llama : refactor tensor offloading as callback

ggml-ci

.devops/full-cuda.Dockerfile

@@ -14,8 +14,7 @@ ARG CUDA_DOCKER_ARCH=all
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
+COPY requirements.txt requirements.txt
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/full-rocm.Dockerfile

@@ -23,8 +23,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1101 \
     gfx1102
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
+COPY requirements.txt requirements.txt
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/full.Dockerfile

@@ -5,8 +5,7 @@ FROM ubuntu:$UBUNTU_VERSION as build
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
+COPY requirements.txt requirements.txt
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/main-intel.Dockerfile

@@ -1,26 +0,0 @@
-ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
-ARG UBUNTU_VERSION=22.04
-
-FROM intel/hpckit:$ONEAPI_VERSION as build
-
-RUN apt-get update && \
-    apt-get install -y git
-
-WORKDIR /app
-
-COPY . .
-
-# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
-RUN mkdir build && \
-    cd build && \
-    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
-    cmake --build . --config Release --target main server
-
-FROM ubuntu:$UBUNTU_VERSION as runtime
-
-COPY --from=build /app/build/bin/main /main
-COPY --from=build /app/build/bin/server /server
-
-ENV LC_ALL=C.utf8
-
-ENTRYPOINT [ "/main" ]

.devops/main-rocm.Dockerfile

@@ -23,8 +23,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1101 \
     gfx1102
 
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
+COPY requirements.txt requirements.txt
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/nix/apps.nix

@@ -1,22 +0,0 @@
-{
-  perSystem =
-    { config, lib, ... }:
-    {
-      apps =
-        let
-          inherit (config.packages) default;
-          binaries = [
-            "llama"
-            "llama-embedding"
-            "llama-server"
-            "quantize"
-            "train-text-from-scratch"
-          ];
-          mkApp = name: {
-            type = "app";
-            program = "${default}/bin/${name}";
-          };
-        in
-        lib.genAttrs binaries mkApp;
-    };
-}

.devops/nix/devshells.nix

@@ -1,13 +0,0 @@
-{
-  perSystem =
-    { config, lib, ... }:
-    {
-      devShells =
-        lib.concatMapAttrs
-          (name: package: {
-            ${name} = package.passthru.shell;
-            ${name + "-extra"} = package.passthru.shell-extra;
-          })
-          config.packages;
-    };
-}

.devops/nix/jetson-support.nix

@@ -1,39 +0,0 @@
-{ inputs, ... }:
-{
-  perSystem =
-    {
-      config,
-      system,
-      lib,
-      pkgsCuda,
-      ...
-    }:
-    {
-      legacyPackages =
-        let
-          caps.llamaPackagesXavier = "7.2";
-          caps.llamaPackagesOrin = "8.7";
-          caps.llamaPackagesTX2 = "6.2";
-          caps.llamaPackagesNano = "5.3";
-
-          pkgsFor =
-            cap:
-            import inputs.nixpkgs {
-              inherit system;
-              config = {
-                cudaSupport = true;
-                cudaCapabilities = [ cap ];
-                cudaEnableForwardCompat = false;
-                inherit (pkgsCuda.config) allowUnfreePredicate;
-              };
-            };
-        in
-        builtins.mapAttrs (name: cap: (pkgsFor cap).callPackage ./scope.nix { }) caps;
-
-      packages = lib.optionalAttrs (system == "aarch64-linux") {
-        jetson-xavier = config.legacyPackages.llamaPackagesXavier.llama-cpp;
-        jetson-orin = config.legacyPackages.llamaPackagesOrin.llama-cpp;
-        jetson-nano = config.legacyPackages.llamaPackagesNano.llama-cpp;
-      };
-    };
-}

.devops/nix/nixpkgs-instances.nix

@@ -1,47 +0,0 @@
-{ inputs, ... }:
-{
-  # The _module.args definitions are passed on to modules as arguments. E.g.
-  # the module `{ pkgs ... }: { /* config */ }` implicitly uses
-  # `_module.args.pkgs` (defined in this case by flake-parts).
-  perSystem =
-    { system, ... }:
-    {
-      _module.args = {
-        # Note: bringing up https://zimbatm.com/notes/1000-instances-of-nixpkgs
-        # again, the below creates several nixpkgs instances which the
-        # flake-centric CLI will be forced to evaluate e.g. on `nix flake show`.
-        #
-        # This is currently "slow" and "expensive", on a certain scale.
-        # This also isn't "right" in that this hinders dependency injection at
-        # the level of flake inputs. This might get removed in the foreseeable
-        # future.
-        #
-        # Note that you can use these expressions without Nix
-        # (`pkgs.callPackage ./devops/nix/scope.nix { }` is the entry point).
-
-        pkgsCuda = import inputs.nixpkgs {
-          inherit system;
-          # Ensure dependencies use CUDA consistently (e.g. that openmpi, ucc,
-          # and ucx are built with CUDA support)
-          config.cudaSupport = true;
-          config.allowUnfreePredicate =
-            p:
-            builtins.all
-              (
-                license:
-                license.free
-                || builtins.elem license.shortName [
-                  "CUDA EULA"
-                  "cuDNN EULA"
-                ]
-              )
-              (p.meta.licenses or [ p.meta.license ]);
-        };
-        # Ensure dependencies use ROCm consistently
-        pkgsRocm = import inputs.nixpkgs {
-          inherit system;
-          config.rocmSupport = true;
-        };
-      };
-    };
-}

.devops/nix/package.nix

@@ -1,277 +0,0 @@
-{
-  lib,
-  config,
-  stdenv,
-  mkShell,
-  cmake,
-  ninja,
-  pkg-config,
-  git,
-  python3,
-  mpi,
-  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
-  cudaPackages,
-  darwin,
-  rocmPackages,
-  clblast,
-  useBlas ? builtins.all (x: !x) [
-    useCuda
-    useMetalKit
-    useOpenCL
-    useRocm
-  ],
-  useCuda ? config.cudaSupport,
-  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
-  useMpi ? false, # Increases the runtime closure size by ~700M
-  useOpenCL ? false,
-  useRocm ? config.rocmSupport,
-  llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
-}@inputs:
-
-let
-  inherit (lib)
-    cmakeBool
-    cmakeFeature
-    optionals
-    strings
-    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" ]
-    ++ lib.optionals useCuda [ "CUDA" ]
-    ++ lib.optionals useMetalKit [ "MetalKit" ]
-    ++ lib.optionals useMpi [ "MPI" ]
-    ++ lib.optionals useOpenCL [ "OpenCL" ]
-    ++ lib.optionals useRocm [ "ROCm" ];
-
-  pnameSuffix =
-    strings.optionalString (suffices != [ ])
-      "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
-  descriptionSuffix =
-    strings.optionalString (suffices != [ ])
-      ", accelerated with ${strings.concatStringsSep ", " suffices}";
-
-  # TODO: package the Python in this repository in a Nix-like way.
-  # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
-  # is PEP 517-compatible, and ensure the correct .dist-info is generated.
-  # https://peps.python.org/pep-0517/
-  llama-python = python3.withPackages (
-    ps: [
-      ps.numpy
-      ps.sentencepiece
-    ]
-  );
-
-  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
-  llama-python-extra = python3.withPackages (
-    ps: [
-      ps.numpy
-      ps.sentencepiece
-      ps.tiktoken
-      ps.torchWithoutCuda
-      ps.transformers
-    ]
-  );
-
-  # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
-  # separately
-  darwinBuildInputs =
-    with darwin.apple_sdk.frameworks;
-    [
-      Accelerate
-      CoreVideo
-      CoreGraphics
-    ]
-    ++ optionals useMetalKit [ MetalKit ];
-
-  cudaBuildInputs = with cudaPackages; [
-    cuda_cccl.dev # <nv/target>
-
-    # A temporary hack for reducing the closure size, remove once cudaPackages
-    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
-    cuda_cudart.dev
-    cuda_cudart.lib
-    cuda_cudart.static
-    libcublas.dev
-    libcublas.lib
-    libcublas.static
-  ];
-
-  rocmBuildInputs = with rocmPackages; [
-    clr
-    hipblas
-    rocblas
-  ];
-in
-
-effectiveStdenv.mkDerivation (
-  finalAttrs: {
-    pname = "llama-cpp${pnameSuffix}";
-    version = llamaVersion;
-
-    # Note: none of the files discarded here are visible in the sandbox or
-    # affect the output hash. This also means they can be modified without
-    # triggering a rebuild.
-    src = lib.cleanSourceWith {
-      filter =
-        name: type:
-        let
-          noneOf = builtins.all (x: !x);
-          baseName = baseNameOf name;
-        in
-        noneOf [
-          (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
-          (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
-          (lib.hasPrefix "." baseName) # Skip hidden files and directories
-          (baseName == "flake.lock")
-        ];
-      src = lib.cleanSource ../../.;
-    };
-
-    postPatch = ''
-      substituteInPlace ./ggml-metal.m \
-        --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-
-      # TODO: Package up each Python script or service appropriately.
-      # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
-      # we could make those *.py into setuptools' entrypoints
-      substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
-    '';
-
-    nativeBuildInputs =
-      [
-        cmake
-        ninja
-        pkg-config
-        git
-      ]
-      ++ optionals useCuda [
-        cudaPackages.cuda_nvcc
-
-        # TODO: Replace with autoAddDriverRunpath
-        # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
-        cudaPackages.autoAddOpenGLRunpathHook
-      ];
-
-    buildInputs =
-      optionals effectiveStdenv.isDarwin darwinBuildInputs
-      ++ optionals useCuda cudaBuildInputs
-      ++ optionals useMpi [ mpi ]
-      ++ optionals useOpenCL [ clblast ]
-      ++ optionals useRocm rocmBuildInputs;
-
-    cmakeFlags =
-      [
-        (cmakeBool "LLAMA_NATIVE" false)
-        (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" true)
-        (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
-        (cmakeBool "LLAMA_BLAS" useBlas)
-        (cmakeBool "LLAMA_CLBLAST" useOpenCL)
-        (cmakeBool "LLAMA_CUBLAS" useCuda)
-        (cmakeBool "LLAMA_HIPBLAS" useRocm)
-        (cmakeBool "LLAMA_METAL" useMetalKit)
-        (cmakeBool "LLAMA_MPI" useMpi)
-      ]
-      ++ optionals useCuda [
-        (
-          with cudaPackages.flags;
-          cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
-            builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
-          )
-        )
-      ]
-      ++ optionals useRocm [
-        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
-        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
-
-        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-        # and select the line that matches the current nixpkgs version of rocBLAS.
-        # Should likely use `rocmPackages.clr.gpuTargets`.
-        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
-      ]
-      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
-      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
-
-    # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
-    # if they haven't been added yet.
-    postInstall = ''
-      mv $out/bin/main $out/bin/llama
-      mv $out/bin/server $out/bin/llama-server
-      mkdir -p $out/include
-      cp $src/llama.h $out/include/
-    '';
-
-    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
-    passthru = {
-      inherit
-        useBlas
-        useCuda
-        useMetalKit
-        useMpi
-        useOpenCL
-        useRocm
-        ;
-
-      shell = mkShell {
-        name = "shell-${finalAttrs.finalPackage.name}";
-        description = "contains numpy and sentencepiece";
-        buildInputs = [ llama-python ];
-        inputsFrom = [ finalAttrs.finalPackage ];
-        shellHook = ''
-          addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib"
-        '';
-      };
-
-      shell-extra = mkShell {
-        name = "shell-extra-${finalAttrs.finalPackage.name}";
-        description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
-        buildInputs = [ llama-python-extra ];
-        inputsFrom = [ finalAttrs.finalPackage ];
-      };
-    };
-
-    meta = {
-      # Configurations we don't want even the CI to evaluate. Results in the
-      # "unsupported platform" messages. This is mostly a no-op, because
-      # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
-
-      # Configurations that are known to result in build failures. Can be
-      # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
-
-      description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
-      homepage = "https://github.com/ggerganov/llama.cpp/";
-      license = lib.licenses.mit;
-
-      # Accommodates `nix run` and `lib.getExe`
-      mainProgram = "llama";
-
-      # These people might respond, on the best effort basis, if you ping them
-      # in case of Nix-specific regressions or for reviewing Nix-specific PRs.
-      # Consider adding yourself to this list if you want to ensure this flake
-      # stays maintained and you're willing to invest your time. Do not add
-      # other people without their consent. Consider removing people after
-      # they've been unreachable for long periods of time.
-
-      # Note that lib.maintainers is defined in Nixpkgs, but you may just add
-      # an attrset following the same format as in
-      # https://github.com/NixOS/nixpkgs/blob/f36a80e54da29775c78d7eff0e628c2b4e34d1d7/maintainers/maintainer-list.nix
-      maintainers = with lib.maintainers; [
-        philiptaron
-        SomeoneSerge
-      ];
-
-      # Extend `badPlatforms` instead
-      platforms = lib.platforms.all;
-    };
-  }
-)

.devops/nix/scope.nix

@@ -1,16 +0,0 @@
-{
-  lib,
-  newScope,
-  llamaVersion ? "0.0.0",
-}:
-
-# We're using `makeScope` instead of just writing out an attrset
-# because it allows users to apply overlays later using `overrideScope'`.
-# Cf. https://noogle.dev/f/lib/makeScope
-
-lib.makeScope newScope (
-  self: {
-    inherit llamaVersion;
-    llama-cpp = self.callPackage ./package.nix { };
-  }
-)

.devops/server-cuda.Dockerfile

@@ -1,32 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-# This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
-# Target the CUDA build image
-ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
-# Target the CUDA runtime image
-ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
-
-FROM ${BASE_CUDA_DEV_CONTAINER} as build
-
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
-
-RUN apt-get update && \
-    apt-get install -y build-essential git
-
-WORKDIR /app
-
-COPY . .
-
-# Set nvcc architecture
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
-
-RUN make
-
-FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
-
-COPY --from=build /app/server /server
-
-ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -1,25 +0,0 @@
-ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
-ARG UBUNTU_VERSION=22.04
-
-FROM intel/hpckit:$ONEAPI_VERSION as build
-
-RUN apt-get update && \
-    apt-get install -y git
-
-WORKDIR /app
-
-COPY . .
-
-# for some reasons, "-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DLLAMA_NATIVE=ON" give worse performance
-RUN mkdir build && \
-    cd build && \
-    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx && \
-    cmake --build . --config Release --target main server
-
-FROM ubuntu:$UBUNTU_VERSION as runtime
-
-COPY --from=build /app/build/bin/server /server
-
-ENV LC_ALL=C.utf8
-
-ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -1,45 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-
-# This needs to generally match the container host's environment.
-ARG ROCM_VERSION=5.6
-
-# Target the CUDA build image
-ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
-
-FROM ${BASE_ROCM_DEV_CONTAINER} as build
-
-# Unless otherwise specified, we make a fat build.
-# List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
-# This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
-    gfx803 \
-    gfx900 \
-    gfx906 \
-    gfx908 \
-    gfx90a \
-    gfx1010 \
-    gfx1030 \
-    gfx1100 \
-    gfx1101 \
-    gfx1102
-
-COPY requirements.txt   requirements.txt
-COPY requirements       requirements
-
-RUN pip install --upgrade pip setuptools wheel \
-    && pip install -r requirements.txt
-
-WORKDIR /app
-
-COPY . .
-
-# Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
-# Enable ROCm
-ENV LLAMA_HIPBLAS=1
-ENV CC=/opt/rocm/llvm/bin/clang
-ENV CXX=/opt/rocm/llvm/bin/clang++
-
-RUN make
-
-ENTRYPOINT [ "/app/server" ]

.devops/server.Dockerfile

@@ -1,20 +0,0 @@
-ARG UBUNTU_VERSION=22.04
-
-FROM ubuntu:$UBUNTU_VERSION as build
-
-RUN apt-get update && \
-    apt-get install -y build-essential git
-
-WORKDIR /app
-
-COPY . .
-
-RUN make
-
-FROM ubuntu:$UBUNTU_VERSION as runtime
-
-COPY --from=build /app/server /server
-
-ENV LC_ALL=C.utf8
-
-ENTRYPOINT [ "/server" ]

.devops/tools.sh

@@ -13,8 +13,6 @@ elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     ./main "$@"
-elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
-    ./finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -36,8 +34,6 @@ else
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
-    echo "  --finetune (-f): Run finetune command to create a lora finetune of the model"
-    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"

.editorconfig

@@ -15,14 +15,8 @@ indent_size = 4
 [Makefile]
 indent_style = tab
 
-[scripts/*.mk]
-indent_style = tab
-
 [prompts/*.txt]
 insert_final_newline = unset
 
 [examples/server/public/*]
 indent_size = 2
-
-[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
-indent_style = tab

.github/ISSUE_TEMPLATE/bug.md

@@ -6,4 +6,179 @@ 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.
+# Prerequisites
+
+Please answer the following questions for yourself before submitting an issue.
+
+- [ ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
+- [ ] I carefully followed the [README.md](https://github.com/ggerganov/llama.cpp/blob/master/README.md).
+- [ ] I [searched using keywords relevant to my issue](https://docs.github.com/en/issues/tracking-your-work-with-issues/filtering-and-searching-issues-and-pull-requests) to make sure that I am creating a new issue that is not already open (or closed).
+- [ ] I reviewed the [Discussions](https://github.com/ggerganov/llama.cpp/discussions), and have a new bug or useful enhancement to share.
+
+# Expected Behavior
+
+Please provide a detailed written description of what you were trying to do, and what you expected `llama.cpp` to do.
+
+# Current Behavior
+
+Please provide a detailed written description of what `llama.cpp` did, instead.
+
+# Environment and Context
+
+Please provide detailed information about your computer setup. This is important in case the issue is not reproducible except for under certain specific conditions.
+
+* Physical (or virtual) hardware you are using, e.g. for Linux:
+
+`$ lscpu`
+
+* Operating System, e.g. for Linux:
+
+`$ uname -a`
+
+* SDK version, e.g. for Linux:
+
+```
+$ python3 --version
+$ make --version
+$ g++ --version
+```
+
+# Failure Information (for bugs)
+
+Please help provide information about the failure / bug.
+
+# Steps to Reproduce
+
+Please provide detailed steps for reproducing the issue. We are not sitting in front of your screen, so the more detail the better.
+
+1. step 1
+2. step 2
+3. step 3
+4. etc.
+
+# Failure Logs
+
+Please include any relevant log snippets or files. If it works under one configuration but not under another, please provide logs for both configurations and their corresponding outputs so it is easy to see where behavior changes.
+
+Also, please try to **avoid using screenshots** if at all possible. Instead, copy/paste the console output and use [Github's markdown](https://docs.github.com/en/get-started/writing-on-github/getting-started-with-writing-and-formatting-on-github/basic-writing-and-formatting-syntax) to cleanly format your logs for easy readability.
+
+Example environment info:
+```
+llama.cpp$ git log | head -1
+commit 2af23d30434a677c6416812eea52ccc0af65119c
+
+llama.cpp$ lscpu | egrep "AMD|Flags"
+Vendor ID:                       AuthenticAMD
+Model name:                      AMD Ryzen Threadripper 1950X 16-Core Processor
+Flags:                           fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid amd_dcm aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb hw_pstate ssbd ibpb vmmcall fsgsbase bmi1 avx2 smep bmi2 rdseed adx smap clflushopt sha_ni xsaveopt xsavec xgetbv1 xsaves clzero irperf xsaveerptr arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif overflow_recov succor smca sme sev
+Virtualization:                  AMD-V
+
+llama.cpp$ python3 --version
+Python 3.10.9
+
+llama.cpp$ pip list | egrep "torch|numpy|sentencepiece"
+numpy                         1.24.2
+numpydoc                      1.5.0
+sentencepiece                 0.1.97
+torch                         1.13.1
+torchvision                   0.14.1
+
+llama.cpp$ make --version | head -1
+GNU Make 4.3
+
+$ md5sum ./models/65B/ggml-model-q4_0.bin
+dbdd682cce80e2d6e93cefc7449df487  ./models/65B/ggml-model-q4_0.bin
+```
+
+Example run with the Linux command [perf](https://www.brendangregg.com/perf.html)
+```
+llama.cpp$ perf stat ./main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p "Please close your issue when it has been answered."
+main: seed = 1679149377
+llama_model_load: loading model from './models/65B/ggml-model-q4_0.bin' - please wait ...
+llama_model_load: n_vocab = 32000
+llama_model_load: n_ctx   = 512
+llama_model_load: n_embd  = 8192
+llama_model_load: n_mult  = 256
+llama_model_load: n_head  = 64
+llama_model_load: n_layer = 80
+llama_model_load: n_rot   = 128
+llama_model_load: f16     = 2
+llama_model_load: n_ff    = 22016
+llama_model_load: n_parts = 8
+llama_model_load: ggml ctx size = 41477.73 MB
+llama_model_load: memory_size =  2560.00 MB, n_mem = 40960
+llama_model_load: loading model part 1/8 from './models/65B/ggml-model-q4_0.bin'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 2/8 from './models/65B/ggml-model-q4_0.bin.1'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 3/8 from './models/65B/ggml-model-q4_0.bin.2'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 4/8 from './models/65B/ggml-model-q4_0.bin.3'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 5/8 from './models/65B/ggml-model-q4_0.bin.4'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 6/8 from './models/65B/ggml-model-q4_0.bin.5'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 7/8 from './models/65B/ggml-model-q4_0.bin.6'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+llama_model_load: loading model part 8/8 from './models/65B/ggml-model-q4_0.bin.7'
+llama_model_load: .......................................................................................... done
+llama_model_load: model size =  4869.09 MB / num tensors = 723
+
+system_info: n_threads = 16 / 32 | AVX = 1 | AVX2 = 1 | AVX512 = 0 | FMA = 1 | NEON = 0 | ARM_FMA = 0 | F16C = 1 | FP16_VA = 0 | WASM_SIMD = 0 | BLAS = 0 | SSE3 = 1 | VSX = 0 |
+
+main: prompt: 'Please close your issue when it has been answered.'
+main: number of tokens in prompt = 11
+     1 -> ''
+ 12148 -> 'Please'
+  3802 -> ' close'
+   596 -> ' your'
+  2228 -> ' issue'
+   746 -> ' when'
+   372 -> ' it'
+   756 -> ' has'
+  1063 -> ' been'
+  7699 -> ' answered'
+ 29889 -> '.'
+
+sampling parameters: temp = 0.800000, top_k = 40, top_p = 0.950000, repeat_last_n = 64, repeat_penalty = 1.300000
+
+
+Please close your issue when it has been answered.
+@duncan-donut: I'm trying to figure out what kind of "support" you need for this script and why, exactly? Is there a question about how the code works that hasn't already been addressed in one or more comments below this ticket, or are we talking something else entirely like some sorta bugfixing job because your server setup is different from mine??
+I can understand if your site needs to be running smoothly and you need help with a fix of sorts but there should really be nothing wrong here that the code itself could not handle. And given that I'm getting reports about how it works perfectly well on some other servers, what exactly are we talking? A detailed report will do wonders in helping us get this resolved for ya quickly so please take your time and describe the issue(s) you see as clearly & concisely as possible!!
+@duncan-donut: I'm not sure if you have access to cPanel but you could try these instructions. It is worth a shot! Let me know how it goes (or what error message, exactly!) when/if ya give that code a go? [end of text]
+
+
+main: mem per token = 71159620 bytes
+main:     load time = 19309.95 ms
+main:   sample time =   168.62 ms
+main:  predict time = 223895.61 ms / 888.47 ms per token
+main:    total time = 246406.42 ms
+
+ Performance counter stats for './main -m ./models/65B/ggml-model-q4_0.bin -t 16 -n 1024 -p Please close your issue when it has been answered.':
+
+        3636882.89 msec task-clock                #   14.677 CPUs utilized
+             13509      context-switches          #    3.714 /sec
+              2436      cpu-migrations            #    0.670 /sec
+          10476679      page-faults               #    2.881 K/sec
+    13133115082869      cycles                    #    3.611 GHz                      (16.77%)
+       29314462753      stalled-cycles-frontend   #    0.22% frontend cycles idle     (16.76%)
+    10294402631459      stalled-cycles-backend    #   78.39% backend cycles idle      (16.74%)
+    23479217109614      instructions              #    1.79  insn per cycle
+                                                  #    0.44  stalled cycles per insn  (16.76%)
+     2353072268027      branches                  #  647.002 M/sec                    (16.77%)
+        1998682780      branch-misses             #    0.08% of all branches          (16.76%)
+
+     247.802177522 seconds time elapsed
+
+    3618.573072000 seconds user
+      18.491698000 seconds sys
+```

