sync : ggml

* ggml : try fix 32-bit arm compat

* ggml : fix cont

.devops/nix/docker.nix

@@ -0,0 +1,37 @@
+{
+  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

@@ -1,5 +1,6 @@
 {
   lib,
+  glibc,
   config,
   stdenv,
   mkShell,
@@ -30,6 +31,11 @@
   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
@@ -41,10 +47,7 @@ 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" ]
@@ -167,6 +170,9 @@ 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 =
@@ -181,7 +187,7 @@ effectiveStdenv.mkDerivation (
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" true)
+        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
@@ -190,6 +196,7 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
+        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -255,11 +262,11 @@ 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 || useVulkan) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL) 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) || (useVulkan && effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,5 +12,8 @@ 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

@@ -0,0 +1,27 @@
+{
+  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

@@ -7,3 +7,5 @@ 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

@@ -37,6 +37,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -65,7 +67,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -100,7 +102,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
@@ -143,6 +145,28 @@ 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
 
@@ -244,6 +268,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -277,7 +303,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_METAL=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -665,8 +691,7 @@ jobs:
         run: |
           cd examples/llama.android
 
-          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
-          ./gradlew build --no-daemon -Pskip-armeabi-v7a
+          ./gradlew build --no-daemon
 
 #  freeBSD-latest:
 #    runs-on: macos-12

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

@@ -19,7 +19,6 @@ on:
 
 jobs:
   nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -37,8 +36,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -46,7 +45,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - 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: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -37,7 +37,6 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
-    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -51,8 +50,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -60,7 +59,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
+        name: llama-cpp
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

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

@@ -3,12 +3,14 @@ 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'
@@ -26,4 +28,4 @@ jobs:
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh nocleanup
+        run:  bash scripts/check-requirements.sh

.github/workflows/server.yml

@@ -0,0 +1,136 @@
+# 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

@@ -23,11 +23,13 @@
 .clang-tidy
 .vs/
 .vscode/
+.idea/
 
 lcov-report/
 gcovr-report/
 
 build*
+cmake-build-*
 out/
 tmp/
 
@@ -43,6 +45,7 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/gritlm
 /imatrix
 /infill
 /libllama.so

CMakeLists.txt

@@ -55,6 +55,9 @@ option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"
 option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
 option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
+# build
+option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
+
 # sanitizers
 option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
 option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
@@ -107,22 +110,20 @@ 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)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
+option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
-
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
-if (LLAMA_PERF)
-    add_definitions(-DGGML_PERF)
-endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -130,6 +131,7 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
+
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -140,6 +142,7 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
+
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
@@ -151,17 +154,17 @@ endif()
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
-        link_libraries(-fsanitize=thread)
+        link_libraries     (-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries(-fsanitize=address)
+        link_libraries     (-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
-        link_libraries(-fsanitize=undefined)
+        link_libraries     (-fsanitize=undefined)
     endif()
 endif()
 
@@ -196,9 +199,33 @@ if (LLAMA_METAL)
     # get full path to the file
     #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
 
-    # copy ggml-metal.metal to bin directory
+    # copy ggml-common.h and ggml-metal.metal to bin directory
+    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
     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
@@ -298,14 +325,17 @@ if (LLAMA_BLAS)
         endif()
 
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
+
         add_compile_options(${BLAS_LINKER_FLAGS})
+
         add_compile_definitions(GGML_USE_OPENBLAS)
+
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
-        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
 
+        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -330,9 +360,6 @@ if (LLAMA_CUBLAS)
         set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
         add_compile_definitions(GGML_USE_CUBLAS)
-#        if (LLAMA_CUDA_CUBLAS)
-#            add_compile_definitions(GGML_CUDA_CUBLAS)
-#        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -387,15 +414,20 @@ if (LLAMA_MPI)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
+
         set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
+        set(GGML_SOURCES_MPI ggml-mpi.c)
+
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
+
         if (NOT MSVC)
             add_compile_options(-Wno-cast-qual)
         endif()
+
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
+
         # Even if you're only using the C header, C++ programs may bring in MPI
         # C++ functions, so more linkage is needed
         if (MPI_CXX_FOUND)
@@ -427,31 +459,28 @@ if (LLAMA_VULKAN)
     if (Vulkan_FOUND)
         message(STATUS "Vulkan found")
 
-        add_library(ggml-vulkan OBJECT ggml-vulkan.cpp ggml-vulkan.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
+        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
+        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
         if (LLAMA_VULKAN_CHECK_RESULTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
+            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
         endif()
 
         if (LLAMA_VULKAN_DEBUG)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
+            add_compile_definitions(GGML_VULKAN_DEBUG)
         endif()
 
         if (LLAMA_VULKAN_VALIDATE)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
+            add_compile_definitions(GGML_VULKAN_VALIDATE)
         endif()
 
         if (LLAMA_VULKAN_RUN_TESTS)
-            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
+            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -463,43 +492,45 @@ if (LLAMA_HIPBLAS)
     if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
+
     if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
 
-    find_package(hip)
-    find_package(hipblas)
-    find_package(rocblas)
+    find_package(hip     REQUIRED)
+    find_package(hipblas REQUIRED)
+    find_package(rocblas REQUIRED)
 
-    if (${hipblas_FOUND} AND ${hip_FOUND})
-        message(STATUS "HIP and hipBLAS found")
-        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
-        if (LLAMA_HIP_UMA)
-            add_compile_definitions(GGML_HIP_UMA)
-        endif()
-        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
-        if (BUILD_SHARED_LIBS)
-            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
-        endif()
-        if (LLAMA_CUDA_FORCE_DMMV)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
-        endif()
-        if (LLAMA_CUDA_FORCE_MMQ)
-            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
-        endif()
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
-        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
+    message(STATUS "HIP and hipBLAS found")
 
-        if (LLAMA_STATIC)
-            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-        endif()
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
-    else()
-        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
+    set(GGML_HEADERS_ROCM ggml-cuda.h)
+    set(GGML_SOURCES_ROCM ggml-cuda.cu)
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+
+    if (LLAMA_HIP_UMA)
+        add_compile_definitions(GGML_HIP_UMA)
     endif()
+
+    if (LLAMA_CUDA_FORCE_DMMV)
+        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
+    endif()
+
+    if (LLAMA_CUDA_FORCE_MMQ)
+        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
+    endif()
+
+    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+
+    set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+
+    if (LLAMA_STATIC)
+        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+    endif()
+
+    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
@@ -509,10 +540,14 @@ if (LLAMA_SYCL)
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
+
+    message(STATUS "SYCL found")
+
+    add_compile_definitions(GGML_USE_SYCL)
+
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
-    add_compile_definitions(GGML_USE_SYCL)
 
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
@@ -521,7 +556,7 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
 
-    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
@@ -540,61 +575,61 @@ if (LLAMA_KOMPUTE)
     endif()
 
     function(compile_shader)
-      set(options)
-      set(oneValueArgs)
-      set(multiValueArgs SOURCES)
-      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-      foreach(source ${compile_shader_SOURCES})
-        get_filename_component(filename ${source} NAME)
-        set(spv_file ${filename}.spv)
-        add_custom_command(
-            OUTPUT ${spv_file}
-            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-            COMMENT "Compiling ${source} to ${spv_file}"
-        )
+        set(options)
+        set(oneValueArgs)
+        set(multiValueArgs SOURCES)
+        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+        foreach(source ${compile_shader_SOURCES})
+            get_filename_component(filename ${source} NAME)
+            set(spv_file ${filename}.spv)
+            add_custom_command(
+                OUTPUT ${spv_file}
+                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+                COMMENT "Compiling ${source} to ${spv_file}"
+                )
 
-        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-        set(FILE_NAME "shader${RAW_FILE_NAME}")
-        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-        if(CMAKE_GENERATOR MATCHES "Visual Studio")
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-            )
-        else()
-            add_custom_command(
-              OUTPUT ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-              DEPENDS ${spv_file} xxd
-              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-            )
-        endif()
-      endforeach()
+            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+            set(FILE_NAME "shader${RAW_FILE_NAME}")
+            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+            if(CMAKE_GENERATOR MATCHES "Visual Studio")
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+                    )
+            else()
+                add_custom_command(
+                    OUTPUT ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+                    DEPENDS ${spv_file} xxd
+                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+                    )
+            endif()
+        endforeach()
     endfunction()
 
     if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@@ -604,66 +639,66 @@ if (LLAMA_KOMPUTE)
 
         # Compile our shaders
         compile_shader(SOURCES
-          kompute-shaders/op_scale.comp
-          kompute-shaders/op_scale_8.comp
-          kompute-shaders/op_add.comp
-          kompute-shaders/op_addrow.comp
-          kompute-shaders/op_mul.comp
-          kompute-shaders/op_silu.comp
-          kompute-shaders/op_relu.comp
-          kompute-shaders/op_gelu.comp
-          kompute-shaders/op_softmax.comp
-          kompute-shaders/op_norm.comp
-          kompute-shaders/op_rmsnorm.comp
-          kompute-shaders/op_diagmask.comp
-          kompute-shaders/op_mul_mat_mat_f32.comp
-          kompute-shaders/op_mul_mat_f16.comp
-          kompute-shaders/op_mul_mat_q8_0.comp
-          kompute-shaders/op_mul_mat_q4_0.comp
-          kompute-shaders/op_mul_mat_q4_1.comp
-          kompute-shaders/op_mul_mat_q6_k.comp
-          kompute-shaders/op_getrows_f16.comp
-          kompute-shaders/op_getrows_q4_0.comp
-          kompute-shaders/op_getrows_q4_1.comp
-          kompute-shaders/op_getrows_q6_k.comp
-          kompute-shaders/op_rope_f16.comp
-          kompute-shaders/op_rope_f32.comp
-          kompute-shaders/op_cpy_f16_f16.comp
-          kompute-shaders/op_cpy_f16_f32.comp
-          kompute-shaders/op_cpy_f32_f16.comp
-          kompute-shaders/op_cpy_f32_f32.comp
+            kompute-shaders/op_scale.comp
+            kompute-shaders/op_scale_8.comp
+            kompute-shaders/op_add.comp
+            kompute-shaders/op_addrow.comp
+            kompute-shaders/op_mul.comp
+            kompute-shaders/op_silu.comp
+            kompute-shaders/op_relu.comp
+            kompute-shaders/op_gelu.comp
+            kompute-shaders/op_softmax.comp
+            kompute-shaders/op_norm.comp
+            kompute-shaders/op_rmsnorm.comp
+            kompute-shaders/op_diagmask.comp
+            kompute-shaders/op_mul_mat_mat_f32.comp
+            kompute-shaders/op_mul_mat_f16.comp
+            kompute-shaders/op_mul_mat_q8_0.comp
+            kompute-shaders/op_mul_mat_q4_0.comp
+            kompute-shaders/op_mul_mat_q4_1.comp
+            kompute-shaders/op_mul_mat_q6_k.comp
+            kompute-shaders/op_getrows_f16.comp
+            kompute-shaders/op_getrows_q4_0.comp
+            kompute-shaders/op_getrows_q4_1.comp
+            kompute-shaders/op_getrows_q6_k.comp
+            kompute-shaders/op_rope_f16.comp
+            kompute-shaders/op_rope_f32.comp
+            kompute-shaders/op_cpy_f16_f16.comp
+            kompute-shaders/op_cpy_f16_f32.comp
+            kompute-shaders/op_cpy_f32_f16.comp
+            kompute-shaders/op_cpy_f32_f32.comp
         )
 
