flags & GGML_TENSOR_FLAG_INPUT) {
+                ggml_gallocr_allocate_node(galloc, graph->nodes[i]->src[j], get_node_buffer_id(node_buffer_ids, i));
+            }
+        }
+    }
+
     // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
@@ -558,37 +568,14 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
 
-        if (node->flags & GGML_TENSOR_FLAG_INPUT) {
-            ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
-        }
-
         for (int j = 0; j < GGML_MAX_SRC; j++) {
-            struct ggml_tensor * src = node->src[j];
-            if (src == NULL) {
-                continue;
-            }
-
-            ggml_gallocr_hash_get(galloc, src)->n_children += 1;
-
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
-                ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
+            struct ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                break;
             }
+            ggml_gallocr_hash_get(galloc, parent)->n_children += 1;
         }
-    }
-
-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
+   }
 
     // allocate tensors
     for (int i = 0; i < graph->n_nodes; i++) {
@@ -599,7 +586,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             ggml_gallocr_allocate_node(galloc, parent, buffer_id);
         }
@@ -611,7 +598,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             AT_PRINTF("%s", parent->name);
             if (j < GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
@@ -624,7 +611,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * parent = node->src[j];
             if (parent == NULL) {
-                continue;
+                break;
             }
             struct hash_node * p_hn = ggml_gallocr_hash_get(galloc, parent);
             p_hn->n_children -= 1;
@@ -709,18 +696,6 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
             }
         }
     }
-    if (galloc->n_leafs < graph->n_leafs) {
-        free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
-        GGML_ASSERT(galloc->leaf_allocs != NULL);
-    }
-    galloc->n_leafs = graph->n_leafs;
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
-    }
 
     // reallocate buffers if needed
     for (int i = 0; i < galloc->n_buffers; i++) {
@@ -747,8 +722,8 @@ bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
     return ggml_gallocr_reserve_n(galloc, graph, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, struct node_alloc * node_alloc, struct tensor_alloc * tensor_alloc) {
+    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
 
     if (node->view_src != NULL) {
         if (node->buffer == NULL) {
@@ -757,20 +732,29 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[node_alloc->buffer_id], node);
         }
     } else {
         if (node->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
-            void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[node_alloc->buffer_id], node) <= tensor_alloc->size_max);
+            void * base = ggml_backend_buffer_get_base(galloc->buffers[node_alloc->buffer_id]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[node_alloc->buffer_id], node, addr);
         } else {
             if (node->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
+
+#ifndef NDEBUG
+            size_t offset =
+                (char *)node->data -
+                (char *)ggml_backend_buffer_get_base(node->buffer);
+            size_t size = ggml_backend_buffer_get_alloc_size(node->buffer, node);
+            assert(tensor_alloc->offset == SIZE_MAX || offset == tensor_alloc->offset);
+            assert(tensor_alloc->offset == SIZE_MAX || size <= tensor_alloc->size_max);
+#endif
         }
     }
 }
@@ -789,13 +773,6 @@ static bool ggml_gallocr_needs_realloc(ggml_gallocr_t galloc, struct ggml_cgraph
         return true;
     }
 
-    if (galloc->n_leafs != graph->n_leafs) {
-#ifndef NDEBUG
-        fprintf(stderr, "%s: graph has different number of leafs\n", __func__);
-#endif
-        return true;
-    }
-
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
@@ -850,7 +827,6 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
     }
 
     // allocate the graph tensors from the previous assignments
-    // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         struct node_alloc * node_alloc = &galloc->node_allocs[i];
@@ -859,15 +835,9 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
             if (src == NULL) {
                 continue;
             }
-            ggml_gallocr_init_tensor(galloc, src, node_alloc->buffer_id, &node_alloc->src[j]);
+            ggml_gallocr_init_tensor(galloc, src, node_alloc, &node_alloc->src[j]);
         }
-        ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
-    }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
+        ggml_gallocr_init_tensor(galloc, node, node_alloc, &node_alloc->dst);
     }
 
     return true;

ggml-backend-impl.h

@@ -91,22 +91,19 @@ extern "C" {
         // (optional) complete all pending operations
         void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        // create a plan for ggml_cgraph and free it
+        // compute graph with a plan
         ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+        void                      (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-        // compute graph with a plan
-        enum ggml_status (*GGML_CALL graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
         // compute graph without a plan (async)
-        enum ggml_status (*GGML_CALL graph_compute)     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        bool (*GGML_CALL graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
     };
 
     struct ggml_backend {
-        ggml_guid_t guid;
-
         struct ggml_backend_i iface;
 
         ggml_backend_context_t context;

ggml-backend.c

@@ -12,6 +12,7 @@
 
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 
+
 // backend buffer type
 
 const char * ggml_backend_buft_name(ggml_backend_buffer_type_t buft) {
@@ -158,13 +159,6 @@ bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml
 
 // backend
 
-ggml_guid_t ggml_backend_guid(ggml_backend_t backend) {
-    if (backend == NULL) {
-        return NULL;
-    }
-    return backend->guid;
-}
-
 const char * ggml_backend_name(ggml_backend_t backend) {
     if (backend == NULL) {
         return "NULL";
@@ -262,11 +256,11 @@ void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_pla
     backend->iface.graph_plan_free(backend, plan);
 }
 
-enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
-    return backend->iface.graph_plan_compute(backend, plan);
+void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    backend->iface.graph_plan_compute(backend, plan);
 }
 
-enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return backend->iface.graph_compute(backend, cgraph);
 }
 
@@ -732,15 +726,15 @@ GGML_CALL static void ggml_backend_cpu_graph_plan_free(ggml_backend_t backend, g
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static enum ggml_status ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+GGML_CALL static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
     struct ggml_backend_plan_cpu * cpu_plan = (struct ggml_backend_plan_cpu *)plan;
 
-    return ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
+    ggml_graph_compute(&cpu_plan->cgraph, &cpu_plan->cplan);
 
     GGML_UNUSED(backend);
 }
 
-GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
 
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
@@ -755,7 +749,8 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
     cplan.abort_callback      = cpu_ctx->abort_callback;
     cplan.abort_callback_data = cpu_ctx->abort_callback_data;
 
-    return ggml_graph_compute(cgraph, &cplan);
+    ggml_graph_compute(cgraph, &cplan);
+    return true;
 }
 
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -786,11 +781,6 @@ static struct ggml_backend_i cpu_backend_i = {
     /* .supports_op             = */ ggml_backend_cpu_supports_op,
 };
 
-static ggml_guid_t ggml_backend_cpu_guid(void) {
-    static ggml_guid guid = { 0xaa, 0x67, 0xc7, 0x43, 0x96, 0xe6, 0xa3, 0x8a, 0xe3, 0xaf, 0xea, 0x92, 0x36, 0xbc, 0xfc, 0x89 };
-    return &guid;
-}
-
 ggml_backend_t ggml_backend_cpu_init(void) {
     struct ggml_backend_cpu_context * ctx = malloc(sizeof(struct ggml_backend_cpu_context));
     if (ctx == NULL) {
@@ -810,7 +800,6 @@ ggml_backend_t ggml_backend_cpu_init(void) {
     }
 
     *cpu_backend = (struct ggml_backend) {
-        /* .guid      = */ ggml_backend_cpu_guid(),
         /* .interface = */ cpu_backend_i,
         /* .context   = */ ctx
     };
@@ -818,7 +807,7 @@ ggml_backend_t ggml_backend_cpu_init(void) {
 }
 
 GGML_CALL bool ggml_backend_is_cpu(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_cpu_guid());
+    return backend && backend->iface.get_name == ggml_backend_cpu_name;
 }
 
 void ggml_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads) {
@@ -1436,7 +1425,7 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
     return true;
 }
 
-static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
+static bool ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
     uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
     uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
 
@@ -1471,9 +1460,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
         uint64_t compute_start_us = ggml_time_us();
         if (!sched->callback_eval) {
-            enum ggml_status ec = ggml_backend_graph_compute(split_backend, &split->graph);
-            if (ec != GGML_STATUS_SUCCESS) {
-                return ec;
+            if (!ggml_backend_graph_compute(split_backend, &split->graph)) {
+                return false;
             }
             //ggml_backend_synchronize(split_backend); // necessary to measure compute time
         } else {
@@ -1494,9 +1482,8 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
                 struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
 
-                enum ggml_status ec = ggml_backend_graph_compute(split_backend, &gv);
-                if (ec != GGML_STATUS_SUCCESS) {
-                    return ec;
+                if (!ggml_backend_graph_compute(split_backend, &gv)) {
+                    return false;
                 }
 
                 if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
@@ -1520,7 +1507,7 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
     }
 #endif
 
-    return GGML_STATUS_SUCCESS;
+    return true;
 }
 
 ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size) {
@@ -1582,7 +1569,7 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
     return true;
 }
 
-enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+bool ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
 
     if (!sched->is_reset) {
@@ -1591,10 +1578,14 @@ enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, st
 
     ggml_backend_sched_split_graph(sched, graph);
     if (!ggml_backend_sched_alloc_splits(sched)) {
-        return GGML_STATUS_ALLOC_FAILED;
+        return false;
     }
 
-    return ggml_backend_sched_compute_splits(sched);
+    if (!ggml_backend_sched_compute_splits(sched)) {
+        return false;
+    }
+
+    return true;
 }
 
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {

ggml-backend.h

@@ -49,7 +49,7 @@ extern "C" {
     // Backend
     //
 
-    GGML_API ggml_guid_t  ggml_backend_guid(ggml_backend_t backend);
+
     GGML_API const char * ggml_backend_name(ggml_backend_t backend);
     GGML_API void         ggml_backend_free(ggml_backend_t backend);
 
@@ -66,13 +66,12 @@ extern "C" {
 
     GGML_API void ggml_backend_synchronize(ggml_backend_t backend);
 
-    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
-    GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create (ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);
@@ -158,26 +157,26 @@ extern "C" {
     typedef bool (*ggml_backend_sched_eval_callback)(struct ggml_tensor * t, bool ask, void * user_data);
 
     // Initialize a backend scheduler
-    GGML_API ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
-    GGML_API void                 ggml_backend_sched_free(ggml_backend_sched_t sched);
+    GGML_API ggml_backend_sched_t  ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size);
+    GGML_API void                  ggml_backend_sched_free(ggml_backend_sched_t sched);
     // Initialize backend buffers from a measure graph
-    GGML_API bool                 ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
+    GGML_API bool                  ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph);
     // Get the number of splits of the last graph
-    GGML_API int                  ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
+    GGML_API int                   ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched);
 
-    GGML_API size_t               ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
+    GGML_API size_t                ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend);
 
-    GGML_API void                 ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
-    GGML_API ggml_backend_t       ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
+    GGML_API void                  ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend);
+    GGML_API ggml_backend_t        ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node);
 
     // Allocate and compute graph on the backend scheduler
-    GGML_API enum ggml_status     ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
+    GGML_API bool                  ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph);
 
     // Reset all assignments and allocators - must be called before changing the node backends
-    GGML_API void                 ggml_backend_sched_reset(ggml_backend_sched_t sched);
+    GGML_API void                  ggml_backend_sched_reset(ggml_backend_sched_t sched);
 
     // Set a callback to be called for each resulting node during graph compute
-    GGML_API void                 ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
+    GGML_API void                  ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data);
 