.github/workflows/build.yml

@@ -72,7 +72,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -107,7 +107,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -141,52 +141,8 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose
+          ctest --verbose
 
-  ubuntu-22-cmake-sycl:
-    runs-on: ubuntu-22.04
-
-    continue-on-error: true
-
-    steps:
-      - uses: actions/checkout@v2
-
-      - name: add oneAPI to apt
-        shell: bash
-        run: |
-          cd /tmp
-          wget https://apt.repos.intel.com/intel-gpg-keys/GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo apt-key add GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          rm GPG-PUB-KEY-INTEL-SW-PRODUCTS.PUB
-          sudo add-apt-repository "deb https://apt.repos.intel.com/oneapi all main"
-
-      - name: install oneAPI dpcpp compiler
-        shell: bash
-        run: |
-          sudo apt update
-          sudo apt install intel-oneapi-compiler-dpcpp-cpp
-
-      - name: install oneAPI MKL library
-        shell: bash
-        run: |
-          sudo apt install intel-oneapi-mkl-devel
-
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Build
-        id: cmake_build
-        run: |
-          source /opt/intel/oneapi/setvars.sh
-          mkdir build
-          cd build
-          cmake -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ..
-          cmake --build . --config Release -j $(nproc)
-
-  # TODO: build with LLAMA_NO_METAL because test-backend-ops fail on "Apple Paravirtual device" and I don't know
-  #       how to debug it.
-  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7131777249/job/19420981052#step:5:1124
   macOS-latest-make:
     runs-on: macos-latest
 
@@ -204,18 +160,14 @@ jobs:
       - name: Build
         id: make_build
         run: |
-          LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
+          make -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: make_test
         run: |
-          LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
-          LLAMA_NO_METAL=1 make test  -j $(sysctl -n hw.logicalcpu)
+          make tests -j $(sysctl -n hw.logicalcpu)
+          make test -j $(sysctl -n hw.logicalcpu)
 
-  # TODO: build with LLAMA_METAL=OFF because test-backend-ops fail on "Apple Paravirtual device" and I don't know
-  #       how to debug it.
-  #       ref: https://github.com/ggerganov/llama.cpp/actions/runs/7132125951/job/19422043567?pr=4359#step:5:6584
-  #       would be great if we fix these
   macOS-latest-cmake:
     runs-on: macos-latest
 
@@ -236,14 +188,14 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_METAL=OFF ..
+          cmake ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest --verbose --timeout 900
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -336,7 +288,6 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
-      SDE_VERSION: 9.33.0-2024-01-07
 
     strategy:
       matrix:
@@ -432,22 +383,10 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # not all machines have native AVX-512
+        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
         run: |
           cd build
-          ctest -L main -C Release --verbose --timeout 900
-
-      - name: Test (Intel SDE)
-        id: cmake_test_sde
-        if: ${{ matrix.build == 'avx512' && env.HAS_AVX512F == '0' }} # use Intel SDE for AVX-512 emulation
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/sde.tar.xz -L "https://downloadmirror.intel.com/813591/sde-external-${env:SDE_VERSION}-win.tar.xz"
-          # for some weird reason windows tar doesn't like sde tar.xz
-          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar.xz
-          7z x "-o${env:RUNNER_TEMP}" $env:RUNNER_TEMP/sde.tar
-          $sde = $(join-path $env:RUNNER_TEMP sde-external-${env:SDE_VERSION}-win/sde.exe)
-          cd build
-          & $sde -future -- ctest -L main -C Release --verbose --timeout 900
+          ctest -C Release --verbose --timeout 900
 
       - name: Determine tag name
         id: tag
@@ -546,41 +485,6 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
-  ios-xcode-build:
-    runs-on: macos-latest
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v3
-
-      - name: Build Xcode project
-        run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
-
-  android-build:
-    runs-on: ubuntu-latest
-
-    steps:
-      - name: Clone
-        uses: actions/checkout@v3
-
-      - name: Set up JDK
-        uses: actions/setup-java@v3
-        with:
-          java-version: 17
-          distribution: zulu
-
-      - name: Setup Android SDK
-        uses: android-actions/setup-android@v3
-        with:
-          log-accepted-android-sdk-licenses: false
-
-      - name: Build
-        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
-
 #  freeBSD-latest:
 #    runs-on: macos-12
 #    steps:

.github/workflows/docker.yml

@@ -28,18 +28,13 @@ jobs:
         config:
           - { tag: "light", dockerfile: ".devops/main.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full", dockerfile: ".devops/full.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server", dockerfile: ".devops/server.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           # NOTE(canardletter): The CUDA builds on arm64 are very slow, so I
           #                     have disabled them for now until the reason why
           #                     is understood.
           - { tag: "light-cuda", dockerfile: ".devops/main-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-cuda", dockerfile: ".devops/server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "light-rocm", dockerfile: ".devops/main-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "server-rocm", dockerfile: ".devops/server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          - { tag: "light-intel", dockerfile: ".devops/main-intel.Dockerfile", platforms: "linux/amd64" }
-          - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v3
@@ -57,36 +52,6 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
-      - name: Free Disk Space (Ubuntu)
-        uses: jlumbroso/free-disk-space@main
-        with:
-          # this might remove tools that are actually needed,
-          # if set to "true" but frees about 6 GB
-          tool-cache: false
-
-          # all of these default to true, but feel free to set to
-          # "false" if necessary for your workflow
-          android: true
-          dotnet: true
-          haskell: true
-          large-packages: true
-          docker-images: true
-          swap-storage: true
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -94,7 +59,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -103,5 +68,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
           file: ${{ matrix.config.dockerfile }}

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

@@ -1,62 +0,0 @@
-name: Nix aarch64 builds
-
-on:
-  workflow_dispatch: # allows manual triggering
-  schedule:
-    # Rebuild daily rather than on every push because QEMU is expensive (e.g.
-    # 1.5h instead of minutes with the cold cache).
-    #
-    # randint(0, 59), randint(0, 23)
-    - cron: '26 12 * * *'
-  # But also rebuild if we touched any of the Nix expressions:
-  push:
-    branches:
-      - master
-    paths: ['**/*.nix', 'flake.lock']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['**/*.nix', 'flake.lock']
-
-jobs:
-  nix-build-aarch64:
-    if: ${{ vars.CACHIX_NAME != '' }}
-    runs-on: ubuntu-latest
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install QEMU
-      # Copy-paste from https://github.com/orgs/community/discussions/8305#discussioncomment-5888654
-      run: |
-        sudo apt-get update
-        sudo apt-get install -y qemu-user-static qemu-system-aarch64
-        sudo usermod -a -G kvm $USER
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      with:
-        github-token: ${{ secrets.GITHUB_TOKEN }}
-        extra-conf: |
-          extra-platforms = aarch64-linux
-          extra-system-features = nixos-test kvm
-          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.aarch64-linux"
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --systems aarch64-linux
-          --flake
-          ".#checks.aarch64-linux"

.github/workflows/nix-ci.yml

@@ -1,69 +0,0 @@
-name: Nix CI
-
-on:
-  workflow_dispatch: # allows manual triggering
-  push:
-    branches:
-      - master
-  pull_request:
-    types: [opened, synchronize, reopened]
-
-jobs:
-  nix-eval:
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ ubuntu-latest, macos-latest ]
-    runs-on: ${{ matrix.os }}
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      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 = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: List all flake outputs
-      run: nix flake show --all-systems
-    - name: Show all output paths
-      run: >
-          nix run github:nix-community/nix-eval-jobs
-          -- --gc-roots-dir gcroot
-          --flake
-          ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
-  nix-build:
-    if: ${{ vars.CACHIX_NAME != '' }}
-    strategy:
-      fail-fast: false
-      matrix:
-        os: [ ubuntu-latest, macos-latest ]
-    runs-on: ${{ matrix.os }}
-    steps:
-    - name: Checkout repository
-      uses: actions/checkout@v4
-    - name: Install Nix
-      uses: DeterminateSystems/nix-installer-action@v9
-      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 = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
-    - uses: DeterminateSystems/magic-nix-cache-action@v2
-      with:
-        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
-    - name: Set-up cachix to push the results to
-      uses: cachix/cachix-action@v13
-      with:
-        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: ${{ vars.CACHIX_NAME }}
-    - name: Build
-      run: >
-          nix run github:Mic92/nix-fast-build
-          -- --skip-cached --no-nom
-          --flake
-          ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"

.github/workflows/nix-flake-update.yml

@@ -1,22 +0,0 @@
-name: update-flake-lock
-on:
-  workflow_dispatch:
-  schedule:
-    - cron: '0 0 * * 0' # runs weekly on Sunday at 00:00
-
-jobs:
-  lockfile:
-    runs-on: ubuntu-latest
-    steps:
-      - name: Checkout repository
-        uses: actions/checkout@v4
-      - name: Install Nix
-        uses: DeterminateSystems/nix-installer-action@main
-      - name: Update flake.lock
-        uses: DeterminateSystems/update-flake-lock@main
-        with:
-          pr-title: "nix: update flake.lock"
-          pr-labels: |
-            nix
-          pr-reviewers: philiptaron,SomeoneSerge
-          token: ${{ secrets.FLAKE_TOKEN }}

.github/workflows/nix-publish-flake.yml

@@ -1,36 +0,0 @@
-# Make the flake discoverable on https://flakestry.dev and https://flakehub.com/flakes
-name: "Publish a flake to flakestry & flakehub"
-on:
-    push:
-        tags:
-        - "*"
-    workflow_dispatch:
-        inputs:
-            tag:
-                description: "The existing tag to publish"
-                type: "string"
-                required: true
-jobs:
-    flakestry-publish:
-        runs-on: ubuntu-latest
-        permissions:
-            id-token: "write"
-            contents: "read"
-        steps:
-            - uses: flakestry/flakestry-publish@main
-              with:
-                version: "${{ inputs.tag || github.ref_name }}"
-    flakehub-publish:
-      runs-on: "ubuntu-latest"
-      permissions:
-        id-token: "write"
-        contents: "read"
-      steps:
-        - uses: "actions/checkout@v4"
-          with:
-            ref: "${{ (inputs.tag != null) && format('refs/tags/{0}', inputs.tag) || '' }}"
-        - uses: "DeterminateSystems/nix-installer-action@main"
-        - uses: "DeterminateSystems/flakehub-push@main"
-          with:
-            visibility: "public"
-            tag: "${{ inputs.tag }}"

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

@@ -1,29 +0,0 @@
-name: Python check requirements.txt
-
-on:
-  push:
-    paths:
-      - 'scripts/check-requirements.sh'
-      - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
-  pull_request:
-    paths:
-      - 'scripts/check-requirements.sh'
-      - 'convert*.py'
-      - 'requirements.txt'
-      - 'requirements/*.txt'
-
-jobs:
-  python-check-requirements:
-    runs-on: ubuntu-latest
-    name: check-requirements
-    steps:
-      - name: Check out source repository
-        uses: actions/checkout@v3
-      - name: Set up Python environment
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.11"
-      - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh nocleanup

.github/workflows/python-lint.yml

@@ -1,20 +0,0 @@
-name: flake8 Lint
-
-on: [push, pull_request]
-
-jobs:
-  flake8-lint:
-    runs-on: ubuntu-latest
-    name: Lint
-    steps:
-      - name: Check out source repository
-        uses: actions/checkout@v3
-      - name: Set up Python environment
-        uses: actions/setup-python@v4
-        with:
-          python-version: "3.11"
-      - name: flake8 Lint
-        uses: py-actions/flake8@v2
-        with:
-            ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704"
-            exclude: "examples/*,examples/*/**,*/**/__init__.py"

.gitignore

@@ -15,7 +15,6 @@
 .DS_Store
 .build/
 .cache/
-.ccls-cache/
 .direnv/
 .envrc
 .swiftpm
@@ -27,7 +26,7 @@
 lcov-report/
 gcovr-report/
 
-build*
+build*/
 out/
 tmp/
 
@@ -43,16 +42,12 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
-/imatrix
 /infill
 /libllama.so
 /llama-bench
-/llava-cli
-/lookahead
-/lookup
+/llava
 /main
 /metal
-/passkey
 /perplexity
 /q8dot
 /quantize
@@ -68,9 +63,8 @@ models-mnt
 /speculative
 /parallel
 /train-text-from-scratch
-/tokenize
 /vdot
-/common/build-info.cpp
+build-info.h
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
@@ -89,3 +83,17 @@ examples/jeopardy/results.txt
 
 poetry.lock
 poetry.toml
+
+# Test binaries
+tests/test-grammar-parser
+tests/test-llama-grammar
+tests/test-double-float
+tests/test-grad0
+tests/test-opt
+tests/test-quantize-fns
+tests/test-quantize-perf
+tests/test-sampling
+tests/test-tokenizer-0-llama
+tests/test-tokenizer-0-falcon
+tests/test-tokenizer-1-llama
+tests/test-tokenizer-1-bpe

CMakeLists.txt

@@ -1,6 +1,5 @@
-cmake_minimum_required(VERSION 3.14)  # for add_link_options and implicit target directories.
+cmake_minimum_required(VERSION 3.13)  # for add_link_options
 project("llama.cpp" C CXX)
-include(CheckIncludeFileCXX)
 
 set(CMAKE_EXPORT_COMPILE_COMMANDS ON)
 
@@ -11,7 +10,7 @@ endif()
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
-if (CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
+if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(LLAMA_STANDALONE ON)
 
     # configure project version
@@ -44,11 +43,9 @@ else()
 endif()
 
 # general
-option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
 option(LLAMA_NATIVE                     "llama: enable -march=native flag"                      ON)
 option(LLAMA_LTO                        "llama: enable link time optimization"                  OFF)
-option(LLAMA_CCACHE                     "llama: use ccache if available"                        ON)
 
 # debug
 option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"                   ON)
@@ -78,10 +75,6 @@ if (NOT MSVC)
     option(LLAMA_F16C                        "llama: enable F16C"                               ${INS_ENB})
 endif()
 
-if (WIN32)
-    option(LLAMA_WIN_VER                     "llama: Windows Version"                           0x602)
-endif()
-
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
@@ -97,40 +90,55 @@ set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
-option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
-option(LLAMA_VULKAN                          "llama: use Vulkan"                                OFF)
 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_MPI                             "llama: use MPI"                                   OFF)
+option(LLAMA_K_QUANTS                        "llama: use k-quants"                              ON)
 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_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
-option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
+option(LLAMA_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)
 
+#
+# Build info header
+#
 
-# add perf arguments
-option(LLAMA_PERF                            "llama: enable perf"                               OFF)
-if (LLAMA_PERF)
-    add_definitions(-DGGML_PERF)
-endif()
-
-# Required for relocatable CMake package
+# Generate initial build-info.h
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git")
+
+    # Is git submodule
+    if(NOT IS_DIRECTORY "${GIT_DIR}")
+        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
+        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
+        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
+    endif()
+
+    # Add a custom target for build-info.h
+    add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
+
+    # Add a custom command to rebuild build-info.h when .git/index changes
+    add_custom_command(
+        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
+        COMMENT "Generating build details from Git"
+        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
+        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
+        DEPENDS "${GIT_DIR}/index"
+        VERBATIM
+    )
+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.")
+endif()
+
 #
 # Compile flags
 #
-if (LLAMA_SYCL)
-    set(CMAKE_CXX_STANDARD 17)
-else()
-    set(CMAKE_CXX_STANDARD 11)
-endif()
 
+set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
@@ -138,11 +146,6 @@ set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
-# enable libstdc++ assertions for debug builds
-if (CMAKE_SYSTEM_NAME MATCHES "Linux")
-    add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
-endif()
-
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
@@ -175,9 +178,9 @@ if (APPLE AND LLAMA_ACCELERATE)
 endif()
 
 if (LLAMA_METAL)
-    find_library(FOUNDATION_LIBRARY Foundation REQUIRED)
-    find_library(METAL_FRAMEWORK    Metal      REQUIRED)
-    find_library(METALKIT_FRAMEWORK MetalKit   REQUIRED)
+    find_library(FOUNDATION_LIBRARY         Foundation              REQUIRED)
+    find_library(METAL_FRAMEWORK            Metal                   REQUIRED)
+    find_library(METALKIT_FRAMEWORK         MetalKit                REQUIRED)
 
     message(STATUS "Metal framework found")
     set(GGML_HEADERS_METAL ggml-metal.h)
@@ -192,36 +195,7 @@ if (LLAMA_METAL)
     #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
 