         # Create a custom target for our generated shaders
         add_custom_target(generated_shaders DEPENDS
-          shaderop_scale.h
-          shaderop_scale_8.h
-          shaderop_add.h
-          shaderop_addrow.h
-          shaderop_mul.h
-          shaderop_silu.h
-          shaderop_relu.h
-          shaderop_gelu.h
-          shaderop_softmax.h
-          shaderop_norm.h
-          shaderop_rmsnorm.h
-          shaderop_diagmask.h
-          shaderop_mul_mat_mat_f32.h
-          shaderop_mul_mat_f16.h
-          shaderop_mul_mat_q8_0.h
-          shaderop_mul_mat_q4_0.h
-          shaderop_mul_mat_q4_1.h
-          shaderop_mul_mat_q6_k.h
-          shaderop_getrows_f16.h
-          shaderop_getrows_q4_0.h
-          shaderop_getrows_q4_1.h
-          shaderop_getrows_q6_k.h
-          shaderop_rope_f16.h
-          shaderop_rope_f32.h
-          shaderop_cpy_f16_f16.h
-          shaderop_cpy_f16_f32.h
-          shaderop_cpy_f32_f16.h
-          shaderop_cpy_f32_f32.h
+            shaderop_scale.h
+            shaderop_scale_8.h
+            shaderop_add.h
+            shaderop_addrow.h
+            shaderop_mul.h
+            shaderop_silu.h
+            shaderop_relu.h
+            shaderop_gelu.h
+            shaderop_softmax.h
+            shaderop_norm.h
+            shaderop_rmsnorm.h
+            shaderop_diagmask.h
+            shaderop_mul_mat_mat_f32.h
+            shaderop_mul_mat_f16.h
+            shaderop_mul_mat_q8_0.h
+            shaderop_mul_mat_q4_0.h
+            shaderop_mul_mat_q4_1.h
+            shaderop_mul_mat_q6_k.h
+            shaderop_getrows_f16.h
+            shaderop_getrows_q4_0.h
+            shaderop_getrows_q4_1.h
+            shaderop_getrows_q6_k.h
+            shaderop_rope_f16.h
+            shaderop_rope_f32.h
+            shaderop_cpy_f16_f16.h
+            shaderop_cpy_f16_f32.h
+            shaderop_cpy_f32_f16.h
+            shaderop_cpy_f32_f32.h
         )
 
         # Create a custom command that depends on the generated_shaders
@@ -676,8 +711,10 @@ if (LLAMA_KOMPUTE)
 
         # Add the stamp to the main sources to ensure dependency tracking
         set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
+
         add_compile_definitions(GGML_USE_KOMPUTE)
+
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
     else()
@@ -685,6 +722,18 @@ if (LLAMA_KOMPUTE)
     endif()
 endif()
 
+if (LLAMA_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+
+    add_compile_definitions(GGML_USE_CPU_HBM)
+
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
+
+if (LLAMA_PERF)
+    add_compile_definitions(GGML_PERF)
+endif()
+
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -709,28 +758,30 @@ function(get_flags CCID CCVER)
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
-    elseif (CCID MATCHES "Intel")
-        if (NOT LLAMA_SYCL)
-            # enable max optimization level when using Intel compiler
-            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
-            add_link_options(-fuse-ld=lld -static-intel)
-        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     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)
-        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 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(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
-        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+        list(APPEND C_FLAGS   ${WARNING_FLAGS})
+        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -746,9 +797,10 @@ endif()
 set(CUDA_CXX_FLAGS "")
 
 if (LLAMA_CUBLAS)
-    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
-    if (NOT MSVC)
-        list(APPEND CUDA_FLAGS -Wno-pedantic)
+    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)
@@ -782,7 +834,11 @@ if (LLAMA_CUBLAS)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
+    endif()
+
+    if (NOT MSVC)
+        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
     endif()
 endif()
 
@@ -821,6 +877,7 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
+
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -830,10 +887,10 @@ endif()
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 if (MSVC)
-  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
 else ()
-  set(CMAKE_GENERATOR_PLATFORM_LWR "")
+    set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
 
 if (NOT MSVC)
@@ -855,11 +912,21 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
      CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
     message(STATUS "ARM detected")
     if (MSVC)
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
-        add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
+
+        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
+        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
+        if (GGML_COMPILER_SUPPORT_DOTPROD)
+            add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        endif ()
+        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
+            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+        endif ()
+        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@@ -870,10 +937,16 @@ 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 (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            # Raspberry Pi 2
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            if ("${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()
         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()
@@ -1017,11 +1090,6 @@ endif()
 
 # ggml
 
-if (GGML_USE_CPU_HBM)
-    add_definitions(-DGGML_USE_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-endif()
-
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -1038,16 +1106,17 @@ add_library(ggml OBJECT
             ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
             ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
             ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features(ggml PUBLIC c_std_11) # don't bump
+target_compile_features   (ggml PUBLIC c_std_11) # don't bump
+
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
-if (GGML_USE_CPU_HBM)
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
+
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@@ -1063,7 +1132,8 @@ add_library(llama
             )
 
 target_include_directories(llama PUBLIC .)
-target_compile_features(llama PUBLIC cxx_std_11) # don't bump
+target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
+
 target_link_libraries(llama PRIVATE
     ggml
     ${LLAMA_EXTRA_LIBS}
@@ -1114,7 +1184,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")

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 tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -97,9 +97,10 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = -std=c11   -fPIC
-MK_CXXFLAGS = -std=c++11 -fPIC
+MK_CPPFLAGS  = -I. -Icommon
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -172,7 +173,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -200,6 +201,10 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(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 ''
@@ -215,6 +220,11 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -375,10 +385,18 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	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
+	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
 	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
@@ -435,11 +453,11 @@ 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-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.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 $(NVCCFLAGS) -Xcompiler "$(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) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -524,11 +542,29 @@ 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
@@ -540,9 +576,10 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-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 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 NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -550,7 +587,7 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
 #
@@ -569,7 +606,7 @@ $(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
@@ -593,7 +630,7 @@ 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-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
 OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
@@ -687,14 +724,17 @@ 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/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)
+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)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(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)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -862,3 +902,7 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-chat-template: tests/test-chat-template.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)

README-sycl.md

@@ -1,6 +1,7 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
+- [News](#news)
 - [OS](#os)
 - [Intel GPU](#intel-gpu)
 - [Docker](#docker)
@@ -25,6 +26,21 @@ 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|
@@ -272,7 +288,7 @@ 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: **/opt/intel/oneapi**.
+Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
 
 Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 
@@ -449,6 +465,7 @@ 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
 
@@ -458,6 +475,10 @@ 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

@@ -8,16 +8,20 @@
 
 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
 
-- 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
+- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
 
 ----
 
@@ -61,7 +65,7 @@ variety of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@@ -107,16 +111,21 @@ 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)
+- [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] [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:**
 
@@ -145,6 +154,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
+- [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)
@@ -156,6 +166,9 @@ 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)
 
 ---
 
@@ -768,7 +781,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```
@@ -781,7 +794,7 @@ 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
 
@@ -845,7 +858,7 @@ 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.
@@ -958,7 +971,7 @@ We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 

awq-py/README.md

@@ -1,116 +0,0 @@
-# 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

@@ -1,254 +0,0 @@
-"""
-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

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

build.zig

@@ -123,6 +123,7 @@ 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 });
@@ -131,7 +132,7 @@ 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 });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -33,7 +33,7 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA=""
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
@@ -45,7 +45,8 @@ 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:\n source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
+        echo "source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 
@@ -219,7 +220,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -272,19 +273,19 @@ 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 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/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
@@ -343,17 +344,17 @@ 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 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
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
     compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 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
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
     compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 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 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
     compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     set +e
@@ -401,7 +402,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"
@@ -580,6 +581,10 @@ function gg_run_embd_bge_small {
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+
+    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
 
     path_models="../models-mnt/bge-small"
 

common/CMakeLists.txt

@@ -19,7 +19,12 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    set(GIT_INDEX "${GIT_DIR}/index")
+    if(EXISTS "${GIT_DIR}/index")
+        set(GIT_INDEX "${GIT_DIR}/index")
+    else()
+        message(WARNING "Git index not found in git repository.")
+        set(GIT_INDEX "")
+    endif()
 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_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; }
+            /**/ 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; }
             else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
@@ -335,13 +335,29 @@ 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;
                 break;
             }
             const auto sampler_names = string_split(argv[i], ';');
-            sparams.samplers_sequence = sampler_types_from_names(sampler_names);
+            sparams.samplers_sequence = sampler_types_from_names(sampler_names, true);
         } else if (arg == "--sampling-seq") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -497,12 +513,6 @@ 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;
@@ -630,11 +640,15 @@ 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_NONE;
+                params.split_mode = LLAMA_SPLIT_MODE_NONE;
             } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_LAYER;
+                params.split_mode = LLAMA_SPLIT_MODE_LAYER;
             } else if (arg_next == "row") {
-                params.split_mode = LLAMA_SPLIT_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;
             } else {
                 invalid_param = true;
                 break;
@@ -671,7 +685,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
-            params.numa = true;
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            std::string value(argv[i]);
+            /**/ if (value == "distribute" || value == "") { params.numa = GGML_NUMA_STRATEGY_DISTRIBUTE; }
+            else if (value == "isolate") { params.numa = GGML_NUMA_STRATEGY_ISOLATE; }
+            else if (value == "numactl") { params.numa = GGML_NUMA_STRATEGY_NUMACTL; }
+            else { invalid_param = true; break; }
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
         } else if (arg == "--no-display-prompt") {
@@ -829,15 +851,15 @@ 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_INT;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
                 kvo.int_value = std::atol(sep);
             } else if (strncmp(sep, "float:", 6) == 0) {
                 sep += 6;
-                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
                 kvo.float_value = std::atof(sep);
             } else if (strncmp(sep, "bool:", 5) == 0) {
                 sep += 5;
-                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
+                kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
                 if (std::strcmp(sep, "true") == 0) {
                     kvo.bool_value = true;
                 } else if (std::strcmp(sep, "false") == 0) {
@@ -935,7 +957,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -tb N, --threads-batch N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads)\n");
     printf("  -td N, --threads-draft N");
-    printf("                        number of threads to use during generation (default: same as --threads)");
+    printf("                        number of threads to use during generation (default: same as --threads)\n");
     printf("  -tbd N, --threads-batch-draft N\n");
     printf("                        number of threads to use during batch and prompt processing (default: same as --threads-draft)\n");
     printf("  -p PROMPT, --prompt PROMPT\n");
@@ -956,7 +978,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -n N, --n-predict N   number of tokens to predict (default: %d, -1 = infinity, -2 = until context filled)\n", params.n_predict);
     printf("  -c N, --ctx-size N    size of the prompt context (default: %d, 0 = loaded from model)\n", params.n_ctx);
     printf("  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
-    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\' (default: %s)\n", sampler_type_names.c_str());
+    printf("  --samplers            samplers that will be used for generation in the order, separated by \';\'\n");
+    printf("                        (default: %s)\n", sampler_type_names.c_str());
     printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sampler_type_chars.c_str());
     printf("  --top-k N             top-k sampling (default: %d, 0 = disabled)\n", sparams.top_k);
     printf("  --top-p N             top-p sampling (default: %.1f, 1.0 = disabled)\n", (double)sparams.top_p);
@@ -995,23 +1018,26 @@ 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("  --logits-all          return logits for all tokens in the batch (default: disabled)\n");
+    printf("  --all-logits          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");
     printf("  --winogrande-tasks N  number of tasks to use when computing the Winogrande score (default: %zu)\n", params.winogrande_tasks);
     printf("  --multiple-choice     compute multiple choice score over random tasks from datafile supplied with -f\n");
     printf("  --multiple-choice-tasks N number of tasks to use when computing the multiple choice score (default: %zu)\n", params.winogrande_tasks);
-    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base");
+    printf("  --kl-divergence       computes KL-divergence to logits provided via --kl-divergence-base\n");
     printf("  --keep N              number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
     printf("  --draft N             number of tokens to draft for speculative decoding (default: %d)\n", params.n_draft);
     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");
@@ -1022,7 +1048,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_supports_mmap()) {
         printf("  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
-    printf("  --numa                attempt optimizations that help on some NUMA systems\n");
+    printf("  --numa TYPE           attempt optimizations that help on some NUMA systems\n");
+    printf("                          - distribute: spread execution evenly over all nodes\n");
+    printf("                          - isolate: only spawn threads on CPUs on the node that execution started on\n");
+    printf("                          - numactl: use the CPU map provided by numactl\n");
     printf("                        if run without this previously, it is recommended to drop the system page cache before using this\n");
     printf("                        see https://github.com/ggerganov/llama.cpp/issues/1437\n");
     if (llama_supports_gpu_offload()) {
@@ -1122,34 +1151,50 @@ std::vector<std::string> string_split(std::string input, char separator) {
     return parts;
 }
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names) {
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
+    };
+
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, llama_sampler_type> sampler_name_map {
-        {"top_k",       llama_sampler_type::TOP_K},
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
         {"top-k",       llama_sampler_type::TOP_K},
-        {"top_p",       llama_sampler_type::TOP_P},
         {"top-p",       llama_sampler_type::TOP_P},
         {"nucleus",     llama_sampler_type::TOP_P},
-        {"typical_p",   llama_sampler_type::TYPICAL_P},
         {"typical-p",   llama_sampler_type::TYPICAL_P},
         {"typical",     llama_sampler_type::TYPICAL_P},
-        {"min_p",       llama_sampler_type::MIN_P},
         {"min-p",       llama_sampler_type::MIN_P},
-        {"tfs_z",       llama_sampler_type::TFS_Z},
         {"tfs-z",       llama_sampler_type::TFS_Z},
         {"tfs",         llama_sampler_type::TFS_Z},
-        {"temp",        llama_sampler_type::TEMP},
-        {"temperature", llama_sampler_type::TEMP}
+        {"temp",        llama_sampler_type::TEMPERATURE}
     };
 
     std::vector<llama_sampler_type> sampler_types;
     sampler_types.reserve(names.size());
-    for (const auto& name : names) {
-        const auto sampler_item = sampler_name_map.find(name);
-        if (sampler_item != sampler_name_map.end()) {
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
             sampler_types.push_back(sampler_item->second);
         }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
     }
     return sampler_types;
 }
@@ -1161,7 +1206,7 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
         {'y', llama_sampler_type::TYPICAL_P},
         {'m', llama_sampler_type::MIN_P},
         {'f', llama_sampler_type::TFS_Z},
-        {'t', llama_sampler_type::TEMP}
+        {'t', llama_sampler_type::TEMPERATURE}
     };
 
     std::vector<llama_sampler_type> sampler_types;
@@ -1177,12 +1222,12 @@ std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & nam
 