     //
     // Utils

ggml-common.h

@@ -1,779 +0,0 @@
-#pragma once
-
-#if defined(GGML_COMMON_IMPL_C)
-#include <stdint.h>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_METAL)
-#include <metal_stdlib>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const constant type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_CUDA)
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static const __device__ __constant__ type name[size] = {
-#define GGML_TABLE_END() };
-
-#define GGML_COMMON_IMPL
-#elif defined(GGML_COMMON_IMPL_SYCL)
-#include <cstdint>
-
-#define GGML_TABLE_BEGIN(type, name, size) static dpct::global_memory<const type, 1> name(sycl::range<1>(size), {
-#define GGML_TABLE_END() });
-
-#define GGML_COMMON_IMPL
-#endif
-
-#if defined(GGML_COMMON_IMPL)
-
-GGML_TABLE_BEGIN(uint8_t, kmask_iq2xs, 8)
-    1, 2, 4, 8, 16, 32, 64, 128
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint8_t, ksigns_iq2xs, 128)
-      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
-    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
-    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
-     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
-    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
-     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
-     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
-    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
-GGML_TABLE_END()
-
-//#if __CUDA_ARCH__ >= MIN_CC_DP4A // lowest compute capability for integer intrinsics
-GGML_TABLE_BEGIN(uint64_t, ksigns64, 128)
-    0x0000000000000000, 0xff000000000000ff, 0xff0000000000ff00, 0x000000000000ffff,
-    0xff00000000ff0000, 0x0000000000ff00ff, 0x0000000000ffff00, 0xff00000000ffffff,
-    0xff000000ff000000, 0x00000000ff0000ff, 0x00000000ff00ff00, 0xff000000ff00ffff,
-    0x00000000ffff0000, 0xff000000ffff00ff, 0xff000000ffffff00, 0x00000000ffffffff,
-    0xff0000ff00000000, 0x000000ff000000ff, 0x000000ff0000ff00, 0xff0000ff0000ffff,
-    0x000000ff00ff0000, 0xff0000ff00ff00ff, 0xff0000ff00ffff00, 0x000000ff00ffffff,
-    0x000000ffff000000, 0xff0000ffff0000ff, 0xff0000ffff00ff00, 0x000000ffff00ffff,
-    0xff0000ffffff0000, 0x000000ffffff00ff, 0x000000ffffffff00, 0xff0000ffffffffff,
-    0xff00ff0000000000, 0x0000ff00000000ff, 0x0000ff000000ff00, 0xff00ff000000ffff,
-    0x0000ff0000ff0000, 0xff00ff0000ff00ff, 0xff00ff0000ffff00, 0x0000ff0000ffffff,
-    0x0000ff00ff000000, 0xff00ff00ff0000ff, 0xff00ff00ff00ff00, 0x0000ff00ff00ffff,
-    0xff00ff00ffff0000, 0x0000ff00ffff00ff, 0x0000ff00ffffff00, 0xff00ff00ffffffff,
-    0x0000ffff00000000, 0xff00ffff000000ff, 0xff00ffff0000ff00, 0x0000ffff0000ffff,
-    0xff00ffff00ff0000, 0x0000ffff00ff00ff, 0x0000ffff00ffff00, 0xff00ffff00ffffff,
-    0xff00ffffff000000, 0x0000ffffff0000ff, 0x0000ffffff00ff00, 0xff00ffffff00ffff,
-    0x0000ffffffff0000, 0xff00ffffffff00ff, 0xff00ffffffffff00, 0x0000ffffffffffff,
-    0xffff000000000000, 0x00ff0000000000ff, 0x00ff00000000ff00, 0xffff00000000ffff,
-    0x00ff000000ff0000, 0xffff000000ff00ff, 0xffff000000ffff00, 0x00ff000000ffffff,
-    0x00ff0000ff000000, 0xffff0000ff0000ff, 0xffff0000ff00ff00, 0x00ff0000ff00ffff,
-    0xffff0000ffff0000, 0x00ff0000ffff00ff, 0x00ff0000ffffff00, 0xffff0000ffffffff,
-    0x00ff00ff00000000, 0xffff00ff000000ff, 0xffff00ff0000ff00, 0x00ff00ff0000ffff,
-    0xffff00ff00ff0000, 0x00ff00ff00ff00ff, 0x00ff00ff00ffff00, 0xffff00ff00ffffff,
-    0xffff00ffff000000, 0x00ff00ffff0000ff, 0x00ff00ffff00ff00, 0xffff00ffff00ffff,
-    0x00ff00ffffff0000, 0xffff00ffffff00ff, 0xffff00ffffffff00, 0x00ff00ffffffffff,
-    0x00ffff0000000000, 0xffffff00000000ff, 0xffffff000000ff00, 0x00ffff000000ffff,
-    0xffffff0000ff0000, 0x00ffff0000ff00ff, 0x00ffff0000ffff00, 0xffffff0000ffffff,
-    0xffffff00ff000000, 0x00ffff00ff0000ff, 0x00ffff00ff00ff00, 0xffffff00ff00ffff,
-    0x00ffff00ffff0000, 0xffffff00ffff00ff, 0xffffff00ffffff00, 0x00ffff00ffffffff,
-    0xffffffff00000000, 0x00ffffff000000ff, 0x00ffffff0000ff00, 0xffffffff0000ffff,
-    0x00ffffff00ff0000, 0xffffffff00ff00ff, 0xffffffff00ffff00, 0x00ffffff00ffffff,
-    0x00ffffffff000000, 0xffffffffff0000ff, 0xffffffffff00ff00, 0x00ffffffff00ffff,
-    0xffffffffffff0000, 0x00ffffffffff00ff, 0x00ffffffffffff00, 0xffffffffffffffff,
-GGML_TABLE_END()
-//#endif
-
-
-GGML_TABLE_BEGIN(uint64_t, iq2xxs_grid, 256)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
-    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
-    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
-    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
-    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
-    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
-    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
-    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
-    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
-    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
-    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
-    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
-    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
-    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
-    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
-    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
-    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
-    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
-    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
-    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
-    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
-    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
-    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
-    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
-    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
-    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
-    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
-    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
-    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
-    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
-    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
-    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
-    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
-    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
-    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
-    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
-    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
-    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
-    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
-    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
-    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
-    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
-    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
-    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
-    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
-    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
-    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
-    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
-    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
-    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
-    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
-    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
-    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
-    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
-    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
-    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
-    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
-    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
-    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2xs_grid, 512)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
-    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
-    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
-    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
-    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
-    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
-    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
-    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
-    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
-    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
-    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
-    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
-    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
-    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
-    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
-    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
-    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
-    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
-    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
-    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
-    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
-    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
-    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
-    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
-    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
-    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
-    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
-    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
-    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
-    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
-    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
-    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
-    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
-    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
-    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
-    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
-    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
-    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
-    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
-    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
-    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
-    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
-    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
-    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
-    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
-    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
-    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
-    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
-    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
-    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
-    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
-    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
-    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
-    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
-    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
-    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
-    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
-    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
-    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
-    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
-    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
-    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
-    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
-    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
-    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
-    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
-    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
-    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
-    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
-    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
-    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
-    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
-    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
-    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
-    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
-    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
-    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
-    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
-    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
-    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
-    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
-    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
-    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
-    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
-    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
-    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
-    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
-    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
-    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
-    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
-    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
-    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
-    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
-    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
-    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
-    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
-    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
-    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
-    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
-    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
-    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
-    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
-    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
-    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
-    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
-    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
-    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
-    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
-    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
-    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