     # copy ggml-metal.metal to bin directory
-    configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
-
-    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
-        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-        #
-        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-        #       disabling fast math is needed in order to pass tests/test-backend-ops
-        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-        set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        if (LLAMA_QKK_64)
-            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
-        endif()
-
-        add_custom_command(
-            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            DEPENDS ggml-metal.metal
-            COMMENT "Compiling Metal kernels"
-        )
-
-        add_custom_target(
-            ggml-metal ALL
-            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        )
-    endif()
+    configure_file(ggml-metal.metal bin/ggml-metal.metal COPYONLY)
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
@@ -250,11 +224,7 @@ if (LLAMA_BLAS)
             if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
                 pkg_check_modules(DepBLAS REQUIRED blas)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
-                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
-                pkg_check_modules(DepBLAS openblas64)
-                if (NOT DepBLAS_FOUND)
-                    pkg_check_modules(DepBLAS REQUIRED openblas)
-                endif()
+                pkg_check_modules(DepBLAS REQUIRED openblas)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
                 pkg_check_modules(DepBLAS REQUIRED blis)
             elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
@@ -308,8 +278,13 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
-if (LLAMA_QKK_64)
-    add_compile_definitions(GGML_QKK_64)
+if (LLAMA_K_QUANTS)
+    set(GGML_HEADERS_EXTRA k_quants.h)
+    set(GGML_SOURCES_EXTRA k_quants.c)
+    add_compile_definitions(GGML_USE_K_QUANTS)
+    if (LLAMA_QKK_64)
+        add_compile_definitions(GGML_QKK_64)
+    endif()
 endif()
 
 if (LLAMA_CUBLAS)
@@ -346,18 +321,11 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
-            if (WIN32)
-                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
-            else ()
-                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
-            endif()
+            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
         else()
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cuda_driver)
-
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
         # 52 == lowest CUDA 12 standard
         # 60 == f16 CUDA intrinsics
@@ -417,28 +385,6 @@ if (LLAMA_CLBLAST)
     endif()
 endif()
 
-if (LLAMA_VULKAN)
-    find_package(Vulkan)
-    if (Vulkan_FOUND)
-        message(STATUS "Vulkan found")
-
-        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
-        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
-
-        add_library(ggml-vulkan STATIC 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)
-
-        add_compile_definitions(GGML_USE_VULKAN)
-
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
-    else()
-        message(WARNING "Vulkan not found")
-    endif()
-endif()
-
 if (LLAMA_HIPBLAS)
     list(APPEND CMAKE_PREFIX_PATH /opt/rocm)
 
@@ -456,9 +402,6 @@ if (LLAMA_HIPBLAS)
     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)
@@ -484,132 +427,57 @@ if (LLAMA_HIPBLAS)
     endif()
 endif()
 
-
-if (LLAMA_SYCL)
-    if ( NOT DEFINED ENV{ONEAPI_ROOT})
-        message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
-    endif()
-    #todo: AOT
-
-    find_package(IntelSYCL REQUIRED)
-    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/${SYCL_INCLUDE_DIR})
-
-    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
-    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_SOURCES_SYCL ggml-sycl.cpp)
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} sycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-endif()
-
-function(get_flags CCID CCVER)
-    set(C_FLAGS "")
-    set(CXX_FLAGS "")
-
-    if (CCID MATCHES "Clang")
-        set(C_FLAGS   -Wunreachable-code-break -Wunreachable-code-return)
-        set(CXX_FLAGS -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi)
-
-        if (
-            (CCID STREQUAL "Clang"      AND CCVER VERSION_GREATER_EQUAL 3.8.0) OR
-            (CCID STREQUAL "AppleClang" AND CCVER VERSION_GREATER_EQUAL 7.3.0)
-        )
-            list(APPEND C_FLAGS -Wdouble-promotion)
-        endif()
-    elseif (CCID STREQUAL "GNU")
-        set(C_FLAGS   -Wdouble-promotion)
-        set(CXX_FLAGS -Wno-array-bounds)
-
-        if (CCVER VERSION_GREATER_EQUAL 7.1.0)
-            list(APPEND CXX_FLAGS -Wno-format-truncation)
-        endif()
-        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_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)
+        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)
+        set(host_cxx_flags "")
 
-        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
-        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
+        if (CMAKE_C_COMPILER_ID MATCHES "Clang")
+            set(warning_flags ${warning_flags} -Wunreachable-code-break -Wunreachable-code-return)
+            set(host_cxx_flags ${host_cxx_flags} -Wmissing-prototypes -Wextra-semi)
 
-        get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
+            if (
+                (CMAKE_C_COMPILER_ID STREQUAL "Clang"      AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 3.8.0) OR
+                (CMAKE_C_COMPILER_ID STREQUAL "AppleClang" AND CMAKE_C_COMPILER_VERSION VERSION_GREATER_EQUAL 7.3.0)
+            )
+                set(c_flags ${c_flags} -Wdouble-promotion)
+            endif()
+        elseif (CMAKE_C_COMPILER_ID STREQUAL "GNU")
+            set(c_flags ${c_flags} -Wdouble-promotion)
+            set(host_cxx_flags ${host_cxx_flags} -Wno-array-bounds)
 
-        add_compile_options("$<$<COMPILE_LANGUAGE:C>:${C_FLAGS};${GF_C_FLAGS}>"
-                            "$<$<COMPILE_LANGUAGE:CXX>:${CXX_FLAGS};${GF_CXX_FLAGS}>")
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 7.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wno-format-truncation)
+            endif()
+            if (CMAKE_CXX_COMPILER_VERSION VERSION_GREATER_EQUAL 8.1.0)
+                set(host_cxx_flags ${host_cxx_flags} -Wextra-semi)
+            endif()
+        endif()
     else()
         # todo : msvc
-        set(C_FLAGS   "")
-        set(CXX_FLAGS "")
     endif()
+
+    set(c_flags   ${c_flags}   ${warning_flags})
+    set(cxx_flags ${cxx_flags} ${warning_flags})
+    add_compile_options("$<$<COMPILE_LANGUAGE:C>:${c_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${cxx_flags}>"
+                        "$<$<COMPILE_LANGUAGE:CXX>:${host_cxx_flags}>")
+
 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)
-    endif()
-
-    if (LLAMA_ALL_WARNINGS AND NOT MSVC)
-        set(NVCC_CMD ${CMAKE_CUDA_COMPILER} .c)
-        if (NOT CMAKE_CUDA_HOST_COMPILER STREQUAL "")
-            list(APPEND NVCC_CMD -ccbin ${CMAKE_CUDA_HOST_COMPILER})
-        endif()
-
-        execute_process(
-            COMMAND ${NVCC_CMD} -Xcompiler --version
-            OUTPUT_VARIABLE CUDA_CCFULLVER
-            ERROR_QUIET
-        )
-
-        if (NOT CUDA_CCFULLVER MATCHES clang)
-            set(CUDA_CCID "GNU")
-            execute_process(
-                COMMAND ${NVCC_CMD} -Xcompiler "-dumpfullversion -dumpversion"
-                OUTPUT_VARIABLE CUDA_CCVER
-                ERROR_QUIET
-            )
-        else()
-            if (CUDA_CCFULLVER MATCHES Apple)
-                set(CUDA_CCID "AppleClang")
-            else()
-                set(CUDA_CCID "Clang")
-            endif()
-            string(REGEX REPLACE "^.* version ([0-9.]*).*$" "\\1" CUDA_CCVER ${CUDA_CCFULLVER})
-        endif()
-
-        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
-    endif()
+if (NOT MSVC)
+    set(cuda_flags -Wno-pedantic)
 endif()
+set(cuda_flags ${cxx_flags} -use_fast_math ${cuda_flags})
+
+list(JOIN host_cxx_flags " " cuda_host_flags)  # pass host compiler flags as a single argument
+if (NOT cuda_host_flags STREQUAL "")
+    set(cuda_flags ${cuda_flags} -Xcompiler ${cuda_host_flags})
+endif()
+
+add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${cuda_flags}>")
 
 if (WIN32)
     add_compile_definitions(_CRT_SECURE_NO_WARNINGS)
@@ -629,27 +497,6 @@ if (LLAMA_LTO)
     endif()
 endif()
 
-if (LLAMA_CCACHE)
-    find_program(LLAMA_CCACHE_FOUND ccache)
-    if (LLAMA_CCACHE_FOUND)
-        set_property(GLOBAL PROPERTY RULE_LAUNCH_COMPILE ccache)
-        set(ENV{CCACHE_SLOPPINESS} time_macros)
-        message(STATUS "Using ccache")
-    else()
-        message(STATUS "Warning: ccache not found - consider installing it or use LLAMA_CCACHE=OFF")
-    endif ()
-endif()
-
-# this version of Apple ld64 is buggy
-execute_process(
-    COMMAND ${CMAKE_C_COMPILER} ${CMAKE_EXE_LINKER_FLAGS} -Wl,-v
-    ERROR_VARIABLE output
-    OUTPUT_QUIET
-)
-if (output MATCHES "dyld-1015\.7")
-    add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
-endif()
-
 # Architecture specific
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
@@ -673,8 +520,6 @@ if (NOT MSVC)
     endif()
 endif()
 
-set(ARCH_FLAGS "")
-
 if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64") OR ("${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "arm64"))
     message(STATUS "ARM detected")
     if (MSVC)
@@ -686,30 +531,27 @@ if ((${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm") OR (${CMAKE_SYSTEM_PROCESSOR} MATC
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
-            list(APPEND ARCH_FLAGS -mfp16-format=ieee)
+            add_compile_options(-mfp16-format=ieee)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
             # Raspberry Pi 1, Zero
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
             # Raspberry Pi 2
-            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
+            add_compile_options(-mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
-            list(APPEND ARCH_FLAGS -mno-unaligned-access)
+            add_compile_options(-mno-unaligned-access)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GENERATOR_PLATFORM_LWR}" MATCHES "^(x86_64|i686|amd64|x64)$" )
     message(STATUS "x86 detected")
     if (MSVC)
-        # instruction set detection for MSVC only
-        if (LLAMA_NATIVE)
-            include(cmake/FindSIMD.cmake)
-        endif ()
         if (LLAMA_AVX512)
-            list(APPEND ARCH_FLAGS /arch:AVX512)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX512>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX512>)
             # MSVC has no compile-time flags enabling specific
             # AVX512 extensions, neither it defines the
             # macros corresponding to the extensions.
@@ -723,66 +565,47 @@ elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$" OR "${CMAKE_GE
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
             endif()
         elseif (LLAMA_AVX2)
-            list(APPEND ARCH_FLAGS /arch:AVX2)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX2>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX2>)
         elseif (LLAMA_AVX)
-            list(APPEND ARCH_FLAGS /arch:AVX)
+            add_compile_options($<$<COMPILE_LANGUAGE:C>:/arch:AVX>)
+            add_compile_options($<$<COMPILE_LANGUAGE:CXX>:/arch:AVX>)
         endif()
     else()
         if (LLAMA_NATIVE)
-            list(APPEND ARCH_FLAGS -march=native)
+            add_compile_options(-march=native)
         endif()
         if (LLAMA_F16C)
-            list(APPEND ARCH_FLAGS -mf16c)
+            add_compile_options(-mf16c)
         endif()
         if (LLAMA_FMA)
-            list(APPEND ARCH_FLAGS -mfma)
+            add_compile_options(-mfma)
         endif()
         if (LLAMA_AVX)
-            list(APPEND ARCH_FLAGS -mavx)
+            add_compile_options(-mavx)
         endif()
         if (LLAMA_AVX2)
-            list(APPEND ARCH_FLAGS -mavx2)
+            add_compile_options(-mavx2)
         endif()
         if (LLAMA_AVX512)
-            list(APPEND ARCH_FLAGS -mavx512f)
-            list(APPEND ARCH_FLAGS -mavx512bw)
+            add_compile_options(-mavx512f)
+            add_compile_options(-mavx512bw)
         endif()
         if (LLAMA_AVX512_VBMI)
-            list(APPEND ARCH_FLAGS -mavx512vbmi)
+            add_compile_options(-mavx512vbmi)
         endif()
         if (LLAMA_AVX512_VNNI)
-            list(APPEND ARCH_FLAGS -mavx512vnni)
+            add_compile_options(-mavx512vnni)
         endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
-    if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
-        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)
-    else()
-        list(APPEND ARCH_FLAGS -mcpu=native -mtune=native)
-        #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
-    endif()
+    add_compile_options(-mcpu=native -mtune=native)
+    #TODO: Add  targets for Power8/Power9 (Altivec/VSX) and Power10(MMA) and query for big endian systems (ppc64/le/be)
 else()
     message(STATUS "Unknown architecture")
 endif()
 
-add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
-add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
-
-if (LLAMA_CUBLAS)
-    list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
-    list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
-    if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
-        list(APPEND CUDA_FLAGS -Xcompiler ${CUDA_CXX_FLAGS_JOINED})
-    endif()
-    add_compile_options("$<$<COMPILE_LANGUAGE:CUDA>:${CUDA_FLAGS}>")
-endif()
-
-if (MINGW)
-    # Target Windows 8 for PrefetchVirtualMemory
-    add_compile_definitions(_WIN32_WINNT=${LLAMA_WIN_VER})
-endif()
-
 #
 # POSIX conformance
 #
@@ -850,15 +673,11 @@ add_library(ggml OBJECT
             ggml-alloc.h
             ggml-backend.c
             ggml-backend.h
-            ggml-quants.c
-            ggml-quants.h
-            ${GGML_SOURCES_CUDA}   ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_CUDA} ${GGML_HEADERS_CUDA}
             ${GGML_SOURCES_OPENCL} ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_VULKAN} ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_METAL}  ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}    ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA}  ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}   ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_METAL} ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI} ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA} ${GGML_HEADERS_EXTRA}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
@@ -934,8 +753,8 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
               ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfigVersion.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}" "${GGML_HEADERS_VULKAN}"
+set(GGML_PUBLIC_HEADERS "ggml.h"
+        "${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

@@ -1,15 +1,14 @@
 # Define the default target now so that it is always the first target
 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
+	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	simple batched batched-bench save-load-state server gguf llama-bench llava baby-llama beam-search  \
+	speculative infill benchmark-matmult parallel finetune export-lora tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
-	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
+	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -26,6 +25,20 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
+ifeq '' '$(findstring clang,$(shell $(CC) --version))'
+	CC_IS_GCC=1
+	CC_VER := $(shell $(CC) -dumpfullversion -dumpversion | awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+else
+	CC_IS_CLANG=1
+	ifeq '' '$(findstring Apple LLVM,$(shell $(CC) --version))'
+		CC_IS_LLVM_CLANG=1
+	else
+		CC_IS_APPLE_CLANG=1
+	endif
+	CC_VER := $(shell $(CC) --version | sed -n 's/^.* version \([0-9.]*\).*$$/\1/p' \
+				| awk -F. '{ printf("%02d%02d%02d", $$1, $$2, $$3) }')
+endif
+
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@@ -43,6 +56,10 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
+ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
+BUILD_TARGETS += metal
+endif
+
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -61,7 +78,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test %s FAILED!\n\n' $$test_target; \
+			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \
@@ -103,12 +120,12 @@ MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-MK_CFLAGS     += -Ofast
-HOST_CXXFLAGS += -Ofast
-MK_NVCCFLAGS  += -O3
+MK_CFLAGS        += -Ofast
+MK_HOST_CXXFLAGS += -Ofast
+MK_CUDA_CXXFLAGS += -O3
 else
-MK_CFLAGS     += -O3
-MK_CXXFLAGS   += -O3
+MK_CFLAGS        += -O3
+MK_CXXFLAGS      += -O3
 endif
 
 # clock_gettime came in POSIX.1b (1993)
@@ -157,10 +174,6 @@ ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
-
-	ifeq ($(UNAME_S),Linux)
-		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
-	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
 endif
@@ -202,9 +215,28 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
-# this version of Apple ld64 is buggy
-ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
-	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
+ifeq ($(CC_IS_CLANG), 1)
+	# clang options
+	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
+	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+
+	ifneq '' '$(and $(CC_IS_LLVM_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 030800)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+	ifneq '' '$(and $(CC_IS_APPLE_CLANG),$(filter 1,$(shell expr $(CC_VER) \>= 070300)))'
+		MK_CFLAGS += -Wdouble-promotion
+	endif
+else
+	# gcc options
+	MK_CFLAGS        += -Wdouble-promotion
+	MK_HOST_CXXFLAGS += -Wno-array-bounds
+
+	ifeq ($(shell expr $(CC_VER) \>= 070100), 1)
+		MK_HOST_CXXFLAGS += -Wno-format-truncation
+	endif
+	ifeq ($(shell expr $(CC_VER) \>= 080100), 1)
+		MK_HOST_CXXFLAGS += -Wextra-semi
+	endif
 endif
 
 # OS specific
@@ -252,8 +284,8 @@ ifndef RISCV
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
-	MK_CFLAGS     += -march=native -mtune=native
-	HOST_CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS   += -march=native -mtune=native
+	MK_HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -264,31 +296,19 @@ ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	#MK_CXXFLAGS += -mssse3
 endif
 
+# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
+# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
+# https://github.com/ggerganov/llama.cpp/issues/2922
 ifneq '' '$(findstring mingw,$(shell $(CC) -dumpmachine))'
-	# The stack is only 16-byte aligned on Windows, so don't let gcc emit aligned moves.
-	# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=54412
-	# https://github.com/ggerganov/llama.cpp/issues/2922
 	MK_CFLAGS   += -Xassembler -muse-unaligned-vector-move
 	MK_CXXFLAGS += -Xassembler -muse-unaligned-vector-move
-
-	# Target Windows 8 for PrefetchVirtualMemory
-	MK_CPPFLAGS += -D_WIN32_WINNT=0x602
 endif
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)
-	# Nvidia Jetson
 	MK_CFLAGS   += -mcpu=native
 	MK_CXXFLAGS += -mcpu=native
-	JETSON_RELEASE_INFO = $(shell jetson_release)
-	ifdef JETSON_RELEASE_INFO
-		ifneq ($(filter TX2%,$(JETSON_RELEASE_INFO)),)
-			JETSON_EOL_MODULE_DETECT = 1
-			CC = aarch64-unknown-linux-gnu-gcc
-			cxx = aarch64-unknown-linux-gnu-g++
-		endif
-	endif
 endif
 
 ifneq ($(filter armv6%,$(UNAME_M)),)
@@ -317,20 +337,18 @@ ifneq ($(filter ppc64%,$(UNAME_M)),)
 	endif
 endif
 
-ifneq ($(filter ppc64le%,$(UNAME_M)),)
-	MK_CFLAGS   += -mcpu=powerpc64le
-	MK_CXXFLAGS += -mcpu=powerpc64le
-	CUDA_POWER_ARCH = 1
-endif
-
 else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
 
+ifndef LLAMA_NO_K_QUANTS
+	MK_CPPFLAGS += -DGGML_USE_K_QUANTS
+	OBJS     += k_quants.o
 ifdef LLAMA_QKK_64
 	MK_CPPFLAGS += -DGGML_QKK_64
 endif
+endif
 
 ifndef LLAMA_NO_ACCELERATE
 	# Mac OS - include Accelerate framework.
@@ -347,7 +365,7 @@ ifdef LLAMA_MPI
 	MK_CPPFLAGS += -DGGML_USE_MPI
 	MK_CFLAGS   += -Wno-cast-qual
 	MK_CXXFLAGS += -Wno-cast-qual
-	OBJS        += ggml-mpi.o
+	OBJS     += ggml-mpi.o
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
@@ -362,72 +380,62 @@ 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
-	OBJS         += ggml-cuda.o
-	MK_NVCCFLAGS  = -use_fast_math
-ifndef JETSON_EOL_MODULE_DETECT
-	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
-endif # JETSON_EOL_MODULE_DETECT
-ifdef LLAMA_DEBUG
-	MK_NVCCFLAGS += -lineinfo
-endif # LLAMA_DEBUG
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
+	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
+	OBJS      += ggml-cuda.o
+	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
 ifdef LLAMA_CUDA_NVCC
 	NVCC = $(LLAMA_CUDA_NVCC)
 else
 	NVCC = nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else ifndef CUDA_POWER_ARCH
-	MK_NVCCFLAGS += -arch=native
+	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else
+	NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_FORCE_MMQ
-	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
 ifdef LLAMA_CUDA_MMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 else ifdef LLAMA_CUDA_DMMV_Y
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	MK_NVCCFLAGS += -DGGML_CUDA_F16
+	NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
 else
-	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
-	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-else
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endif # JETSON_EOL_MODULE_DETECT
+	$(NVCC) $(NVCCFLAGS) -c $< -o $@
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -448,36 +456,14 @@ ggml-opencl.o: ggml-opencl.cpp ggml-opencl.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # LLAMA_CLBLAST
 