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type) {
     switch (sampler_type) {
-        case llama_sampler_type::TOP_K:     return "top_k";
-        case llama_sampler_type::TFS_Z:     return "tfs_z";
-        case llama_sampler_type::TYPICAL_P: return "typical_p";
-        case llama_sampler_type::TOP_P:     return "top_p";
-        case llama_sampler_type::MIN_P:     return "min_p";
-        case llama_sampler_type::TEMP:      return "temp";
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
         default : return "";
     }
 }
@@ -1243,12 +1288,12 @@ 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.embedding         = params.embedding;
+    cparams.embeddings        = 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;
@@ -1257,6 +1302,8 @@ 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);
@@ -1676,6 +1723,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     }
     fprintf(stream, "lora_base: %s\n", params.lora_base.c_str());
     fprintf(stream, "main_gpu: %d # default: 0\n", params.main_gpu);
+    fprintf(stream, "min_keep: %d # default: 0 (disabled)\n", sparams.min_keep);
     fprintf(stream, "mirostat: %d # default: 0 (disabled)\n", sparams.mirostat);
     fprintf(stream, "mirostat_ent: %f # default: 5.0\n", sparams.mirostat_tau);
     fprintf(stream, "mirostat_lr: %f # default: 0.1\n", sparams.mirostat_eta);
@@ -1687,9 +1735,7 @@ 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, "numa: %s # default: false\n", params.numa ? "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);
     fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);
@@ -1714,7 +1760,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
 
     fprintf(stream, "rope_freq_base: %f # default: 10000.0\n", params.rope_freq_base);
     fprintf(stream, "rope_freq_scale: %f # default: 1.0\n", params.rope_freq_scale);
-    fprintf(stream, "seed: %d # default: -1 (random seed)\n", params.seed);
+    fprintf(stream, "seed: %u # default: -1 (random seed)\n", params.seed);
     fprintf(stream, "simple_io: %s # default: false\n", params.simple_io ? "true" : "false");
     fprintf(stream, "cont_batching: %s # default: false\n", params.cont_batching ? "true" : "false");
     fprintf(stream, "temp: %f # default: 0.8\n", sparams.temp);
@@ -1723,7 +1769,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     dump_vector_float_yaml(stream, "tensor_split", tensor_split_vector);
 
     fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
-    fprintf(stream, "threads: %d # default: %d\n", params.n_threads, std::thread::hardware_concurrency());
+    fprintf(stream, "threads: %d # default: %u\n", params.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
@@ -1774,7 +1820,8 @@ void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
             if (cs_curr[j] < 0) { continue; }
             if (seqs.find(cs_curr[j]) == seqs.end()) {
                 if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-                seqs[cs_curr[j]] = seqs.size();
+                const size_t sz = seqs.size();
+                seqs[cs_curr[j]] = sz;
             }
         }
         if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
@@ -1805,3 +1852,18 @@ 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                 = -1;    // RNG seed
+    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
     int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;
@@ -53,15 +53,14 @@ 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               = 8;     // number of tokens to draft during speculative decoding
+    int32_t n_draft               = 5;     // 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_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_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.
@@ -75,7 +74,12 @@ 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
-    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
+    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
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -113,7 +117,6 @@ 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
@@ -134,7 +137,6 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
@@ -165,7 +167,7 @@ void process_escapes(std::string& input);
 // String utils
 //
 
-std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names);
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
@@ -258,3 +260,10 @@ 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

@@ -278,6 +278,22 @@ 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

@@ -121,7 +121,7 @@ static void sampler_queue(
                    struct llama_context * ctx_main,
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
-                                 size_t & min_keep) {
+                                 size_t   min_keep) {
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -139,7 +139,7 @@ static void sampler_queue(
             case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
             case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
             case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case llama_sampler_type::TEMP:
+            case llama_sampler_type::TEMPERATURE:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
@@ -249,7 +249,7 @@ static llama_token llama_sampling_sample_impl(
             id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
             // temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
+            size_t min_keep = std::max(1, params.min_keep);
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 
@@ -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());
         }
     }
 
@@ -295,6 +295,77 @@ 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,
@@ -304,6 +375,14 @@ 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

@@ -10,18 +10,19 @@
 
 // sampler types
 enum class llama_sampler_type : char {
-    TOP_K     = 'k',
-    TOP_P     = 'p',
-    MIN_P     = 'm',
-    TFS_Z     = 'f',
-    TYPICAL_P = 'y',
-    TEMP      = 't'
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
 };
 
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
     int32_t     top_k                 = 40;       // <= 0 to use vocab size
     float       top_p                 = 0.95f;    // 1.0 = disabled
     float       min_p                 = 0.05f;    // 0.0 = disabled
@@ -45,7 +46,7 @@ typedef struct llama_sampling_params {
         llama_sampler_type::TYPICAL_P,
         llama_sampler_type::TOP_P,
         llama_sampler_type::MIN_P,
-        llama_sampler_type::TEMP
+        llama_sampler_type::TEMPERATURE
     };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
@@ -130,6 +131,13 @@ 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_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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_ADAM;
+        opt->params.type = GGML_OPT_TYPE_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_LBFGS;
+        opt->params.type = GGML_OPT_TYPE_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_ADAM:
+        case GGML_OPT_TYPE_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_LBFGS:
+        case GGML_OPT_TYPE_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

@@ -8,9 +8,10 @@ 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, ContextManager, Iterator, cast
+from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
 
 import numpy as np
 import torch
@@ -25,15 +26,6 @@ import gguf
 from convert import HfVocab
 
 
-# check for any of the given keys in the dictionary and return the value of the first key found
-def get_key_opts(d, keys):
-    for k in keys:
-        if k in d:
-            return d[k]
-    print(f"Could not find any of {keys}")
-    sys.exit()
-
-
 ###### MODEL DEFINITIONS ######
 
 class SentencePieceTokenTypes(IntEnum):
@@ -45,7 +37,12 @@ class SentencePieceTokenTypes(IntEnum):
     BYTE = 6
 
 
-class Model:
+AnyModel = TypeVar("AnyModel", bound="type[Model]")
+
+
+class Model(ABC):
+    _model_classes: dict[str, type[Model]] = {}
+
     def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
@@ -56,8 +53,21 @@ class Model:
         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:
+            return self.hparams[key]
+        if optional:
+            return None
+        raise KeyError(f"could not find any of: {keys}")
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -79,28 +89,35 @@ class Model:
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_block_count(self.hparams.get(
-            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
-        ))
-        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
+        self.gguf_writer.add_block_count(self.block_count)
+
+        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
-        if (n_embd := self.hparams.get("hidden_size")) is not None:
-            self.gguf_writer.add_embedding_length(n_embd)
-        if (n_ff := self.hparams.get("intermediate_size")) is not None:
+
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        self.gguf_writer.add_embedding_length(n_embd)
+
+        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-        if (n_head := self.hparams.get("num_attention_heads")) is not None:
-            self.gguf_writer.add_head_count(n_head)
+
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_head_count(n_head)
+
         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 (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
-            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
+        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:
+            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)
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
 
-        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
+        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")))
@@ -169,49 +186,22 @@ class Model:
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             return json.load(f)
 
-    @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
-        return Model
+    @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
 
     def _is_model_safetensors(self) -> bool:
         return Model.count_model_parts(self.dir_model, ".safetensors") > 0
@@ -226,51 +216,6 @@ class Model:
             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
-
-        raise NotImplementedError(f'Architecture "{arch}" not supported!')
-
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
@@ -438,7 +383,10 @@ class Model:
         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"]
 
@@ -455,7 +403,10 @@ 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"))
@@ -547,7 +498,10 @@ 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)
@@ -609,13 +563,11 @@ 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()
 
@@ -646,6 +598,8 @@ 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):
@@ -692,7 +646,10 @@ 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()
 
@@ -807,7 +764,10 @@ 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:
@@ -900,7 +860,10 @@ 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"]
 
@@ -915,7 +878,10 @@ 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
@@ -999,7 +965,10 @@ 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"]
@@ -1022,7 +991,6 @@ 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()
@@ -1048,7 +1016,10 @@ 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()
@@ -1065,18 +1036,25 @@ 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"])
-        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        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_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(1e-5)
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
 
+@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")
@@ -1153,7 +1131,10 @@ 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
@@ -1233,7 +1214,15 @@ 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"])
@@ -1295,29 +1284,35 @@ class GPT2Model(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
+@Model.register("PhiForCausalLM")
 class Phi2Model(Model):
-    def set_gguf_parameters(self):
-        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
+    model_arch = gguf.MODEL_ARCH.PHI2
 
-        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
-        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
-        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
+    def set_gguf_parameters(self):
+        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+
+        rot_pct = self.find_hparam(["partial_rotary_factor"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         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()
 
@@ -1396,7 +1391,10 @@ 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"]
 
@@ -1461,7 +1459,10 @@ 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
@@ -1633,23 +1634,40 @@ 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.block_count = self.hparams["num_hidden_layers"]
+        self.vocab_size = None
 
     def set_gguf_parameters(self):
-        # TODO(cebtenzzre): merge with parent class
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
-        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-        self.gguf_writer.add_block_count(self.block_count)
-        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        super().set_gguf_parameters()
         self.gguf_writer.add_causal_attention(False)
-        self.gguf_writer.add_pooling_layer(True)
-        self.gguf_writer.add_file_type(self.ftype)
+
+        # get pooling path
+        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
+
+        # get pooling type
+        if pooling_path is not None:
+            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
+                pooling = json.load(f)
+            if pooling["pooling_mode_mean_tokens"]:
+                pooling_type = gguf.PoolingType.MEAN
+            elif pooling["pooling_mode_cls_token"]:
+                pooling_type = gguf.PoolingType.CLS
+            else:
+                raise NotImplementedError("Only MEAN and CLS pooling types supported")
+            self.gguf_writer.add_pooling_type(pooling_type)
 
     def set_vocab(self):
         path = self.dir_model
@@ -1659,6 +1677,7 @@ class BertModel(Model):
         vocab = HfVocab(path, added_tokens_path)
         tokens, scores, toktypes = zip(*vocab.all_tokens())
         assert len(tokens) == vocab.vocab_size
+        self.vocab_size = vocab.vocab_size
 
         # we need this to validate the size of the token_type embeddings
         # though currently we are passing all zeros to the token_type embeddings
@@ -1672,7 +1691,7 @@ class BertModel(Model):
             if tok.startswith(b"##"):
                 return tok[2:]
             return b"\xe2\x96\x81" + tok
-        tokens = [phantom(t, y) for t, y in zip(tokens, toktypes)]
+        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
 
         # set up bos and eos tokens (cls and sep)
         self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
@@ -1724,6 +1743,228 @@ 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)
+
+        # the HF config claims n_ctx=8192, but it uses RoPE scaling
+        self.hparams["n_ctx"] = 2048
+
+        # SwigLU activation
+        assert self.hparams["activation_function"] == "swiglu"
+        # this doesn't do anything in the HF version
+        assert self.hparams["causal"] is False
+        # no bias tensors
+        assert self.hparams["qkv_proj_bias"] is False
+        assert self.hparams["mlp_fc1_bias"] is False
+        assert self.hparams["mlp_fc2_bias"] is False
+        # norm at end of layer
+        assert self.hparams["prenorm"] is False
+        # standard RoPE
+        assert self.hparams["rotary_emb_fraction"] == 1.0
+        assert self.hparams["rotary_emb_interleaved"] is False
+        assert self.hparams["rotary_emb_scale_base"] is None
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
+
+    def get_tensors(self):
+        assert self.vocab_size is not None
+        for name, data in super().get_tensors():
+            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
+            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
+                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
+                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
+                data = data[:self.vocab_size, :]
+            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, cfg.model_metadata_dir)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), 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", choices=["spm", "bpe"], default="spm",
-                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
+    parser.add_argument("--vocabtype", default="spm,hfft",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
     return parser.parse_args()
 