-    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
-    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
-    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
-    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
-    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
-    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
-    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint64_t, iq2s_grid, 1024)
-    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
-    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
-    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
-    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
-    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
-    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x08080808192b192b,
-    0x08080808192b2b19, 0x080808082b080808, 0x080808082b08082b, 0x080808082b081919,
-    0x080808082b082b08, 0x080808082b190819, 0x080808082b191908, 0x080808082b2b0808,
-    0x080808082b2b1919, 0x080808082b2b2b2b, 0x0808081908080819, 0x0808081908081908,
-    0x080808190808192b, 0x0808081908082b19, 0x0808081908190808, 0x080808190819082b,
-    0x0808081908191919, 0x0808081908192b08, 0x08080819082b0819, 0x08080819082b1908,
-    0x0808081919080808, 0x080808191908082b, 0x0808081919081919, 0x0808081919082b08,
-    0x0808081919190819, 0x0808081919191908, 0x080808191919192b, 0x0808081919192b19,
-    0x08080819192b0808, 0x08080819192b1919, 0x08080819192b2b08, 0x080808192b080819,
-    0x080808192b081908, 0x080808192b190808, 0x080808192b19082b, 0x080808192b191919,
-    0x080808192b2b0819, 0x080808192b2b1908, 0x0808082b08080808, 0x0808082b0808082b,
-    0x0808082b08081919, 0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908,
-    0x0808082b082b0808, 0x0808082b082b2b2b, 0x0808082b19080819, 0x0808082b19081908,
-    0x0808082b1908192b, 0x0808082b19082b19, 0x0808082b19190808, 0x0808082b19191919,
-    0x0808082b2b080808, 0x0808082b2b081919, 0x0808082b2b082b2b, 0x0808082b2b191908,
-    0x0808082b2b2b082b, 0x0808190808080819, 0x0808190808081908, 0x080819080808192b,
-    0x0808190808082b19, 0x0808190808190808, 0x080819080819082b, 0x0808190808191919,
-    0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908, 0x08081908082b192b,
-    0x08081908082b2b19, 0x0808190819080808, 0x080819081908082b, 0x0808190819081919,
-    0x0808190819082b08, 0x0808190819082b2b, 0x0808190819190819, 0x0808190819191908,
-    0x080819081919192b, 0x0808190819192b19, 0x08081908192b0808, 0x08081908192b082b,
-    0x08081908192b1919, 0x080819082b080819, 0x080819082b081908, 0x080819082b08192b,
-    0x080819082b082b19, 0x080819082b190808, 0x080819082b191919, 0x080819082b192b08,
-    0x080819082b2b0819, 0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b,
-    0x0808191908081919, 0x0808191908082b08, 0x0808191908082b2b, 0x0808191908190819,
-    0x0808191908191908, 0x080819190819192b, 0x0808191908192b19, 0x08081919082b0808,
-    0x08081919082b1919, 0x08081919082b2b08, 0x0808191919080819, 0x0808191919081908,
-    0x080819191908192b, 0x0808191919082b19, 0x0808191919190808, 0x080819191919082b,
-    0x0808191919191919, 0x0808191919192b08, 0x08081919192b0819, 0x08081919192b1908,
-    0x080819192b080808, 0x080819192b08082b, 0x080819192b081919, 0x080819192b082b08,
-    0x080819192b190819, 0x080819192b191908, 0x080819192b2b0808, 0x0808192b08080819,
-    0x0808192b08081908, 0x0808192b0808192b, 0x0808192b08082b19, 0x0808192b08190808,
-    0x0808192b08191919, 0x0808192b19080808, 0x0808192b19081919, 0x0808192b19082b08,
-    0x0808192b19190819, 0x0808192b19191908, 0x0808192b192b0808, 0x0808192b2b080819,
-    0x0808192b2b081908, 0x0808192b2b190808, 0x08082b0808080808, 0x08082b080808082b,
-    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808190819, 0x08082b0808191908,
-    0x08082b080819192b, 0x08082b0808192b19, 0x08082b08082b0808, 0x08082b08082b1919,
-    0x08082b08082b2b2b, 0x08082b0819080819, 0x08082b0819081908, 0x08082b081908192b,
-    0x08082b0819082b19, 0x08082b0819190808, 0x08082b081919082b, 0x08082b0819191919,
-    0x08082b0819192b08, 0x08082b08192b0819, 0x08082b08192b1908, 0x08082b082b080808,
-    0x08082b082b081919, 0x08082b082b191908, 0x08082b082b2b2b2b, 0x08082b1908080819,
-    0x08082b1908081908, 0x08082b1908190808, 0x08082b190819082b, 0x08082b1908191919,
-    0x08082b1908192b08, 0x08082b19082b0819, 0x08082b1919080808, 0x08082b1919081919,
-    0x08082b1919082b08, 0x08082b1919190819, 0x08082b1919191908, 0x08082b19192b0808,
-    0x08082b192b080819, 0x08082b192b190808, 0x08082b2b08080808, 0x08082b2b08190819,
-    0x08082b2b08191908, 0x08082b2b082b082b, 0x08082b2b082b2b08, 0x08082b2b082b2b2b,
-    0x08082b2b19190808, 0x08082b2b2b192b19, 0x0819080808080819, 0x0819080808081908,
-    0x081908080808192b, 0x0819080808082b19, 0x0819080808190808, 0x081908080819082b,
-    0x0819080808191919, 0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908,
-    0x08190808082b192b, 0x0819080819080808, 0x081908081908082b, 0x0819080819081919,
-    0x0819080819082b08, 0x0819080819190819, 0x0819080819191908, 0x081908081919192b,
-    0x0819080819192b19, 0x08190808192b0808, 0x08190808192b082b, 0x08190808192b1919,
-    0x08190808192b2b08, 0x081908082b080819, 0x081908082b081908, 0x081908082b08192b,
-    0x081908082b190808, 0x081908082b191919, 0x081908082b192b08, 0x081908082b2b0819,
-    0x081908082b2b1908, 0x0819081908080808, 0x081908190808082b, 0x0819081908081919,
-    0x0819081908082b08, 0x0819081908082b2b, 0x0819081908190819, 0x0819081908191908,
-    0x081908190819192b, 0x0819081908192b19, 0x08190819082b0808, 0x08190819082b082b,
-    0x08190819082b1919, 0x08190819082b2b08, 0x0819081919080819, 0x0819081919081908,
-    0x081908191908192b, 0x0819081919082b19, 0x0819081919190808, 0x081908191919082b,
-    0x0819081919191919, 0x0819081919192b08, 0x08190819192b0819, 0x08190819192b1908,
-    0x081908192b080808, 0x081908192b08082b, 0x081908192b081919, 0x081908192b082b08,
-    0x081908192b190819, 0x081908192b191908, 0x0819082b08080819, 0x0819082b08081908,
-    0x0819082b08082b19, 0x0819082b08190808, 0x0819082b08191919, 0x0819082b082b0819,
-    0x0819082b082b1908, 0x0819082b19080808, 0x0819082b19081919, 0x0819082b19190819,
-    0x0819082b19191908, 0x0819082b2b080819, 0x0819082b2b081908, 0x0819082b2b190808,
-    0x0819190808080808, 0x081919080808082b, 0x0819190808081919, 0x0819190808082b08,
-    0x0819190808190819, 0x0819190808191908, 0x081919080819192b, 0x0819190808192b19,
-    0x08191908082b0808, 0x08191908082b1919, 0x08191908082b2b08, 0x0819190819080819,
-    0x0819190819081908, 0x081919081908192b, 0x0819190819082b19, 0x0819190819190808,
-    0x081919081919082b, 0x0819190819191919, 0x0819190819192b08, 0x08191908192b0819,
-    0x08191908192b1908, 0x081919082b080808, 0x081919082b08082b, 0x081919082b081919,
-    0x081919082b082b08, 0x081919082b190819, 0x081919082b191908, 0x081919082b2b0808,
-    0x0819191908080819, 0x0819191908081908, 0x081919190808192b, 0x0819191908082b19,
-    0x0819191908190808, 0x081919190819082b, 0x0819191908191919, 0x0819191908192b08,
-    0x08191919082b0819, 0x08191919082b1908, 0x0819191919080808, 0x081919191908082b,
-    0x0819191919081919, 0x0819191919082b08, 0x0819191919190819, 0x0819191919191908,
-    0x08191919192b0808, 0x081919192b080819, 0x081919192b081908, 0x081919192b190808,
-    0x0819192b08080808, 0x0819192b08081919, 0x0819192b08082b08, 0x0819192b08190819,
-    0x0819192b08191908, 0x0819192b082b0808, 0x0819192b19080819, 0x0819192b19081908,
-    0x0819192b19190808, 0x0819192b2b080808, 0x0819192b2b2b2b2b, 0x08192b0808080819,
-    0x08192b0808081908, 0x08192b080808192b, 0x08192b0808082b19, 0x08192b0808190808,
-    0x08192b0808191919, 0x08192b0808192b08, 0x08192b08082b0819, 0x08192b0819080808,
-    0x08192b081908082b, 0x08192b0819081919, 0x08192b0819082b08, 0x08192b0819190819,
-    0x08192b0819191908, 0x08192b08192b0808, 0x08192b082b080819, 0x08192b082b081908,
-    0x08192b1908080808, 0x08192b190808082b, 0x08192b1908081919, 0x08192b1908082b08,
-    0x08192b1908190819, 0x08192b1908191908, 0x08192b19082b0808, 0x08192b1919080819,
-    0x08192b1919081908, 0x08192b1919190808, 0x08192b19192b2b19, 0x08192b192b2b082b,
-    0x08192b2b08081908, 0x08192b2b08190808, 0x08192b2b19080808, 0x08192b2b1919192b,
-    0x082b080808080808, 0x082b08080808082b, 0x082b080808081919, 0x082b080808082b08,
-    0x082b080808190819, 0x082b080808191908, 0x082b08080819192b, 0x082b080808192b19,
-    0x082b0808082b0808, 0x082b0808082b1919, 0x082b0808082b2b2b, 0x082b080819080819,
-    0x082b080819081908, 0x082b080819190808, 0x082b08081919082b, 0x082b080819191919,
-    0x082b0808192b1908, 0x082b08082b080808, 0x082b08082b082b2b, 0x082b08082b191908,
-    0x082b08082b2b2b2b, 0x082b081908080819, 0x082b081908081908, 0x082b081908190808,
-    0x082b08190819082b, 0x082b081908191919, 0x082b0819082b0819, 0x082b081919080808,
-    0x082b08191908082b, 0x082b081919081919, 0x082b081919190819, 0x082b081919191908,
-    0x082b0819192b0808, 0x082b08192b080819, 0x082b08192b081908, 0x082b08192b190808,
-    0x082b082b08080808, 0x082b082b08082b2b, 0x082b082b082b082b, 0x082b082b082b2b08,
-    0x082b082b082b2b2b, 0x082b082b19081908, 0x082b082b19190808, 0x082b082b2b082b08,
-    0x082b082b2b082b2b, 0x082b082b2b2b2b08, 0x082b190808080819, 0x082b190808081908,
-    0x082b19080808192b, 0x082b190808082b19, 0x082b190808190808, 0x082b190808191919,
-    0x082b190808192b08, 0x082b1908082b0819, 0x082b1908082b1908, 0x082b190819080808,
-    0x082b19081908082b, 0x082b190819081919, 0x082b190819082b08, 0x082b190819190819,
-    0x082b190819191908, 0x082b1908192b0808, 0x082b19082b080819, 0x082b19082b081908,
-    0x082b19082b190808, 0x082b191908080808, 0x082b191908081919, 0x082b191908082b08,
-    0x082b191908190819, 0x082b191908191908, 0x082b1919082b0808, 0x082b191919080819,
-    0x082b191919081908, 0x082b191919190808, 0x082b1919192b192b, 0x082b19192b080808,
-    0x082b192b08080819, 0x082b192b08081908, 0x082b192b08190808, 0x082b192b19080808,
-    0x082b192b19192b19, 0x082b2b0808080808, 0x082b2b0808081919, 0x082b2b0808190819,
-    0x082b2b0808191908, 0x082b2b0819080819, 0x082b2b0819081908, 0x082b2b0819190808,
-    0x082b2b082b082b2b, 0x082b2b082b2b2b2b, 0x082b2b1908080819, 0x082b2b1908081908,
-    0x082b2b1908190808, 0x082b2b192b191919, 0x082b2b2b08082b2b, 0x082b2b2b082b082b,
-    0x082b2b2b192b1908, 0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819,
-    0x1908080808081908, 0x190808080808192b, 0x1908080808082b19, 0x1908080808190808,
-    0x190808080819082b, 0x1908080808191919, 0x1908080808192b08, 0x1908080808192b2b,
-    0x19080808082b0819, 0x19080808082b1908, 0x19080808082b192b, 0x1908080819080808,
-    0x190808081908082b, 0x1908080819081919, 0x1908080819082b08, 0x1908080819082b2b,
-    0x1908080819190819, 0x1908080819191908, 0x190808081919192b, 0x1908080819192b19,
-    0x19080808192b0808, 0x19080808192b082b, 0x19080808192b1919, 0x190808082b080819,
-    0x190808082b081908, 0x190808082b190808, 0x190808082b191919, 0x190808082b192b08,
-    0x190808082b2b0819, 0x190808082b2b1908, 0x1908081908080808, 0x190808190808082b,
-    0x1908081908081919, 0x1908081908082b08, 0x1908081908190819, 0x1908081908191908,
-    0x190808190819192b, 0x1908081908192b19, 0x19080819082b0808, 0x19080819082b082b,
-    0x19080819082b1919, 0x1908081919080819, 0x1908081919081908, 0x190808191908192b,
-    0x1908081919082b19, 0x1908081919190808, 0x190808191919082b, 0x1908081919191919,
-    0x1908081919192b08, 0x19080819192b0819, 0x19080819192b1908, 0x190808192b080808,
-    0x190808192b08082b, 0x190808192b081919, 0x190808192b082b08, 0x190808192b190819,
-    0x190808192b191908, 0x190808192b2b0808, 0x1908082b08080819, 0x1908082b08081908,
-    0x1908082b08190808, 0x1908082b0819082b, 0x1908082b08191919, 0x1908082b08192b08,
-    0x1908082b082b1908, 0x1908082b19080808, 0x1908082b19081919, 0x1908082b19082b08,
-    0x1908082b19190819, 0x1908082b19191908, 0x1908082b192b0808, 0x1908082b2b080819,
-    0x1908082b2b081908, 0x1908190808080808, 0x190819080808082b, 0x1908190808081919,
-    0x1908190808082b08, 0x1908190808082b2b, 0x1908190808190819, 0x1908190808191908,
-    0x190819080819192b, 0x1908190808192b19, 0x19081908082b0808, 0x19081908082b082b,
-    0x19081908082b1919, 0x19081908082b2b08, 0x1908190819080819, 0x1908190819081908,
-    0x190819081908192b, 0x1908190819082b19, 0x1908190819190808, 0x190819081919082b,
-    0x1908190819191919, 0x1908190819192b08, 0x19081908192b0819, 0x19081908192b1908,
-    0x190819082b080808, 0x190819082b08082b, 0x190819082b081919, 0x190819082b082b08,
-    0x190819082b190819, 0x190819082b191908, 0x190819082b2b0808, 0x1908191908080819,
-    0x1908191908081908, 0x190819190808192b, 0x1908191908082b19, 0x1908191908190808,
-    0x190819190819082b, 0x1908191908191919, 0x1908191908192b08, 0x19081919082b0819,
-    0x19081919082b1908, 0x1908191919080808, 0x190819191908082b, 0x1908191919081919,
-    0x1908191919082b08, 0x1908191919190819, 0x1908191919191908, 0x19081919192b0808,
-    0x19081919192b2b2b, 0x190819192b080819, 0x190819192b081908, 0x190819192b190808,
-    0x1908192b08080808, 0x1908192b0808082b, 0x1908192b08081919, 0x1908192b08082b08,
-    0x1908192b08190819, 0x1908192b08191908, 0x1908192b082b0808, 0x1908192b19080819,
-    0x1908192b19081908, 0x1908192b19190808, 0x1908192b2b080808, 0x1908192b2b2b1919,
-    0x19082b0808080819, 0x19082b0808081908, 0x19082b0808082b19, 0x19082b0808190808,
-    0x19082b080819082b, 0x19082b0808191919, 0x19082b0808192b08, 0x19082b08082b0819,
-    0x19082b08082b1908, 0x19082b0819080808, 0x19082b081908082b, 0x19082b0819081919,
-    0x19082b0819082b08, 0x19082b0819190819, 0x19082b0819191908, 0x19082b08192b0808,
-    0x19082b082b081908, 0x19082b082b190808, 0x19082b1908080808, 0x19082b190808082b,
-    0x19082b1908081919, 0x19082b1908082b08, 0x19082b1908190819, 0x19082b1908191908,
-    0x19082b19082b0808, 0x19082b1919080819, 0x19082b1919081908, 0x19082b1919190808,
-    0x19082b192b080808, 0x19082b192b19192b, 0x19082b2b08080819, 0x19082b2b08081908,
-    0x19082b2b08190808, 0x19082b2b19080808, 0x1919080808080808, 0x191908080808082b,
-    0x1919080808081919, 0x1919080808082b08, 0x1919080808190819, 0x1919080808191908,
-    0x191908080819192b, 0x1919080808192b19, 0x19190808082b0808, 0x19190808082b082b,
-    0x19190808082b1919, 0x19190808082b2b08, 0x1919080819080819, 0x1919080819081908,
-    0x191908081908192b, 0x1919080819082b19, 0x1919080819190808, 0x191908081919082b,