-ifdef LLAMA_VULKAN
-	MK_CPPFLAGS  += -DGGML_USE_VULKAN
-	MK_LDFLAGS += -lvulkan
-	OBJS    += ggml-vulkan.o
-
-ifdef LLAMA_VULKAN_CHECK_RESULTS
-	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
-endif
-
-ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-endif # LLAMA_VULKAN
-
 ifdef LLAMA_HIPBLAS
-
-	ifeq ($(wildcard /opt/rocm),)
-		ROCM_PATH	?= /usr
-		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
-	else
-		ROCM_PATH	?= /opt/rocm
-		GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
-	endif
-	HIPCC                   ?= $(ROCM_PATH)/bin/hipcc
+	ROCM_PATH	?= /opt/rocm
+	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
+	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
-ifdef LLAMA_HIP_UMA
-	MK_CPPFLAGS += -DGGML_HIP_UMA
-endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
@@ -511,22 +497,21 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-GF_CC := $(CC)
-include scripts/get-flags.mk
+ifndef LLAMA_NO_K_QUANTS
+k_quants.o: k_quants.c k_quants.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_NO_K_QUANTS
 
 # 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 NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
-override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
+override CFLAGS        := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(CFLAGS)
+override CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CUDA_CXXFLAGS := $(MK_CUDA_CXXFLAGS) $(CUDA_CXXFLAGS)
+override HOST_CXXFLAGS := $(MK_HOST_CXXFLAGS) $(HOST_CXXFLAGS)
+override LDFLAGS       := $(MK_LDFLAGS) $(LDFLAGS)
 
-# identify CUDA host compiler
-ifdef LLAMA_CUBLAS
-GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
-include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
-endif
+# save CXXFLAGS before we add host-only options
+NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
+override CXXFLAGS += $(HOST_CXXFLAGS)
 
 #
 # Print build information
@@ -557,16 +542,13 @@ 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
-	$(CC) $(CFLAGS)    -c $< -o $@
-
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
+OBJS += ggml-alloc.o ggml-backend.o
 
 llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
+COMMON_H_DEPS = common/common.h common/sampling.h build-info.h common/log.h
+COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -587,55 +569,49 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 
 #
 # Examples
 #
 
-main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tokenize: examples/tokenize/tokenize.cpp                      ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.h ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -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/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 build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual
 
-gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
+gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
@@ -644,48 +620,41 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-libllava.a: examples/llava/llava.cpp examples/llava/llava.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h common/base64.hpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
-
-llava-cli: examples/llava/llava-cli.cpp examples/llava/clip.h examples/llava/clip.cpp examples/llava/llava.h examples/llava/llava.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
 
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+ifdef LLAMA_METAL
+metal: examples/metal/metal.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+endif
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif
 
-common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
+build-info.h: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
@@ -693,16 +662,13 @@ common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi
 
-build-info.o: common/build-info.cpp
-	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
-
 #
 # Tests
 #
 
 tests: $(TEST_TARGETS)
 
-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
@@ -716,53 +682,41 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
 
-tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-opt: tests/test-opt.cpp ggml.o $(OBJS)
+tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
 tests/test-c.o: tests/test-c.c llama.h
 	$(CC) $(CFLAGS) -c $(filter-out %.h,$^) -o $@
-
-tests/test-backend-ops: tests/test-backend-ops.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
-
-tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

Package.swift

@@ -2,41 +2,60 @@
 
 import PackageDescription
 
+#if arch(arm) || arch(arm64)
+let platforms: [SupportedPlatform]? = [
+    .macOS(.v12),
+    .iOS(.v14),
+    .watchOS(.v4),
+    .tvOS(.v14)
+]
+let exclude: [String] = []
+let resources: [Resource] = [
+    .process("ggml-metal.metal")
+]
+let additionalSources: [String] = ["ggml-metal.m"]
+let additionalSettings: [CSetting] = [
+    .unsafeFlags(["-fno-objc-arc"]),
+    .define("GGML_USE_METAL")
+]
+#else
+let platforms: [SupportedPlatform]? = nil
+let exclude: [String] = ["ggml-metal.metal"]
+let resources: [Resource] = []
+let additionalSources: [String] = []
+let additionalSettings: [CSetting] = []
+#endif
+
 let package = Package(
     name: "llama",
-    platforms: [
-        .macOS(.v12),
-        .iOS(.v14),
-        .watchOS(.v4),
-        .tvOS(.v14)
-    ],
+    platforms: platforms,
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
-    dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
-    ],
     targets: [
         .target(
             name: "llama",
-            dependencies: ["ggml"],
             path: ".",
-            exclude: ["ggml-metal.metal"],
+            exclude: exclude,
             sources: [
+                "ggml.c",
                 "llama.cpp",
-            ],
+                "ggml-alloc.c",
+                "ggml-backend.c",
+                "k_quants.c",
+            ] + additionalSources,
+            resources: resources,
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL"),
+                .define("GGML_USE_K_QUANTS"),
+                .define("GGML_USE_ACCELERATE")
                 // NOTE: NEW_LAPACK will required iOS version 16.4+
                 // We should consider add this in the future when we drop support for iOS 14
                 // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
                 // .define("ACCELERATE_NEW_LAPACK"),
                 // .define("ACCELERATE_LAPACK_ILP64")
-            ],
+            ] + additionalSettings,
             linkerSettings: [
                 .linkedFramework("Accelerate")
             ]

README.md

@@ -2,6 +2,7 @@
 
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
+[![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
@@ -10,12 +11,8 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
-- ⚠️ Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
-- 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
+- LLaVA support: https://github.com/ggerganov/llama.cpp/pull/3436
+- ‼️ BPE tokenizer update: existing Falcon and Starcoder `.gguf` models will need to be reconverted: [#3252](https://github.com/ggerganov/llama.cpp/pull/3252)
 
 ----
 
@@ -63,7 +60,7 @@ The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quant
 - AVX, AVX2 and AVX512 support for x86 architectures
 - Mixed F16 / F32 precision
 - 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
-- CUDA, Metal, OpenCL, SYCL GPU backend support
+- CUDA, Metal and OpenCL GPU backend support
 
 The original implementation of `llama.cpp` was [hacked in an evening](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022).
 Since then, the project has improved significantly thanks to many contributions. This project is mainly for educational purposes and serves
@@ -98,21 +95,6 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
-- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
-- [X] [StableLM-3b-4e1t](https://github.com/ggerganov/llama.cpp/pull/3586)
-- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
-- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
-- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
-- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
-- [x] [GPT-2](https://huggingface.co/gpt2)
-
-**Multimodal models:**
-
-- [x] [Llava 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
-- [x] [Bakllava](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
-- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
-- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
-- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 
 
 **Bindings:**
@@ -120,27 +102,19 @@ as the main playground for developing new features for the [ggml](https://github
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
-- JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
-- Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
+- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
-- Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
-- Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 
 **UI:**
 
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [semperai/amica](https://github.com/semperai/amica)
-- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
-- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
-- [iohub/collama](https://github.com/iohub/coLLaMA)
 
 ---
 
@@ -347,7 +321,7 @@ mpirun -hostfile hostfile -n 3 ./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
 
 ### BLAS Build
 
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS and CLBlast. There are currently several different BLAS implementations available for build and use:
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
 
 - #### Accelerate Framework:
 
@@ -391,37 +365,20 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [BLIS.md](docs/BLIS.md) for more information.
 
-- #### Intel oneMKL
-  - Using manual oneAPI installation:
-    By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
-      ```bash
-      mkdir build
-      cd build
-      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-runtime docker image, only required for manual installation
-      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
-      cmake --build . --config Release
-      ```
+- #### Intel MKL
 
-  - Using oneAPI docker image:
-    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-runtime](https://hub.docker.com/r/intel/oneapi-runtime)
+  By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. You may also specify it by:
 
-      ```bash
-      mkdir build
-      cd build
-      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
-      cmake --build . --config Release
-      ```
-
-  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni.
-
-  Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
+  ```bash
+  mkdir build
+  cd build
+  cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+  cmake --build . --config Release
+  ```
 
 - #### cuBLAS
 
   This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
-
-  For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
-
   - Using `make`:
     ```bash
     make LLAMA_CUBLAS=1
@@ -454,39 +411,22 @@ Building the program with BLAS support may lead to some performance improvements
   This provides BLAS acceleration on HIP-supported AMD GPUs.
   Make sure to have ROCm installed.
   You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
+  Windows support is coming soon...
 
   - Using `make`:
     ```bash
     make LLAMA_HIPBLAS=1
     ```
-  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
+  - Using `CMake`:
     ```bash
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ \
-        cmake -H. -Bbuild -DLLAMA_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-        && cmake --build build -- -j 16
-    ```
-    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DLLAMA_HIP_UMA=ON"`.
-    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
-  - Using `make` (example for target gfx1030, build with 16 CPU threads):
-    ```bash
-    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
-    ```
-
-  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
-    ```bash
-    set PATH=%HIP_PATH%\bin;%PATH%
     mkdir build
     cd build
-    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..
+    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
     cmake --build .
     ```
-    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
-    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
-
 
   The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+  If your GPU is not officialy supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
   | Option                  | Legal values           | Default | Description |
@@ -599,15 +539,6 @@ Building the program with BLAS support may lead to some performance improvements
 
   You can get a list of platforms and devices from the `clinfo -l` command, etc.
 
-- #### SYCL
-
-  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
-
-  llama.cpp based on SYCL is used to support Intel GPU (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
-
-  For detailed info, please refer to [llama.cpp for SYCL](README_sycl.md).
-
-
 ### Prepare Data & Run
 
 ```bash
@@ -941,22 +872,19 @@ Place your desired model into the `~/llama.cpp/models/` directory and execute th
 * Create a folder to store big models & intermediate files (ex. /llama/models)
 
 #### Images
-We have three Docker images available for this project:
+We have two 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`)
 
 Additionally, there the following images, similar to the above:
 
 - `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the Gitlab Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
 
 #### Usage
 
@@ -980,12 +908,6 @@ or with a light image:
 docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
 ```
 
-or with a server image:
-
-```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
-```
-
 ### Docker With CUDA
 
 Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
@@ -995,7 +917,6 @@ Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia
 ```bash
 docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
 docker build -t local/llama.cpp:light-cuda -f .devops/main-cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda -f .devops/server-cuda.Dockerfile .
 ```
 
 You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
@@ -1009,7 +930,6 @@ The resulting images, are essentially the same as the non-CUDA images:
 
 1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
 2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
-3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
 