 

convert.py

@@ -1173,7 +1173,7 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 
-def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
+def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
@@ -1199,7 +1199,11 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     for name, lazy_tensor in model.items():
         tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
         if name_new is None:
-            raise Exception(f"Unexpected tensor name: {name}")
+            if skip_unknown:
+                print(f"Unexpected tensor name: {name} - skipping")
+                continue
+            else:
+                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1278,35 +1282,32 @@ 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.files: dict[str, Path | None] = {
-            "tokenizer.model": None,
-            "vocab.json": None,
-            "tokenizer.json": None,
-        }
-        self._detect_files()
+        self.file_paths = self._detect_files()
+        print(f"Found vocab files: {self.file_paths}")
 
-    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}")
+    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 _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}")
+        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 _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
         load_merges = vocabtype == "bpe"
@@ -1318,30 +1319,30 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    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}'")
+    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}")
 
         added_tokens_path = path.parent / "added_tokens.json"
         vocab: Vocab
-        if vocabtype == "bpe":
+        if vocab_type == "bpe":
             vocab = BpeVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocabtype == "spm":
+        elif vocab_type == "spm":
             vocab = SentencePieceVocab(
                 path, added_tokens_path if added_tokens_path.exists() else None
             )
-        elif vocabtype == "hfft":
+        elif vocab_type == "hfft":
             vocab = HfVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
+                path.parent, added_tokens_path if added_tokens_path.exists() else None
             )
         else:
-            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
+            raise ValueError(vocab_type)
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
-            vocabtype,
+            vocab_type,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1375,35 +1376,22 @@ 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")
-    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",  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)")
-    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
-    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
+    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("--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)")
+    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
+    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     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)
@@ -1443,7 +1431,7 @@ 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, model_parent_path)
+    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
 
     if args.vocab_only:
         if not args.outfile:
@@ -1461,7 +1449,7 @@ def main(args_in: list[str] | None = None) -> None:
     print(f"Special vocab info: {special_vocab}")
 
     model   = model_plus.model
-    model   = convert_model_names(model, params)
+    model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype)

examples/CMakeLists.txt

@@ -20,6 +20,7 @@ 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)
@@ -38,6 +39,7 @@ else()
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
+    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)

examples/baby-llama/baby-llama.cpp

@@ -1533,27 +1533,28 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        ggml_cgraph gf = {};
+        struct ggml_cgraph * gf = NULL;
+        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1600,13 +1601,14 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            ggml_cgraph gf = {};
+            struct ggml_cgraph * gf = NULL;
+            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(&gf, logits);
-            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(gf, logits);
+            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/batched-bench/batched-bench.cpp

@@ -32,16 +32,15 @@ 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] [MMQ] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <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 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 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, };
@@ -65,24 +64,21 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        mmq = std::atoi(argv[5]);
+        n_pp = parse_list(argv[5]);
     }
 
     if (argc >= 7) {
-        n_pp = parse_list(argv[6]);
+        n_tg = parse_list(argv[6]);
     }
 
     if (argc >= 8) {
-        n_tg = parse_list(argv[7]);
-    }
-
-    if (argc >= 9) {
-        n_pl = parse_list(argv[8]);
+        n_pl = parse_list(argv[7]);
     }
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -105,11 +101,13 @@ int main(int argc, char ** argv) {
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
     ctx_params.n_batch   = 512;
-    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) {
@@ -158,7 +156,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
+    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("\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");
@@ -179,10 +177,10 @@ int main(int argc, char ** argv) {
 
                 llama_batch_clear(batch);
 
-                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);
+                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);
+                    }
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
@@ -197,7 +195,7 @@ 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, 0, pp);
+                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
 let n_len: Int = 32
 
 // init LLM
-llama_backend_init(false)
+llama_backend_init()
 defer {
     llama_backend_free()
 }

examples/batched/batched.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -79,6 +80,7 @@ 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;
 
@@ -91,7 +93,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
@@ -131,7 +133,7 @@ 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, 0, batch.n_tokens);
+        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
     }
 
     if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -119,7 +119,8 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/benchmark/benchmark-matmult.cpp

@@ -189,12 +189,10 @@ 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], hist_cur.data(), nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -207,7 +205,7 @@ 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], hist_cur.data(), nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -325,14 +325,14 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
-    printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
-    printf("%s: n_layer: %d\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
+    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
+    printf("%s: n_head:  %u\n", __func__, params->n_head);
+    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
+    printf("%s: n_layer: %u\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -350,25 +350,25 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
 
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
+    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
 
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
 
     // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");

examples/embedding/embedding.cpp

@@ -19,18 +19,7 @@ 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 }, 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;
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
     }
 }
 
@@ -44,11 +33,23 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
         fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
-    // 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);
+    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);
     }
 }
 
@@ -74,7 +75,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -131,7 +133,7 @@ int main(int argc, char ** argv) {
 
     // initialize batch
     const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
     // allocate output
     const int n_embd = llama_n_embd(model);
@@ -144,6 +146,7 @@ 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/export-lora/export-lora.cpp

@@ -7,8 +7,6 @@
 #include <string>
 #include <thread>
 
-static const size_t tensor_alignment = 32;
-
 struct lora_info {
     std::string filename;
     float scale;

examples/finetune/README.md

@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-w1 N                LORA rank for w1 tensor (default 4)
-  --rank-w2 N                LORA rank for w2 tensor (default 4)
-  --rank-w3 N                LORA rank for w3 tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

examples/finetune/finetune.cpp

@@ -60,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -85,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_w1 = 4;
-    uint32_t n_rank_w2 = 4;
-    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_ffn_gate = 4;
+    uint32_t n_rank_ffn_down = 4;
+    uint32_t n_rank_ffn_up = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -117,12 +117,12 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1_a;
-    struct ggml_tensor * w1_b;
-    struct ggml_tensor * w2_a;
-    struct ggml_tensor * w2_b;
-    struct ggml_tensor * w3_a;
-    struct ggml_tensor * w3_b;
+    struct ggml_tensor * ffn_gate_a;
+    struct ggml_tensor * ffn_gate_b;
+    struct ggml_tensor * ffn_down_a;
+    struct ggml_tensor * ffn_down_b;
+    struct ggml_tensor * ffn_up_a;
+    struct ggml_tensor * ffn_up_b;
 };
 
 struct my_llama_lora {
@@ -208,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
-    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
-    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
+    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
+    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -319,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -329,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -362,12 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.w1_a);
-        ggml_set_param(ctx, layer.w1_b);
-        ggml_set_param(ctx, layer.w2_a);
-        ggml_set_param(ctx, layer.w2_b);
-        ggml_set_param(ctx, layer.w3_a);
-        ggml_set_param(ctx, layer.w3_b);
+        ggml_set_param(ctx, layer.ffn_gate_a);
+        ggml_set_param(ctx, layer.ffn_gate_b);
+        ggml_set_param(ctx, layer.ffn_down_a);
+        ggml_set_param(ctx, layer.ffn_down_b);
+        ggml_set_param(ctx, layer.ffn_up_a);
+        ggml_set_param(ctx, layer.ffn_up_b);
     }
 }
 
@@ -435,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
-        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
-        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
-        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
-        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
-        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
+        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
+        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
+        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
+        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
+        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -454,12 +454,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
@@ -497,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
         ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.w1_a, rnd);
-        ggml_set_zero(layer.w1_b);
-        randomize_tensor_normal(layer.w2_a, rnd);
-        ggml_set_zero(layer.w2_b);
-        randomize_tensor_normal(layer.w3_a, rnd);
-        ggml_set_zero(layer.w3_b);
+        randomize_tensor_normal(layer.ffn_gate_a, rnd);
+        ggml_set_zero(layer.ffn_gate_b);
+        randomize_tensor_normal(layer.ffn_down_a, rnd);
+        ggml_set_zero(layer.ffn_down_b);
+        randomize_tensor_normal(layer.ffn_up_a, rnd);
+        ggml_set_zero(layer.ffn_up_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -610,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
-        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
-        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
+        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
+        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -659,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -723,9 +723,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -798,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -825,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
-        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
-        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
-        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
-        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
-        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
+        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
+        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
+        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
+        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
+        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
     }
 }
 
@@ -868,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -894,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.w1_a);
-        gguf_add_tensor(fctx, layer.w1_b);
-        gguf_add_tensor(fctx, layer.w2_a);
-        gguf_add_tensor(fctx, layer.w2_b);
-        gguf_add_tensor(fctx, layer.w3_a);
-        gguf_add_tensor(fctx, layer.w3_b);
+        gguf_add_tensor(fctx, layer.ffn_gate_a);
+        gguf_add_tensor(fctx, layer.ffn_gate_b);
+        gguf_add_tensor(fctx, layer.ffn_down_a);
+        gguf_add_tensor(fctx, layer.ffn_down_b);
+        gguf_add_tensor(fctx, layer.ffn_up_a);
+        gguf_add_tensor(fctx, layer.ffn_up_b);
     }
 }
 
@@ -1104,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1139,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_w1;
-    uint32_t n_rank_w2;
-    uint32_t n_rank_w3;
+    uint32_t n_rank_ffn_gate;
+    uint32_t n_rank_ffn_down;
+    uint32_t n_rank_ffn_up;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1152,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_w1;
-    bool custom_n_rank_w2;
-    bool custom_n_rank_w3;
+    bool custom_n_rank_ffn_gate;
+    bool custom_n_rank_ffn_down;
+    bool custom_n_rank_ffn_up;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1186,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_w1             = 4;
-    params.n_rank_w2             = 4;
-    params.n_rank_w3             = 4;
+    params.n_rank_ffn_gate       = 4;
+    params.n_rank_ffn_down       = 4;
+    params.n_rank_ffn_up         = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1199,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_w1             = false;
-    params.custom_n_rank_w2             = false;
-    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_ffn_gate       = false;
+    params.custom_n_rank_ffn_down       = false;
+    params.custom_n_rank_ffn_up         = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1232,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1369,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-w1") {
+        } else if (arg == "--rank-ffn_gate") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w1 = std::stoi(argv[i]);
-            params->custom_n_rank_w1 = true;
-        } else if (arg == "--rank-w2") {
+            params->n_rank_ffn_gate = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_gate = true;
+        } else if (arg == "--rank-ffn_down") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w2 = std::stoi(argv[i]);
-            params->custom_n_rank_w2 = true;
-        } else if (arg == "--rank-w3") {
+            params->n_rank_ffn_down = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_down = true;
+        } else if (arg == "--rank-ffn_up") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w3 = std::stoi(argv[i]);
-            params->custom_n_rank_w3 = true;
+            params->n_rank_ffn_up = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_up = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1452,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.w1_a);
-        nx += ggml_nelements(layer.w1_b);
-        nx += ggml_nelements(layer.w2_a);
-        nx += ggml_nelements(layer.w2_b);
-        nx += ggml_nelements(layer.w3_a);
-        nx += ggml_nelements(layer.w3_b);
+        nx += ggml_nelements(layer.ffn_gate_a);
+        nx += ggml_nelements(layer.ffn_gate_b);
+        nx += ggml_nelements(layer.ffn_down_a);
+        nx += ggml_nelements(layer.ffn_down_b);
+        nx += ggml_nelements(layer.ffn_up_a);
+        nx += ggml_nelements(layer.ffn_up_b);
     }
     return nx;
 }
@@ -1511,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
-    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
-    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
+    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
+    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1523,15 +1523,15 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_w1             = n_rank_w1;
-    lora.hparams.n_rank_w2             = n_rank_w2;
-    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
+    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
+    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;
@@ -1566,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_w1             != n_rank_w1)
-        || (lora.hparams.n_rank_w2             != n_rank_w2)
-        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
+        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
+        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)

examples/gritlm/CMakeLists.txt

@@ -0,0 +1,5 @@
+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

@@ -0,0 +1,229 @@
+#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/imatrix/imatrix.cpp

@@ -568,7 +568,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
 

examples/infill/infill.cpp

@@ -202,7 +202,8 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -377,10 +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";
         }
@@ -446,8 +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_shift(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_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-to-grammar.py