-    0x1919080819191919, 0x1919080819192b08, 0x19190808192b0819, 0x19190808192b1908,
-    0x191908082b080808, 0x191908082b08082b, 0x191908082b081919, 0x191908082b082b08,
-    0x191908082b190819, 0x191908082b191908, 0x1919081908080819, 0x1919081908081908,
-    0x191908190808192b, 0x1919081908082b19, 0x1919081908190808, 0x191908190819082b,
-    0x1919081908191919, 0x1919081908192b08, 0x19190819082b0819, 0x19190819082b1908,
-    0x1919081919080808, 0x191908191908082b, 0x1919081919081919, 0x1919081919082b08,
-    0x1919081919190819, 0x1919081919191908, 0x19190819192b0808, 0x191908192b080819,
-    0x191908192b081908, 0x191908192b190808, 0x1919082b08080808, 0x1919082b08081919,
-    0x1919082b08082b08, 0x1919082b08190819, 0x1919082b08191908, 0x1919082b082b0808,
-    0x1919082b19080819, 0x1919082b19081908, 0x1919082b19190808, 0x1919082b192b2b19,
-    0x1919082b2b080808, 0x1919190808080819, 0x1919190808081908, 0x191919080808192b,
-    0x1919190808082b19, 0x1919190808190808, 0x191919080819082b, 0x1919190808191919,
-    0x1919190808192b08, 0x19191908082b0819, 0x19191908082b1908, 0x1919190819080808,
-    0x191919081908082b, 0x1919190819081919, 0x1919190819082b08, 0x1919190819190819,
-    0x1919190819191908, 0x19191908192b0808, 0x191919082b080819, 0x191919082b081908,
-    0x191919082b190808, 0x1919191908080808, 0x191919190808082b, 0x1919191908081919,
-    0x1919191908082b08, 0x1919191908190819, 0x1919191908191908, 0x19191919082b0808,
-    0x1919191919080819, 0x1919191919081908, 0x1919191919190808, 0x191919192b080808,
-    0x1919192b08080819, 0x1919192b08081908, 0x1919192b08190808, 0x1919192b082b192b,
-    0x1919192b19080808, 0x19192b0808080808, 0x19192b080808082b, 0x19192b0808081919,
-    0x19192b0808082b08, 0x19192b0808190819, 0x19192b0808191908, 0x19192b08082b0808,
-    0x19192b0819080819, 0x19192b0819081908, 0x19192b0819190808, 0x19192b0819192b2b,
-    0x19192b082b080808, 0x19192b1908080819, 0x19192b1908081908, 0x19192b1908190808,
-    0x19192b1919080808, 0x19192b2b08080808, 0x19192b2b08192b19, 0x19192b2b2b081919,
-    0x19192b2b2b2b2b08, 0x192b080808080819, 0x192b080808081908, 0x192b08080808192b,
-    0x192b080808190808, 0x192b08080819082b, 0x192b080808191919, 0x192b080808192b08,
-    0x192b0808082b0819, 0x192b0808082b1908, 0x192b080819080808, 0x192b080819081919,
-    0x192b080819082b08, 0x192b080819190819, 0x192b080819191908, 0x192b0808192b0808,
-    0x192b08082b081908, 0x192b08082b190808, 0x192b081908080808, 0x192b08190808082b,
-    0x192b081908081919, 0x192b081908082b08, 0x192b081908190819, 0x192b081908191908,
-    0x192b0819082b0808, 0x192b081919080819, 0x192b081919081908, 0x192b081919190808,
-    0x192b08192b080808, 0x192b08192b192b19, 0x192b082b08081908, 0x192b082b08190808,
-    0x192b082b19080808, 0x192b082b1919192b, 0x192b082b2b2b0819, 0x192b190808080808,
-    0x192b190808081919, 0x192b190808082b08, 0x192b190808190819, 0x192b190808191908,
-    0x192b1908082b0808, 0x192b190819080819, 0x192b190819081908, 0x192b190819190808,
-    0x192b19082b080808, 0x192b191908080819, 0x192b191908081908, 0x192b191908190808,
-    0x192b191919080808, 0x192b191919082b2b, 0x192b1919192b2b08, 0x192b19192b19082b,
-    0x192b192b08080808, 0x192b192b2b191908, 0x192b2b0808080819, 0x192b2b0808081908,
-    0x192b2b0808190808, 0x192b2b08192b1919, 0x192b2b082b192b08, 0x192b2b1908080808,
-    0x192b2b19082b2b2b, 0x192b2b2b1908082b, 0x192b2b2b2b2b0819, 0x2b08080808080808,
-    0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08, 0x2b08080808190819,
-    0x2b08080808191908, 0x2b08080808192b19, 0x2b080808082b0808, 0x2b080808082b1919,
-    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808081919082b,
-    0x2b08080819191919, 0x2b08080819192b08, 0x2b080808192b0819, 0x2b0808082b080808,
-    0x2b0808082b081919, 0x2b0808082b190819, 0x2b0808082b191908, 0x2b08081908080819,
-    0x2b08081908081908, 0x2b08081908082b19, 0x2b08081908190808, 0x2b0808190819082b,
-    0x2b08081908191919, 0x2b08081908192b08, 0x2b080819082b0819, 0x2b080819082b1908,
-    0x2b08081919080808, 0x2b0808191908082b, 0x2b08081919081919, 0x2b08081919082b08,
-    0x2b08081919190819, 0x2b08081919191908, 0x2b0808192b080819, 0x2b0808192b081908,
-    0x2b0808192b190808, 0x2b0808192b2b2b19, 0x2b08082b08080808, 0x2b08082b08081919,
-    0x2b08082b08082b2b, 0x2b08082b08190819, 0x2b08082b08191908, 0x2b08082b19080819,
-    0x2b08082b19081908, 0x2b08082b19190808, 0x2b08190808080819, 0x2b08190808081908,
-    0x2b0819080808192b, 0x2b08190808082b19, 0x2b08190808190808, 0x2b0819080819082b,
-    0x2b08190808191919, 0x2b08190808192b08, 0x2b081908082b0819, 0x2b08190819080808,
-    0x2b0819081908082b, 0x2b08190819081919, 0x2b08190819082b08, 0x2b08190819190819,
-    0x2b08190819191908, 0x2b081908192b0808, 0x2b0819082b080819, 0x2b0819082b081908,
-    0x2b0819082b190808, 0x2b08191908080808, 0x2b0819190808082b, 0x2b08191908081919,
-    0x2b08191908082b08, 0x2b08191908190819, 0x2b08191908191908, 0x2b081919082b0808,
-    0x2b08191919080819, 0x2b08191919081908, 0x2b08191919190808, 0x2b0819192b080808,
-    0x2b0819192b082b2b, 0x2b08192b08080819, 0x2b08192b08081908, 0x2b08192b08190808,
-    0x2b08192b082b2b19, 0x2b08192b19080808, 0x2b082b0808080808, 0x2b082b0808081919,
-    0x2b082b0808190819, 0x2b082b0808191908, 0x2b082b0819080819, 0x2b082b0819081908,
-    0x2b082b0819190808, 0x2b082b082b2b082b, 0x2b082b1908080819, 0x2b082b1908081908,
-    0x2b082b1919080808, 0x2b082b19192b1919, 0x2b082b2b082b082b, 0x2b082b2b19192b08,
-    0x2b082b2b19192b2b, 0x2b082b2b2b08082b, 0x2b082b2b2b2b082b, 0x2b19080808080819,
-    0x2b19080808081908, 0x2b19080808082b19, 0x2b19080808190808, 0x2b1908080819082b,
-    0x2b19080808191919, 0x2b19080808192b08, 0x2b190808082b1908, 0x2b19080819080808,
-    0x2b1908081908082b, 0x2b19080819081919, 0x2b19080819082b08, 0x2b19080819190819,
-    0x2b19080819191908, 0x2b190808192b0808, 0x2b1908082b080819, 0x2b1908082b081908,
-    0x2b1908082b190808, 0x2b19081908080808, 0x2b19081908081919, 0x2b19081908190819,
-    0x2b19081908191908, 0x2b19081919080819, 0x2b19081919081908, 0x2b19081919190808,
-    0x2b19081919192b2b, 0x2b19082b08080819, 0x2b19082b08081908, 0x2b19082b08190808,
-    0x2b19082b19080808, 0x2b19082b2b2b192b, 0x2b19190808080808, 0x2b1919080808082b,
-    0x2b19190808081919, 0x2b19190808082b08, 0x2b19190808190819, 0x2b19190808191908,
-    0x2b191908082b0808, 0x2b19190819080819, 0x2b19190819081908, 0x2b19190819190808,
-    0x2b1919082b080808, 0x2b1919082b19192b, 0x2b19191908080819, 0x2b19191908081908,
-    0x2b19191908190808, 0x2b19191919080808, 0x2b1919192b192b08, 0x2b1919192b2b0819,
-    0x2b19192b08080808, 0x2b19192b1908192b, 0x2b19192b192b1908, 0x2b192b0808080819,
-    0x2b192b0808081908, 0x2b192b0808190808, 0x2b192b08082b192b, 0x2b192b0819080808,
-    0x2b192b082b2b2b19, 0x2b192b1908080808, 0x2b192b1919082b19, 0x2b192b191919082b,
-    0x2b192b2b2b190808, 0x2b2b080808080808, 0x2b2b080808081919, 0x2b2b080808082b2b,
-    0x2b2b080808191908, 0x2b2b0808082b082b, 0x2b2b0808082b2b2b, 0x2b2b080819080819,
-    0x2b2b080819081908, 0x2b2b080819190808, 0x2b2b08082b2b082b, 0x2b2b08082b2b2b2b,
-    0x2b2b081919080808, 0x2b2b0819192b1919, 0x2b2b082b0808082b, 0x2b2b082b08082b2b,
-    0x2b2b082b082b082b, 0x2b2b082b082b2b08, 0x2b2b082b082b2b2b, 0x2b2b082b2b08082b,
-    0x2b2b082b2b082b08, 0x2b2b082b2b082b2b, 0x2b2b082b2b2b2b08, 0x2b2b190808080819,
-    0x2b2b190808081908, 0x2b2b190808190808, 0x2b2b190819080808, 0x2b2b19082b082b19,
-    0x2b2b19082b2b1908, 0x2b2b191908080808, 0x2b2b191908192b19, 0x2b2b192b19190819,
-    0x2b2b2b0808082b2b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b082b, 0x2b2b2b1919191908,
-    0x2b2b2b192b08192b, 0x2b2b2b2b08082b08, 0x2b2b2b2b08082b2b, 0x2b2b2b2b082b0808,
-    0x2b2b2b2b082b082b, 0x2b2b2b2b082b2b08, 0x2b2b2b2b2b082b08, 0x2b2b2b2b2b2b2b2b,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3xxs_grid, 256)
-    0x04040404, 0x04040414, 0x04040424, 0x04040c0c, 0x04040c1c, 0x04040c3e, 0x04041404, 0x04041414,
-    0x04041c0c, 0x04042414, 0x04043e1c, 0x04043e2c, 0x040c040c, 0x040c041c, 0x040c0c04, 0x040c0c14,
-    0x040c140c, 0x040c142c, 0x040c1c04, 0x040c1c14, 0x040c240c, 0x040c2c24, 0x040c3e04, 0x04140404,
-    0x04140414, 0x04140424, 0x04140c0c, 0x04141404, 0x04141414, 0x04141c0c, 0x04141c1c, 0x04141c3e,
-    0x04142c0c, 0x04142c3e, 0x04143e2c, 0x041c040c, 0x041c043e, 0x041c0c04, 0x041c0c14, 0x041c142c,
-    0x041c3e04, 0x04240c1c, 0x04241c3e, 0x04242424, 0x04242c3e, 0x04243e1c, 0x04243e2c, 0x042c040c,
-    0x042c043e, 0x042c1c14, 0x042c2c14, 0x04341c2c, 0x04343424, 0x043e0c04, 0x043e0c24, 0x043e0c34,
-    0x043e241c, 0x043e340c, 0x0c04040c, 0x0c04041c, 0x0c040c04, 0x0c040c14, 0x0c04140c, 0x0c04141c,
-    0x0c041c04, 0x0c041c14, 0x0c041c24, 0x0c04243e, 0x0c042c04, 0x0c0c0404, 0x0c0c0414, 0x0c0c0c0c,
-    0x0c0c1404, 0x0c0c1414, 0x0c14040c, 0x0c14041c, 0x0c140c04, 0x0c140c14, 0x0c14140c, 0x0c141c04,
-    0x0c143e14, 0x0c1c0404, 0x0c1c0414, 0x0c1c1404, 0x0c1c1c0c, 0x0c1c2434, 0x0c1c3434, 0x0c24040c,
-    0x0c24042c, 0x0c242c04, 0x0c2c1404, 0x0c2c1424, 0x0c2c2434, 0x0c2c3e0c, 0x0c34042c, 0x0c3e1414,
-    0x0c3e2404, 0x14040404, 0x14040414, 0x14040c0c, 0x14040c1c, 0x14041404, 0x14041414, 0x14041434,
-    0x14041c0c, 0x14042414, 0x140c040c, 0x140c041c, 0x140c042c, 0x140c0c04, 0x140c0c14, 0x140c140c,
-    0x140c1c04, 0x140c341c, 0x140c343e, 0x140c3e04, 0x14140404, 0x14140414, 0x14140c0c, 0x14140c3e,
-    0x14141404, 0x14141414, 0x14141c3e, 0x14142404, 0x14142c2c, 0x141c040c, 0x141c0c04, 0x141c0c24,
-    0x141c3e04, 0x141c3e24, 0x14241c2c, 0x14242c1c, 0x142c041c, 0x142c143e, 0x142c240c, 0x142c3e24,
-    0x143e040c, 0x143e041c, 0x143e0c34, 0x143e242c, 0x1c04040c, 0x1c040c04, 0x1c040c14, 0x1c04140c,
-    0x1c04141c, 0x1c042c04, 0x1c04342c, 0x1c043e14, 0x1c0c0404, 0x1c0c0414, 0x1c0c1404, 0x1c0c1c0c,
-    0x1c0c2424, 0x1c0c2434, 0x1c14040c, 0x1c14041c, 0x1c140c04, 0x1c14142c, 0x1c142c14, 0x1c143e14,
-    0x1c1c0c0c, 0x1c1c1c1c, 0x1c241c04, 0x1c24243e, 0x1c243e14, 0x1c2c0404, 0x1c2c0434, 0x1c2c1414,
-    0x1c2c2c2c, 0x1c340c24, 0x1c341c34, 0x1c34341c, 0x1c3e1c1c, 0x1c3e3404, 0x24040424, 0x24040c3e,
-    0x24041c2c, 0x24041c3e, 0x24042c1c, 0x24042c3e, 0x240c3e24, 0x24141404, 0x24141c3e, 0x24142404,
-    0x24143404, 0x24143434, 0x241c043e, 0x241c242c, 0x24240424, 0x24242c0c, 0x24243424, 0x242c142c,
-    0x242c241c, 0x242c3e04, 0x243e042c, 0x243e0c04, 0x243e0c14, 0x243e1c04, 0x2c040c14, 0x2c04240c,
-    0x2c043e04, 0x2c0c0404, 0x2c0c0434, 0x2c0c1434, 0x2c0c2c2c, 0x2c140c24, 0x2c141c14, 0x2c143e14,
-    0x2c1c0414, 0x2c1c2c1c, 0x2c240c04, 0x2c24141c, 0x2c24143e, 0x2c243e14, 0x2c2c0414, 0x2c2c1c0c,
-    0x2c342c04, 0x2c3e1424, 0x2c3e2414, 0x34041424, 0x34042424, 0x34042434, 0x34043424, 0x340c140c,
-    0x340c340c, 0x34140c3e, 0x34143424, 0x341c1c04, 0x341c1c34, 0x34242424, 0x342c042c, 0x342c2c14,
-    0x34341c1c, 0x343e041c, 0x343e140c, 0x3e04041c, 0x3e04042c, 0x3e04043e, 0x3e040c04, 0x3e041c14,
-    0x3e042c14, 0x3e0c1434, 0x3e0c2404, 0x3e140c14, 0x3e14242c, 0x3e142c14, 0x3e1c0404, 0x3e1c0c2c,
-    0x3e1c1c1c, 0x3e1c3404, 0x3e24140c, 0x3e24240c, 0x3e2c0404, 0x3e2c0414, 0x3e2c1424, 0x3e341c04,
-GGML_TABLE_END()
-
-GGML_TABLE_BEGIN(uint32_t, iq3s_grid, 512)
-    0x01010101, 0x01010103, 0x01010105, 0x0101010b, 0x0101010f, 0x01010301, 0x01010303, 0x01010305,
-    0x01010309, 0x0101030d, 0x01010501, 0x01010503, 0x0101050b, 0x01010707, 0x01010901, 0x01010905,
-    0x0101090b, 0x0101090f, 0x01010b03, 0x01010b07, 0x01010d01, 0x01010d05, 0x01010f03, 0x01010f09,
-    0x01010f0f, 0x01030101, 0x01030103, 0x01030105, 0x01030109, 0x01030301, 0x01030303, 0x0103030b,
-    0x01030501, 0x01030507, 0x0103050f, 0x01030703, 0x0103070b, 0x01030909, 0x01030d03, 0x01030d0b,
-    0x01030f05, 0x01050101, 0x01050103, 0x0105010b, 0x0105010f, 0x01050301, 0x01050307, 0x0105030d,
-    0x01050503, 0x0105050b, 0x01050701, 0x01050709, 0x01050905, 0x0105090b, 0x0105090f, 0x01050b03,
-    0x01050b07, 0x01050f01, 0x01050f07, 0x01070107, 0x01070303, 0x0107030b, 0x01070501, 0x01070505,
-    0x01070703, 0x01070707, 0x0107070d, 0x01070909, 0x01070b01, 0x01070b05, 0x01070d0f, 0x01070f03,
-    0x01070f0b, 0x01090101, 0x01090307, 0x0109030f, 0x01090503, 0x01090509, 0x01090705, 0x01090901,
-    0x01090907, 0x01090b03, 0x01090f01, 0x010b0105, 0x010b0109, 0x010b0501, 0x010b0505, 0x010b050d,
-    0x010b0707, 0x010b0903, 0x010b090b, 0x010b090f, 0x010b0d0d, 0x010b0f07, 0x010d010d, 0x010d0303,
-    0x010d0307, 0x010d0703, 0x010d0b05, 0x010d0f03, 0x010f0101, 0x010f0105, 0x010f0109, 0x010f0501,
-    0x010f0505, 0x010f050d, 0x010f0707, 0x010f0b01, 0x010f0b09, 0x03010101, 0x03010103, 0x03010105,
-    0x03010109, 0x03010301, 0x03010303, 0x03010307, 0x0301030b, 0x0301030f, 0x03010501, 0x03010505,
-    0x03010703, 0x03010709, 0x0301070d, 0x03010b09, 0x03010b0d, 0x03010d03, 0x03010f05, 0x03030101,
-    0x03030103, 0x03030107, 0x0303010d, 0x03030301, 0x03030309, 0x03030503, 0x03030701, 0x03030707,
-    0x03030903, 0x03030b01, 0x03030b05, 0x03030f01, 0x03030f0d, 0x03050101, 0x03050305, 0x0305030b,
-    0x0305030f, 0x03050501, 0x03050509, 0x03050705, 0x03050901, 0x03050907, 0x03050b0b, 0x03050d01,
-    0x03050f05, 0x03070103, 0x03070109, 0x0307010f, 0x03070301, 0x03070307, 0x03070503, 0x0307050f,
-    0x03070701, 0x03070709, 0x03070903, 0x03070d05, 0x03070f01, 0x03090107, 0x0309010b, 0x03090305,
-    0x03090309, 0x03090703, 0x03090707, 0x03090905, 0x0309090d, 0x03090b01, 0x03090b09, 0x030b0103,
-    0x030b0301, 0x030b0307, 0x030b0503, 0x030b0701, 0x030b0705, 0x030b0b03, 0x030d0501, 0x030d0509,
-    0x030d050f, 0x030d0909, 0x030d090d, 0x030f0103, 0x030f0107, 0x030f0301, 0x030f0305, 0x030f0503,
-    0x030f070b, 0x030f0903, 0x030f0d05, 0x030f0f01, 0x05010101, 0x05010103, 0x05010107, 0x0501010b,
-    0x0501010f, 0x05010301, 0x05010305, 0x05010309, 0x0501030d, 0x05010503, 0x05010507, 0x0501050f,
-    0x05010701, 0x05010705, 0x05010903, 0x05010907, 0x0501090b, 0x05010b01, 0x05010b05, 0x05010d0f,
-    0x05010f01, 0x05010f07, 0x05010f0b, 0x05030101, 0x05030105, 0x05030301, 0x05030307, 0x0503030f,
-    0x05030505, 0x0503050b, 0x05030703, 0x05030709, 0x05030905, 0x05030b03, 0x05050103, 0x05050109,
-    0x0505010f, 0x05050503, 0x05050507, 0x05050701, 0x0505070f, 0x05050903, 0x05050b07, 0x05050b0f,
-    0x05050f03, 0x05050f09, 0x05070101, 0x05070105, 0x0507010b, 0x05070303, 0x05070505, 0x05070509,