 #### Usage
 
@@ -1018,7 +938,6 @@ After building locally, Usage is similar to the non-CUDA examples, but you'll ne
 ```bash
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
 
 ### Contributing
@@ -1038,8 +957,6 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
-- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
 
 ### Docs
 

README_sycl.md

@@ -1,252 +0,0 @@
-# llama.cpp for SYCL
-
-[Background](#background)
-
-[OS](#os)
-
-[Intel GPU](#intel-gpu)
-
-[Linux](#linux)
-
-[Environment Variable](#environment-variable)
-
-[Known Issue](#known-issue)
-
-[Todo](#todo)
-
-## Background
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
-
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
-
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
-
-To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
-
-The llama.cpp for SYCL is used to support Intel GPUs.
-
-For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
-
-## OS
-
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04|
-|Windows|Ongoing| |
-
-
-## Intel GPU
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1165G7|
-
-
-## Linux
-
-### Setup Environment
-
-1. Install Intel GPU driver.
-
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
-
-Note: for iGPU, please install the client GPU driver.
-
-b. Add user to group: video, render.
-
-```
-sudo usermod -aG render username
-sudo usermod -aG video username
-```
-
-Note: re-login to enable it.
-
-c. Check
-
-```
-sudo apt install clinfo
-sudo clinfo -l
-```
-
-Output (example):
-
-```
-Platform #0: Intel(R) OpenCL Graphics
- `-- Device #0: Intel(R) Arc(TM) A770 Graphics
-
-
-Platform #0: Intel(R) OpenCL HD Graphics
- `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
-```
-
-2. Install Intel® oneAPI Base toolkit.
-
-
-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**.
-
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
-
-b. Check
-
-```
-source /opt/intel/oneapi/setvars.sh
-
-sycl-ls
-```
-
-There should be one or more level-zero devices. Like **[ext_oneapi_level_zero:gpu:0]**.
-
-Output (example):
-```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
-
-```
-
-2. Build locally:
-
-```
-mkdir -p build
-cd build
-source /opt/intel/oneapi/setvars.sh
-
-#for FP16
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON # faster for long-prompt inference
-
-#for FP32
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-#build example/main only
-#cmake --build . --config Release --target main
-
-#build all binary
-cmake --build . --config Release -v
-
-```
-
-or
-
-```
-./examples/sycl/build.sh
-```
-
-Note:
-
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
-
-### Run
-
-1. Put model file to folder **models**
-
-2. Enable oneAPI running environment
-
-```
-source /opt/intel/oneapi/setvars.sh
-```
-
-3. List device ID
-
-Run without parameter:
-
-```
-./build/bin/ls-sycl-device
-
-or
-
-./build/bin/main
-```
-
-Check the ID in startup log, like:
-
-```
-found 4 SYCL devices:
-  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
-    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
-  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
-    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
-  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
-    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
-
-```
-
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
-
-4. Set device ID and execute llama.cpp
-
-Set device ID = 0 by **GGML_SYCL_DEVICE=0**
-
-```
-GGML_SYCL_DEVICE=0 ./build/bin/main -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
-```
-or run by script:
-
-```
-./examples/sycl/run_llama2.sh
-```
-
-Note:
-
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
-
-
-5. Check the device ID in output
-
-Like：
-```
-Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
-```
-
-
-## Environment Variable
-
-#### Build
-
-|Name|Value|Function|
-|-|-|-|
-|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
-|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx|use icpx for SYCL code path|
-
-#### Running
-
-
-|Name|Value|Function|
-|-|-|-|
-|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|
-
-## Known Issue
-
-- Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
-
-  Miss to enable oneAPI running environment.
-
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
-
-
-- Hang during startup
-
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
-
-  Solution: add **--no-mmap**.
-
-## Todo
-
-- Support to build in Windows.
-
-- Support multiple cards.

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

@@ -10,6 +10,7 @@ const Maker = struct {
     builder: *std.build.Builder,
     target: CrossTarget,
     optimize: Mode,
+    config_header: *ConfigHeader,
     enable_lto: bool,
 
     include_dirs: ArrayList([]const u8),
@@ -40,24 +41,26 @@ const Maker = struct {
         const commit_hash = try std.ChildProcess.exec(
             .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
         );
-        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
-            \\int LLAMA_BUILD_NUMBER = {};
-            \\char const *LLAMA_COMMIT = "{s}";
-            \\char const *LLAMA_COMPILER = "Zig {s}";
-            \\char const *LLAMA_BUILD_TARGET = "{s}";
-            \\
-        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
+        const config_header = builder.addConfigHeader(
+            .{ .style = .blank, .include_path = "build-info.h" },
+            .{
+                .BUILD_NUMBER = 0,
+                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
+                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
+                .BUILD_TARGET = try target.allocDescription(builder.allocator),
+            },
+        );
         var m = Maker{
             .builder = builder,
             .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
+            .config_header = config_header,
             .enable_lto = false,
             .include_dirs = ArrayList([]const u8).init(builder.allocator),
             .cflags = ArrayList([]const u8).init(builder.allocator),
             .cxxflags = ArrayList([]const u8).init(builder.allocator),
             .objs = ArrayList(*Compile).init(builder.allocator),
         };
-
         try m.addCFlag("-std=c11");
         try m.addCxxFlag("-std=c++11");
         try m.addProjectInclude(&.{});
@@ -69,7 +72,7 @@ const Maker = struct {
         const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
         if (o.target.getAbi() != .msvc)
             o.defineCMacro("_GNU_SOURCE", null);
-
+        o.addConfigHeader(m.config_header);
         if (std.mem.endsWith(u8, src, ".c")) {
             o.addCSourceFiles(&.{src}, m.cflags.items);
             o.linkLibC();
@@ -82,6 +85,7 @@ const Maker = struct {
                 o.linkLibCpp();
             }
         }
+        o.addConfigHeader(m.config_header);
         for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
         o.want_lto = m.enable_lto;
         return o;
@@ -101,6 +105,7 @@ const Maker = struct {
             // linkLibCpp already add (libc++ + libunwind + libc)
             e.linkLibCpp();
         }
+        e.addConfigHeader(m.config_header);
         m.builder.installArtifact(e);
         e.want_lto = m.enable_lto;
         return e;
@@ -111,12 +116,16 @@ pub fn build(b: *std.build.Builder) !void {
     var make = try Maker.init(b);
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
+    if (b.option(bool, "k-quants", "Enable K-quants, (default: true)") orelse true) {
+        try make.addFlag("-DGGML_USE_K_QUANTS");
+        const k_quants = make.obj("k_quants", "k_quants.c");
+        try make.objs.append(k_quants);
+    }
+
     const ggml = make.obj("ggml", "ggml.c");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
-    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const llama = make.obj("llama", "llama.cpp");
-    const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
@@ -124,14 +133,14 @@ pub fn build(b: *std.build.Builder) !void {
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.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 });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
-    _ = 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 });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, 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, llama, common, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/README.md

@@ -22,8 +22,4 @@ bash ./ci/run.sh ./tmp/results ./tmp/mnt
 
 # with CUDA support
 GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-
-# with SYCL support
-source /opt/intel/oneapi/setvars.sh
-GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 ```

ci/run.sh

@@ -10,9 +10,6 @@
 # # with CUDA support
 # GG_BUILD_CUDA=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with SYCL support
-# GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -25,32 +22,14 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -f "$OUT/*.log"
-rm -f "$OUT/*.exit"
-rm -f "$OUT/*.md"
+rm -v $OUT/*.log
+rm -v $OUT/*.exit
+rm -v $OUT/*.md
 
 sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA=""
-
-if [ ! -z ${GG_BUILD_METAL} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
-fi
-
-if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
-fi
-
-if [ ! -z ${GG_BUILD_SYCL} ]; then
-    if [ -z ${ONEAPI_ROOT} ]; then
-        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
-        exit 1
-    fi
-
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON"
-fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -102,10 +81,10 @@ function gg_run_ctest_debug {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
 
-    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -130,13 +109,13 @@ function gg_run_ctest_release {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -152,61 +131,6 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
-function gg_get_model {
-    local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
-    local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
-    if [[ -s $gguf_3b ]]; then
-        echo -n "$gguf_3b"
-    elif [[ -s $gguf_7b ]]; then
-        echo -n "$gguf_7b"
-    else
-        echo >&2 "No model found. Can't run gg_run_ctest_with_model."
-        exit 1
-    fi
-}
-
-function gg_run_ctest_with_model_debug {
-    cd ${SRC}
-
-    local model; model=$(gg_get_model)
-    cd build-ci-debug
-    set -e
-    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
-    set +e
-    cd ..
-}
-
-function gg_run_ctest_with_model_release {
-    cd ${SRC}
-
-    local model; model=$(gg_get_model)
-    cd build-ci-release
-    set -e
-    (LLAMACPP_TEST_MODELFILE="$model" time ctest --output-on-failure -L model) 2>&1 | tee -a $OUT/${ci}-ctest.log
-    set +e
-    cd ..
-}
-
-function gg_sum_ctest_with_model_debug {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs ctest with model files in debug mode\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
-    gg_printf '```\n'
-}
-
-function gg_sum_ctest_with_model_release {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Runs ctest with model files in release mode\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '```\n'
-    gg_printf '%s\n' "$(cat $OUT/${ci}-ctest.log)"
-    gg_printf '```\n'
-}
-
 # open_llama_3b_v2
 
 function gg_run_open_llama_3b_v2 {
@@ -230,8 +154,8 @@ function gg_run_open_llama_3b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -284,8 +208,6 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
@@ -313,8 +235,6 @@ function gg_run_open_llama_3b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
     # lora
     function compare_ppl {
         qnt="$1"
@@ -356,6 +276,7 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
     compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
+
     set +e
 }
 
@@ -365,7 +286,6 @@ function gg_sum_open_llama_3b_v2 {
     gg_printf 'OpenLLaMA 3B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
     gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
@@ -411,8 +331,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release -DLLAMA_CUBLAS=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                              ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -465,8 +385,6 @@ function gg_run_open_llama_7b_v2 {
     (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
-
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
@@ -494,8 +412,6 @@ function gg_run_open_llama_7b_v2 {
     check_ppl "q5_k" "$(cat $OUT/${ci}-tg-q5_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
     check_ppl "q6_k" "$(cat $OUT/${ci}-tg-q6_k.log | grep "^\[1\]")" | tee -a $OUT/${ci}-ppl.log
 
-    cat $OUT/${ci}-imatrix.log | grep "Final" >> $OUT/${ci}-imatrix-sum.log
-
     # lora
     function compare_ppl {
         qnt="$1"
@@ -547,7 +463,6 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
     gg_printf '- perplexity:\n%s\n' "$(cat $OUT/${ci}-ppl.log)"
-    gg_printf '- imatrix:\n```\n%s\n```\n' "$(cat $OUT/${ci}-imatrix-sum.log)"
     gg_printf '- lora:\n%s\n' "$(cat $OUT/${ci}-lora-ppl.log)"
     gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
     gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
@@ -571,18 +486,14 @@ function gg_sum_open_llama_7b_v2 {
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
+
     mnt_models=${MNT}/models
     mkdir -p ${mnt_models}
     ln -sfn ${mnt_models} ${SRC}/models-mnt
 
-    # Create a fresh python3 venv and enter it
-    python3 -m venv "$MNT/venv"
-    source "$MNT/venv/bin/activate"
-
-    pip install -r ${SRC}/requirements.txt --disable-pip-version-check
-    pip install --editable gguf-py --disable-pip-version-check
+    python3 -m pip install -r ${SRC}/requirements.txt
+    python3 -m pip install --editable gguf-py
 fi
 
 ret=0
@@ -597,8 +508,6 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
         else
             test $ret -eq 0 && gg_run open_llama_7b_v2
         fi
-        test $ret -eq 0 && gg_run ctest_with_model_debug
-        test $ret -eq 0 && gg_run ctest_with_model_release
     fi
 fi
 

cmake/FindSIMD.cmake

@@ -1,100 +0,0 @@
-include(CheckCSourceRuns)
-
-set(AVX_CODE "
-    #include <immintrin.h>
-    int main()
-    {
-        __m256 a;
-        a = _mm256_set1_ps(0);
-        return 0;
-    }
-")
-
-set(AVX512_CODE "
-    #include <immintrin.h>
-    int main()
-    {
-        __m512i a = _mm512_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0,
-                                    0, 0, 0, 0, 0, 0, 0, 0);
-        __m512i b = a;
-        __mmask64 equality_mask = _mm512_cmp_epi8_mask(a, b, _MM_CMPINT_EQ);
-        return 0;
-    }
-")
-
-set(AVX2_CODE "
-    #include <immintrin.h>
-    int main()
-    {
-        __m256i a = {0};
-        a = _mm256_abs_epi16(a);
-        __m256i x;
-        _mm256_extract_epi64(x, 0); // we rely on this in our AVX2 code
-        return 0;
-    }
-")
-
-set(FMA_CODE "
-    #include <immintrin.h>
-    int main()
-    {
-        __m256 acc = _mm256_setzero_ps();
-        const __m256 d = _mm256_setzero_ps();
-        const __m256 p = _mm256_setzero_ps();
-        acc = _mm256_fmadd_ps( d, p, acc );
-        return 0;
-    }
-")
-
-macro(check_sse type flags)
-    set(__FLAG_I 1)
-    set(CMAKE_REQUIRED_FLAGS_SAVE ${CMAKE_REQUIRED_FLAGS})
-    foreach (__FLAG ${flags})
-        if (NOT ${type}_FOUND)
-            set(CMAKE_REQUIRED_FLAGS ${__FLAG})
-            check_c_source_runs("${${type}_CODE}" HAS_${type}_${__FLAG_I})
-            if (HAS_${type}_${__FLAG_I})
-                set(${type}_FOUND TRUE CACHE BOOL "${type} support")
-                set(${type}_FLAGS "${__FLAG}" CACHE STRING "${type} flags")
-            endif()
-            math(EXPR __FLAG_I "${__FLAG_I}+1")
-        endif()
-    endforeach()
-    set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_SAVE})
-
-    if (NOT ${type}_FOUND)
-        set(${type}_FOUND FALSE CACHE BOOL "${type} support")
-        set(${type}_FLAGS "" CACHE STRING "${type} flags")
-    endif()
-
-    mark_as_advanced(${type}_FOUND ${type}_FLAGS)
-endmacro()
-
-# flags are for MSVC only!
-check_sse("AVX" " ;/arch:AVX")
-if (NOT ${AVX_FOUND})
-    set(LLAMA_AVX OFF)
-else()
-    set(LLAMA_AVX ON)
-endif()
-
-check_sse("AVX2" " ;/arch:AVX2")
-check_sse("FMA" " ;/arch:AVX2")
-if ((NOT ${AVX2_FOUND}) OR (NOT ${FMA_FOUND}))
-    set(LLAMA_AVX2 OFF)
-else()
-    set(LLAMA_AVX2 ON)
-endif()
-
-check_sse("AVX512" " ;/arch:AVX512")
-if (NOT ${AVX512_FOUND})
-    set(LLAMA_AVX512 OFF)
-else()
-    set(LLAMA_AVX512 ON)
-endif()

common/CMakeLists.txt

@@ -1,52 +1,8 @@
 # common
 
-
-# Build info header
-#
-
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
-
-    # Is git submodule
-    if(NOT IS_DIRECTORY "${GIT_DIR}")
-        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
-        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        string(FIND "${REAL_GIT_DIR}" "/" SLASH_POS)
-        if (SLASH_POS EQUAL 0)
-            set(GIT_DIR "${REAL_GIT_DIR}")
-        else()
-            set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../${REAL_GIT_DIR}")
-        endif()
-    endif()
-
-    set(GIT_INDEX "${GIT_DIR}/index")
-else()
-    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-    set(GIT_INDEX "")
-endif()
-
-# Add a custom command to rebuild build-info.cpp when .git/index changes
-add_custom_command(
-    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
-    COMMENT "Generating build details from Git"
-    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
-            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
-    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
-    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
-    VERBATIM
-)
-set(TARGET build_info)
-add_library(${TARGET} OBJECT build-info.cpp)
-if (BUILD_SHARED_LIBS)
-    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
-endif()
-
-
 set(TARGET common)
 
-add_library(${TARGET} STATIC
-    base64.hpp
+add_library(${TARGET} OBJECT
     common.h
     common.cpp
     sampling.h
@@ -65,4 +21,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)
+target_link_libraries(${TARGET} PRIVATE llama)

common/base64.hpp

@@ -1,392 +0,0 @@
-/*
-This is free and unencumbered software released into the public domain.
-
-Anyone is free to copy, modify, publish, use, compile, sell, or
-distribute this software, either in source code form or as a compiled
-binary, for any purpose, commercial or non-commercial, and by any
-means.
-
-In jurisdictions that recognize copyright laws, the author or authors
-of this software dedicate any and all copyright interest in the
-software to the public domain. We make this dedication for the benefit
-of the public at large and to the detriment of our heirs and
-successors. We intend this dedication to be an overt act of
-relinquishment in perpetuity of all present and future rights to this
-software under copyright law.
-
-THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
-EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
-MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.
-IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR
-OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
-ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
-OTHER DEALINGS IN THE SOFTWARE.
-
-For more information, please refer to <http://unlicense.org>
-*/
-
-#ifndef PUBLIC_DOMAIN_BASE64_HPP_
-#define PUBLIC_DOMAIN_BASE64_HPP_
-
-#include <cstdint>
-#include <iterator>
-#include <stdexcept>
-#include <string>
-
-class base64_error : public std::runtime_error
-{
-public:
-    using std::runtime_error::runtime_error;
-};
-
-class base64
-{
-public:
-    enum class alphabet
-    {
-        /** the alphabet is detected automatically */
-        auto_,
-        /** the standard base64 alphabet is used */
-        standard,
-        /** like `standard` except that the characters `+` and `/` are replaced by `-` and `_` respectively*/
-        url_filename_safe
-    };
-
-    enum class decoding_behavior
-    {
-        /** if the input is not padded, the remaining bits are ignored */
-        moderate,
-        /** if a padding character is encounter decoding is finished */
-        loose
-    };
-
-    /**
-     Encodes all the elements from `in_begin` to `in_end` to `out`.
-
-     @warning The source and destination cannot overlap. The destination must be able to hold at least
-     `required_encode_size(std::distance(in_begin, in_end))`, otherwise the behavior depends on the output iterator.
-
-     @tparam Input_iterator the source; the returned elements are cast to `std::uint8_t` and should not be greater than
-     8 bits
-     @tparam Output_iterator the destination; the elements written to it are from the type `char`
-     @param in_begin the beginning of the source
-     @param in_end the ending of the source
-     @param out the destination iterator
-     @param alphabet which alphabet should be used
-     @returns the iterator to the next element past the last element copied
-     @throws see `Input_iterator` and `Output_iterator`
-    */
-    template<typename Input_iterator, typename Output_iterator>
-    static Output_iterator encode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
-                                  alphabet alphabet = alphabet::standard)
-    {
-        constexpr auto pad = '=';
-        const char* alpha  = alphabet == alphabet::url_filename_safe
-                                ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
-                                : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
-
-        while (in_begin != in_end) {
-            std::uint8_t i0 = 0, i1 = 0, i2 = 0;
-
-            // first character
-            i0 = static_cast<std::uint8_t>(*in_begin);
-            ++in_begin;
-
-            *out = alpha[i0 >> 2 & 0x3f];
-            ++out;
-
-            // part of first character and second
-            if (in_begin != in_end) {
-                i1 = static_cast<std::uint8_t>(*in_begin);
-                ++in_begin;
-
-                *out = alpha[((i0 & 0x3) << 4) | (i1 >> 4 & 0x0f)];
-                ++out;
-            } else {
-                *out = alpha[(i0 & 0x3) << 4];
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                break;
-            }
-
-            // part of second character and third
-            if (in_begin != in_end) {
-                i2 = static_cast<std::uint8_t>(*in_begin);
-                ++in_begin;
-
-                *out = alpha[((i1 & 0xf) << 2) | (i2 >> 6 & 0x03)];
-                ++out;
-            } else {
-                *out = alpha[(i1 & 0xf) << 2];
-                ++out;
-
-                // last padding
-                *out = pad;
-                ++out;
-
-                break;
-            }
-
-            // rest of third
-            *out = alpha[i2 & 0x3f];
-            ++out;
-        }
-
-        return out;
-    }
-    /**
-     Encodes a string.
-
-     @param str the string that should be encoded
-     @param alphabet which alphabet should be used
-     @returns the encoded base64 string
-     @throws see base64::encode()
-    */
-    static std::string encode(const std::string& str, alphabet alphabet = alphabet::standard)
-    {
-        std::string result;
-
-        result.reserve(required_encode_size(str.length()) + 1);
-
-        encode(str.begin(), str.end(), std::back_inserter(result), alphabet);
-
-        return result;
-    }
-    /**
-     Encodes a char array.
-
-     @param buffer the char array
-     @param size the size of the array
-     @param alphabet which alphabet should be used
-     @returns the encoded string
-    */
-    static std::string encode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::standard)
-    {
-        std::string result;
-
-        result.reserve(required_encode_size(size) + 1);
-
-        encode(buffer, buffer + size, std::back_inserter(result), alphabet);
-
-        return result;
-    }
-    /**
-     Decodes all the elements from `in_begin` to `in_end` to `out`. `in_begin` may point to the same location as `out`,
-     in other words: inplace decoding is possible.
-
-     @warning The destination must be able to hold at least `required_decode_size(std::distance(in_begin, in_end))`,
-     otherwise the behavior depends on the output iterator.
-
-     @tparam Input_iterator the source; the returned elements are cast to `char`
-     @tparam Output_iterator the destination; the elements written to it are from the type `std::uint8_t`
-     @param in_begin the beginning of the source
-     @param in_end the ending of the source
-     @param out the destination iterator
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the iterator to the next element past the last element copied
-     @throws base64_error depending on the set behavior
-     @throws see `Input_iterator` and `Output_iterator`
-    */
-    template<typename Input_iterator, typename Output_iterator>
-    static Output_iterator decode(Input_iterator in_begin, Input_iterator in_end, Output_iterator out,
-                                  alphabet alphabet          = alphabet::auto_,
-                                  decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        //constexpr auto pad = '=';
-        std::uint8_t last  = 0;
-        auto bits          = 0;
-
-        while (in_begin != in_end) {
-            auto c = *in_begin;
-            ++in_begin;
-
-            if (c == '=') {
-                break;
-            }
-
-            auto part = _base64_value(alphabet, c);
-
-            // enough bits for one byte
-            if (bits + 6 >= 8) {
-                *out = (last << (8 - bits)) | (part >> (bits - 2));
-                ++out;
-
-                bits -= 2;
-            } else {
-                bits += 6;
-            }
-
-            last = part;
-        }
-
-        // check padding
-        if (behavior != decoding_behavior::loose) {
-            while (in_begin != in_end) {
-                auto c = *in_begin;
-                ++in_begin;
-
-                if (c != '=') {
-                    throw base64_error("invalid base64 character.");
-                }
-            }
-        }
-
-        return out;
-    }
-    /**
-     Decodes a string.
-
-     @param str the base64 encoded string
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the decoded string
-     @throws see base64::decode()
-    */
-    static std::string decode(const std::string& str, alphabet alphabet = alphabet::auto_,
-                              decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        std::string result;
-
-        result.reserve(max_decode_size(str.length()));
-
-        decode(str.begin(), str.end(), std::back_inserter(result), alphabet, behavior);
-
-        return result;
-    }
-    /**
-     Decodes a string.
-
-     @param buffer the base64 encoded buffer
-     @param size the size of the buffer
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the decoded string
-     @throws see base64::decode()
-    */
-    static std::string decode(const char* buffer, std::size_t size, alphabet alphabet = alphabet::auto_,
-                              decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        std::string result;
-
-        result.reserve(max_decode_size(size));
-
-        decode(buffer, buffer + size, std::back_inserter(result), alphabet, behavior);
-
-        return result;
-    }
-    /**
-     Decodes a string inplace.
-
-     @param[in,out] str the base64 encoded string
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @throws base64::decode_inplace()
-    */
-    static void decode_inplace(std::string& str, alphabet alphabet = alphabet::auto_,