@@ -87,7 +87,21 @@ 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')
-            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            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'
             return self._add_rule(rule_name, rule)
 
         else:

examples/llama-bench/README.md

@@ -35,7 +35,6 @@ 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

@@ -123,20 +123,15 @@ static std::string get_gpu_info() {
     }
 #endif
 #ifdef GGML_USE_SYCL
-    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;
+    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) {
             id += "/";
         }
     }
-    if (id.length() >2 ) {
-        id.pop_back();
-    }
 #endif
     // TODO: other backends
     return id;
@@ -157,9 +152,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_NONE:  return "none";
-        case LLAMA_SPLIT_LAYER: return "layer";
-        case LLAMA_SPLIT_ROW:   return "row";
+        case LLAMA_SPLIT_MODE_NONE:  return "none";
+        case LLAMA_SPLIT_MODE_LAYER: return "layer";
+        case LLAMA_SPLIT_MODE_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -176,9 +171,9 @@ 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;
@@ -193,12 +188,12 @@ 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_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_MODE_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
@@ -221,7 +216,7 @@ 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("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").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));
@@ -358,11 +353,11 @@ 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_NONE;
+                    mode = LLAMA_SPLIT_MODE_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_LAYER;
+                    mode = LLAMA_SPLIT_MODE_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_ROW;
+                    mode = LLAMA_SPLIT_MODE_ROW;
                 } else {
                     invalid_param = true;
                     break;
@@ -383,13 +378,6 @@ 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;
@@ -397,6 +385,13 @@ 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;
@@ -466,9 +461,9 @@ 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;
@@ -486,9 +481,9 @@ 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();
@@ -518,8 +513,8 @@ 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;
     }
@@ -535,10 +530,10 @@ 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) {
@@ -557,9 +552,9 @@ 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);
         }
@@ -580,9 +575,9 @@ 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);
         }
@@ -616,9 +611,9 @@ 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;
@@ -639,9 +634,9 @@ 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
@@ -713,7 +708,7 @@ struct test {
             "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
-            "mul_mat_q", "tensor_split", "use_mmap",
+            "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -733,7 +728,7 @@ 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 == "mul_mat_q" || field == "use_mmap") {
+            field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -767,7 +762,7 @@ 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),
-            std::to_string(mul_mat_q), tensor_split_str, std::to_string(use_mmap),
+            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
             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())
@@ -931,15 +926,15 @@ 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";
         }
@@ -974,9 +969,6 @@ 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");
         }
@@ -986,6 +978,9 @@ 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");
 
@@ -1151,8 +1146,7 @@ int main(int argc, char ** argv) {
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
     }
-    bool numa = false;
-    llama_backend_init(numa);
+    llama_backend_init();
 
     // initialize printer
     std::unique_ptr<printer> p;

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

@@ -21,12 +21,8 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // 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")
-            }
+            // Add NDK properties if wanted, e.g.
+            // abiFilters += listOf("arm64-v8a")
         }
         externalNativeBuild {
             cmake {

examples/llama.android/app/src/main/cpp/llama-android.cpp

@@ -274,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
-    llama_backend_init(numa);
+Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
+    llama_backend_init();
 }
 
 extern "C"

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -51,7 +51,7 @@ actor LlamaContext {
     }
 
     static func create_context(path: String) throws -> LlamaContext {
-        llama_backend_init(false)
+        llama_backend_init()
         var model_params = llama_model_default_params()
 
 #if targetEnvironment(simulator)

examples/llava/README.md

@@ -1,10 +1,12 @@
 # LLaVA
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
+as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
 
 The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
 and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
 models are available.
+For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
 After API is confirmed, more models will be supported / uploaded.
 
@@ -18,10 +20,11 @@ After building, run: `./llava-cli` to see the usage. For example:
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
+**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-## Model conversion
+## LLaVA 1.5
 
-- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally:
+- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@@ -50,13 +53,79 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-pa
 5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b
+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/
+```
+- 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
+```
+- 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
+```
+
+**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)
+
+## llava-cli templating and llava-1.6 prompting
+
+llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
+For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
+
+**For Mistral and using llava-cli binary:**
+Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
+The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
+
+**For the 34B this should work:**
+Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
+
+
+## How to know if you are running in llava-1.5 or llava-1.6 mode
+
+When running llava-cli you will see a visual information right before the prompt is being processed:
+
+**Llava-1.5:**
+`encode_image_with_clip: image embedding created: 576 tokens`
+
+**Llava-1.6 (anything above 576):**
+`encode_image_with_clip: image embedding created: 2880 tokens`
+
+
+Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
+
+
+
+
 ## TODO
 
-- [ ] Support non-CPU backend for the image encoding part.
+- [x] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.h

@@ -24,25 +24,7 @@ struct clip_ctx;
 extern "C" {
 #endif
 
-struct clip_vision_hparams {
-    int32_t image_size;
-    int32_t patch_size;
-    int32_t hidden_size;
-    int32_t n_intermediate;
-    int32_t projection_dim;
-    int32_t n_head;
-    int32_t n_layer;
-    float eps;
-};
-
-CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
-
-CLIP_API void clip_free(struct clip_ctx * ctx);
-
-CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+struct clip_ctx;
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -54,18 +36,43 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
+CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
+CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
+CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
+
+// TODO: should be enum, not string
+CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
+
+CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
-CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
+CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  & batch);
+CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch & batch);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
+/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
+CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch & res_imgs );
+
+CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
+
 CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
 CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 

examples/llava/convert-image-encoder-to-gguf.py

@@ -78,18 +78,19 @@ ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
-ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
+ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
                 help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
+ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
+                help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
 ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
-ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
-ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
+# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
 default_image_mean = [0.48145466, 0.4578275, 0.40821073]
 default_image_std = [0.26862954, 0.26130258, 0.27577711]
-ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
-ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
 # with proper
 args = ap.parse_args()
@@ -105,7 +106,7 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
     vocab = None
     tokens = None
 else:
@@ -133,7 +134,7 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-if args.clip_model_is_vision:
+if args.clip_model_is_vision or args.clip_model_is_openclip:
     model = CLIPVisionModel.from_pretrained(dir_model)
     processor = None
 else:
@@ -202,6 +203,57 @@ if has_vision_encoder:
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
+                            #     /**
+                            #      "image_grid_pinpoints": [
+                            #         [
+                            #         336,
+                            #         672
+                            #         ],
+                            #         [
+                            #         672,
+                            #         336
+                            #         ],
+                            #         [
+                            #         672,
+                            #         672
+                            #         ],
+                            #         [
+                            #         1008,
+                            #         336
+                            #         ],
+                            #         [
+                            #         336,
+                            #         1008
+                            #         ]
+                            #     ],
+                            #     Flattened:
+                            #     [
+                            #         336, 672,
+                            #         672, 336,
+                            #         672, 672,
+                            #         1008, 336,
+                            #         336, 1008
+                            #     ]
+                            #  *
+                            #  */
+    if "image_grid_pinpoints" in v_hparams:
+        # flatten it
+        image_grid_pinpoints = []
+        for pinpoint in v_hparams["image_grid_pinpoints"]:
+            for p in pinpoint:
+                image_grid_pinpoints.append(p)
+        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
+    if "image_crop_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
+    if "image_aspect_ratio" in v_hparams:
+        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
+    if "image_split_resolution" in v_hparams:
+        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
+    if "mm_patch_merge_type" in v_hparams:
+        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
+    if "mm_projector_type" in v_hparams:
+        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
+
 
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean

examples/llava/llava-cli.cpp

@@ -155,11 +155,29 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
         printf("system_prompt: %s\n", system_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
         printf("user_prompt: %s\n", user_prompt.c_str());
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     } else {
         // llava-1.5 native mode
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
+        if (params->verbose_prompt) {
+            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            for (int i = 0; i < (int) tmp.size(); i++) {
+                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+            }
+        }
     }
 
     eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@@ -171,13 +189,17 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-
+    std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
-
         printf("%s", tmp);
+        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
+        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
+        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
+
         fflush(stdout);
     }
 
@@ -196,7 +218,8 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
-    llama_backend_init(params->numa);
+    llama_backend_init();
+    llama_numa_init(params->numa);
 
     llama_model_params model_params = llama_model_params_from_gpt_params(*params);
 

examples/llava/llava-surgery-v2.py

@@ -0,0 +1,155 @@
+import argparse
+import glob
+import os
+import torch
+from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
+
+# Function to determine if file is a SafeTensor file
+def is_safetensor_file(file_path):
+    return file_path.endswith('.safetensors')
+
+
+# Unified loading function
+def load_model(file_path):
+    if is_safetensor_file(file_path):
+        tensors = {}
+        with safe_open(file_path, framework="pt", device="cpu") as f:
+            for key in f.keys():
+                tensors[key] = f.get_tensor(key).clone()
+                # output shape
+                print(f"{key} : {tensors[key].shape}")
+        return tensors, 'safetensor'
+    else:
+        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
+
+
+# Unified saving function
+def save_model(model, file_path, file_type):
+    if file_type == 'safetensor':
+        # safe_save(model, file_path)
+        save_file(model, file_path)
+    else:
+        torch.save(model, file_path)
+
+
+# Adapted function to clean vision tower from checkpoint
+def clean_vision_tower_from_checkpoint(checkpoint_path):
+    checkpoint, file_type = load_model(checkpoint_path)
+    # file_type = 'pytorch'
+    model_path = os.path.dirname(checkpoint_path)
+    print(f"Searching for vision tower tensors in {checkpoint_path}")
+    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
+
+    if len(clip_tensors) > 0:
+        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
+        # Adapted for file type
+        clip_path = os.path.join(model_path, "llava.clip")
+
+        if os.path.exists(clip_path):
+            print(f"Loading existing llava.clip from {clip_path}")
+            existing_clip, _ = load_model(clip_path)
+        else:
+            print(f"Creating new llava.clip at {clip_path}")
+            existing_clip = {}
+        # Update existing_clip with new tensors, avoid duplicates
+        for name in clip_tensors:
+            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
+            print(f"Adding {simple_name} to llava.clip")
+            if simple_name not in existing_clip:
+                existing_clip[simple_name] = checkpoint[name]
+
+        # Save the updated clip tensors back to llava.clip
+        save_model(existing_clip, clip_path, 'pytorch')
+
+        # Remove the tensors from the original checkpoint
+        for name in clip_tensors:
+            del checkpoint[name]
+
+        checkpoint_path = checkpoint_path
+        return True
+    return False
+
+def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
+    newline_checkpoint_path = None
+    projector_checkpoint_path = None
+
+    for path in checkpoint_paths:
+        checkpoint, _ = load_model(path)
+        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
+            newline_checkpoint_path = path
+        if projector(checkpoint):
+            projector_checkpoint_path = path
+
+    return newline_checkpoint_path, projector_checkpoint_path
+
+def newline_criteria(checkpoint):
+    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
+
+def proj_criteria(checkpoint):
+    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
+
+
+# Command-line interface setup
+ap = argparse.ArgumentParser()
+ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
+ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
+args = ap.parse_args()
+
+if args.clean_vision_tower:
+    # Generalized to handle both PyTorch and SafeTensors models
+    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
+    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+    for projector_checkpoint_path in checkpoint_paths:
+        print(f"Cleaning {projector_checkpoint_path}")
+        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
+            print(f"No vision tower found in {projector_checkpoint_path}")
+            # we break once none is found, so far all models append them at the end
+            # break
+    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
+
+# Now we look for the projector in the last checkpoint
+model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
+checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
+# last_checkpoint_path = checkpoint_paths[0]
+# first_checkpoint_path = checkpoint_paths[-1]
+newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
+
+print(f"Taking projector from {projector_checkpoint_path}")
+first_mm_tensors = []
+first_checkpoint = None
+if newline_checkpoint_path is not None:
+    print(f"Taking newline from {newline_checkpoint_path}")
+    first_checkpoint, file_type = load_model(newline_checkpoint_path)
+    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
+
+# Load the checkpoint
+mm_tensors = []
+last_checkpoint = None
+if projector_checkpoint_path is not None:
+    last_checkpoint, file_type = load_model(projector_checkpoint_path)
+    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
+
+if len(mm_tensors) == 0:
+    if last_checkpoint is not None:
+        for k, v in last_checkpoint.items():
+            print(k)
+    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
+    print("No tensors found. Is this a LLaVA model?")
+    exit()
+
+print(f"Found {len(mm_tensors)} tensors to extract.")
+print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
+# projector = {name: checkpoint.[name].float() for name in mm_tensors}
+projector = {}
+for name in mm_tensors:
+    projector[name] = last_checkpoint[name].float()
+for name in first_mm_tensors:
+    projector[name] = first_checkpoint[name].float()
+
+if len(projector) > 0:
+    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
+
+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-surgery.py