-    0x05070703, 0x05070707, 0x05070905, 0x05070b01, 0x05070d0d, 0x05090103, 0x0509010f, 0x05090501,
-    0x05090507, 0x05090705, 0x0509070b, 0x05090903, 0x05090f05, 0x05090f0b, 0x050b0109, 0x050b0303,
-    0x050b0505, 0x050b070f, 0x050b0901, 0x050b0b07, 0x050b0f01, 0x050d0101, 0x050d0105, 0x050d010f,
-    0x050d0503, 0x050d0b0b, 0x050d0d03, 0x050f010b, 0x050f0303, 0x050f050d, 0x050f0701, 0x050f0907,
-    0x050f0b01, 0x07010105, 0x07010303, 0x07010307, 0x0701030b, 0x0701030f, 0x07010505, 0x07010703,
-    0x07010707, 0x0701070b, 0x07010905, 0x07010909, 0x0701090f, 0x07010b03, 0x07010d07, 0x07010f03,
-    0x07030103, 0x07030107, 0x0703010b, 0x07030309, 0x07030503, 0x07030507, 0x07030901, 0x07030d01,
-    0x07030f05, 0x07030f0d, 0x07050101, 0x07050305, 0x07050501, 0x07050705, 0x07050709, 0x07050b01,
-    0x07070103, 0x07070301, 0x07070309, 0x07070503, 0x07070507, 0x0707050f, 0x07070701, 0x07070903,
-    0x07070907, 0x0707090f, 0x07070b0b, 0x07070f07, 0x07090107, 0x07090303, 0x0709030d, 0x07090505,
-    0x07090703, 0x07090b05, 0x07090d01, 0x07090d09, 0x070b0103, 0x070b0301, 0x070b0305, 0x070b050b,
-    0x070b0705, 0x070b0909, 0x070b0b0d, 0x070b0f07, 0x070d030d, 0x070d0903, 0x070f0103, 0x070f0107,
-    0x070f0501, 0x070f0505, 0x070f070b, 0x09010101, 0x09010109, 0x09010305, 0x09010501, 0x09010509,
-    0x0901050f, 0x09010705, 0x09010903, 0x09010b01, 0x09010f01, 0x09030105, 0x0903010f, 0x09030303,
-    0x09030307, 0x09030505, 0x09030701, 0x0903070b, 0x09030907, 0x09030b03, 0x09030b0b, 0x09050103,
-    0x09050107, 0x09050301, 0x0905030b, 0x09050503, 0x09050707, 0x09050901, 0x09050b0f, 0x09050d05,
-    0x09050f01, 0x09070109, 0x09070303, 0x09070307, 0x09070501, 0x09070505, 0x09070703, 0x0907070b,
-    0x09090101, 0x09090105, 0x09090509, 0x0909070f, 0x09090901, 0x09090f03, 0x090b010b, 0x090b010f,
-    0x090b0503, 0x090b0d05, 0x090d0307, 0x090d0709, 0x090d0d01, 0x090f0301, 0x090f030b, 0x090f0701,
-    0x090f0907, 0x090f0b03, 0x0b010105, 0x0b010301, 0x0b010309, 0x0b010505, 0x0b010901, 0x0b010909,
-    0x0b01090f, 0x0b010b05, 0x0b010d0d, 0x0b010f09, 0x0b030103, 0x0b030107, 0x0b03010b, 0x0b030305,
-    0x0b030503, 0x0b030705, 0x0b030f05, 0x0b050101, 0x0b050303, 0x0b050507, 0x0b050701, 0x0b05070d,
-    0x0b050b07, 0x0b070105, 0x0b07010f, 0x0b070301, 0x0b07050f, 0x0b070909, 0x0b070b03, 0x0b070d0b,
-    0x0b070f07, 0x0b090103, 0x0b090109, 0x0b090501, 0x0b090705, 0x0b09090d, 0x0b0b0305, 0x0b0b050d,
-    0x0b0b0b03, 0x0b0b0b07, 0x0b0d0905, 0x0b0f0105, 0x0b0f0109, 0x0b0f0505, 0x0d010303, 0x0d010307,
-    0x0d01030b, 0x0d010703, 0x0d010707, 0x0d010d01, 0x0d030101, 0x0d030501, 0x0d03050f, 0x0d030d09,
-    0x0d050305, 0x0d050709, 0x0d050905, 0x0d050b0b, 0x0d050d05, 0x0d050f01, 0x0d070101, 0x0d070309,
-    0x0d070503, 0x0d070901, 0x0d09050b, 0x0d090907, 0x0d090d05, 0x0d0b0101, 0x0d0b0107, 0x0d0b0709,
-    0x0d0b0d01, 0x0d0d010b, 0x0d0d0901, 0x0d0f0303, 0x0d0f0307, 0x0f010101, 0x0f010109, 0x0f01010f,
-    0x0f010501, 0x0f010505, 0x0f01070d, 0x0f010901, 0x0f010b09, 0x0f010d05, 0x0f030105, 0x0f030303,
-    0x0f030509, 0x0f030907, 0x0f03090b, 0x0f050103, 0x0f050109, 0x0f050301, 0x0f05030d, 0x0f050503,
-    0x0f050701, 0x0f050b03, 0x0f070105, 0x0f070705, 0x0f07070b, 0x0f070b07, 0x0f090103, 0x0f09010b,
-    0x0f090307, 0x0f090501, 0x0f090b01, 0x0f0b0505, 0x0f0b0905, 0x0f0d0105, 0x0f0d0703, 0x0f0f0101,
-GGML_TABLE_END()
-
-#define NGRID_IQ2XXS 512
-GGML_TABLE_BEGIN(uint64_t, iq1s_grid, NGRID_IQ2XXS)
-    0xffffffffffff0101, 0xffffffffff01ff00, 0xffffffffff010100, 0xffffffff00000000,
-    0xffffffff01ff00ff, 0xffffffff01ff0001, 0xffffffff0101ffff, 0xffffffff0101ff01,
-    0xffffff00ff000000, 0xffffff000000ff00, 0xffffff00000000ff, 0xffffff0000000100,
-    0xffffff0000010000, 0xffffff0001000000, 0xffffff01ffff00ff, 0xffffff01ff01ff00,
-    0xffffff01ff010100, 0xffffff0100000001, 0xffffff0101ffff00, 0xffffff0101ff0101,
-    0xffffff0101010100, 0xffff00ffff00ff01, 0xffff00ffff0000ff, 0xffff00ff00ff0100,
-    0xffff00ff0100ff00, 0xffff00ff010001ff, 0xffff0000ff0101ff, 0xffff000000ffff00,
-    0xffff000000000000, 0xffff00000001ff01, 0xffff000001000101, 0xffff0000010100ff,
-    0xffff0001ffff0100, 0xffff00010000ff00, 0xffff000100010101, 0xffff000101000000,
-    0xffff01ffffff0000, 0xffff01ffff01ffff, 0xffff01ffff010100, 0xffff01ff00000000,
-    0xffff01ff01ffffff, 0xffff01ff01ff0001, 0xffff01ff0101ffff, 0xffff01ff01010001,
-    0xffff0100ffffff01, 0xffff01000000ffff, 0xffff010000000100, 0xffff010001ff01ff,
-    0xffff010001000000, 0xffff0101ff000000, 0xffff0101000101ff, 0xffff010101ffff01,
-    0xffff01010101ff00, 0xff00ffffff000000, 0xff00ffff00ffff00, 0xff00ffff00000001,
-    0xff00ffff000001ff, 0xff00ffff01010000, 0xff00ff00ffff0000, 0xff00ff00ff00ff00,
-    0xff00ff00ff0000ff, 0xff00ff00ff000100, 0xff00ff00ff010001, 0xff00ff0000ff0001,
-    0xff00ff000000ffff, 0xff00ff0000000000, 0xff00ff000001ff00, 0xff00ff0000010100,
-    0xff00ff0001ff0000, 0xff00ff000100ff00, 0xff00ff0001000100, 0xff00ff01ff000000,
-    0xff00ff0100ff0000, 0xff00ff01000001ff, 0xff00ff0101010001, 0xff0000ff00000000,
-    0xff0000ff0001ff00, 0xff0000ff00010100, 0xff000000ffff0101, 0xff000000ff000000,
-    0xff000000ff01ff00, 0xff00000000ff0000, 0xff0000000000ff00, 0xff000000000000ff,
-    0xff00000000000000, 0xff00000000000001, 0xff00000000000100, 0xff0000000001ffff,
-    0xff00000000010000, 0xff00000001000000, 0xff00000001010100, 0xff000001ff00ff01,
-    0xff000001ff0100ff, 0xff00000100000000, 0xff0000010001ff00, 0xff00000101ff0100,
-    0xff0000010100ff00, 0xff0001ff00ff00ff, 0xff0001ff00000101, 0xff0001ff000100ff,
-    0xff0001ff01000000, 0xff000100ff0001ff, 0xff0001000000ff01, 0xff00010000000000,
-    0xff00010000010001, 0xff00010000010100, 0xff00010001ffff00, 0xff00010001ff0101,
-    0xff00010001010000, 0xff000101ffffffff, 0xff000101ff000101, 0xff00010101ff00ff,
-    0xff00010101000001, 0xff000101010100ff, 0xff01ffffff000101, 0xff01ffffff01ffff,
-    0xff01ffffff01ff01, 0xff01ffffff0101ff, 0xff01ffff00000000, 0xff01ffff01ff0001,
-    0xff01ffff0101ff01, 0xff01ff00ff000000, 0xff01ff0000ff0100, 0xff01ff000000ff01,
-    0xff01ff0000010000, 0xff01ff00010000ff, 0xff01ff01ff01ff00, 0xff01ff0100000101,
-    0xff0100ffffff0000, 0xff0100ffff010000, 0xff0100ff01ff00ff, 0xff0100ff01000100,
-    0xff0100ff010100ff, 0xff010000ffffff01, 0xff01000000000000, 0xff0100000101ff00,
-    0xff010001ffff00ff, 0xff010001ff000100, 0xff01000100ffff00, 0xff01000100010001,
-    0xff01000101ff0001, 0xff010001010001ff, 0xff0101ffffffffff, 0xff0101ffff01ffff,
-    0xff0101ffff010101, 0xff0101ff0000ff00, 0xff0101ff01010001, 0xff010100ff000000,
-    0xff010100ff01ff01, 0xff01010000ff0001, 0xff01010000000100, 0xff01010001000000,
-    0xff0101010100ffff, 0x00ffffff0000ff01, 0x00ffffff000000ff, 0x00ffffff00000100,
-    0x00ffffff00010000, 0x00ffff00ffff0001, 0x00ffff00ff0000ff, 0x00ffff00ff000100,
-    0x00ffff0000000000, 0x00ffff0001000100, 0x00ffff0001010001, 0x00ffff01ff00ff01,
-    0x00ffff0100ff0100, 0x00ffff010000ff00, 0x00ffff01000100ff, 0x00ffff0101ff00ff,
-    0x00ffff010101ff00, 0x00ff00ffffffffff, 0x00ff00ffffff01ff, 0x00ff00ffff000101,
-    0x00ff00ff00000000, 0x00ff00ff000101ff, 0x00ff00ff01010101, 0x00ff0000ff000000,
-    0x00ff0000ff01ffff, 0x00ff000000ff0000, 0x00ff00000000ff00, 0x00ff0000000000ff,
-    0x00ff000000000000, 0x00ff000000000001, 0x00ff000000000100, 0x00ff000000010000,
-    0x00ff000001ffff01, 0x00ff000001000000, 0x00ff0001ff000101, 0x00ff000100ffffff,
-    0x00ff000100000000, 0x00ff0001010001ff, 0x00ff01ffff000000, 0x00ff01ff0001ff00,
-    0x00ff01ff01ff0100, 0x00ff0100ff01ff01, 0x00ff010000ff00ff, 0x00ff010000ff0101,
-    0x00ff010000000000, 0x00ff010000010101, 0x00ff01000100ff00, 0x00ff010001010000,
-    0x00ff0101ffffff00, 0x00ff01010000ff01, 0x00ff010100000100, 0x00ff010101ff0000,
-    0x0000ffffffff0100, 0x0000ffffff00ff00, 0x0000ffffff0000ff, 0x0000ffffff010000,
-    0x0000ffff00000000, 0x0000ffff00010101, 0x0000ffff01ffff01, 0x0000ffff01000100,
-    0x0000ff00ff000000, 0x0000ff00ff01ff00, 0x0000ff00ff0101ff, 0x0000ff0000ff0000,
-    0x0000ff000000ff00, 0x0000ff00000000ff, 0x0000ff0000000000, 0x0000ff0000000001,
-    0x0000ff0000000100, 0x0000ff0000010000, 0x0000ff0001ffffff, 0x0000ff0001ff01ff,
-    0x0000ff0001000000, 0x0000ff000101ffff, 0x0000ff01ffff0101, 0x0000ff01ff010000,
-    0x0000ff0100000000, 0x0000ff0101000101, 0x000000ffffff0001, 0x000000ffff000000,
-    0x000000ff00ff0000, 0x000000ff0000ff00, 0x000000ff000000ff, 0x000000ff00000000,
-    0x000000ff00000001, 0x000000ff00000100, 0x000000ff00010000, 0x000000ff01000000,
-    0x000000ff0101ff00, 0x00000000ffff0000, 0x00000000ff00ff00, 0x00000000ff0000ff,
-    0x00000000ff000000, 0x00000000ff000001, 0x00000000ff000100, 0x00000000ff010000,
-    0x0000000000ffff00, 0x0000000000ff00ff, 0x0000000000ff0000, 0x0000000000ff0001,
-    0x0000000000ff0100, 0x000000000000ffff, 0x000000000000ff00, 0x000000000000ff01,
-    0x00000000000000ff, 0x0000000000000001, 0x00000000000001ff, 0x0000000000000100,
-    0x0000000000000101, 0x000000000001ff00, 0x00000000000100ff, 0x0000000000010000,
-    0x0000000000010001, 0x0000000000010100, 0x0000000001ff0000, 0x000000000100ff00,
-    0x00000000010000ff, 0x0000000001000000, 0x0000000001000001, 0x0000000001000100,
-    0x0000000001010000, 0x00000001ffff01ff, 0x00000001ff000000, 0x0000000100ff0000,
-    0x000000010000ff00, 0x00000001000000ff, 0x0000000100000000, 0x0000000100000001,
-    0x0000000100000100, 0x0000000100010000, 0x0000000101000000, 0x000001ffff00ff00,
-    0x000001ffff010001, 0x000001ffff0101ff, 0x000001ff00ffff01, 0x000001ff0000ffff,
-    0x000001ff00000000, 0x000001ff010000ff, 0x000001ff01010100, 0x00000100ffff0100,
-    0x00000100ff000000, 0x0000010000ff0000, 0x000001000000ff00, 0x00000100000000ff,
-    0x0000010000000000, 0x0000010000000001, 0x0000010000000100, 0x0000010000010000,
-    0x0000010001000000, 0x000001000101ff01, 0x00000101ffff0001, 0x00000101ff01ffff,
-    0x0000010100000000, 0x0000010101010100, 0x0001ffffff000000, 0x0001ffff00ffffff,
-    0x0001ffff00000100, 0x0001ffff0001ff00, 0x0001ffff01000000, 0x0001ff00ffffff00,
-    0x0001ff00ffff01ff, 0x0001ff00ff010000, 0x0001ff0000000000, 0x0001ff0000010001,
-    0x0001ff0001ff0000, 0x0001ff0001010100, 0x0001ff01ff0000ff, 0x0001ff01ff000001,
-    0x0001ff0100ffffff, 0x0001ff010001ffff, 0x0001ff01000101ff, 0x0001ff010100ff01,
-    0x000100ffff00ffff, 0x000100ffff00ff01, 0x000100ffff000100, 0x000100ff00000000,
-    0x000100ff000101ff, 0x000100ff01ff0101, 0x000100ff0100ffff, 0x000100ff01010101,
-    0x00010000ff000000, 0x00010000ff010100, 0x0001000000ff0000, 0x000100000000ff00,
-    0x00010000000000ff, 0x0001000000000000, 0x0001000000000001, 0x0001000000000100,
-    0x0001000000010000, 0x0001000001ffff01, 0x0001000001000000, 0x0001000100ff0101,
-    0x0001000100000000, 0x00010001010100ff, 0x000101ffffff01ff, 0x000101ffffff0101,
-    0x000101ff00010000, 0x000101ff01ff0000, 0x000101ff0100ff01, 0x00010100ffff0000,
-    0x0001010000000000, 0x000101000001ffff, 0x0001010000010101, 0x00010100010001ff,
-    0x00010101ff00ff00, 0x00010101ff010001, 0x0001010100ffffff, 0x0001010100ff01ff,
-    0x00010101000101ff, 0x0001010101ff0000, 0x000101010100ff01, 0x0001010101000101,
-    0x01ffffffffff0101, 0x01ffffffff01ffff, 0x01ffffffff01ff01, 0x01ffffffff0101ff,
-    0x01ffffffff010101, 0x01ffffff00000000, 0x01ffffff01ff01ff, 0x01ffffff01000101,
-    0x01ffffff0101ff01, 0x01ffffff010100ff, 0x01ffff000000ff00, 0x01ffff0000000001,
-    0x01ffff00000001ff, 0x01ffff0000010000, 0x01ffff0001ff0000, 0x01ffff01ffffffff,
-    0x01ffff01ffff01ff, 0x01ffff01ff000000, 0x01ffff01ff01ffff, 0x01ffff01ff0101ff,
-    0x01ffff010100ffff, 0x01ff00ffffff0000, 0x01ff00ffff010000, 0x01ff00ff00ffff01,
-    0x01ff0000ff0000ff, 0x01ff000000000000, 0x01ff00000001ff01, 0x01ff000001ffffff,
-    0x01ff000001010100, 0x01ff0001ffffff01, 0x01ff0001ff010001, 0x01ff000101ff0100,
-    0x01ff000101000001, 0x01ff0001010100ff, 0x01ff01ffff00ffff, 0x01ff01ff00010001,
-    0x01ff01ff01000000, 0x01ff01ff010101ff, 0x01ff0100ff000001, 0x01ff010000ffff00,
-    0x01ff010000000100, 0x01ff010001ff01ff, 0x01ff01000101ffff, 0x01ff0101ffff00ff,
-    0x01ff0101ffff0101, 0x01ff0101ff0101ff, 0x01ff010100010000, 0x0100ffff00ff00ff,
-    0x0100ffff00ff0001, 0x0100ffff00000100, 0x0100ffff0100ff00, 0x0100ff00ffff0000,
-    0x0100ff00ff00ffff, 0x0100ff00ff00ff01, 0x0100ff00ff000100, 0x0100ff00ff010000,
-    0x0100ff0000000000, 0x0100ff00000100ff, 0x0100ff0001ff0101, 0x0100ff0001010101,
-    0x0100ff0100ff00ff, 0x0100ff0100ff0001, 0x0100ff0100000100, 0x0100ff0100010001,
-    0x0100ff0101000000, 0x010000ffff00ff00, 0x010000ff0000ffff, 0x010000ff00000000,
-    0x010000ff010001ff, 0x010000ff01010001, 0x01000000ffffff00, 0x01000000ffff0101,
-    0x01000000ff000000, 0x01000000ff0100ff, 0x01000000ff010101, 0x0100000000ff0000,
-    0x010000000000ff00, 0x01000000000000ff, 0x0100000000000000, 0x0100000000000001,
-    0x0100000000000100, 0x0100000000010000, 0x0100000001000000, 0x0100000100000000,
-    0x01000001000101ff, 0x0100000101ffff01, 0x010001ffff000101, 0x010001ff00ff0100,
-    0x010001ff0000ff00, 0x010001ff000100ff, 0x010001ff01ffffff, 0x01000100ffff0000,
-    0x01000100ff0001ff, 0x0100010000000000, 0x010001000001ff00, 0x0100010001ff0000,
-    0x01000100010000ff, 0x0100010001000101, 0x01000101ff00ff01, 0x0100010100ff0100,
-    0x010001010000ffff, 0x0100010101010001, 0x0101ffffffff0101, 0x0101ffffff0001ff,
-    0x0101ffffff01ffff, 0x0101ffffff010101, 0x0101ffff00000000, 0x0101ffff0101ffff,
-    0x0101ffff010101ff, 0x0101ff00ff000000, 0x0101ff0000ff0100, 0x0101ff000000ff00,
-    0x0101ff0000010000, 0x0101ff00010000ff, 0x0101ff0001000001, 0x0101ff01ff010101,
-    0x0101ff0100000000, 0x0101ff010101ff00, 0x010100ffffff0000, 0x010100ffff010000,
-    0x010100ff00ff01ff, 0x010100ff000000ff, 0x010100ff00000101, 0x010100ff01ffff00,
-    0x01010000ffffff01, 0x01010000ff000100, 0x01010000ff01ff01, 0x0101000000000000,
-    0x01010000000100ff, 0x010100000101ff01, 0x01010001ffff0000, 0x01010001ff00ffff,
-    0x01010001ff010000, 0x0101000101ffffff, 0x0101000101ff01ff, 0x0101000101010101,
-    0x010101ffff01ffff, 0x010101ff00000000, 0x010101ff0001ff01, 0x010101ff0101ffff,
-    0x010101ff010101ff, 0x01010100ffffffff, 0x01010100ff000001, 0x010101000000ff00,
-    0x0101010001010000, 0x0101010100ff0001, 0x010101010001ff01, 0x010101010101ffff,
-GGML_TABLE_END()
-
-#endif // GGML_COMMON_IMPL