-                               decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        str.resize(decode(str.begin(), str.end(), str.begin(), alphabet, behavior) - str.begin());
-    }
-    /**
-     Decodes a char array inplace.
-
-     @param[in,out] str the string array
-     @param size the length of the array
-     @param alphabet which alphabet should be used
-     @param behavior the behavior when an error was detected
-     @returns the pointer to the next element past the last element decoded
-     @throws base64::decode_inplace()
-    */
-    static char* decode_inplace(char* str, std::size_t size, alphabet alphabet = alphabet::auto_,
-                                decoding_behavior behavior = decoding_behavior::moderate)
-    {
-        return decode(str, str + size, str, alphabet, behavior);
-    }
-    /**
-     Returns the required decoding size for a given size. The value is calculated with the following formula:
-
-     $$
-     \lceil \frac{size}{4} \rceil \cdot 3
-     $$
-
-     @param size the size of the encoded input
-     @returns the size of the resulting decoded buffer; this the absolute maximum
-    */
-    static std::size_t max_decode_size(std::size_t size) noexcept
-    {
-        return (size / 4 + (size % 4 ? 1 : 0)) * 3;
-    }
-    /**
-     Returns the required encoding size for a given size. The value is calculated with the following formula:
-
-     $$
-     \lceil \frac{size}{3} \rceil \cdot 4
-     $$
-
-     @param size the size of the decoded input
-     @returns the size of the resulting encoded buffer
-    */
-    static std::size_t required_encode_size(std::size_t size) noexcept
-    {
-        return (size / 3 + (size % 3 ? 1 : 0)) * 4;
-    }
-
-private:
-    static std::uint8_t _base64_value(alphabet& alphabet, char c)
-    {
-        if (c >= 'A' && c <= 'Z') {
-            return c - 'A';
-        } else if (c >= 'a' && c <= 'z') {
-            return c - 'a' + 26;
-        } else if (c >= '0' && c <= '9') {
-            return c - '0' + 52;
-        }
-
-        // comes down to alphabet
-        if (alphabet == alphabet::standard) {
-            if (c == '+') {
-                return 62;
-            } else if (c == '/') {
-                return 63;
-            }
-        } else if (alphabet == alphabet::url_filename_safe) {
-            if (c == '-') {
-                return 62;
-            } else if (c == '_') {
-                return 63;
-            }
-        } // auto detect
-        else {
-            if (c == '+') {
-                alphabet = alphabet::standard;
-
-                return 62;
-            } else if (c == '/') {
-                alphabet = alphabet::standard;
-
-                return 63;
-            } else if (c == '-') {
-                alphabet = alphabet::url_filename_safe;
-
-                return 62;
-            } else if (c == '_') {
-                alphabet = alphabet::url_filename_safe;
-
-                return 63;
-            }
-        }
-
-        throw base64_error("invalid base64 character.");
-    }
-};
-
-#endif // !PUBLIC_DOMAIN_BASE64_HPP_

common/build-info.cpp.in

@@ -1,4 +0,0 @@
-int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
-char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
-char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
-char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";

common/common.cpp

@@ -1,4 +1,5 @@
 #include "common.h"
+#include "build-info.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -12,7 +13,6 @@
 #include <regex>
 #include <sstream>
 #include <string>
-#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
@@ -42,10 +42,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL))
-#define GGML_USE_CUBLAS_SYCL
-#endif
-
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -95,19 +91,6 @@ void process_escapes(std::string& input) {
                 case '\'': input[output_idx++] = '\''; break;
                 case '\"': input[output_idx++] = '\"'; break;
                 case '\\': input[output_idx++] = '\\'; break;
-                case 'x':
-                    // Handle \x12, etc
-                    if (input_idx + 2 < input_len) {
-                        const char x[3] = { input[input_idx + 1], input[input_idx + 2], 0 };
-                        char *err_p = nullptr;
-                        const long val = std::strtol(x, &err_p, 16);
-                        if (err_p == x + 2) {
-                            input_idx += 2;
-                            input[output_idx++] = char(val);
-                            break;
-                        }
-                    }
-                    // fall through
                 default:   input[output_idx++] = '\\';
                            input[output_idx++] = input[input_idx]; break;
             }
@@ -120,24 +103,9 @@ void process_escapes(std::string& input) {
 }
 
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
-    bool result = true;
-    try {
-        if (!gpt_params_parse_ex(argc, argv, params)) {
-            gpt_print_usage(argc, argv, gpt_params());
-            exit(0);
-        }
-    }
-    catch (const std::invalid_argument & ex) {
-        fprintf(stderr, "%s\n", ex.what());
-        gpt_print_usage(argc, argv, gpt_params());
-        exit(1);
-    }
-    return result;
-}
-
-bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
+    gpt_params default_params;
     const std::string arg_prefix = "--";
     llama_sampling_params & sparams = params.sparams;
 
@@ -171,24 +139,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             if (params.n_threads_batch <= 0) {
                 params.n_threads_batch = std::thread::hardware_concurrency();
             }
-        } else if (arg == "-td" || arg == "--threads-draft") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_threads_draft = std::stoi(argv[i]);
-            if (params.n_threads_draft <= 0) {
-                params.n_threads_draft = std::thread::hardware_concurrency();
-            }
-        } else if (arg == "-tbd" || arg == "--threads-batch-draft") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_threads_batch_draft = std::stoi(argv[i]);
-            if (params.n_threads_batch_draft <= 0) {
-                params.n_threads_batch_draft = std::thread::hardware_concurrency();
-            }
         } else if (arg == "-p" || arg == "--prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -207,23 +157,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.prompt_cache_all = true;
         } else if (arg == "--prompt-cache-ro") {
             params.prompt_cache_ro = true;
-        } else if (arg == "-bf" || arg == "--binary-file") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::ifstream file(argv[i], std::ios::binary);
-            if (!file) {
-                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-            // store the external file name in params
-            params.prompt_file = argv[i];
-            std::ostringstream ss;
-            ss << file.rdbuf();
-            params.prompt = ss.str();
-            fprintf(stderr, "Read %zu bytes from binary file %s\n", params.prompt.size(), argv[i]);
         } else if (arg == "-f" || arg == "--file") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -259,20 +192,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
-        } else if (arg == "--grp-attn-n" || arg == "-gan") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-
-            params.grp_attn_n = std::stoi(argv[i]);
-        } else if (arg == "--grp-attn-w" || arg == "-gaw") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-
-            params.grp_attn_w = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -285,76 +204,20 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.rope_freq_scale = std::stof(argv[i]);
-        } else if (arg == "--rope-scaling") {
-            if (++i >= argc) {
-                invalid_param = true;
-                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; }
-            else { invalid_param = true; break; }
         } else if (arg == "--rope-scale") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rope_freq_scale = 1.0f/std::stof(argv[i]);
-        } else if (arg == "--yarn-orig-ctx") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_orig_ctx = std::stoi(argv[i]);
-        } else if (arg == "--yarn-ext-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_ext_factor = std::stof(argv[i]);
-        } else if (arg == "--yarn-attn-factor") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_attn_factor = std::stof(argv[i]);
-        } else if (arg == "--yarn-beta-fast") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_beta_fast = std::stof(argv[i]);
-        } else if (arg == "--yarn-beta-slow") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.yarn_beta_slow = std::stof(argv[i]);
-        } else if (arg == "--samplers") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.samplers_sequence = parse_samplers_input(argv[i]);
-        } else if (arg == "--sampling-seq") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.samplers_sequence = argv[i];
+        } else if (arg == "--memory-f32") {
+            params.memory_f16 = false;
         } else if (arg == "--top-p") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             sparams.top_p = std::stof(argv[i]);
-        } else if (arg == "--min-p") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            sparams.min_p = std::stof(argv[i]);
         } else if (arg == "--temp") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -480,18 +343,6 @@ 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;
-                break;
-            }
-            params.p_split = std::stof(argv[i]);
         } else if (arg == "-m" || arg == "--model") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -555,18 +406,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.interactive_first = true;
         } else if (arg == "-ins" || arg == "--instruct") {
             params.instruct = true;
-        } else if (arg == "-cml" || arg == "--chatml") {
-            params.chatml = true;
         } else if (arg == "--infill") {
             params.infill = true;
-        } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
-            params.dump_kv_cache = true;
-        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
-            params.no_kv_offload = true;
-        } else if (arg == "-ctk" || arg == "--cache-type-k") {
-            params.cache_type_k = argv[++i];
-        } else if (arg == "-ctv" || arg == "--cache-type-v") {
-            params.cache_type_v = argv[++i];
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -582,8 +423,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers = std::stoi(argv[i]);
-#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
+#else
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -592,8 +434,9 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
             params.n_gpu_layers_draft = std::stoi(argv[i]);
-#ifndef LLAMA_SUPPORTS_GPU_OFFLOAD
+#else
             fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers-draft option will be ignored\n");
             fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
 #endif
@@ -602,45 +445,25 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
+#ifdef GGML_USE_CUBLAS
             params.main_gpu = std::stoi(argv[i]);
-#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the main GPU has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
-        } else if (arg == "--split-mode" || arg == "-sm") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            std::string arg_next = argv[i];
-            if (arg_next == "none") {
-                params.split_mode = LLAMA_SPLIT_NONE;
-            } else if (arg_next == "layer") {
-                params.split_mode = LLAMA_SPLIT_LAYER;
-            } else if (arg_next == "row") {
-                params.split_mode = LLAMA_SPLIT_ROW;
-            } else {
-                invalid_param = true;
-                break;
-            }
-#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
-
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a main GPU.\n");
+#endif
         } else if (arg == "--tensor-split" || arg == "-ts") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
+#ifdef GGML_USE_CUBLAS
             std::string arg_next = argv[i];
 
             // split string by , and /
             const std::regex regex{R"([,/]+)"};
             std::sregex_token_iterator it{arg_next.begin(), arg_next.end(), regex, -1};
             std::vector<std::string> split_arg{it, {}};
-            if (split_arg.size() >= LLAMA_MAX_DEVICES) {
-                invalid_param = true;
-                break;
-            }
+            GGML_ASSERT(split_arg.size() <= LLAMA_MAX_DEVICES);
+
             for (size_t i = 0; i < LLAMA_MAX_DEVICES; ++i) {
                 if (i < split_arg.size()) {
                     params.tensor_split[i] = std::stof(split_arg[i]);
@@ -648,17 +471,21 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                     params.tensor_split[i] = 0.0f;
                 }
             }
-#ifndef GGML_USE_CUBLAS_SYCL
-            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting a tensor split has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. It is not possible to set a tensor split.\n");
+#endif // GGML_USE_CUBLAS
+        } else if (arg == "--no-mul-mat-q" || arg == "-nommq") {
+#ifdef GGML_USE_CUBLAS
+            params.mul_mat_q = false;
+#else
+            fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS. Disabling mul_mat_q kernels has no effect.\n");
+#endif // GGML_USE_CUBLAS
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--numa") {
             params.numa = true;
         } else if (arg == "--verbose-prompt") {
             params.verbose_prompt = true;
-        } else if (arg == "--no-display-prompt") {
-            params.display_prompt = false;
         } else if (arg == "-r" || arg == "--reverse-prompt") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -675,12 +502,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             if (params.logdir.back() != DIRECTORY_SEPARATOR) {
                 params.logdir += DIRECTORY_SEPARATOR;
             }
-        } else if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.logits_file = argv[i];
         } else if (arg == "--perplexity" || arg == "--all-logits") {
             params.logits_all = true;
         } else if (arg == "--ppl-stride") {
@@ -689,12 +510,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.ppl_stride = std::stoi(argv[i]);
-        } else if (arg == "-ptc" || arg == "--print-token-count") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_print = std::stoi(argv[i]);
         } else if (arg == "--ppl-output-type") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -709,24 +524,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.hellaswag_tasks = std::stoi(argv[i]);
-        } else if (arg == "--winogrande") {
-            params.winogrande = true;
-        } else if (arg == "--winogrande-tasks") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.winogrande_tasks = std::stoi(argv[i]);
-        } else if (arg == "--multiple-choice") {
-            params.multiple_choice = true;
-        } else if (arg == "--multiple-choice-tasks") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.multiple_choice_tasks = std::stoi(argv[i]);
-        } else if (arg == "--kl-divergence") {
-            params.kl_divergence = true;
         } else if (arg == "--ignore-eos") {
             params.ignore_eos = true;
         } else if (arg == "--no-penalize-nl") {
@@ -751,11 +548,10 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
         } else if (arg == "-h" || arg == "--help") {
-            return false;
-
-        } else if (arg == "--version") {
-            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+            gpt_print_usage(argc, argv, default_params);
+#ifndef LOG_DISABLE_LOGS
+            log_print_usage();
+#endif // LOG_DISABLE_LOGS
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -795,47 +591,6 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 std::istreambuf_iterator<char>(),
                 std::back_inserter(sparams.grammar)
             );
-        } else if (arg == "--override-kv") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            char * sep = strchr(argv[i], '=');
-            if (sep == nullptr || sep - argv[i] >= 128) {
-                fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-            struct llama_model_kv_override kvo;
-            std::strncpy(kvo.key, argv[i], sep - argv[i]);
-            kvo.key[sep - argv[i]] = 0;
-            sep++;
-            if (strncmp(sep, "int:", 4) == 0) {
-                sep += 4;
-                kvo.tag = LLAMA_KV_OVERRIDE_INT;
-                kvo.int_value = std::atol(sep);
-            } else if (strncmp(sep, "float:", 6) == 0) {
-                sep += 6;
-                kvo.tag = LLAMA_KV_OVERRIDE_FLOAT;
-                kvo.float_value = std::atof(sep);
-            } else if (strncmp(sep, "bool:", 5) == 0) {
-                sep += 5;
-                kvo.tag = LLAMA_KV_OVERRIDE_BOOL;
-                if (std::strcmp(sep, "true") == 0) {
-                    kvo.bool_value = true;
-                } else if (std::strcmp(sep, "false") == 0) {
-                    kvo.bool_value = false;
-                } else {
-                    fprintf(stderr, "error: Invalid boolean value for KV override: %s\n", argv[i]);
-                    invalid_param = true;
-                    break;
-                }
-            } else {
-                fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
-                invalid_param = true;
-                break;
-            }
-            params.kv_overrides.push_back(kvo);
 #ifndef LOG_DISABLE_LOGS
         // Parse args for logging parameters
         } else if ( log_param_single_parse( argv[i] ) ) {
@@ -856,17 +611,22 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         // End of Parse args for logging parameters
 #endif // LOG_DISABLE_LOGS
         } else {
-            throw std::invalid_argument("error: unknown argument: " + arg);
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            gpt_print_usage(argc, argv, default_params);
+            exit(1);
         }
     }
     if (invalid_param) {
-        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        gpt_print_usage(argc, argv, default_params);
+        exit(1);
     }
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
              params.instruct)) {
-
-        throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
+        fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
+        gpt_print_usage(argc, argv, default_params);
+        exit(1);
     }
 
     if (params.escape) {
@@ -879,27 +639,19 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         }
     }
 
-    if (!params.kv_overrides.empty()) {
-        params.kv_overrides.emplace_back(llama_model_kv_override());
-        params.kv_overrides.back().key[0] = 0;
-    }
-
     return true;
 }
 
 void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     const llama_sampling_params & sparams = params.sparams;
 
-    printf("\n");
     printf("usage: %s [options]\n", argv[0]);
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
-    printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
-    printf("  -cml, --chatml        run in chatml mode (use with ChatML-compatible models)\n");
     printf("  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
     printf("  -r PROMPT, --reverse-prompt PROMPT\n");
     printf("                        halt generation at PROMPT, return control in interactive mode\n");
@@ -909,10 +661,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -t N, --threads N     number of threads to use during generation (default: %d)\n", params.n_threads);
     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("  -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");
     printf("                        prompt to start generation with (default: empty)\n");
     printf("  -e, --escape          process prompt escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
@@ -926,16 +674,11 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --in-suffix STRING    string to suffix after user inputs with (default: empty)\n");
     printf("  -f FNAME, --file FNAME\n");
     printf("                        prompt file to start generation.\n");
-    printf("  -bf FNAME, --binary-file FNAME\n");
-    printf("                        binary file containing multiple choice tasks.\n");
     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 \';\', for example: \"top_k;tfs;typical;top_p;min_p;temp\"\n");
-    printf("  --sampling-seq        simplified sequence for samplers that will be used (default: %s)\n", sparams.samplers_sequence.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);
-    printf("  --min-p N             min-p sampling (default: %.1f, 0.0 = disabled)\n", (double)sparams.min_p);
     printf("  --tfs N               tail free sampling, parameter z (default: %.1f, 1.0 = disabled)\n", (double)sparams.tfs_z);
     printf("  --typical N           locally typical sampling, parameter p (default: %.1f, 1.0 = disabled)\n", (double)sparams.typical_p);
     printf("  --repeat-last-n N     last n tokens to consider for penalize (default: %d, 0 = disabled, -1 = ctx_size)\n", sparams.penalty_last_n);
@@ -958,34 +701,22 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --cfg-negative-prompt-file FNAME\n");
     printf("                        negative prompt file to use for guidance. (default: empty)\n");
     printf("  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", sparams.cfg_scale);
-    printf("  --rope-scaling {none,linear,yarn}\n");
-    printf("                        RoPE frequency scaling method, defaults to linear unless specified by the model\n");
-    printf("  --rope-scale N        RoPE context scaling factor, expands context by a factor of N\n");
+    printf("  --rope-scale N        RoPE context linear scaling factor, inverse of --rope-freq-scale\n");
     printf("  --rope-freq-base N    RoPE base frequency, used by NTK-aware scaling (default: loaded from model)\n");
-    printf("  --rope-freq-scale N   RoPE frequency scaling factor, expands context by a factor of 1/N\n");
-    printf("  --yarn-orig-ctx N     YaRN: original context size of model (default: 0 = model training context size)\n");
-    printf("  --yarn-ext-factor N   YaRN: extrapolation mix factor (default: 1.0, 0.0 = full interpolation)\n");
-    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("  --rope-freq-scale N   RoPE frequency linear scaling factor (default: loaded from model)\n");
     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("  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
+    printf("                        not recommended: doubles context memory required and no measurable increase in quality\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("  --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("  --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");
     printf("  --image IMAGE_FILE    path to an image file. use with multimodal models\n");
@@ -1003,30 +734,16 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        number of layers to store in VRAM\n");
     printf("  -ngld N, --n-gpu-layers-draft N\n");
     printf("                        number of layers to store in VRAM for the draft model\n");
-    printf("  -sm SPLIT_MODE, --split-mode SPLIT_MODE\n");
-    printf("                        how to split the model across multiple GPUs, one of:\n");
-    printf("                          - none: use one GPU only\n");
-    printf("                          - layer (default): split layers and KV across GPUs\n");
-    printf("                          - row: split rows across GPUs\n");
-    printf("  -ts SPLIT, --tensor-split SPLIT\n");
-    printf("                        fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1\n");
-    printf("  -mg i, --main-gpu i   the GPU to use for the model (with split-mode = none),\n");
-    printf("                        or for intermediate results and KV (with split-mode = row) (default: %d)\n", params.main_gpu);
-#endif // LLAMA_SUPPORTS_GPU_OFFLOAD
-    printf("  --verbose-prompt      print a verbose prompt before generation (default: %s)\n", params.verbose_prompt ? "true" : "false");
-    printf("  --no-display-prompt   don't print prompt at generation (default: %s)\n", !params.display_prompt ? "true" : "false");
-    printf("  -gan N, --grp-attn-n N\n");
-    printf("                        group-attention factor (default: %d)\n", params.grp_attn_n);
-    printf("  -gaw N, --grp-attn-w N\n");
-    printf("                        group-attention width (default: %.1f)\n", (double)params.grp_attn_w);
-    printf("  -dkvc, --dump-kv-cache\n");
-    printf("                        verbose print of the KV cache\n");
-    printf("  -nkvo, --no-kv-offload\n");
-    printf("                        disable KV offload\n");
-    printf("  -ctk TYPE, --cache-type-k TYPE\n");
-    printf("                        KV cache data type for K (default: %s)\n", params.cache_type_k.c_str());
-    printf("  -ctv TYPE, --cache-type-v TYPE\n");
-    printf("                        KV cache data type for V (default: %s)\n", params.cache_type_v.c_str());
+    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
+    printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
+#ifdef GGML_USE_CUBLAS
+    printf("  -nommq, --no-mul-mat-q\n");
+    printf("                        use " GGML_CUBLAS_NAME " instead of custom mul_mat_q " GGML_CUDA_NAME " kernels.\n");
+    printf("                        Not recommended since this is both slower and uses more VRAM.\n");
+#endif // GGML_USE_CUBLAS
+#endif
+    printf("  --verbose-prompt      print prompt before generation\n");
     printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@@ -1034,18 +751,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding\n");
+    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
-    printf("  --override-kv KEY=TYPE:VALUE\n");
-    printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
-    printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
-    printf("  -ptc N, --print-token-count N\n");
-    printf("                        print token count every N tokens (default: %d)\n", params.n_print);
     printf("\n");
-#ifndef LOG_DISABLE_LOGS
-    log_print_usage();
-#endif // LOG_DISABLE_LOGS
 }
 
 std::string get_system_info(const gpt_params & params) {
@@ -1078,48 +787,6 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
     GGML_UNREACHABLE();
 }
 
-//
-// String parsing
-//
-
-std::string parse_samplers_input(std::string input) {
-    std::string output = "";
-    // since samplers names are written multiple ways
-    // make it ready for both system names and input names
-    std::unordered_map<std::string, char> samplers_symbols {
-        {"top_k",      'k'},
-        {"top-k",      'k'},
-        {"top_p",      'p'},
-        {"top-p",      'p'},
-        {"nucleus",    'p'},
-        {"typical_p",  'y'},
-        {"typical-p",  'y'},
-        {"typical",    'y'},
-        {"min_p",      'm'},
-        {"min-p",      'm'},
-        {"tfs_z",      'f'},
-        {"tfs-z",      'f'},
-        {"tfs",        'f'},
-        {"temp",       't'},
-        {"temperature",'t'}
-    };
-    // expected format example: "temp;top_k;tfs_z;typical_p;top_p;min_p"
-    size_t separator = input.find(';');
-    while (separator != input.npos) {