@@ -19,19 +19,12 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
 projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
-# remove these tensors from the checkpoint and save it again
-for name in mm_tensors:
-    del checkpoint[name]
-
 # BakLLaVA models contain CLIP tensors in it
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
 if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
-    # remove these tensors
-    for name in clip_tensors:
-        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -39,7 +32,6 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
-torch.save(checkpoint, path)
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

examples/llava/llava.cpp

@@ -2,32 +2,296 @@
 #include "common.h"
 #include "llama.h"
 #include "llava.h"
+#include "base64.hpp"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
+#include <numeric>
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
+struct clip_image_grid_shape {
+    int first;
+    int second;
+};
+
+/**
+ * Selects the best resolution from a list of possible resolutions based on the original size.
+ *
+ * @param original_size The original size of the image in the format (width, height).
+ * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+ * @return The best fit resolution in the format (width, height).
+ */
+static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
+    int original_width  = original_size.first;
+    int original_height = original_size.second;
+
+    std::pair<int, int> best_fit;
+    int max_effective_resolution = 0;
+    int min_wasted_resolution = std::numeric_limits<int>::max();
+
+    for (const auto& resolution : possible_resolutions) {
+        int width = resolution.first;
+        int height = resolution.second;
+        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
+        int downscaled_width  = static_cast<int>(original_width * scale);
+        int downscaled_height = static_cast<int>(original_height * scale);
+        int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
+        int wasted_resolution = (width * height) - effective_resolution;
+        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
+            max_effective_resolution = effective_resolution;
+            min_wasted_resolution = wasted_resolution;
+            best_fit = resolution;
+        }
+    }
+
+    return best_fit;
+}
+
+/**
+ * @brief Get the anyres image grid shape object
+ *
+ * @param image_size
+ * @param grid_pinpoints
+ * @param image_patch_size
+ * @return <int, int>
+ */
+static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
+    /**
+        Conversion from gguf flat array to vector:
+        std::vector<std::pair<int, int>> possible_resolutions;
+        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
+            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
+        }
+     */
+    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
+    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
+}
+
+// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
+static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
+    struct {
+        struct ggml_tensor * newline;
+        struct ggml_context * ctx;
+    } model;
+
+    const int32_t image_size = clip_image_size(ctx_clip);
+    const int32_t patch_size = clip_patch_size(ctx_clip);
+
+    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
+
+    int num_patches_width  = grid_shape.first;  // grid 1-4
+    int num_patches_height = grid_shape.second; // grid 1-4
+
+    const size_t num_images = num_patches_width * num_patches_height + 1;
+
+    // TODO: size calculation is not calculated - it's only tens of MB
+    size_t ctx_size = 0;
+
+    {
+        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
+        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
+    }
+
+    struct ggml_init_params params {
+        /*.mem_size   =*/ ctx_size,
+        /*.mem_buffer =*/ NULL,
+        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
+    };
+
+    // Python reference code for full unpad:
+    /*
+        base_image_feature = image_feature[0]
+        image_feature = image_feature[1:]
+        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
+        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
+        image_feature = unpad_image(image_feature, image_sizes[image_idx])
+        image_feature = torch.cat((
+            image_feature,
+            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
+        ), dim=-1)
+        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
+        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
+    */
+    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
+    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
+    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
+    // Once all images are processed to prepended the base_image_features without any changes.
+
+    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
+    /*
+        image_feature = image_feature.view(2, 2, 24, 24, 4096)
+        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
+        image_feature = image_feature.view(2, 24, 2, 24, 4096)
+        image_feature = image_feature.flatten(0, 3)
+
+        // Reshape to 4D tensor by merging the last two dimensions
+        image_feature = image_feature.view(2, 2, 24, 24*4096)
+        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
+        image_feature = image_feature.view(-1, 4096)
+    */
+
+    model.ctx = ggml_init(params);
+
+    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
+    model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
+    if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
+        if (newline_tmp->buffer == NULL) {
+            printf("newline_tmp tensor buffer is NULL\n");
+        }
+        ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
+    } else {
+        model.newline->data = newline_tmp->data;
+        if (model.newline->data == NULL) {
+            printf("newline_tmp tensor data is NULL\n");
+        }
+    }
+
+    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
+    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
+    // fill it with the image embeddings, ignoring the base
+    for (size_t i = 1; i < num_images; i++) {
+        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
+        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
+    }
+
+    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
+    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
+
+    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
+                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                num_patches_per_side,
+                                                                num_patches_width,
+                                                                num_patches_height,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
+                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
+    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
+    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
+    /**
+     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
+         image_feature = torch.cat((
+        image_feature,
+        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
+    ), dim=-1)
+     *
+     */
+
+    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
+    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
+    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
+    ggml_build_forward_expand(gf, flatten);
+    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
+    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+
+    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
+    // append without newline tokens (default behavior in llava_arch when not using unpad ):
+    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
+    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
+
+    // Debug: Test single segments
+    // Current findings: sending base image, sending a segment embedding all works similar to python
+    // However, permuted embeddings do not work yet (stride issue?)
+    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
+    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
+    // *n_img_pos_out=576;
+
+    ggml_free(model.ctx);
+    return true;
+}
 
-#include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = clip_image_f32_init();
-    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
+    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
+    clip_image_f32_batch img_res_v;
+    img_res_v.size = 0;
+    img_res_v.data = nullptr;
+    if (!clip_image_preprocess(ctx_clip, img, img_res_v)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
-        clip_image_f32_free(img_res);
+        delete[] img_res_v.data;
         return false;
     }
 
-    *n_img_pos = clip_n_patches(ctx_clip);
-
     const int64_t t_img_enc_start_us = ggml_time_us();
-    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
-    clip_image_f32_free(img_res);
-    if (!encoded) {
-        fprintf(stderr, "Unable to encode image\n");
 
-        return false;
+    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
+
+    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
+        // flat / default llava-1.5 type embedding
+        *n_img_pos = clip_n_patches(ctx_clip);
+        bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
+        delete[] img_res_v.data;
+        if (!encoded) {
+            fprintf(stderr, "Unable to encode image\n");
+
+            return false;
+        }
+    } else {
+        // spatial_unpad llava-1.6 type embedding
+        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
+        std::vector<float *> image_embd_v;
+        image_embd_v.resize(img_res_v.size);
+        for (size_t i = 0; i < img_res_v.size; i++) {
+            image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
+            const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
+            if (!encoded) {
+                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                return false;
+            }
+        }
+        const int64_t t_img_enc_batch_us = ggml_time_us();
+        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+
+        const int32_t * image_grid = clip_image_grid(ctx_clip);
+
+        std::vector<std::pair<int, int>> grid_pinpoints;
+        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
+            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
+        }
+
+        // free all img_res_v - not needed anymore
+        delete[] img_res_v.data;
+        img_res_v.size = 0;
+        img_res_v.data = nullptr;
+
+        const int32_t image_size = clip_image_size(ctx_clip);
+
+        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
+
+        int n_img_pos_out;
+        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
+        *n_img_pos = n_img_pos_out;
+
+        for (size_t i = 0; i < image_embd_v.size(); i++) {
+            free(image_embd_v[i]);
+        }
+        image_embd_v.clear();
+
+        // debug image/segment/normalization content:
+        // clip_image_u8 * tmp = clip_image_u8_init();
+        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
+        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
+    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
@@ -47,11 +311,10 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-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));
+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");
-        free(image_embd);
         return false;
     }
 
@@ -85,7 +348,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
     return true;
 }
 
-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) {
+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) {
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
@@ -142,7 +405,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     return true;
 }
 
-LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
     unsigned char* image_bytes;
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@@ -151,13 +414,13 @@ LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct
         return NULL;
     }
 
-    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
     free(image_bytes);
 
     return embed;
 }
 
-LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
+void llava_image_embed_free(struct llava_image_embed * embed) {
     free(embed->embed);
     free(embed);
 }

examples/llava/llava.h

@@ -3,7 +3,6 @@
 
 #include "ggml.h"
 
-
 #ifdef LLAMA_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
@@ -32,6 +31,8 @@ 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 */
@@ -42,7 +43,6 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
 
-
 #ifdef __cplusplus
 }
 #endif

examples/lookahead/lookahead.cpp

@@ -54,7 +54,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/lookup/lookup.cpp

@@ -31,7 +31,8 @@ int main(int argc, char ** argv){
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/main/README.md

@@ -283,7 +283,11 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### NUMA support
 
--   `--numa`: Attempt optimizations that help on some systems with non-uniform memory access. This currently consists of pinning an equal proportion of the threads to the cores on each NUMA node, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
+-   `--numa distribute`: Pin an equal proportion of the threads to the cores on each NUMA node. This will spread the load amongst all cores on the system, utilitizing all memory channels at the expense of potentially requiring memory to travel over the slow links between nodes.
+-   `--numa isolate`: Pin all threads to the NUMA node that the program starts on. This limits the number of cores and amount of memory that can be used, but guarantees all memory access remains local to the NUMA node.
+-   `--numa numactl`: Pin threads to the CPUMAP that is passed to the program by starting it with the numactl utility. This is the most flexible mode, and allow arbitraty core usage patterns, for example a map that uses all the cores on one NUMA nodes, and just enough cores on a second node to saturate the inter-node memory bus.
+
+ These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
 ### Memory Float 32
 

examples/main/main.cpp

@@ -185,7 +185,8 @@ int main(int argc, char ** argv) {
     }
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -333,6 +334,8 @@ 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
@@ -382,8 +385,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        if (params.n_keep > add_bos) {
+            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());
             }
@@ -508,6 +511,14 @@ 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) {
@@ -539,14 +550,14 @@ int main(int argc, char ** argv) {
                         break;
                     }
 
-                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_left    = n_past - params.n_keep;
                     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 + 1            , params.n_keep + n_discard + 1);
-                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -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);
 
                     n_past -= n_discard;
 
@@ -573,9 +584,9 @@ 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_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);
+                    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);
 
                     n_past -= bd;
 
@@ -766,6 +777,18 @@ 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

@@ -107,6 +107,9 @@ 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;
 
@@ -122,7 +125,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;
@@ -195,8 +199,8 @@ 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, 0, n_tokens_system);
+        for (int32_t i = 1; i <= n_clients; ++i) {
+            llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
         }
 
         LOG_TEE("\n");
@@ -220,15 +224,17 @@ 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 }, true);
+            llama_batch_add(batch, client.sampled, n_tokens_system + client.n_prompt + client.n_decoded, { client.id + 1 }, true);
 
             client.n_decoded += 1;
         }
 
         if (batch.n_tokens == 0) {
             // all sequences have ended - clear the entire KV cache
-            for (int i = 0; i < n_clients; ++i) {
-                llama_kv_cache_seq_rm(ctx, i, n_tokens_system, -1);
+            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);
             }
 
             LOG_TEE("%s: clearing the KV cache\n", __func__);
@@ -254,7 +260,7 @@ 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 }, false);
+                        llama_batch_add(batch, tokens_prompt[i], i + n_tokens_system, { client.id + 1 }, false);
                     }
 
                     // extract the logits only for the last token
@@ -365,7 +371,8 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id, n_tokens_system, -1);
+                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
 

examples/passkey/passkey.cpp

@@ -71,7 +71,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -125,7 +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", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+    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);
 
     // print the prompt token-by-token
 
@@ -145,10 +146,11 @@ 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_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);
+            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);
 
-            n_past -= bd;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
 
         llama_batch_clear(batch);
@@ -178,10 +180,12 @@ 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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-        n_past -= n_discard;
+        n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
 
         llama_batch_clear(batch);
 
@@ -207,10 +211,12 @@ 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_shift(ctx, 0, n_keep + n_discard, n_ctx,  -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);
 
-            n_past -= n_discard;
+            n_past = llama_kv_cache_seq_pos_max(ctx, 0) + 1;
         }
     }
 

examples/perplexity/perplexity.cpp

@@ -309,7 +309,7 @@ static void process_logits(int n_vocab, const float * logits, const int * tokens
 }
 
 static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params & params) {
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // BOS tokens will be added for each chunk before eval
@@ -442,18 +442,17 @@ 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) {
+static results_perplexity perplexity(llama_context * ctx, const gpt_params & params, const int32_t n_ctx) {
     if (params.ppl_stride > 0) {
         return perplexity_v2(ctx, params);
     }
 