ggml-impl.h

@@ -53,23 +53,11 @@ extern "C" {
 //
 #include <arm_neon.h>
 
-#define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-#define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
+#define GGML_COMPUTE_FP16_TO_FP32(x) ((float) (x))
+#define GGML_COMPUTE_FP32_TO_FP16(x) (x)
 
-#define GGML_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
-
-static inline float ggml_compute_fp16_to_fp32(ggml_fp16_t h) {
-    __fp16 tmp;
-    memcpy(&tmp, &h, sizeof(ggml_fp16_t));
-    return (float)tmp;
-}
-
-static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
-    ggml_fp16_t res;
-    __fp16 tmp = f;
-    memcpy(&res, &tmp, sizeof(ggml_fp16_t));
-    return res;
-}
+#define GGML_FP16_TO_FP32(x) ((float) (x))
+#define GGML_FP32_TO_FP16(x) (x)
 
 #else
 
@@ -226,7 +214,8 @@ extern float ggml_table_f32_f16[1 << 16];
 // On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
 // so we define GGML_FP16_TO_FP32 and GGML_FP32_TO_FP16 elsewhere for NEON.
 // This is also true for POWER9.
-#if !defined(GGML_FP16_TO_FP32)
+#if !defined(GGML_FP16_TO_FP32) || !defined(GGML_FP32_TO_FP16)
+
 inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
     uint16_t s;
     memcpy(&s, &f, sizeof(uint16_t));
@@ -234,10 +223,8 @@ inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
 }
 
 #define GGML_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
-#endif
-
-#if !defined(GGML_FP32_TO_FP16)
 #define GGML_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+
 #endif
 
 #define GGML_HASHTABLE_FULL ((size_t)-1)

ggml-kompute.cpp

@@ -1927,10 +1927,10 @@ static ggml_backend_buffer_type_t ggml_backend_kompute_get_default_buffer_type(g
     return ggml_backend_kompute_buffer_type(ctx->device);
 }
 
-static ggml_status ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static bool ggml_backend_kompute_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     auto * ctx = static_cast<ggml_kompute_context *>(backend->context);
     ggml_vk_graph_compute(ctx, cgraph);
-    return GGML_STATUS_SUCCESS;
+    return true;
 }
 
 static bool ggml_backend_kompute_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -1953,17 +1953,11 @@ static struct ggml_backend_i kompute_backend_i = {
     /* .supports_op             = */ ggml_backend_kompute_supports_op,
 };
 
-static ggml_guid_t ggml_backend_kompute_guid() {
-    static ggml_guid guid = { 0x7b, 0x57, 0xdc, 0xaf, 0xde, 0x12, 0x1d, 0x49, 0xfb, 0x35, 0xfa, 0x9b, 0x18, 0x31, 0x1d, 0xca };
-    return &guid;
-}
-
 ggml_backend_t ggml_backend_kompute_init(int device) {
     GGML_ASSERT(s_kompute_context == nullptr);
     s_kompute_context = new ggml_kompute_context(device);
 
     ggml_backend_t kompute_backend = new ggml_backend {
-        /* .guid      = */ ggml_backend_kompute_guid(),
         /* .interface = */ kompute_backend_i,
         /* .context   = */ s_kompute_context,
     };
@@ -1972,7 +1966,7 @@ ggml_backend_t ggml_backend_kompute_init(int device) {
 }
 
 bool ggml_backend_is_kompute(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_kompute_guid());
+    return backend && backend->iface.get_name == ggml_backend_kompute_name;
 }
 
 static ggml_backend_t ggml_backend_reg_kompute_init(const char * params, void * user_data) {

ggml-metal.m

@@ -61,11 +61,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
-    GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
     GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
     GGML_METAL_KERNEL_TYPE_RMS_NORM,
     GGML_METAL_KERNEL_TYPE_GROUP_NORM,
@@ -88,11 +84,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
   //GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
@@ -111,11 +103,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
@@ -131,11 +119,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
@@ -151,11 +135,7 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
     GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
-    GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F32,
     GGML_METAL_KERNEL_TYPE_ROPE_F16,
     GGML_METAL_KERNEL_TYPE_ALIBI_F32,
@@ -163,11 +143,15 @@ enum ggml_metal_kernel_type {
     GGML_METAL_KERNEL_TYPE_IM2COL_F32,
     GGML_METAL_KERNEL_TYPE_UPSCALE_F32,
     GGML_METAL_KERNEL_TYPE_PAD_F32,
-    GGML_METAL_KERNEL_TYPE_ARANGE_F32,
-    GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,
     GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,
     GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,
+    GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F16,
     GGML_METAL_KERNEL_TYPE_CPY_F32_F32,
     GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,
@@ -299,14 +283,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 return NULL;
             }
         } else {
-#if GGML_METAL_EMBED_LIBRARY
-            GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
-
-            extern const char ggml_metallib_start[];
-            extern const char ggml_metallib_end[];
-
-            NSString * src  = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
-#else
             GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath;
@@ -329,7 +305,6 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
                 GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
                 return NULL;
             }
-#endif
 
             @autoreleasepool {
                 // dictionary of preprocessor macros
@@ -426,6 +401,9 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
             id<MTLFunction> metal_function = [metal_library newFunctionWithName:@"kernel_"#name]; \
             kernel->pipeline = [ctx->device newComputePipelineStateWithFunction:metal_function error:&error]; \
             [metal_function release]; \
+            GGML_METAL_LOG_INFO("%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) kernel->pipeline, \
+                    (int) kernel->pipeline.maxTotalThreadsPerThreadgroup, \
+                    (int) kernel->pipeline.threadExecutionWidth); \
             if (error) { \
                 GGML_METAL_LOG_ERROR("%s: error: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
                 [metal_library release]; \
@@ -437,157 +415,141 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
 
         // simd_sum and simd_max requires MTLGPUFamilyApple7
 
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                       add,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                   add_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                       mul,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                   mul_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                       div,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                  true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,           diag_mask_inf_8,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,              get_rows_f32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,              get_rows_f16,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,             get_rows_q4_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,             get_rows_q4_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,             get_rows_q5_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,             get_rows_q5_1,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,             get_rows_q8_0,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,             get_rows_q2_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,             get_rows_q3_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,             get_rows_q4_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,             get_rows_q5_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,          true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,       true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,            get_rows_iq3_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,            get_rows_iq2_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,           get_rows_iq4_nl,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,           get_rows_iq4_xs,        true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,               ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,             ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                      norm,                   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,            mul_mv_f32_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,            mul_mv_f16_f16,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,            mul_mv_f16_f32,         ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,       mul_mv_f16_f32_1row,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,         mul_mv_f16_f32_l4,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,           mul_mv_q4_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,           mul_mv_q4_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,           mul_mv_q5_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,           mul_mv_q5_1_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,           mul_mv_q8_0_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,           mul_mv_q2_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,           mul_mv_q3_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,           mul_mv_q4_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,           mul_mv_q5_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,        ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,          mul_mv_iq3_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,          mul_mv_iq2_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,       ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,         mul_mv_iq4_nl_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,         mul_mv_iq4_xs_f32,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,      ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,      ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,    mul_mv_id_f16_f32_1row, ctx->support_simdgroup_reduction);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,      mul_mv_id_f16_f32_l4,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,        mul_mv_id_q4_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,        mul_mv_id_q4_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,        mul_mv_id_q5_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,        mul_mv_id_q5_1_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,        mul_mv_id_q8_0_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,        mul_mv_id_q2_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,        mul_mv_id_q3_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,        mul_mv_id_q4_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,        mul_mv_id_q5_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,     ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,  ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,       mul_mv_id_iq3_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,       mul_mv_id_iq2_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,    ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,      mul_mv_id_iq4_nl_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,      mul_mv_id_iq4_xs_f32,   ctx->support_simdgroup_reduction);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,         ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,           mul_mm_q4_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,           mul_mm_q5_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,           mul_mm_q5_1_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,           mul_mm_q8_0_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,           mul_mm_q2_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,           mul_mm_q3_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,           mul_mm_q4_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,           mul_mm_q5_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,        ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,          mul_mm_iq3_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,          mul_mm_iq2_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,       ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,         mul_mm_iq4_nl_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,         mul_mm_iq4_xs_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,      ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,        mul_mm_id_q4_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,        mul_mm_id_q5_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,        mul_mm_id_q5_1_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,        mul_mm_id_q8_0_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,        mul_mm_id_q2_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,        mul_mm_id_q3_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,        mul_mm_id_q4_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,        mul_mm_id_q5_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,     ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,  ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,       mul_mm_id_iq3_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,       mul_mm_id_iq2_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,    ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,      mul_mm_id_iq4_nl_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,      mul_mm_id_iq4_xs_f32,   ctx->support_simdgroup_mm);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,               true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,              true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                im2col_f16,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                im2col_f32,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,               upscale_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                   pad_f32,                true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32,    timestep_embedding_f32, true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARANGE_F32,                arange_f32,             true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,       argsort_f32_i32_asc,    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,      argsort_f32_i32_desc,   true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,         true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,           true);
-      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,           true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,            true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                 true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                       sqr,                    true);
-        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD,                       add,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ADD_ROW,                   add_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL,                       mul,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_ROW,                   mul_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV,                       div,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIV_ROW,                   div_row,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE,                     scale,                   true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SCALE_4,                   scale_4,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_TANH,                      tanh,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RELU,                      relu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU,                      gelu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GELU_QUICK,                gelu_quick,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SILU,                      silu,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX,                  soft_max,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SOFT_MAX_4,                soft_max_4,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF,             diag_mask_inf,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_DIAG_MASK_INF_8,           diag_mask_inf_8,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F32,              get_rows_f32,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_F16,              get_rows_f16,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_0,             get_rows_q4_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_1,             get_rows_q4_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_0,             get_rows_q5_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_1,             get_rows_q5_1,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q8_0,             get_rows_q8_0,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q2_K,             get_rows_q2_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q3_K,             get_rows_q3_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q4_K,             get_rows_q4_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q5_K,             get_rows_q5_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_Q6_K,             get_rows_q6_K,           true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,          get_rows_iq2_xxs,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,           get_rows_iq2_xs,         true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,          get_rows_iq3_xxs,        true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,            get_rows_iq1_s,          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,              get_rows_i32,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM,                  rms_norm,                ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM,                group_norm,              ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM,                      norm,                    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,            mul_mv_f32_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,            mul_mv_f16_f16,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,            mul_mv_f16_f32,          ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_1ROW,       mul_mv_f16_f32_1row,     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32_L4,         mul_mv_f16_f32_l4,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_0_F32,           mul_mv_q4_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_1_F32,           mul_mv_q4_1_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_0_F32,           mul_mv_q5_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_1_F32,           mul_mv_q5_1_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q8_0_F32,           mul_mv_q8_0_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q2_K_F32,           mul_mv_q2_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q3_K_F32,           mul_mv_q3_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q4_K_F32,           mul_mv_q4_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q5_K_F32,           mul_mv_q5_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_Q6_K_F32,           mul_mv_q6_K_f32,         ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,        mul_mv_iq2_xxs_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,         mul_mv_iq2_xs_f32,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,        mul_mv_iq3_xxs_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,          mul_mv_iq1_s_f32,        ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,         mul_mv_id_f32_f32,       ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,         mul_mv_id_f16_f16,       ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,         mul_mv_id_f16_f32,       ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_1ROW,    mul_mv_id_f16_f32_1row,  ctx->support_simdgroup_reduction);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32_L4,      mul_mv_id_f16_f32_l4,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_0_F32,        mul_mv_id_q4_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_1_F32,        mul_mv_id_q4_1_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_0_F32,        mul_mv_id_q5_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_1_F32,        mul_mv_id_q5_1_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q8_0_F32,        mul_mv_id_q8_0_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q2_K_F32,        mul_mv_id_q2_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q3_K_F32,        mul_mv_id_q3_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q4_K_F32,        mul_mv_id_q4_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q5_K_F32,        mul_mv_id_q5_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_Q6_K_F32,        mul_mv_id_q6_K_f32,      ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,     mul_mv_id_iq2_xxs_f32,   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,      mul_mv_id_iq2_xs_f32,    ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,     mul_mv_id_iq3_xxs_f32,   ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,       mul_mv_id_iq1_s_f32,     ctx->support_simdgroup_reduction);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,            mul_mm_f32_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,            mul_mm_f16_f32,          ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,           mul_mm_q4_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_1_F32,           mul_mm_q4_1_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_0_F32,           mul_mm_q5_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_1_F32,           mul_mm_q5_1_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q8_0_F32,           mul_mm_q8_0_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q2_K_F32,           mul_mm_q2_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q3_K_F32,           mul_mm_q3_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_K_F32,           mul_mm_q4_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q5_K_F32,           mul_mm_q5_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q6_K_F32,           mul_mm_q6_K_f32,         ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,        mul_mm_iq2_xxs_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,         mul_mm_iq2_xs_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,        mul_mm_iq3_xxs_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,          mul_mm_iq1_s_f32,        ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,         mul_mm_id_f32_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,         mul_mm_id_f16_f32,       ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,        mul_mm_id_q4_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_1_F32,        mul_mm_id_q4_1_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_0_F32,        mul_mm_id_q5_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_1_F32,        mul_mm_id_q5_1_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q8_0_F32,        mul_mm_id_q8_0_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q2_K_F32,        mul_mm_id_q2_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q3_K_F32,        mul_mm_id_q3_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_K_F32,        mul_mm_id_q4_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q5_K_F32,        mul_mm_id_q5_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q6_K_F32,        mul_mm_id_q6_K_f32,      ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,     mul_mm_id_iq2_xxs_f32,   ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,      mul_mm_id_iq2_xs_f32,    ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,     mul_mm_id_iq3_xxs_f32,   ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,       mul_mm_id_iq1_s_f32,     ctx->support_simdgroup_mm);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32,                  rope_f32,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16,                  rope_f16,                true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32,                 alibi_f32,               true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F16,                im2col_f16,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_IM2COL_F32,                im2col_f32,              true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_UPSCALE_F32,               upscale_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_PAD_F32,                   pad_f32,                 true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC,       argsort_f32_i32_asc,     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC,      argsort_f32_i32_desc,    true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_LEAKY_RELU_F32,            leaky_relu_f32,          true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64,    flash_attn_ext_f16_h64,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80,    flash_attn_ext_f16_h80,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96,    flash_attn_ext_f16_h96,  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112,   flash_attn_ext_f16_h112, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128,   flash_attn_ext_f16_h128, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256,   flash_attn_ext_f16_h256, true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F16,               cpy_f32_f16,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_F32,               cpy_f32_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q8_0,              cpy_f32_q8_0,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_0,              cpy_f32_q4_0,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q4_1,              cpy_f32_q4_1,            true);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_0,              cpy_f32_q5_0,            true);
+      //GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F32_Q5_1,              cpy_f32_q5_1,            true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F16,               cpy_f16_f16,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CPY_F16_F32,               cpy_f16_f32,             true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_CONCAT,                    concat,                  true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SQR,                       sqr,                     true);
+        GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SUM_ROWS,                  sum_rows,                true);
     }
 