-        std::string name = input.substr(0,separator);
-        input = input.substr(separator+1);
-        separator = input.find(';');
-
-        if (samplers_symbols.find(name) != samplers_symbols.end()) {
-            output += samplers_symbols[name];
-        }
-    }
-    if (samplers_symbols.find(input) != samplers_symbols.end()) {
-        output += samplers_symbols[input];
-    }
-    return output;
-}
-
 //
 // Model utils
 //
@@ -1131,69 +798,27 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
         mparams.n_gpu_layers = params.n_gpu_layers;
     }
     mparams.main_gpu        = params.main_gpu;
-    mparams.split_mode      = params.split_mode;
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
-    if (params.kv_overrides.empty()) {
-        mparams.kv_overrides = NULL;
-    } else {
-        GGML_ASSERT(params.kv_overrides.back().key[0] == 0 && "KV overrides not terminated with empty key");
-        mparams.kv_overrides = params.kv_overrides.data();
-    }
 
     return mparams;
 }
 
-static ggml_type kv_cache_type_from_str(const std::string & s) {
-    if (s == "f32") {
-        return GGML_TYPE_F32;
-    }
-    if (s == "f16") {
-        return GGML_TYPE_F16;
-    }
-    if (s == "q8_0") {
-        return GGML_TYPE_Q8_0;
-    }
-    if (s == "q4_0") {
-        return GGML_TYPE_Q4_0;
-    }
-    if (s == "q4_1") {
-        return GGML_TYPE_Q4_1;
-    }
-    if (s == "q5_0") {
-        return GGML_TYPE_Q5_0;
-    }
-    if (s == "q5_1") {
-        return GGML_TYPE_Q5_1;
-    }
-
-    throw std::runtime_error("Invalid cache type: " + s);
-}
-
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params) {
     auto cparams = llama_context_default_params();
 
-    cparams.n_ctx             = params.n_ctx;
-    cparams.n_batch           = params.n_batch;
-    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.rope_scaling_type = params.rope_scaling_type;
-    cparams.rope_freq_base    = params.rope_freq_base;
-    cparams.rope_freq_scale   = params.rope_freq_scale;
-    cparams.yarn_ext_factor   = params.yarn_ext_factor;
-    cparams.yarn_attn_factor  = params.yarn_attn_factor;
-    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.offload_kqv       = !params.no_kv_offload;
-
-    cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
-    cparams.type_v = kv_cache_type_from_str(params.cache_type_v);
+    cparams.n_ctx           = params.n_ctx;
+    cparams.n_batch         = params.n_batch;
+    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.f16_kv          = params.memory_f16;
+    cparams.logits_all      = params.logits_all;
+    cparams.embedding       = params.embedding;
+    cparams.rope_freq_base  = params.rope_freq_base;
+    cparams.rope_freq_scale = params.rope_freq_scale;
 
     return cparams;
 }
@@ -1209,7 +834,7 @@ void llama_batch_add(
     const std::vector<llama_seq_id> & seq_ids,
                                bool   logits) {
     batch.token   [batch.n_tokens] = id;
-    batch.pos     [batch.n_tokens] = pos;
+    batch.pos     [batch.n_tokens] = pos,
     batch.n_seq_id[batch.n_tokens] = seq_ids.size();
     for (size_t i = 0; i < seq_ids.size(); ++i) {
         batch.seq_id[batch.n_tokens][i] = seq_ids[i];
@@ -1264,7 +889,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
         llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
-        llama_kv_cache_clear(lctx);
+        llama_kv_cache_tokens_rm(lctx, -1, -1);
         llama_reset_timings(lctx);
     }
 
@@ -1350,12 +975,6 @@ std::string llama_detokenize_bpe(llama_context * ctx, const std::vector<llama_to
     return result;
 }
 
-bool llama_should_add_bos_token(const llama_model * model) {
-    const int add_bos = llama_add_bos_token(model);
-
-    return add_bos != -1 ? bool(add_bos) : (llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM);
-}
-
 //
 // YAML utils
 //
@@ -1472,7 +1091,6 @@ void dump_string_yaml_multiline(FILE * stream, const char * prop_name, const cha
     if (!data_str.empty() && (std::isspace(data_str[0]) || std::isspace(data_str.back()))) {
         data_str = std::regex_replace(data_str, std::regex("\n"), "\\n");
         data_str = std::regex_replace(data_str, std::regex("\""), "\\\"");
-        data_str = std::regex_replace(data_str, std::regex(R"(\\[^n"])"), R"(\$&)");
         data_str = "\"" + data_str + "\"";
         fprintf(stream, "%s: %s\n", prop_name, data_str.c_str());
         return;
@@ -1510,15 +1128,15 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
     const llama_sampling_params & sparams = params.sparams;
 
-    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
-    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
+    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
+    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
     fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
     fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
-    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
     fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
     fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
+    fprintf(stream, "cpu_has_blas: %s\n",        ggml_cpu_has_blas()        ? "true" : "false");
     fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
     fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
     fprintf(stream, "cpu_has_fma: %s\n",         ggml_cpu_has_fma()         ? "true" : "false");
@@ -1606,6 +1224,7 @@ 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, "memory_f32: %s # default: false\n", !params.memory_f16 ? "true" : "false");
     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);
@@ -1656,82 +1275,6 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "threads: %d # default: %d\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);
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
-    fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");
-}
-
-//
-// KV cache utils
-//
-
-void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        int seq_count = 0;
-        for (int j = 0; j < view.n_max_seq; j++) {
-            if (cs_curr[j] >= 0) { seq_count++; }
-        }
-        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
-    }
-
-    printf("\n=== Done dumping\n");
-}
-
-void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
-    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
-
-    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
-        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
-
-    std::unordered_map<llama_seq_id, size_t> seqs;
-    llama_kv_cache_view_cell * c_curr = view.cells;
-    llama_seq_id * cs_curr = view.cells_sequences;
-
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
-        for (int j = 0; j < view.n_max_seq; j++) {
-            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();
-            }
-        }
-        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
-    }
-
-    printf("=== Sequence legend: ");
-    for (const auto & it : seqs) {
-        printf("%zu=%d, ", it.second, it.first);
-    }
-    printf("'+'=other sequence ids");
-
-    c_curr = view.cells;
-    cs_curr = view.cells_sequences;
-    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
-        if (i % row_size == 0) {
-            printf("\n%5d: ", i);
-        }
-        for (int j = 0; j < view.n_max_seq; j++) {
-            if (cs_curr[j] >= 0) {
-                const auto & it = seqs.find(cs_curr[j]);
-                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
-            } else {
-                putchar('.');
-            }
-        }
-        putchar(' ');
-    }
-
-    printf("\n=== Done dumping\n");
 }

common/common.h

@@ -9,7 +9,6 @@
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
-#include <cmath>
 #include <string>
 #include <vector>
 #include <random>
@@ -26,57 +25,35 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
 
-#define print_build_info() do {                                                                     \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
+#define print_build_info() do {                                                             \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
 } while(0)
 
-// build info
-extern int LLAMA_BUILD_NUMBER;
-extern char const *LLAMA_COMMIT;
-extern char const *LLAMA_COMPILER;
-extern char const *LLAMA_BUILD_TARGET;
-
 //
 // CLI argument parsing
 //
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                           = -1;    // RNG seed
-
+    uint32_t seed                           = -1;   // RNG seed
     int32_t n_threads                       = get_num_physical_cores();
-    int32_t n_threads_draft                 = -1;
-    int32_t n_threads_batch                 = -1;    // number of threads to use for batch processing (-1 = use n_threads)
-    int32_t n_threads_batch_draft           = -1;
-    int32_t n_predict                       = -1;    // new tokens to predict
-    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_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
-    int32_t main_gpu                        = 0;     // the GPU that is used for scratch and small tensors
-    float   tensor_split[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
-    int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
-    int32_t grp_attn_n                      = 1;     // group-attention factor
-    int32_t grp_attn_w                      = 512;   // group-attention width
-    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
-    float   rope_freq_base                  = 0.0f;  // RoPE base frequency
-    float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
-    float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
-    float   yarn_attn_factor                = 1.0f;  // YaRN magnitude scaling factor
-    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
-    int8_t  rope_scaling_type               = LLAMA_ROPE_SCALING_UNSPECIFIED; // TODO: better to be int32_t for alignment
-                                                                              //       pinging @cebtenzzre
+    int32_t n_threads_batch                 = -1;   // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_predict                       = -1;   // new tokens to predict
+    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                         = 16;   // 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
+    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)
+    int32_t main_gpu                        = 0;    // the GPU that is used for scratch and small tensors
+    float   tensor_split[LLAMA_MAX_DEVICES] = {0};  // how split tensors should be distributed across GPUs
+    int32_t n_beams                         = 0;    // if non-zero then use beam search of given width.
+    float   rope_freq_base                  = 0.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 0.0f; // RoPE frequency scaling factor
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -91,9 +68,6 @@ struct gpt_params {
     std::string input_suffix      = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
     std::string logdir            = "";  // directory in which to save YAML log files
-    std::string logits_file       = "";  // file for saving *all* logits
-
-    std::vector<llama_model_kv_override> kv_overrides;
 
     // TODO: avoid tuple, use struct
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
@@ -103,22 +77,14 @@ struct gpt_params {
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                     //                                       (which is more convenient to use for plotting)
                                     //
-    bool   hellaswag       = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
+    bool hellaswag         = false; // compute HellaSwag score over random tasks from datafile supplied in prompt
     size_t hellaswag_tasks = 400;   // number of tasks to use when computing the HellaSwag score
 
-    bool   winogrande      = false; // compute Winogrande score over random tasks from datafile supplied in prompt
-    size_t winogrande_tasks= 0;     // number of tasks to use when computing the Winogrande score. If 0, all tasks will be computed
-
-    bool   multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
-    size_t multiple_choice_tasks = 0;     // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed
-
-    bool   kl_divergence   = false; // compute KL-divergence
-
     bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
+    bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
-    bool chatml            = false; // chatml mode (used for models trained on chatml syntax)
     bool prompt_cache_all  = false; // save user input and generations to prompt cache
     bool prompt_cache_ro   = false; // open the prompt cache read-only and do not update it
 
@@ -137,21 +103,13 @@ struct gpt_params {
     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
-    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
-    bool no_kv_offload     = false; // disable KV offloading
-
-    std::string cache_type_k = "f16"; // KV cache data type for the K
-    std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
-    std::string image  = ""; // path to an image file
+    std::string image = ""; // path to an image file
 };
 
-bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
-
 bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
@@ -162,12 +120,6 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
-//
-// String parsing
-//
-
-std::string parse_samplers_input(std::string input);
-
 //
 // Model utils
 //
@@ -229,10 +181,6 @@ std::string llama_detokenize_bpe(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // YAML utils
 //
@@ -246,13 +194,3 @@ std::string get_sortable_timestamp();
 void dump_non_result_info_yaml(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
-
-//
-// KV cache utils
-//
-
-// Dump the KV cache view with the number of sequences per cell.
-void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
-
-// 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);

common/grammar-parser.cpp

@@ -190,7 +190,7 @@ namespace grammar_parser {
                 pos = parse_space(pos + 1, is_nested);
             } else if (*pos == '*' || *pos == '+' || *pos == '?') { // repetition operator
                 if (last_sym_start == out_elements.size()) {
-                    throw std::runtime_error(std::string("expecting preceding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceeding item to */+/? at ") + pos);
                 }
 
                 // apply transformation to previous symbol (last_sym_start to end) according to

common/log.h

@@ -61,13 +61,13 @@
 //  #define LOG_TARGET stderr
 //  #include "log.h"
 //
-//  The log target can also be redirected to a different function
+//  The log target can also be redirected to a diffrent function
 //  like so:
 //
-//  #define LOG_TARGET log_handler_different()
+//  #define LOG_TARGET log_handler_diffrent()
 //  #include "log.h"
 //
-//  FILE* log_handler_different()
+//  FILE* log_handler_diffrent()
 //  {
 //      return stderr;
 //  }
@@ -97,56 +97,38 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
-// Utility for synchronizing log configuration state
-//  since std::optional was introduced only in c++17
-enum LogTriState
-{
-    LogTriStateSame,
-    LogTriStateFalse,
-    LogTriStateTrue
-};
-
+// NOTE: currently disabled as it produces too many log files
 // Utility to obtain "pid" like unique process id and use it when creating log files.
-inline std::string log_get_pid()
-{
-   static std::string pid;
-   if (pid.empty())
-   {
-       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
-       //  it's not the same as "pid" but is unique enough to solve multiple instances
-       //  trying to write to the same log.
-       std::stringstream ss;
-       ss << std::this_thread::get_id();
-       pid = ss.str();
-   }
-
-   return pid;
-}
+//inline std::string log_get_pid()
+//{
+//    static std::string pid;
+//    if (pid.empty())
+//    {
+//        // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+//        //  it's not the same as "pid" but is unique enough to solve multiple instances
+//        //  trying to write to the same log.
+//        std::stringstream ss;
+//        ss << std::this_thread::get_id();
+//        pid = ss.str();
+//    }
+//
+//    return pid;
+//}
 
 // Utility function for generating log file names with unique id based on thread id.
 //  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
 //  where the number is a runtime id of the current thread.
 
-#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(log_file_basename, log_file_extension)
 
 // INTERNAL, DO NOT USE
-inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
+inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
 {
-    static bool _multilog = false;
-
-    if (multilog != LogTriStateSame)
-    {
-        _multilog = multilog == LogTriStateTrue;
-    }
-
     std::stringstream buf;
 
     buf << log_file_basename;
-    if (_multilog)
-    {
-        buf << ".";
-        buf << log_get_pid();
-    }
+    //buf << ".";
+    //buf << log_get_pid();
     buf << ".";
     buf << log_file_extension;
 
@@ -231,6 +213,15 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
     #define LOG_TEE_FLF_VAL ,""
 #endif
 
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
@@ -324,23 +315,16 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #endif
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
 {
-    static bool _initialized = false;
-    static bool _append = false;
-    static bool _disabled = filename.empty() && target == nullptr;
+    static bool _initialized{false};
+    static bool _disabled{(filename.empty() && target == nullptr)};
     static std::string log_current_filename{filename};
     static FILE *log_current_target{target};
     static FILE *logfile = nullptr;
 
     if (change)
     {
-        if (append != LogTriStateSame)
-        {
-            _append = append == LogTriStateTrue;
-            return logfile;
-        }
-
         if (disable == LogTriStateTrue)
         {
             // Disable primary target
@@ -393,7 +377,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriS
             }
         }
 
-        logfile = fopen(filename.c_str(), _append ? "a" : "w");
+        logfile = fopen(filename.c_str(), "w");
     }
 
     if (!logfile)
@@ -414,20 +398,20 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriS
 }
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
 {
-    return log_handler1_impl(change, append, disable, filename, target);
+    return log_handler1_impl(change, disable, filename, target);
 }
 
 // Disables logs entirely at runtime.
 //  Makes LOG() and LOG_TEE() produce no output,
-//  until enabled back.
+//  untill enabled back.
 #define log_disable() log_disable_impl()
 
 // INTERNAL, DO NOT USE
 inline FILE *log_disable_impl()
 {
-    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
+    return log_handler1_impl(true, LogTriStateTrue);
 }
 
 // Enables logs at runtime.
@@ -436,31 +420,19 @@ inline FILE *log_disable_impl()
 // INTERNAL, DO NOT USE
 inline FILE *log_enable_impl()
 {
-    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
+    return log_handler1_impl(true, LogTriStateFalse);
 }
 
 // Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
 #define log_set_target(target) log_set_target_impl(target)
 
 // INTERNAL, DO NOT USE
-inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
-inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
 
 // INTERNAL, DO NOT USE
 inline FILE *log_handler() { return log_handler1_impl(); }
 
-// Enable or disable creating separate log files for each run.
-//  can ONLY be invoked BEFORE first log use.
-#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
-// Enable or disable append mode for log file.
-//  can ONLY be invoked BEFORE first log use.
-#define log_append(enable) log_append_impl(enable)
-// INTERNAL, DO NOT USE
-inline FILE *log_append_impl(bool enable)
-{
-    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
-}
-
 inline void log_test()
 {
     log_disable();
@@ -522,18 +494,6 @@ inline bool log_param_single_parse(const std::string & param)
         return true;
     }
 
-    if (param == "--log-new")
-    {
-        log_multilog(true);
-        return true;
-    }
-
-    if (param == "--log-append")
-    {
-        log_append(true);
-        return true;
-    }
-
     return false;
 }
 
@@ -563,9 +523,7 @@ inline void log_print_usage()
     printf("  --log-disable         Disable trace logs\n");
     printf("  --log-enable          Enable trace logs\n");
     printf("  --log-file            Specify a log filename (without extension)\n");
-    printf("  --log-new             Create a separate new log file on start. "
-                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
-    printf("  --log-append          Don't truncate the old log file.\n");
+    printf("                        Log file will be tagged with unique ID and written as \"<name>.<ID>.log\"\n"); /*  */
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)

common/sampling.cpp

@@ -13,7 +13,6 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
         // will be empty (default) if there are parse errors
         if (result->parsed_grammar.rules.empty()) {
             fprintf(stderr, "%s: failed to parse grammar\n", __func__);
-            delete result;
             return nullptr;
         }
 
@@ -40,7 +39,6 @@ void llama_sampling_free(struct llama_sampling_context * ctx) {
 void llama_sampling_reset(llama_sampling_context * ctx) {
     if (ctx->grammar != NULL) {
         llama_grammar_free(ctx->grammar);
-        ctx->grammar = NULL;
     }
 
     if (!ctx->parsed_grammar.rules.empty()) {
@@ -91,86 +89,29 @@ std::string llama_sampling_print(const llama_sampling_params & params) {
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
-            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
+            params.top_k, params.tfs_z, params.top_p, params.typical_p, params.temp,
             params.mirostat, params.mirostat_eta, params.mirostat_tau);
 
     return std::string(result);
 }
 
-std::string llama_sampling_order_print(const llama_sampling_params & params) {
-    std::string result = "CFG -> Penalties ";
-    if (params.mirostat == 0) {
-        for (auto s : params.samplers_sequence) {
-            switch (s) {
-                case 'k': result += "-> top_k "; break;
-                case 'f': result += "-> tfs_z "; break;
-                case 'y': result += "-> typical_p "; break;
-                case 'p': result += "-> top_p "; break;
-                case 'm': result += "-> min_p "; break;
-                case 't': result += "-> temp "; break;
-                default : break;
-            }
-        }
-    } else {
-        result += "-> mirostat ";
-    }
-
-    return result;
-}
-
-// no reasons to expose this function in header
-static void sampler_queue(
-                   struct llama_context * ctx_main,
-            const llama_sampling_params & params,
-                 llama_token_data_array & cur_p,
-                                 size_t & min_keep) {
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
-    const float         temp              = params.temp;
-    const float         dynatemp_range    = params.dynatemp_range;
-    const float         dynatemp_exponent = params.dynatemp_exponent;
-    const int32_t       top_k             = params.top_k <= 0 ? n_vocab : params.top_k;
-    const float         top_p             = params.top_p;
-    const float         min_p             = params.min_p;
-    const float         tfs_z             = params.tfs_z;
-    const float         typical_p         = params.typical_p;
-    const std::string & samplers_sequence = params.samplers_sequence;
-
-    for (auto s : samplers_sequence) {
-        switch (s){
-            case 'k': llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
-            case 'f': llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
-            case 'y': llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
-            case 'p': llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
-            case 'm': llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case 't':
-                if (dynatemp_range > 0) {
-                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
-                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
-                    llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
-                } else {
-                    llama_sample_temp(ctx_main, &cur_p, temp);
-                }
-                break;
-            default : break;
-        }
-    }
-}
-
-static llama_token llama_sampling_sample_impl(
+llama_token llama_sampling_sample(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
+                  const int idx) {
     const llama_sampling_params & params = ctx_sampling->params;
 
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
     const float   temp            = params.temp;
+    const int32_t top_k           = params.top_k <= 0 ? n_vocab : params.top_k;
+    const float   top_p           = params.top_p;
+    const float   tfs_z           = params.tfs_z;
+    const float   typical_p       = params.typical_p;
     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;
@@ -185,27 +126,13 @@ static llama_token llama_sampling_sample_impl(
 
     llama_token id = 0;
 
-    // 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;
-
-    if (!is_resampling) {
-        // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
-        original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
-    }
-
     // 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++) {
@@ -214,15 +141,17 @@ static llama_token llama_sampling_sample_impl(
 
     llama_token_data_array cur_p = { cur.data(), cur.size(), false };
 
+    if (ctx_cfg) {
+        llama_sample_classifier_free_guidance(ctx_main, &cur_p, ctx_cfg, params.cfg_scale);
+    }
+
     // 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) {
+    if (!prev.empty()) {
         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);
+                prev.data() + prev.size() - penalty_last_n,
+                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
@@ -234,8 +163,7 @@ static llama_token llama_sampling_sample_impl(
         }
     }
 
-    // If we are in the resampling phase, apply grammar checks before sampling logic
-    if (is_resampling && ctx_sampling->grammar != NULL) {
+    if (ctx_sampling->grammar != NULL) {
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
@@ -258,7 +186,11 @@ static llama_token llama_sampling_sample_impl(
             // temperature sampling
             size_t min_keep = std::max(1, params.n_probs);
 
-            sampler_queue(ctx_main, params, cur_p, min_keep);
+            llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep);
+            llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep);
+            llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep);
+            llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep);
+            llama_sample_temp     (ctx_main, &cur_p, temp);
 
             id = llama_sample_token(ctx_main, &cur_p);
 
@@ -277,40 +209,9 @@ static llama_token llama_sampling_sample_impl(
         }
     }
 
-    if (ctx_sampling->grammar != NULL && !is_resampling) {
-        // Create an array with a single token data element for the sampled id
-        llama_token_data single_token_data = {id, logits[id], 0.0f};
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