-    // Download: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+    // Download: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     // Run `./perplexity -m models/7B/ggml-model-q4_0.bin -f wiki.test.raw`
     // Output: `perplexity: 13.5106 [114/114]`
     // 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()) {
@@ -499,13 +498,19 @@ 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, batch_size=%d\n", __func__, n_chunk, n_batch);
+    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);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -518,10 +523,26 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         log_probs.resize(n_ctx * nv);
     }
 
-    for (int i = 0; i < n_chunk; ++i) {
+    // 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) {
         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
@@ -531,22 +552,37 @@ 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);
 
-            // save original token and restore it after eval
-            const auto token_org = tokens[batch_start];
+            batch.n_tokens = 0;
+            for (int seq = 0; seq < n_seq_batch; seq++) {
+                int seq_start = batch_start + seq*n_ctx;
 
-            // 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));
+                // 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;
             }
 
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            if (llama_decode(ctx, batch)) {
                 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);
@@ -558,7 +594,7 @@ 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);
+            int total_seconds = (int)(t_total*n_chunk/n_seq);
             if (total_seconds >= 60*60) {
                 fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
@@ -566,37 +602,31 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
-        // 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;
+        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;
 
-        // 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);
+            // 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);
+            }
         }
         fflush(stdout);
 
@@ -615,6 +645,8 @@ 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};
 }
 
@@ -809,7 +841,7 @@ 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 = 4*max_tasks_per_batch;
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1086,7 +1118,7 @@ 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 = 2*max_tasks_per_batch;
+    const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1438,7 +1470,7 @@ 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 = 4*max_tasks_per_batch;
+    const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_max_seq(ctx));
 
     llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
 
@@ -1623,7 +1655,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     uint32_t n_ctx;
     in.read((char *)&n_ctx, sizeof(n_ctx));
     if (n_ctx > llama_n_ctx(ctx)) {
-        fprintf(stderr, "%s: %s has been computed with %d, while the current context is %d. Increase it with -c and retry\n",
+        fprintf(stderr, "%s: %s has been computed with %u, while the current context is %d. Increase it with -c and retry\n",
                 __func__, params.logits_file.c_str(), n_ctx, params.n_ctx);
     }
 
@@ -1782,13 +1814,24 @@ 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;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    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);
+    }
 
     if (params.ppl_stride > 0) {
         fprintf(stderr, "Will perform strided perplexity calculation -> adjusting context size from %d to %d\n",
@@ -1809,11 +1852,15 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     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) {
@@ -1843,7 +1890,7 @@ int main(int argc, char ** argv) {
     } else if (params.kl_divergence) {
         kl_divergence(ctx, params);
     } else {
-        results = perplexity(ctx, params);
+        results = perplexity(ctx, params, n_ctx);
     }
 
     llama_print_timings(ctx);

examples/quantize/quantize.cpp

@@ -23,14 +23,21 @@ 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" },
-    { "Q3_K_XS",LLAMA_FTYPE_MOSTLY_Q3_K_XS,"3-bit extra small quantization"   , },
+    { "IQ3_XS", LLAMA_FTYPE_MOSTLY_IQ3_XS, " 3.3 bpw 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", },
@@ -237,7 +244,7 @@ int main(int argc, char ** argv) {
         params.imatrix = &imatrix_data;
     }
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     // parse command line arguments
     const std::string fname_inp = argv[arg_idx];
@@ -287,9 +294,11 @@ 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_Q2_K_S) && imatrix_data.empty()) {
+    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 IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
         fprintf(stderr, "===============================================================================================\n\n\n");
         return 1;
     }

examples/server-embd.py

@@ -0,0 +1,34 @@
+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

@@ -1,12 +1,18 @@
 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 oai.hpp utils.hpp json.hpp httplib.h)
+add_executable(${TARGET} server.cpp 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 llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${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(${TARGET} PRIVATE CPPHTTPLIB_OPENSSL_SUPPORT)
+endif()
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/README.md

@@ -1,11 +1,24 @@
-# llama.cpp/example/server
+# LLaMA.cpp HTTP Server
 
-This example demonstrates a simple HTTP API server and a simple web front end to interact with llama.cpp.
+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**.
 
-Command line options:
+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:**
 
 - `--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.
@@ -16,13 +29,20 @@ Command line options:
 - `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 - `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 - `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
+- `--numa STRATEGY`: Attempt one of the below optimization strategies  that help on some NUMA systems
+- `--numa distribute`: Spread execution evenly over all nodes
+- `--numa isolate`: Only spawn threads on CPUs on the node that execution started on
+- `--numa numactl`: Use the CPU map provided by numactl
+if run without this previously, it is recommended to drop the system page cache before using this
+see https://github.com/ggerganov/llama.cpp/issues/1437
+
 - `--numa`: Attempt optimizations that help on some NUMA systems.
 - `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 - `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 - `-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 examples/server/public)
+- `--path`: path from which to serve static files (default: disabled)
 - `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
 - `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
 - `--embedding`: Enable embedding extraction, Default: disabled.
@@ -32,6 +52,16 @@ Command line options:
 - `--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)
+- `--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
 
@@ -49,6 +79,28 @@ 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:
@@ -88,6 +140,12 @@ 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.
@@ -125,19 +183,27 @@ node index.js
 ## API Endpoints
 
 - **GET** `/health`: Returns the current state of the server:
-  - `{"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.
+  - 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.
 
 - **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, 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.
+    `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
 
     `temperature`: Adjust the randomness of the generated text (default: 0.8).
 
-    `dynatemp_range`: Dynamic temperature range (default: 0.0, 0.0 = disabled).
+    `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_exponent`: Dynamic temperature exponent (default: 1.0).
 
@@ -189,14 +255,18 @@ node index.js
 
     `n_probs`: If greater than 0, the response also contains the probabilities of top N tokens for each generated token (default: 0)
 
+    `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum (default: 0)
+
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `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`: Save the prompt and generation for avoid reprocess entire prompt if a part of this isn't change (default: false)
+    `cache_prompt`: Re-use previously cached prompt from the last request if possible. This may prevent re-caching the prompt from scratch. (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)
 
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. (default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values)
+
 ### Result JSON
 
 - Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
@@ -224,7 +294,7 @@ 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`
+- `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.).
 - `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
@@ -242,7 +312,7 @@ Notice that each `probs` is an array of length `n_probs`.
 
     `content`: Set the text to tokenize.
 
-    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`.
+    Note that a special `BOS` token is never inserted.
 
 - **POST** `/detokenize`: Convert tokens to text.
 
@@ -286,7 +356,7 @@ 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. Compared to `api_like_OAI.py` this API implementation does not require a wrapper to be served.
+- **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.
 
     *Options:*
 
@@ -370,6 +440,81 @@ Notice that each `probs` is an array of length `n_probs`.
     }'
     ```
 
+- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+
+### Result JSON
+
+```json
+[
+    {
+        "dynatemp_exponent": 1.0,
+        "dynatemp_range": 0.0,
+        "frequency_penalty": 0.0,
+        "grammar": "",
+        "id": 0,
+        "ignore_eos": false,
+        "logit_bias": [],
+        "min_p": 0.05000000074505806,
+        "mirostat": 0,
+        "mirostat_eta": 0.10000000149011612,
+        "mirostat_tau": 5.0,
+        "model": "llama-2-7b-32k-instruct.Q2_K.gguf",
+        "n_ctx": 2048,
+        "n_keep": 0,
+        "n_predict": 100000,
+        "n_probs": 0,
+        "next_token": {
+            "has_next_token": true,
+            "n_remain": -1,
+            "n_decoded": 0,
+            "stopped_eos": false,
+            "stopped_limit": false,
+            "stopped_word": false,
+            "stopping_word": ""
+        },
+        "penalize_nl": true,
+        "penalty_prompt_tokens": [],
+        "presence_penalty": 0.0,
+        "prompt": "Say hello to llama.cpp",
+        "repeat_last_n": 64,
+        "repeat_penalty": 1.100000023841858,
+        "samplers": [
+            "top_k",
+            "tfs_z",
+            "typical_p",
+            "top_p",
+            "min_p",
+            "temperature"
+        ],
+        "seed": 42,
+        "state": 1,
+        "stop": [
+            "\n"
+        ],
+        "stream": false,
+        "task_id": 0,
+        "temperature": 0.0,
+        "tfs_z": 1.0,
+        "top_k": 40,
+        "top_p": 0.949999988079071,
+        "typical_p": 1.0,
+        "use_penalty_prompt_tokens": false
+    }
+]
+```
+
+- **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
@@ -413,24 +558,11 @@ bash chat.sh
 
 ### API like OAI
 
-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
+The HTTP server supports OAI-like API
 
 ### Extending or building alternative Web Front End
 
-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.
+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

@@ -1,228 +0,0 @@
-#!/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

@@ -0,0 +1,88 @@
+### 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

@@ -0,0 +1,120 @@
+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

@@ -1,227 +0,0 @@
-#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/public/index.html

@@ -234,6 +234,7 @@
       mirostat_eta: 0.1, // learning rate
       grammar: '',
       n_probs: 0, // no completion_probabilities,
+      min_keep: 0, // min probs from each sampler,
       image_data: [],
       cache_prompt: true,
       api_key: ''
@@ -791,6 +792,9 @@
             <fieldset>
               ${IntField({ label: "Show Probabilities", max: 10, min: 0, name: "n_probs", value: params.value.n_probs })}
             </fieldset>
+            <fieldset>
+              ${IntField({ label: "Min Probabilities from each Sampler", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
+            </fieldset>
             <fieldset>
               <label for="api_key">API Key</label>
               <input type="text" name="api_key" value="${params.value.api_key}" placeholder="Enter API key" oninput=${updateParams} />

examples/server/tests/README.md

@@ -0,0 +1,67 @@
+# 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

@@ -0,0 +1,94 @@
+@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

@@ -0,0 +1,100 @@
+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

@@ -0,0 +1,5 @@
+# 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

@@ -0,0 +1,101 @@
+@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

@@ -0,0 +1,55 @@
+# 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

@@ -0,0 +1,52 @@
+@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

@@ -0,0 +1,83 @@
+@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

@@ -0,0 +1,22 @@
+# 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

@@ -0,0 +1,6 @@
+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

@@ -0,0 +1,12 @@
+#!/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

@@ -1,19 +1,21 @@
 #pragma once
 
-#include <string>
-#include <vector>
-#include <set>
-#include <mutex>
-#include <condition_variable>
-#include <unordered_map>
+#include "llama.h"
+#include "common.h"
 
 #include "json.hpp"
 
-#include "../llava/clip.h"
+#include <string>
+#include <vector>
+#include <sstream>
+#include <random>
+
+#define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::json;
 
 extern bool server_verbose;
+extern bool server_log_json;
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -27,414 +29,110 @@ extern bool server_verbose;
     {                                                                    \
         if (server_verbose)                                              \
         {                                                                \
-            server_log("VERBOSE", __func__, __LINE__, MSG, __VA_ARGS__); \
+            server_log("VERB", __func__, __LINE__, MSG, __VA_ARGS__); \
         }                                                                \
     } while (0)
 #endif
 
-#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
-//
+#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__)
 
 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;
 }
 
-inline std::string format_llama2(std::vector<json> messages)
-{
-    std::ostringstream output;
-    bool is_inside_turn = false;
+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)},
+    };
 
-    for (auto it = messages.begin(); it != messages.end(); ++it) {
-        if (!is_inside_turn) {
-            output << "[INST] ";
+    if (server_log_json) {
+        log.merge_patch( {
+            {"level",    level},
+            {"function", function},
+            {"line",     line},
+            {"msg",      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;
+
+        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::stringstream ss;
+        ss << buf << " |";
+        for (const auto& el : log.items())
+        {
+            const std::string value = el.value().dump(-1, ' ', false, json::error_handler_t::replace);
+            ss << " " << el.key() << "=" << value;
+        }
+
+        const std::string str = ss.str();
+        printf("%.*s\n", (int)str.size(), str.data());
+        fflush(stdout);
     }
-
-    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
+// chat template utils
 //
 
-struct llama_server_queue {
-    int id = 0;
-    std::mutex mutex_tasks;
-    // 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;
+// 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;
+}
 
-    // 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;
+// 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();
     }
 
-    // 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));
+    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());
     }
 
-    // Get the next id for creating anew task
-    int get_new_id() {
-        std::unique_lock<std::mutex> lock(mutex_tasks);
-        return id++;
-    }
+    const std::string formatted_chat(buf.data(), res);
 
-    // Register function to process a new task
-    void on_new_task(std::function<void(task_server&)> callback) {
-        callback_new_task = callback;
-    }
+    LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
 