     [metal_library release];
@@ -701,10 +663,9 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
             return false;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
-        case GGML_OP_ARANGE:
-        case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
         case GGML_OP_LEAKY_RELU:
+        case GGML_OP_FLASH_ATTN_EXT:
             return true;
         case GGML_OP_MUL_MAT:
         case GGML_OP_MUL_MAT_ID:
@@ -748,7 +709,7 @@ static bool ggml_metal_supports_op(const struct ggml_metal_context * ctx, const
     }
 }
 
-static enum ggml_status ggml_metal_graph_compute(
+static bool ggml_metal_graph_compute(
         struct ggml_metal_context * ctx,
                struct ggml_cgraph * gf) {
 
@@ -1097,8 +1058,7 @@ static enum ggml_status ggml_metal_graph_compute(
                     {
                         GGML_ASSERT(ggml_is_contiguous(src0));
 
-                        float scale;
-                        memcpy(&scale, dst->op_params, sizeof(scale));
+                        const float scale = *(const float *) dst->op_params;
 
                         int64_t n = ggml_nelements(dst);
 
@@ -1241,6 +1201,8 @@ static enum ggml_status ggml_metal_graph_compute(
                     } break;
                 case GGML_OP_SOFT_MAX:
                     {
+                        GGML_ASSERT(!src1 || src1->type == GGML_TYPE_F16);
+
                         int nth = 32; // SIMD width
 
                         id<MTLComputePipelineState> pipeline = nil;
@@ -1257,15 +1219,11 @@ static enum ggml_status ggml_metal_graph_compute(
                             pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SOFT_MAX].pipeline;
                         }
 
-                        float scale;
-                        float max_bias;
-
-                        memcpy(&scale,    ((int32_t *) dst->op_params) + 0, sizeof(scale));
-                        memcpy(&max_bias, ((int32_t *) dst->op_params) + 1, sizeof(max_bias));
+                        const float scale    = ((float *) dst->op_params)[0];
+                        const float max_bias = ((float *) dst->op_params)[1];
 
                         const int64_t nrows_x = ggml_nrows(src0);
                         const int64_t nrows_y = src0->ne[1];
-
                         const uint32_t n_head_kv   = nrows_x/nrows_y;
                         const uint32_t n_head_log2 = 1u << (uint32_t) floorf(log2f((float) n_head_kv));
 
@@ -1388,11 +1346,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                             }
 
@@ -1527,36 +1481,12 @@ static enum ggml_status ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
-                                    } break;
                                 case GGML_TYPE_IQ1_S:
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1589,9 +1519,9 @@ static enum ggml_status ggml_metal_graph_compute(
                             [encoder setBytes:&r2   length:sizeof(r2)   atIndex:17];
                             [encoder setBytes:&r3   length:sizeof(r3)   atIndex:18];
 
-                            if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1  ||
-                                src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1  || src0t == GGML_TYPE_Q8_0 ||
-                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ2_S) {
+                            if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
+                                src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
+                                src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
@@ -1599,16 +1529,11 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                            else if (src0t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
                             else if (src0t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1702,11 +1627,7 @@ static enum ggml_status ggml_metal_graph_compute(
                                 case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
                                 case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
                                 case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
-                                case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32  ].pipeline; break;
                                 case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32  ].pipeline; break;
-                                case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
-                                case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
                                 default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                             }
 
@@ -1844,36 +1765,12 @@ static enum ggml_status ggml_metal_graph_compute(
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ3_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ2_S:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
-                                    } break;
                                 case GGML_TYPE_IQ1_S:
                                     {
                                         nth0 = 4;
                                         nth1 = 16;
                                         pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
                                     } break;
-                                case GGML_TYPE_IQ4_NL:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
-                                    } break;
-                                case GGML_TYPE_IQ4_XS:
-                                    {
-                                        nth0 = 4;
-                                        nth1 = 16;
-                                        pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
-                                    } break;
                                 default:
                                     {
                                         GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1922,9 +1819,9 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
                             }
 
-                            if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1  ||
-                                src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1  || src2t == GGML_TYPE_Q8_0 ||
-                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S || src2t == GGML_TYPE_IQ2_S) {
+                            if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
+                                src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
+                                src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
                             else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
@@ -1932,16 +1829,11 @@ static enum ggml_status ggml_metal_graph_compute(
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ3_XXS || src2t == GGML_TYPE_IQ3_S) {
-                                const int mem_size = src2t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
+                            else if (src2t == GGML_TYPE_IQ3_XXS) {
+                                const int mem_size = 256*4+128;
                                 [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
-                            else if (src2t == GGML_TYPE_IQ4_NL || src2t == GGML_TYPE_IQ4_XS) {
-                                const int mem_size = 32*sizeof(float);
-                                [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
-                                [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
-                            }
                             else if (src2t == GGML_TYPE_Q4_K) {
                                 [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                             }
@@ -1983,11 +1875,7 @@ static enum ggml_status ggml_metal_graph_compute(
                             case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
                             case GGML_TYPE_IQ2_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
                             case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
-                            case GGML_TYPE_IQ3_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S  ].pipeline; break;
-                            case GGML_TYPE_IQ2_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S  ].pipeline; break;
                             case GGML_TYPE_IQ1_S:   pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S  ].pipeline; break;
-                            case GGML_TYPE_IQ4_NL:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
-                            case GGML_TYPE_IQ4_XS:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
                             case GGML_TYPE_I32:     pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32    ].pipeline; break;
                             default: GGML_ASSERT(false && "not implemented");
                         }
@@ -2097,7 +1985,6 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         //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));
 
@@ -2312,50 +2199,6 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(ne1, ne2, ne3) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                     } break;
-                case GGML_OP_ARANGE:
-                    {
-                        GGML_ASSERT(dst->type == GGML_TYPE_F32);
-
-                        float start;
-                        float step;
-
-                        memcpy(&start, ((int32_t *) dst->op_params) + 0, sizeof(float));
-                        memcpy(&step,  ((int32_t *) dst->op_params) + 2, sizeof(float));
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARANGE_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_dst  offset:offs_dst    atIndex:0];
-                        [encoder setBytes:&ne0   length:sizeof(ne0)   atIndex:1];
-                        [encoder setBytes:&start length:sizeof(start) atIndex:2];
-                        [encoder setBytes:&step  length:sizeof(step)  atIndex:3];
-
-                        const int nth = MIN(1024, ne0);
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(1, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                    } break;
-                case GGML_OP_TIMESTEP_EMBEDDING:
-                    {
-                        GGML_ASSERT(src0->type == GGML_TYPE_F32);
-
-                        const int dim        = dst->op_params[0];
-                        const int max_period = dst->op_params[1];
-
-                        const int half = dim / 2;
-
-                        id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_TIMESTEP_EMBEDDING_F32].pipeline;
-
-                        [encoder setComputePipelineState:pipeline];
-                        [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
-                        [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
-                        [encoder setBytes:&nb1   length:sizeof(nb1) atIndex:2];
-                        [encoder setBytes:&dim   length:sizeof(dim) atIndex:3];
-                        [encoder setBytes:&max_period length:sizeof(max_period) atIndex:4];
-
-                        const int nth = MIN(1024, half);
-
-                        [encoder dispatchThreadgroups:MTLSizeMake(ne00, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
-                    } break;
                 case GGML_OP_ARGSORT:
                     {
                         GGML_ASSERT(src0->type == GGML_TYPE_F32);
@@ -2368,8 +2211,8 @@ static enum ggml_status ggml_metal_graph_compute(
                         id<MTLComputePipelineState> pipeline = nil;
 
                         switch (order) {
-                            case GGML_SORT_ORDER_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
-                            case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
+                            case GGML_SORT_ASC:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline;  break;
+                            case GGML_SORT_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
                             default: GGML_ASSERT(false);
                         };
 
@@ -2398,6 +2241,106 @@ static enum ggml_status ggml_metal_graph_compute(
 
                         [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
                     } break;
+                case GGML_OP_FLASH_ATTN_EXT:
+                    {
+                        GGML_ASSERT(ne00 % 4 == 0);
+                        GGML_ASSERT(src0->type == GGML_TYPE_F32);
+
+                        struct ggml_tensor * src3 = gf->nodes[i]->src[3];
+
+                        GGML_ASSERT(ggml_are_same_shape(src1, src2));
+                        GGML_ASSERT(src3);
+
+                        size_t offs_src3 = 0;
+
+                        id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
+
+                        GGML_ASSERT(!src3 || src3->type == GGML_TYPE_F16);
+                        GGML_ASSERT(!src3 || src3->ne[1] >= GGML_PAD(src0->ne[1], 8) &&
+                                "the Flash-Attention Metal kernel requires the mask to be padded to 8 and at least n_queries big");
+
+                        const int64_t  ne30 = src3 ? src3->ne[0] : 0; GGML_UNUSED(ne30);
+                        const int64_t  ne31 = src3 ? src3->ne[1] : 0;
+                        const int64_t  ne32 = src3 ? src3->ne[2] : 0; GGML_UNUSED(ne32);
+                        const int64_t  ne33 = src3 ? src3->ne[3] : 0; GGML_UNUSED(ne33);
+
+                        const uint64_t nb30 = src3 ? src3->nb[0] : 0; GGML_UNUSED(nb30);
+                        const uint64_t nb31 = src3 ? src3->nb[1] : 0;
+                        const uint64_t nb32 = src3 ? src3->nb[2] : 0; GGML_UNUSED(nb32);
+                        const uint64_t nb33 = src3 ? src3->nb[3] : 0; GGML_UNUSED(nb33);
+
+                        const enum ggml_type src2t = src2 ? src2->type : GGML_TYPE_COUNT; GGML_UNUSED(src2t);
+
+                        float scale;
+                        memcpy(&scale, dst->op_params, sizeof(float));
+
+                        id<MTLComputePipelineState> pipeline = nil;
+
+                        switch (ne00) {
+                            case 64:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H64 ].pipeline; break;
+                            case 80:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H80 ].pipeline; break;
+                            case 96:  pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H96 ].pipeline; break;
+                            case 112: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H112].pipeline; break;
+                            case 128: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H128].pipeline; break;
+                            case 256: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_FLASH_ATTN_EXT_F16_H256].pipeline; break;
+                            default:
+                                {
+                                    GGML_METAL_LOG_ERROR("unsupported size: %lld\n", ne00);
+                                    GGML_METAL_LOG_ERROR("add template specialization for this size\n");
+                                    GGML_ASSERT(false && "add template specialization for this size");
+                                }
+                        }
+
+                        // TODO: extend if necessary
+                        [encoder setComputePipelineState:pipeline];
+                        [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                        [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
+                        [encoder setBuffer:id_src2 offset:offs_src2        atIndex:2];
+                        [encoder setBuffer:id_src3 offset:offs_src3        atIndex:3];
+                        [encoder setBuffer:id_dst  offset:offs_dst         atIndex:4];
+                        [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:5];
+                        [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:6];
+                        [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:7];
+                        [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:8];
+                        [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:9];
+                        [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:10];
+                        [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:11];
+                        [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:12];
+                        [encoder setBytes:&ne10    length:sizeof( int64_t) atIndex:13];
+                        [encoder setBytes:&ne11    length:sizeof( int64_t) atIndex:14];
+                        [encoder setBytes:&ne12    length:sizeof( int64_t) atIndex:15];
+                        [encoder setBytes:&ne13    length:sizeof( int64_t) atIndex:16];
+                        [encoder setBytes:&nb10    length:sizeof(uint64_t) atIndex:17];
+                        [encoder setBytes:&nb11    length:sizeof(uint64_t) atIndex:18];
+                        [encoder setBytes:&nb12    length:sizeof(uint64_t) atIndex:19];
+                        [encoder setBytes:&nb13    length:sizeof(uint64_t) atIndex:20];
+                        [encoder setBytes:&ne31    length:sizeof( int64_t) atIndex:21];
+                        [encoder setBytes:&nb31    length:sizeof(uint64_t) atIndex:22];
+                        [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:23];
+                        [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:24];
+                        [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:25];
+                        [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:26];
+                        [encoder setBytes:&scale   length:sizeof(   float) atIndex:27];
+
+                        const int64_t nqptg = 8;  // queries per threadgroup    !! sync with kernel template arguments !!
+                        const int64_t ncpsg = 32; // cache values per simdgroup !! sync with kernel template arguments !!
+
+                        GGML_ASSERT(nqptg <= 32);
+                        GGML_ASSERT(nqptg  % 8  == 0);
+                        GGML_ASSERT(ncpsg  % 32 == 0);
+
+                        // simdgroups per threadgroup (a.k.a. warps)
+                        // for small batches use more simdgroups (needs more tests, to confirm if it's worth it)
+                        const int64_t nsg = ne01 <= nqptg ? MAX(4, MIN(ne11/ncpsg, (int64_t) pipeline.maxTotalThreadsPerThreadgroup/32)) : 4;
+
+                        const size_t smem = nqptg*(ne00 + nsg*(ncpsg + nqptg))*(sizeof(float)/2);
+
+                        //printf("smem: %zu, max: %zu\n", smem, ctx->device.maxThreadgroupMemoryLength);
+                        GGML_ASSERT(smem <= ctx->device.maxThreadgroupMemoryLength);
+                        [encoder setThreadgroupMemoryLength:smem atIndex:0];
+
+                        [encoder dispatchThreadgroups:MTLSizeMake((ne01 + nqptg - 1)/nqptg, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(32, nsg, 1)];
+                    } break;
                 case GGML_OP_DUP:
                 case GGML_OP_CPY:
                 case GGML_OP_CONT:
@@ -2484,7 +2427,7 @@ static enum ggml_status ggml_metal_graph_compute(
         MTLCommandBufferStatus status = [command_buffer status];
         if (status != MTLCommandBufferStatusCompleted) {
             GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
-            return GGML_STATUS_FAILED;
+            return false;
         }
     }
 
@@ -2493,7 +2436,7 @@ static enum ggml_status ggml_metal_graph_compute(
     }
 
     }
-    return GGML_STATUS_SUCCESS;
+    return true;
 }
 
 ////////////////////////////////////////////////////////////////////////////////
@@ -2600,10 +2543,13 @@ GGML_CALL static const char * ggml_backend_metal_buffer_type_get_name(ggml_backe
     UNUSED(buft);
 }
 
-static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
+static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device, size_t size_aligned) {
+#ifndef GGML_METAL_NDEBUG
 #if TARGET_OS_OSX || (TARGET_OS_IOS && __clang_major__ >= 15)
     if (@available(macOS 10.12, iOS 16.0, *)) {
-        GGML_METAL_LOG_INFO(", (%8.2f / %8.2f)",
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f / %8.2f)",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
                 device.currentAllocatedSize / 1024.0 / 1024.0,
                 device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
 
@@ -2613,10 +2559,15 @@ static void ggml_backend_metal_log_allocated_size(id<MTLDevice> device) {
             GGML_METAL_LOG_INFO("\n");
         }
     } else {
-        GGML_METAL_LOG_INFO(", (%8.2f)\n", device.currentAllocatedSize / 1024.0 / 1024.0);
+        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, (%8.2f)\n",
+                __func__,
+                size_aligned / 1024.0 / 1024.0,
+                device.currentAllocatedSize / 1024.0 / 1024.0);
     }
+#endif
 #endif
     UNUSED(device);
+    UNUSED(size_aligned);
 }
 
 GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
@@ -2650,8 +2601,7 @@ GGML_CALL static ggml_backend_buffer_t ggml_backend_metal_buffer_type_alloc_buff
         return NULL;
     }
 
-    GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
-    ggml_backend_metal_log_allocated_size(device);
+    ggml_backend_metal_log_allocated_size(device, size_aligned);
 
     return ggml_backend_buffer_init(buft, ggml_backend_metal_buffer_i, ctx, size);
 }
@@ -2738,7 +2688,7 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
             return false;
         }
 
-        GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB", __func__, size_aligned / 1024.0 / 1024.0);
+        ggml_backend_metal_log_allocated_size(device, size_aligned);
 
         ++ctx->n_buffers;
     } else {
@@ -2761,7 +2711,8 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
                 return false;
             }
 
-            GGML_METAL_LOG_INFO("%s: allocated buffer, size = %8.2f MiB, offs = %12ld", __func__, size_step_aligned / 1024.0 / 1024.0, i);
+            ggml_backend_metal_log_allocated_size(device, size_step_aligned);
+
             if (i + size_step < size) {
                 GGML_METAL_LOG_INFO("\n");
             }
@@ -2770,8 +2721,6 @@ GGML_CALL ggml_backend_buffer_t ggml_backend_metal_buffer_from_ptr(void * data,
         }
     }
 
-    ggml_backend_metal_log_allocated_size(device);
-
     return ggml_backend_buffer_init(ggml_backend_metal_buffer_type(), ggml_backend_metal_buffer_i, ctx, size);
 }
 
@@ -2795,7 +2744,7 @@ GGML_CALL static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffe
     UNUSED(backend);
 }
 
-GGML_CALL static enum ggml_status ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+GGML_CALL static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;
 
     return ggml_metal_graph_compute(metal_ctx, cgraph);
@@ -2827,11 +2776,6 @@ void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void *
     ggml_metal_log_user_data = user_data;
 }
 
-static ggml_guid_t ggml_backend_metal_guid(void) {
-    static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
-    return &guid;
-}
-
 ggml_backend_t ggml_backend_metal_init(void) {
     struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
 
@@ -2842,7 +2786,6 @@ ggml_backend_t ggml_backend_metal_init(void) {
     ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
 
     *metal_backend = (struct ggml_backend) {
-        /* .guid      = */ ggml_backend_metal_guid(),
         /* .interface = */ ggml_backend_metal_i,
         /* .context   = */ ctx,
     };
@@ -2851,7 +2794,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
 }
 
 bool ggml_backend_is_metal(ggml_backend_t backend) {
-    return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
+    return backend && backend->iface.get_name == ggml_backend_metal_name;
 }
 
 void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {

ggml-opencl.cpp

@@ -1354,7 +1354,7 @@ static void ggml_cl_pool_free(cl_mem mem, size_t size) {
 }
 
 void ggml_cl_free_data(const struct ggml_tensor* tensor) {
-    if (tensor->backend != GGML_BACKEND_TYPE_GPU) {
+    if (tensor->backend != GGML_BACKEND_GPU) {
         return;
     }
 
@@ -1412,7 +1412,7 @@ static cl_int ggml_cl_h2d_tensor_2d(cl_command_queue queue, cl_mem dst, size_t o
 }
 
 static void ggml_cl_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
@@ -1476,7 +1476,7 @@ void ggml_cl_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src
 }
 
 static void ggml_cl_add_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
-    GGML_ASSERT(src1->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(src1->backend == GGML_BACKEND_GPU);
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
     const int64_t ne02 = src0->ne[2];
@@ -1566,13 +1566,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(float) * x_ne, &x_size);
     }
-    cl_mem d_Y = src1->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
-    cl_mem d_D =  dst->backend == GGML_BACKEND_TYPE_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
+    cl_mem d_Y = src1->backend == GGML_BACKEND_GPU ? (cl_mem) src1->extra : ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
+    cl_mem d_D =  dst->backend == GGML_BACKEND_GPU ? (cl_mem)  dst->extra : ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
 
     size_t x_offset = 0;
 
@@ -1580,7 +1580,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                if (src0->backend == GGML_BACKEND_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
@@ -1589,7 +1589,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
 
                 for (int64_t i12 = i02 * r2, e12 = i12 + r2; i12 < e12; i12++) {
                     // copy src1 to device
-                    if (src1->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (src1->backend == GGML_BACKEND_CPU) {
                         CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Y, 0, src1, i13, i12, NULL));
                     }
 
@@ -1612,7 +1612,7 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host
-                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (dst->backend == GGML_BACKEND_CPU) {
                         float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                         CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(float) * d_ne, d, 1, &ev_sgemm, NULL));
                     }
@@ -1621,13 +1621,13 @@ static void ggml_cl_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
-    if (src1->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src1->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_Y, y_size);
     }
-    if (dst->backend != GGML_BACKEND_TYPE_GPU) {
+    if (dst->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_D, d_size);
     }
 }
@@ -1670,7 +1670,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
     size_t y_size;
     size_t d_size;
     cl_mem d_X;
-    if (src0->backend == GGML_BACKEND_TYPE_GPU) { // NOLINT
+    if (src0->backend == GGML_BACKEND_GPU) { // NOLINT
         d_X = (cl_mem) src0->extra;
     } else {
         d_X = ggml_cl_pool_malloc(sizeof(ggml_fp16_t) * x_ne, &x_size);
@@ -1687,7 +1687,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         // TODO: copy src0 here when r3>1
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
-                if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                if (src0->backend == GGML_BACKEND_GPU) {
                     x_offset = (i03 * ne02 + i02) * x_ne;
                 } else {
                     // copy src0 to device
@@ -1741,7 +1741,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
                     }
 
                     // copy dst to host, then convert to float
-                    if (dst->backend == GGML_BACKEND_TYPE_CPU) {
+                    if (dst->backend == GGML_BACKEND_CPU) {
                         CL_CHECK(clEnqueueReadBuffer(queue, d_D, true, 0, sizeof(ggml_fp16_t) * d_ne, tmp, 1, &ev_sgemm, NULL));
                         float * d = (float *) ((char *) dst->data + i12*nb2 + i13*nb3);
                         ggml_fp16_to_fp32_row(tmp, d, d_ne);
@@ -1753,7 +1753,7 @@ static void ggml_cl_mul_mat_f16(const ggml_tensor * src0, const ggml_tensor * sr
         }
     }
 
-    if (src0->backend != GGML_BACKEND_TYPE_GPU) {
+    if (src0->backend != GGML_BACKEND_GPU) {
         ggml_cl_pool_free(d_X, x_size);
     }
     ggml_cl_pool_free(d_Y, y_size);
@@ -1798,7 +1798,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     cl_mem d_Y = ggml_cl_pool_malloc(sizeof(float) * y_ne, &y_size);
     cl_mem d_D = ggml_cl_pool_malloc(sizeof(float) * d_ne, &d_size);
     cl_mem d_Q;
-    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+    if (src0->backend == GGML_BACKEND_CPU) {
         d_Q = ggml_cl_pool_malloc(q_sz, &q_size);
     }
 
@@ -1817,10 +1817,10 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
         for (int64_t i13 = i03 * r3, e13 = i13 + r3; i13 < e13; i13++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
                 // copy src0 to device if necessary
-                if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+                if (src0->backend == GGML_BACKEND_CPU) {
                     events.emplace_back();
                     CL_CHECK(ggml_cl_h2d_tensor_2d(queue, d_Q, 0, src0, i03, i02, events.data() + ev_idx++));
-                } else if (src0->backend == GGML_BACKEND_TYPE_GPU) {
+                } else if (src0->backend == GGML_BACKEND_GPU) {
                     d_Q = (cl_mem) src0->extra;
                 } else {
                     GGML_ASSERT(false);
@@ -1829,7 +1829,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
                 if (!mul_mat_vec) {
                     // convert src0 to fp32 on device
                     const size_t global = x_ne / global_denom;
-                    const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                    const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 0, sizeof(cl_mem), &d_Q));
                     CL_CHECK(clSetKernelArg(*to_fp32_cl, 1, sizeof(cl_mem), &d_X));
                     CL_CHECK(clEnqueueNDRangeKernel(queue, *to_fp32_cl, 1, &offset, &global, local > 0 ? &local : NULL, events.size(), !events.empty() ? events.data() : NULL, NULL));
@@ -1843,7 +1843,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
 
                         // compute
                         const size_t global = ne01 * local;
-                        const size_t offset = src0->backend == GGML_BACKEND_TYPE_GPU ? (i03 * ne02 + i02) * x_bps : 0;
+                        const size_t offset = src0->backend == GGML_BACKEND_GPU ? (i03 * ne02 + i02) * x_bps : 0;
                         const cl_int ncols = ne00;
                         events.emplace_back();
                         CL_CHECK(clSetKernelArg(*dmmv, 0, sizeof(cl_mem), &d_Q));
@@ -1895,7 +1895,7 @@ static void ggml_cl_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor *
     }
     ggml_cl_pool_free(d_Y, y_size);
     ggml_cl_pool_free(d_D, d_size);
-    if (src0->backend == GGML_BACKEND_TYPE_CPU) {
+    if (src0->backend == GGML_BACKEND_CPU) {
         ggml_cl_pool_free(d_Q, q_size);
     }
 }
@@ -1911,7 +1911,7 @@ bool ggml_cl_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_TYPE_GPU)) {
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_GPU)) {
         return true;
     }
 
@@ -1993,7 +1993,7 @@ void ggml_cl_transform_tensor(void * data, ggml_tensor * tensor) {
     CL_CHECK(clFinish(queue));
 
     tensor->extra = dst;
-    GGML_ASSERT(tensor->backend == GGML_BACKEND_TYPE_GPU);
+    GGML_ASSERT(tensor->backend == GGML_BACKEND_GPU);
 }
 
 // ggml-backend
@@ -2045,7 +2045,7 @@ static void ggml_backend_opencl_buffer_init_tensor(ggml_backend_buffer_t buffer,
         ctx->sub_buffers.push_back(sub_buffer);
         tensor->extra = sub_buffer;
     }
-    tensor->backend = GGML_BACKEND_TYPE_GPU;
+    tensor->backend = GGML_BACKEND_GPU;
 }
 
 static void ggml_backend_opencl_buffer_set_tensor(ggml_backend_buffer_t buffer, ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
@@ -2231,7 +2231,7 @@ static ggml_backend_buffer_type_t ggml_backend_opencl_get_default_buffer_type(gg
     GGML_UNUSED(backend);
 }
 
-static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
+static bool ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggml_cgraph * graph) {
     for (int i = 0; i < graph->n_nodes; ++i) {
         ggml_tensor * node = graph->nodes[i];
         switch (node->op) {
@@ -2246,7 +2246,7 @@ static ggml_status ggml_backend_opencl_graph_compute(ggml_backend_t backend, ggm
         }
     }
 
-    return GGML_STATUS_SUCCESS;
+    return true;
 
     GGML_UNUSED(backend);
 }

ggml-quants.h

@@ -1,9 +1,9 @@
 #pragma once
 
-// GGML internal header
-
 #include "ggml-impl.h"
 
+// GGML internal header
+
 #include <stdint.h>
 #include <stddef.h>
 
@@ -182,15 +182,6 @@ typedef struct {
 } block_iq2_xs;
 static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
 
-// 2.5625 bpw quants
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t scales[QK_K/32];
-} block_iq2_s;
-static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
-
 // (Almost) "true" 3-bit quantization.
 // Due to the need to use blocks as per ggml design, it ends up using
 // 3.0625 bpw because of the 16-bit scale for each block of 256.
@@ -200,21 +191,6 @@ typedef struct {
 } block_iq3_xxs;
 static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 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_fp16_t d;
-    uint8_t qs[QK_K/4];
-    uint8_t qh[QK_K/32];
-    uint8_t signs[QK_K/8];
-    uint8_t scales[IQ3S_N_SCALE];
-} block_iq3_s;
-static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
-
 typedef struct {
     ggml_fp16_t d;
     uint8_t qs[QK_K/8];
@@ -222,27 +198,6 @@ typedef struct {
 } block_iq1_s;
 static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
 
-// Non-linear quants
-#define QK4_NL 32
-typedef struct {
-    ggml_fp16_t d;
-    uint8_t qs[QK4_NL/2];
-} block_iq4_nl;
-static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
-
-#if QK_K == 64
-#define block_iq4_xs block_iq4_nl
-//typedef struct block_iq4_nl block_iq4_xs;
-#else
-typedef struct {
-    ggml_fp16_t d;
-    uint16_t scales_h;
-    uint8_t  scales_l[QK_K/64];
-    uint8_t  qs[QK_K/2];
-} block_iq4_xs;
-static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
-#endif
-
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -261,12 +216,7 @@ void quantize_row_q4_K_reference(const float * GGML_RESTRICT x, block_q4_K * GGM
 void quantize_row_q5_K_reference(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int k);
 void quantize_row_q6_K_reference(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int k);
 void quantize_row_q8_K_reference(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int k);
-
 void quantize_row_iq3_xxs_reference(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl_reference (const float * GGML_RESTRICT x, block_iq4_nl  * GGML_RESTRICT y, int k);
-void quantize_row_iq4_xs_reference (const float * GGML_RESTRICT x, block_iq4_xs  * GGML_RESTRICT y, int k);
-void quantize_row_iq3_s_reference  (const float * GGML_RESTRICT x, block_iq3_s   * GGML_RESTRICT y, int k);
-void quantize_row_iq2_s_reference  (const float * GGML_RESTRICT x, block_iq2_s   * GGML_RESTRICT y, int k);
 
 void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
@@ -281,12 +231,7 @@ void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-
 void quantize_row_iq3_xxs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq3_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
-void quantize_row_iq2_s  (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int k);
 
 // Dequantization
 void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
@@ -302,15 +247,10 @@ void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRI
 void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-
 void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq2_xs (const block_iq2_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq2_s  (const block_iq2_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 void dequantize_row_iq1_s  (const block_iq1_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq4_nl (const block_iq4_nl  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq4_xs (const block_iq4_xs  * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
-void dequantize_row_iq3_s  (const block_iq3_s   * GGML_RESTRICT x, float * GGML_RESTRICT y, int k);
 
 // Dot product
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
@@ -324,36 +264,27 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const voi
 void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-
 void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq2_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_iq1_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
-void ggml_vec_dot_iq3_s_q8_K  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 
+//
 // Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
-size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq2_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq1_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_iq3_s  (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-
-size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
-size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int nrows, int n_per_row, const float * imatrix);
+//
+size_t quantize_iq2_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq2_xs (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq3_xxs(const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_iq1_s  (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q2_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q3_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q6_K   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q4_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_0   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
+size_t quantize_q5_1   (const float * src, void * dst, int nrows, int n_per_row, int64_t * hist, const float * imatrix);
 
 void iq2xs_init_impl(enum ggml_type type);
 void iq2xs_free_impl(enum ggml_type type);

ggml-sycl.h

@@ -24,11 +24,6 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 GGML_API void   ggml_backend_sycl_print_sycl_devices(void);
 GGML_API GGML_CALL void   ggml_sycl_get_gpu_list(int *id_list, int max_len);
 GGML_API GGML_CALL void   ggml_sycl_get_device_description(int device, char *description, size_t description_size);
-GGML_API GGML_CALL int   ggml_backend_sycl_get_device_count();
-GGML_API GGML_CALL ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const float * tensor_split);
-GGML_API GGML_CALL void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
-GGML_API GGML_CALL int ggml_backend_sycl_get_device_index(int device_id);
-
 #ifdef  __cplusplus
 }
 #endif