-    // 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();
-    }
-
-    // Start the main loop. This call is blocking
-    [[noreturn]]
-    void start_loop() {
-        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()) {
-                    condition_tasks.wait(lock, [&]{
-                        return !queue_tasks.empty();
-                    });
-                }
-            }
-        }
-    }
-
-    //
-    // 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;
-                }
-            }
-        }
-
-        // should never reach here
-    }
-
-    // Register the function to update multitask
-    void on_multitask_update(callback_multitask_t callback) {
-        callback_update_multitask = callback;
-    }
-
-    // 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;
-            }
-
-            if (result.id == task_id)
-            {
-                LOG_VERBOSE("queue_results.push_back", {});
-                queue_results.push_back(result);
-                condition_results.notify_one();
-                return;
-            }
-        }
-    }
-};
+    return formatted_chat;
+}
 
 //
 // base64 utils (TODO: move to common in the future)
@@ -445,13 +143,11 @@ 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;
@@ -463,13 +159,10 @@ 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]);
             }
 
@@ -477,23 +170,20 @@ 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]);
         }
 
@@ -501,8 +191,7 @@ 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]);
         }
     }
@@ -514,8 +203,7 @@ 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;
@@ -530,9 +218,327 @@ 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/simple/simple.cpp

@@ -31,7 +31,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 

examples/speculative/README.md

@@ -6,3 +6,4 @@ 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

@@ -5,6 +5,7 @@
 #include <cstdio>
 #include <string>
 #include <vector>
+#include <set>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -18,6 +19,7 @@ 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;
 };
@@ -37,12 +39,15 @@ 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");
@@ -50,7 +55,8 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model_tgt = NULL;
     llama_model * model_dft = NULL;
@@ -165,7 +171,9 @@ 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
-    params.sparams.temp = -1.0f;    // force greedy sampling with probs for the draft model
+    if (params.sparams.temp == 0) {
+        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);
@@ -181,12 +189,15 @@ 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());
@@ -195,48 +206,156 @@ 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 matches = false;
+                bool accept = false;
+                if (params.sparams.temp > 0) {
+                    // stochastic verification
 
-                for (int s = 0; s < n_seq_dft; ++s) {
-                    if (!drafts[s].active) {
-                        continue;
+                    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);
+                                }
+                            }
+                        }
                     }
 
-                    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());
+                    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);
+                    }
 
-                        s_keep = s;
-                        matches = true;
-                    } else {
-                        drafts[s].active = false;
+                } 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;
+                        }
                     }
                 }
 
-                if (matches) {
+                if (token_id == llama_token_eos(model_tgt)) {
+                    has_eos = true;
+                }
+                ++n_predict;
+
+                if (accept) {
                     ++n_accept;
                     ++n_past_tgt;
                     ++n_past_dft;
@@ -244,17 +363,21 @@ 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());
-                        fflush(stdout);
+                    } else {
+                        printf("%s", token_str.c_str());
                     }
+                    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", id, token_str.c_str());
+        {
+            LOG("the sampled target token (%d, '%s') did not match, or we ran out of drafted tokens\n", token_id, token_str.c_str());
 
             // TODO: simplify
             {
@@ -274,21 +397,21 @@ 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(id);
+            drafts[0].tokens.push_back(token_id);
+            drafts[0].dists.push_back(std::vector<llama_token_data>());
             drafts[0].i_batch_tgt.push_back(0);
 
             llama_batch_clear(batch_dft);
-            llama_batch_add  (batch_dft, id, n_past_dft, { 0 }, true);
+            llama_batch_add  (batch_dft, token_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) {
@@ -333,12 +456,6 @@ 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
@@ -366,6 +483,7 @@ 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;
 
@@ -388,6 +506,8 @@ 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);
@@ -439,6 +559,7 @@ 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 argc, char ** argv) {
+int main() {
     ggml_backend_sycl_print_sycl_devices();
     return 0;
 }

examples/sycl/run-llama2.sh

@@ -8,12 +8,19 @@ INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 source /opt/intel/oneapi/setvars.sh
 
 if [ $# -gt 0 ]; then
-    export GGML_SYCL_DEVICE=$1
+    GGML_SYCL_DEVICE=$1
 else
-    export GGML_SYCL_DEVICE=0
+    GGML_SYCL_DEVICE=0
 fi
-echo GGML_SYCL_DEVICE=$GGML_SYCL_DEVICE
+echo "use $GGML_SYCL_DEVICE as main GPU"
 #export GGML_SYCL_DEBUG=1
-./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
-#./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 5 -e -ngl 33 -t 1 -s 0
+
+
+#ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
+
+#use all GPUs with same max compute units
+ZES_ENABLE_SYSMAN=1 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "${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
 

examples/tokenize/tokenize.cpp

@@ -17,7 +17,7 @@ int main(int argc, char ** argv) {
 
     const bool printing_ids = argc > 3 && std::string(argv[3]) == "--ids";
 
-    llama_backend_init(false);
+    llama_backend_init();
 
     llama_model_params model_params = llama_model_default_params();
     model_params.vocab_only = true;

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

@@ -50,9 +50,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -111,13 +111,13 @@ static const char * LLM_TENSOR_FFN_DOWN      = "blk.%d.ffn_down";
 static const char * LLM_TENSOR_FFN_UP        = "blk.%d.ffn_up";
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
-    printf("%s: n_layer: %d\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
+    printf("%s: n_head:  %u\n", __func__, params->n_head);
+    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
+    printf("%s: n_layer: %u\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
 }
 
 static void set_param_model(struct my_llama_model * model) {
@@ -140,9 +140,9 @@ static void set_param_model(struct my_llama_model * model) {
         ggml_set_param(ctx, layer.wv);
         ggml_set_param(ctx, layer.wo);
         ggml_set_param(ctx, layer.ffn_norm);
-        ggml_set_param(ctx, layer.w1);
-        ggml_set_param(ctx, layer.w2);
-        ggml_set_param(ctx, layer.w3);
+        ggml_set_param(ctx, layer.ffn_gate);
+        ggml_set_param(ctx, layer.ffn_down);
+        ggml_set_param(ctx, layer.ffn_up);
     }
 }
 
@@ -198,9 +198,9 @@ static void init_model(struct my_llama_model * model) {
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
-        layer.w1 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
-        layer.w2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
-        layer.w3 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_gate = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
+        layer.ffn_down = ggml_new_tensor_2d(ctx, GGML_TYPE_F32,   n_ff, n_embd);
+        layer.ffn_up   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd,   n_ff);
 
         ggml_set_name(layer.attention_norm, tni(LLM_TENSOR_ATTN_NORM, i));
 
@@ -211,9 +211,9 @@ static void init_model(struct my_llama_model * model) {
 
         ggml_set_name(layer.ffn_norm,       tni(LLM_TENSOR_FFN_NORM, i));
 
-        ggml_set_name(layer.w1,             tni(LLM_TENSOR_FFN_GATE, i));
-        ggml_set_name(layer.w2,             tni(LLM_TENSOR_FFN_DOWN, i));
-        ggml_set_name(layer.w3,             tni(LLM_TENSOR_FFN_UP, i));
+        ggml_set_name(layer.ffn_gate,       tni(LLM_TENSOR_FFN_GATE, i));
+        ggml_set_name(layer.ffn_down,       tni(LLM_TENSOR_FFN_DOWN, i));
+        ggml_set_name(layer.ffn_up,         tni(LLM_TENSOR_FFN_UP, i));
     }
 
     set_param_model(model);
@@ -244,9 +244,9 @@ static void randomize_model(struct my_llama_model * model, int seed, float mean,
 
         randomize_tensor_normal(layer.ffn_norm, rnd);
 
-        randomize_tensor_normal(layer.w1, rnd);
-        randomize_tensor_normal(layer.w2, rnd);
-        randomize_tensor_normal(layer.w3, rnd);
+        randomize_tensor_normal(layer.ffn_gate, rnd);
+        randomize_tensor_normal(layer.ffn_down, rnd);
+        randomize_tensor_normal(layer.ffn_up,   rnd);
     }
 
     free_random_normal_distribution(rnd);
@@ -356,11 +356,11 @@ static struct ggml_tensor * llama_build_train_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, f_norm_rms_eps);                    set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, layer.ffn_norm, t22);                    set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                               set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.w3, t24);                          set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.w1, t24);                          set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, layer.ffn_up, t24);                      set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, layer.ffn_gate, t24);                    set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                    set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                               set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.w2, t28);                          set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, layer.ffn_down, t28);                    set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                               set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         checkpoints.push_back(cur);
@@ -521,9 +521,9 @@ static void load_llama_model_gguf(struct gguf_context * fctx, struct ggml_contex
         copy_tensor_by_name(layer.wv,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_V, i));
         copy_tensor_by_name(layer.wo,             f_ggml_ctx, tni(LLM_TENSOR_ATTN_OUT, i));
         copy_tensor_by_name(layer.ffn_norm,       f_ggml_ctx, tni(LLM_TENSOR_FFN_NORM, i));
-        copy_tensor_by_name(layer.w1,             f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
-        copy_tensor_by_name(layer.w2,             f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
-        copy_tensor_by_name(layer.w3,             f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
+        copy_tensor_by_name(layer.ffn_gate,       f_ggml_ctx, tni(LLM_TENSOR_FFN_GATE, i));
+        copy_tensor_by_name(layer.ffn_down,       f_ggml_ctx, tni(LLM_TENSOR_FFN_DOWN, i));
+        copy_tensor_by_name(layer.ffn_up,         f_ggml_ctx, tni(LLM_TENSOR_FFN_UP, i));
     }
 }
 
@@ -664,9 +664,9 @@ static void save_llama_model_gguf(struct gguf_context * fctx, const char * fn_vo
         gguf_add_tensor(fctx, layer.wv);
         gguf_add_tensor(fctx, layer.wo);
         gguf_add_tensor(fctx, layer.ffn_norm);
-        gguf_add_tensor(fctx, layer.w1);
-        gguf_add_tensor(fctx, layer.w2);
-        gguf_add_tensor(fctx, layer.w3);
+        gguf_add_tensor(fctx, layer.ffn_gate);
+        gguf_add_tensor(fctx, layer.ffn_down);
+        gguf_add_tensor(fctx, layer.ffn_up);
     }
 }
 
@@ -915,9 +915,9 @@ static int64_t get_parameter_count(struct my_llama_model* model) {
         nx += ggml_nelements(layer.wv);
         nx += ggml_nelements(layer.wo);
         nx += ggml_nelements(layer.ffn_norm);
-        nx += ggml_nelements(layer.w1);
-        nx += ggml_nelements(layer.w2);
-        nx += ggml_nelements(layer.w3);
+        nx += ggml_nelements(layer.ffn_gate);
+        nx += ggml_nelements(layer.ffn_down);
+        nx += ggml_nelements(layer.ffn_up);
     }
     return nx;
 }
@@ -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_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_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": 1706830856,
-        "narHash": "sha256-a0NYyp+h9hlb7ddVz4LUn1vT/PLwqfrWYcHMvFB1xYg=",
+        "lastModified": 1709336216,
+        "narHash": "sha256-Dt/wOWeW6Sqm11Yh+2+t0dfEWxoMxGBvv3JpIocFl9E=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "b253292d9c0a5ead9bc98c4e9a26c6312e27d69f",
+        "rev": "f7b3c975cf067e56e7cda6cb098ebe3fb4d74ca2",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1707268954,
-        "narHash": "sha256-2en1kvde3cJVc3ZnTy8QeD2oKcseLFjYPLKhIGDanQ0=",
+        "lastModified": 1709703039,
+        "narHash": "sha256-6hqgQ8OK6gsMu1VtcGKBxKQInRLHtzulDo9Z5jxHEFY=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "f8e2ebd66d097614d51a56a755450d4ae1632df1",
+        "rev": "9df3e30ce24fd28c7b3e2de0d986769db5d6225d",
         "type": "github"
       },
       "original": {
@@ -37,11 +37,11 @@
     "nixpkgs-lib": {
       "locked": {
         "dir": "lib",
-        "lastModified": 1706550542,
-        "narHash": "sha256-UcsnCG6wx++23yeER4Hg18CXWbgNpqNXcHIo5/1Y+hc=",
+        "lastModified": 1709237383,
+        "narHash": "sha256-cy6ArO4k5qTx+l5o+0mL9f5fa86tYUX3ozE1S+Txlds=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "97b17f32362e475016f942bbdfda4a4a72a8a652",
+        "rev": "1536926ef5621b09bba54035ae2bb6d806d72ac8",
         "type": "github"
       },
       "original": {