-
-        // Apply grammar constraints to the single token
-        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
-
-        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
-        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-
-        // If the token is not valid according to the grammar, perform resampling
-        if (!is_valid) {
-            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
-
-            // Restore logits from the copy
-            std::copy(original_logits.begin(), original_logits.end(), logits);
-
-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
-        }
-    }
-
     return id;
 }
 
-llama_token llama_sampling_sample(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx) {
-    // Call the implementation function with is_resampling set to false by default
-    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
-}
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -10,25 +10,21 @@
 
 // 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     top_k                 = 40;       // <= 0 to use vocab size
-    float       top_p                 = 0.95f;    // 1.0 = disabled
-    float       min_p                 = 0.05f;    // 0.0 = disabled
-    float       tfs_z                 = 1.00f;    // 1.0 = disabled
-    float       typical_p             = 1.00f;    // 1.0 = disabled
-    float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float       dynatemp_range        = 0.00f;    // 0.0 = disabled
-    float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.10f;    // 1.0 = disabled
-    float       penalty_freq          = 0.00f;    // 0.0 = disabled
-    float       penalty_present       = 0.00f;    // 0.0 = disabled
-    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float       mirostat_tau          = 5.00f;    // target entropy
-    float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = true;     // consider newlines as a repeatable token
-    std::string samplers_sequence     = "kfypmt"; // top_k, tail_free, typical_p, top_p, min_p, temp
+    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 top_k             = 40;    // <= 0 to use vocab size
+    float   top_p             = 0.95f; // 1.0 = disabled
+    float   tfs_z             = 1.00f; // 1.0 = disabled
+    float   typical_p         = 1.00f; // 1.0 = disabled
+    float   temp              = 0.80f; // 1.0 = disabled
+    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat    = 1.10f; // 1.0 = disabled
+    float   penalty_freq      = 0.00f; // 0.0 = disabled
+    float   penalty_present   = 0.00f; // 0.0 = disabled
+    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau      = 5.00f; // target entropy
+    float   mirostat_eta      = 0.10f; // learning rate
+    bool    penalize_nl       = true;  // consider newlines as a repeatable token
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 
@@ -38,9 +34,6 @@ typedef struct llama_sampling_params {
     float       cfg_scale     = 1.f; // how strong is guidance
 
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
-
-    std::vector<llama_token> penalty_prompt_tokens;
-    bool                     use_penalty_prompt_tokens = false;
 } llama_sampling_params;
 
 // general sampler context
@@ -86,9 +79,6 @@ std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama
 // Print sampling parameters into a string
 std::string llama_sampling_print(const llama_sampling_params & params);
 
-// Print sampling order into a string
-std::string llama_sampling_order_print(const llama_sampling_params & params);
-
 // this is a common sampling function used across the examples for convenience
 // it can serve as a starting point for implementing your own sampling function
 // Note: When using multiple sequences, it is the caller's responsibility to call

common/train.cpp

@@ -32,7 +32,6 @@ 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.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
     return state;
@@ -71,7 +70,7 @@ void free_random_uniform_distribution(struct random_uniform_distribution * rnd)
 
 struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct random_normal_distribution * rnd) {
     float scale = 1.0f; // xavier
-    switch (ggml_n_dims(tensor)) {
+    switch (tensor->n_dims) {
         case 1:
             scale /= sqrtf((float) tensor->ne[0]);
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@@ -119,7 +118,7 @@ struct ggml_tensor * randomize_tensor_normal(struct ggml_tensor * tensor, struct
 }
 
 struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd) {
-    switch (ggml_n_dims(tensor)) {
+    switch (tensor->n_dims) {
         case 1:
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                 float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -183,27 +182,25 @@ float fclamp(const float v, const float min, const float max) {
 }
 
 void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0) {
+    GGML_ASSERT(tensor->n_dims == 1);
     GGML_ASSERT(tensor->ne[0] == ne0);
-    GGML_ASSERT(tensor->ne[1] == 1);
-    GGML_ASSERT(tensor->ne[2] == 1);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1) {
+    GGML_ASSERT(tensor->n_dims == 2);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
-    GGML_ASSERT(tensor->ne[2] == 1);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2) {
+    GGML_ASSERT(tensor->n_dims == 3);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
-    GGML_ASSERT(tensor->ne[3] == 1);
 }
 
 void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3) {
+    GGML_ASSERT(tensor->n_dims == 4);
     GGML_ASSERT(tensor->ne[0] == ne0);
     GGML_ASSERT(tensor->ne[1] == ne1);
     GGML_ASSERT(tensor->ne[2] == ne2);
@@ -227,8 +224,8 @@ int64_t get_example_targets_batch(
     bool                   sample_random_offsets
 ) {
     GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(ggml_is_matrix(tokens_input));
-    GGML_ASSERT(ggml_is_3d(target_probs));
+    GGML_ASSERT(tokens_input->n_dims  == 2);
+    GGML_ASSERT(target_probs->n_dims  == 3);
     int64_t n_vocab  = target_probs->ne[0];
     int64_t n_tokens = tokens_input->ne[0];
     int64_t n_batch  = tokens_input->ne[1];
@@ -1048,7 +1045,6 @@ struct train_params_common get_default_train_params_common() {
     params.n_batch    =    8;
     params.n_gradient_accumulation = 1;
     params.n_epochs   = -1;
-    params.n_gpu_layers = 0;
 
     params.custom_n_ctx = false;
 
@@ -1084,7 +1080,6 @@ struct train_params_common get_default_train_params_common() {
     params.adam_beta2          = 0.999f;
     params.adam_gclip          = 1.0f;
     params.adam_eps_f          = 0.0f;
-
     return params;
 }
 
@@ -1107,7 +1102,7 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train
     fprintf(stderr, "  --sample-start STR         Sets the starting point for samples after the specified pattern. If empty use every token position as sample start. (default '%s')\n", params->sample_start.c_str());
     fprintf(stderr, "  --include-sample-start     Include the sample start in the samples. (default off)\n");
     fprintf(stderr, "  --escape                   process sample start escapes sequences (\\n, \\r, \\t, \\', \\\", \\\\)\n");
-    fprintf(stderr, "  --overlapping-samples      Samples may overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
+    fprintf(stderr, "  --overlapping-samples      Samples my overlap, will include sample-start of second and following samples. When off, samples will end at begin of next sample. (default off)\n");
     fprintf(stderr, "  --fill-with-next-samples   Samples shorter than context length will be followed by the next (shuffled) samples. (default off)\n");
     fprintf(stderr, "  --separate-with-eos        When fill-with-next-samples, insert end-of-sequence token between samples.%s\n", params->separate_with_eos ? " (default)" : "");
     fprintf(stderr, "  --separate-with-bos        When fill-with-next-samples, insert begin-of-sequence token between samples.%s\n", params->separate_with_bos ? " (default)" : "");
@@ -1138,7 +1133,6 @@ void print_common_train_usage(int /*argc*/, char ** /*argv*/, const struct train
     fprintf(stderr, "  --adam-beta2 N             AdamW beta2 in interval [0,1). How much to smooth the second moment of gradients. (default %f)\n", params->adam_beta2);
     fprintf(stderr, "  --adam-gclip N             AdamW gradient clipping. Disabled when zero. (default %f)\n", params->adam_gclip);
     fprintf(stderr, "  --adam-epsf N              AdamW epsilon for convergence test. Disabled when <= zero. (default %f)\n", params->adam_eps_f);
-    fprintf(stderr, "  -ngl N, --n-gpu-layers N   Number of model layers to offload to GPU (default %d)", params->n_gpu_layers);
     fprintf(stderr, "\n");
 }
 
@@ -1358,17 +1352,6 @@ bool consume_common_train_arg(
             return true;
         }
         params->adam_gclip = std::stof(argv[i]);
-    } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
-            if (++i >= argc) {
-                *invalid_param = true;
-                return true;
-            }
-#ifdef LLAMA_SUPPORTS_GPU_OFFLOAD
-            params->n_gpu_layers = std::stoi(argv[i]);
-#else
-            fprintf(stderr, "warning: not compiled with GPU offload support, --n-gpu-layers option will be ignored\n");
-            fprintf(stderr, "warning: see main README.md for information on enabling GPU BLAS support\n");
-#endif
     } else if (arg == "-h" || arg == "--help") {
         params->print_usage = true;
         return true;

common/train.h

@@ -9,8 +9,6 @@
 #include "ggml.h"
 #include "llama.h"
 
-#define LLAMA_TRAIN_MAX_NODES 16384
-
 typedef std::string mt19937_state;
 
 struct train_state {
@@ -46,7 +44,6 @@ struct train_params_common {
     int n_batch;
     int n_gradient_accumulation;
     int n_epochs;
-    int n_gpu_layers;
 
     bool custom_n_ctx;
 

convert-baichuan-hf-to-gguf.py

@@ -0,0 +1,316 @@
+#!/usr/bin/env python3
+# HF baichuan --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import TYPE_CHECKING, Any
+import itertools
+import numpy as np
+import torch
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+if TYPE_CHECKING:
+    from typing import TypeAlias
+
+NDArray: TypeAlias = 'np.ndarray[Any, Any]'
+
+# reverse HF permute back to original pth layout
+
+
+def reverse_hf_permute(weights: NDArray, n_head: int, n_kv_head: int | None = None) -> NDArray:
+    if n_kv_head is not None and n_head != n_kv_head:
+        n_head //= n_kv_head
+
+    return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape))
+
+def reverse_hf_permute_part(weights: NDArray, n_part: int, n_head: int, n_head_kv: int| None = None) -> NDArray:
+        r = weights.shape[0] // 3
+        return (reverse_hf_permute(weights[r * n_part : r * n_part + r, ...], n_head, n_head_kv))
+
+def reverse_hf_part(weights: NDArray, n_part: int) -> NDArray:
+        r = weights.shape[0] // 3
+        return weights[r * n_part : r * n_part + r, ...]
+
+def count_model_parts(dir_model: str) -> int:
+    num_parts = 0
+
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+
+    return num_parts
+
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a HuggingFace LLaMA model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    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 (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+endianess = gguf.GGUFEndian.LITTLE
+if args.bigendian:
+    endianess = gguf.GGUFEndian.BIG
+endianess_str = "Big Endian" if args.bigendian else "Little Endian"
+print(f"gguf: Conversion Endianess {endianess}")
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+print("hello print: ",hparams["architectures"][0])
+if hparams["architectures"][0] != "BaichuanForCausalLM" and hparams["architectures"][0] != "BaiChuanForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+print(f"num_parts:{num_parts}\n")
+ARCH=gguf.MODEL_ARCH.BAICHUAN
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+head_count = hparams["num_attention_heads"]
+
+if "num_key_value_heads" in hparams:
+    head_count_kv = hparams["num_key_value_heads"]
+else:
+    head_count_kv = head_count
+
+if "_name_or_path" in hparams:
+    hf_repo = hparams["_name_or_path"]
+else:
+    hf_repo = ""
+
+if "max_sequence_length" in hparams:
+    ctx_length = hparams["max_sequence_length"]
+elif "max_position_embeddings" in hparams:
+    ctx_length = hparams["max_position_embeddings"]
+elif "model_max_length" in hparams:
+    ctx_length = hparams["model_max_length"]
+else:
+    print("gguf: can not find ctx length parameter.")
+
+    sys.exit()
+
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_source_hf_repo(hf_repo)
+gguf_writer.add_tensor_data_layout("Meta AI original pth")
+gguf_writer.add_context_length(ctx_length)
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+gguf_writer.add_head_count(head_count)
+gguf_writer.add_head_count_kv(head_count_kv)
+gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+
+if "rope_scaling" in hparams and hparams["rope_scaling"] != None and "factor" in hparams["rope_scaling"]:
+    if "type" in hparams["rope_scaling"]:
+        if hparams["rope_scaling"]["type"] == "linear":
+            gguf_writer.add_rope_scale_linear(hparams["rope_scaling"]["factor"])
+
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytes] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+tokenizer_model_file = dir_model / 'tokenizer.model'
+if not tokenizer_model_file.is_file():
+    print(f'Error: Missing {tokenizer_model_file}', file = sys.stderr)
+    sys.exit(1)
+
+# vocab type sentencepiece
+print("gguf: get sentencepiece tokenizer vocab, scores and token types")
+
+tokenizer = SentencePieceProcessor(str(tokenizer_model_file))
+vocab_size = hparams.get('vocab_size')
+if vocab_size is None:
+    vocab_size = tokenizer.vocab_size()
+
+for i in range(vocab_size):
+    text: bytes
+    score: float
+
+    piece = tokenizer.id_to_piece(i)
+    text = piece.encode("utf-8")
+    score = tokenizer.get_score(i)
+
+    toktype = 1  # defualt to normal token type
+    if tokenizer.is_unknown(i):
+        toktype = 2
+    if tokenizer.is_control(i):
+        toktype = 3
+
+    # toktype = 4 is user-defined = tokens from added_tokens.json
+
+    if tokenizer.is_unused(i):
+        toktype = 5
+    if tokenizer.is_byte(i):
+        toktype = 6
+
+    tokens.append(text)
+    scores.append(score)
+    toktypes.append(toktype)
+
+added_tokens_file = dir_model / 'added_tokens.json'
+if added_tokens_file.is_file():
+    with open(added_tokens_file, "r", encoding="utf-8") as f:
+        addtokens_json = json.load(f)
+
+        print("gguf: get added tokens")
+
+        for key in addtokens_json:
+            tokens.append( key.encode("utf-8") )
+            scores.append(-1000.0)
+            toktypes.append(4) # user-defined token type
+
+
+gguf_writer.add_tokenizer_model("llama")
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    tmp=model_part
+    for i in range(block_count):
+        if f"model.layers.{i}.self_attn.W_pack.weight" in model_part:
+            print(f"Unpacking and permuting layer {i}")
+            tmp[f"model.layers.{i}.self_attn.q_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],0,head_count,head_count)
+            tmp[f"model.layers.{i}.self_attn.k_proj.weight"]=reverse_hf_permute_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],1,head_count,head_count_kv)
+            tmp[f"model.layers.{i}.self_attn.v_proj.weight"]=reverse_hf_part(model_part[f"model.layers.{i}.self_attn.W_pack.weight"],2)
+            del tmp[f"model.layers.{i}.self_attn.W_pack.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+        # we don't need these
+        if name.endswith(".rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name + " -> " +  new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-bloom-hf-to-gguf.py

@@ -0,0 +1,247 @@
+#!/usr/bin/env python3
+# HF bloom --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import re
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+# Supported Models:
+#   https://huggingface.co/bigscience/bloom-1b7
+#   https://huggingface.co/bigscience/bloom-3b
+#   https://huggingface.co/bigscience/bloom-7b1
+#   https://huggingface.co/Langboat/bloom-1b4-zh
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Bloom model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    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 (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "BloomForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.BLOOM
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Bloom")
+n_embed = hparams.get("hidden_size", hparams.get("n_embed"))
+n_head = hparams.get("n_head", hparams.get("num_attention_heads"))
+gguf_writer.add_context_length(hparams.get("seq_length", n_embed))
+gguf_writer.add_embedding_length(n_embed)
+gguf_writer.add_feed_forward_length(4 * n_embed)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(n_head)
+gguf_writer.add_head_count_kv(n_head)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+added_vocab = tokenizer.get_added_vocab()
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+# params for qkv transform
+n_head_kv = hparams.get("n_head_kv", n_head)
+head_dim = n_embed // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    has_lm_head = True
+    if "lm_head.weight" not in model_part.keys() and "output.weight" not in model_part.keys():
+        has_lm_head = False
+
+    for original_name in model_part.keys():
+        data = model_part[original_name]
+        name = re.sub(r'transformer\.', '', original_name)
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        if re.match(r"h\.\d+\.self_attention\.query_key_value\.weight", name):
+            # Map bloom-style qkv_linear to gpt-style qkv_linear
+            # bloom: https://github.com/huggingface/transformers/blob/main/src/transformers/models/bloom/modeling_bloom.py#L238-L252  # noqa
+            # gpt-2: https://github.com/huggingface/transformers/blob/main/src/transformers/models/gpt2/modeling_gpt2.py#L312  # noqa
+            qkv_weights = data.reshape((n_head, 3, n_embed // n_head, n_embed))
+            data = np.concatenate(
+                (qkv_weights[:, 0, :, :].reshape((-1, n_embed)),
+                 qkv_weights[:, 1, :, :].reshape((-1, n_embed)),
+                 qkv_weights[:, 2, :, :].reshape((-1, n_embed))),
+                axis=0
+            )
+            print("re-format attention.linear_qkv.weight")
+        elif re.match(r"h\.\d+\.self_attention\.query_key_value\.bias", name):
+            qkv_bias = data.reshape((n_head, 3, n_embed // n_head))
+            data = np.concatenate(
+                (qkv_bias[:, 0, :].reshape((n_embed,)),
+                 qkv_bias[:, 1, :].reshape((n_embed,)),
+                 qkv_bias[:, 2, :].reshape((n_embed,))),
+                axis=0
+            )
+            print("re-format attention.linear_qkv.bias")
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+        if not has_lm_head and name == "word_embeddings.weight":
+            gguf_writer.add_tensor("output.weight", data)
+            print(name, "=>", "output.weight" + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))  # noqa
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-falcon-hf-to-gguf.py

@@ -0,0 +1,253 @@
+#!/usr/bin/env python3
+# HF falcon--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import contextlib
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path, prefix: str) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith(prefix):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a Falcon model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    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 (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] not in ("RWForCausalLM", "FalconForCausalLM"):
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model, "model-00")
+if num_parts:
+    is_safetensors = True
+    from safetensors import safe_open
+else:
+    is_safetensors = False
+    num_parts = count_model_parts(dir_model, "pytorch_model-")
+
+ARCH=gguf.MODEL_ARCH.FALCON
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams.get("num_hidden_layers")
+if block_count is None:
+    block_count = hparams["n_layer"]  # old name
+
+n_head = hparams.get("num_attention_heads")
+if n_head is None:
+    n_head = hparams["n_head"]  # old name
+
+n_head_kv = hparams.get("num_kv_heads")
+if n_head_kv is None:
+    n_head_kv = hparams.get("n_head_kv", 1)  # old name
+
+gguf_writer.add_name("Falcon")
+gguf_writer.add_context_length(2048) # not in config.json
+gguf_writer.add_tensor_data_layout("jploski") # qkv tensor transform
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_feed_forward_length(4 * hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(n_head)
+gguf_writer.add_head_count_kv(n_head_kv)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    tokens.append(reverse_vocab[i])
+    scores.append(0.0) # dummy
+    toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_scores(scores)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+head_dim = hparams["hidden_size"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+elif is_safetensors:
+    part_names = (
+        f"model-{n:05}-of-{num_parts:05}.safetensors" for n in range(1, num_parts + 1)
+    )
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    if is_safetensors:
+        ctx = safe_open(dir_model / part_name, framework="pt", device="cpu")
+    else:
+        ctx = contextlib.nullcontext(torch.load(dir_model / part_name, map_location="cpu"))
+
+    with ctx as model_part:
+        for name in model_part.keys():
+            data = model_part.get_tensor(name) if is_safetensors else model_part[name]
+
+            old_dtype = data.dtype
+
+            # convert any unsupported data types to float32
+            if data.dtype != torch.float16 and data.dtype != torch.float32:
+                data = data.to(torch.float32)
+
+            # QKV tensor transform
+            # The original query_key_value tensor contains n_head_kv "kv groups",
+            # each consisting of n_head/n_head_kv query weights followed by one key
+            # and one value weight (shared by all query heads in the kv group).
+            # This layout makes it a big pain to work with in GGML.
+            # So we rearrange them here,, so that we have n_head query weights
+            # followed by n_head_kv key weights followed by n_head_kv value weights,
+            # in contiguous fashion.
+            # ref: https://github.com/jploski/ggml/blob/falcon40b/examples/falcon/convert-hf-to-ggml.py
+
+            if "query_key_value" in name:
+                qkv = data.view(n_head_kv, n_head // n_head_kv + 2, head_dim, head_dim * n_head)
+                q = qkv[:, :-2 ].reshape(n_head * head_dim, head_dim * n_head)
+                k = qkv[:, [-2]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                v = qkv[:, [-1]].reshape(n_head_kv * head_dim, head_dim * n_head)
+                data = torch.cat((q,k,v)).reshape_as(data)
+
+            data = data.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+            if new_name is None:
+                print("Can not map tensor '" + name + "'")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+            gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-gptneox-hf-to-gguf.py

@@ -0,0 +1,221 @@
+#!/usr/bin/env python3
+# HF gptneox--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a GPT-NeoX model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    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 (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTNeoXForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.GPTNEOX
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["num_hidden_layers"]
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_context_length(hparams["max_position_embeddings"])
+gguf_writer.add_embedding_length(hparams["hidden_size"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+gguf_writer.add_rope_dimension_count(int(hparams["rotary_pct"]*(hparams["hidden_size"]//hparams["num_attention_heads"])))
+gguf_writer.add_head_count(hparams["num_attention_heads"])
+gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_eps"])
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+added_vocab = tokenizer.get_added_vocab()
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        # we don't need these
+        if name.endswith(".attention.masked_bias") or name.endswith(".attention.bias") or name.endswith(".attention.rotary_emb.inv_freq"):
+            continue
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-llama-ggml-to-gguf.py

@@ -2,7 +2,7 @@
 from __future__ import annotations
 
 import argparse
-import os
+import math
 import struct
 import sys
 from enum import IntEnum
@@ -10,17 +10,36 @@ from pathlib import Path
 
 import numpy as np
 
+import os
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
+# Note: Does not support GGML_QKK_64
+QK_K = 256
+# Items here are (block size, type size)
+GGML_QUANT_SIZES = {
+    gguf.GGMLQuantizationType.F32  : (1, 4),
+    gguf.GGMLQuantizationType.F16  : (1, 2),
+    gguf.GGMLQuantizationType.Q4_0 : (32, 2 + 16),
+    gguf.GGMLQuantizationType.Q4_1 : (32, 2 + 2 + 16),
+    gguf.GGMLQuantizationType.Q5_0 : (32, 2 + 4 + 16),
+    gguf.GGMLQuantizationType.Q5_1 : (32, 2 + 2 + 4 + 16),
+    gguf.GGMLQuantizationType.Q8_0 : (32, 2 + 32),
+    gguf.GGMLQuantizationType.Q8_1 : (32, 4 + 4 + 32),
+    gguf.GGMLQuantizationType.Q2_K : (256, 2 + 2 + QK_K // 16 + QK_K // 4),
+    gguf.GGMLQuantizationType.Q3_K : (256, 2 + QK_K // 4 + QK_K // 8 + 12),
+    gguf.GGMLQuantizationType.Q4_K : (256, 2 + 2 + QK_K // 2 + 12),
+    gguf.GGMLQuantizationType.Q5_K : (256, 2 + 2 + QK_K // 2 + QK_K // 8 + 12),
+    gguf.GGMLQuantizationType.Q6_K : (256, 2 + QK_K // 2 + QK_K // 4 + QK_K // 16),
+    gguf.GGMLQuantizationType.Q8_K : (256, 4 + QK_K + QK_K // 8),
+}
 
 class GGMLFormat(IntEnum):
     GGML = 0
     GGMF = 1
     GGJT = 2
 
-
 class GGMLFType(IntEnum):
     ALL_F32              = 0
     MOSTLY_F16           = 1
@@ -40,7 +59,6 @@ class GGMLFType(IntEnum):
     MOSTLY_Q5_K_M        = 17
     MOSTLY_Q6_K          = 18
 
-
 class Hyperparameters:
     def __init__(self):
         self.n_vocab = self.n_embd = self.n_mult = self.n_head = 0
@@ -72,7 +90,6 @@ class Hyperparameters:
     def __str__(self):
         return f'<Hyperparameters: n_vocab={self.n_vocab}, n_embd={self.n_embd}, n_mult={self.n_mult}, n_head={self.n_head}, n_layer={self.n_layer}, n_rot={self.n_rot}, n_ff={self.n_ff}, ftype={self.ftype.name}>'
 
-
 class Vocab:
     def __init__(self, load_scores = True):
         self.items = []
@@ -94,7 +111,6 @@ class Vocab:
             self.items.append((item_text, item_score))
         return offset - orig_offset
 
-
 class Tensor:
     def __init__(self, use_padding = True):
         self.name = None
@@ -109,7 +125,7 @@ class Tensor:
         (n_dims, name_len, dtype) = struct.unpack('<3I', data[offset:offset + 12])
         assert n_dims >= 0 and n_dims <= 4, f'Invalid tensor dimensions {n_dims}'
         assert name_len < 4096, 'Absurd tensor name length'
-        quant = gguf.GGML_QUANT_SIZES.get(dtype)
+        quant = GGML_QUANT_SIZES.get(dtype)
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
@@ -128,7 +144,6 @@ class Tensor:
         # print(n_dims, name_len, dtype, self.dims, self.name, pad)
         return offset - orig_offset
 
-
 class GGMLModel:
     def __init__(self):
         self.hyperparameters = None
@@ -165,8 +180,8 @@ class GGMLModel:
             if ftype not in (GGMLFType.ALL_F32, GGMLFType.MOSTLY_F16):
                 err = 'Quantizations changed in GGJTv2. Can only convert unquantized GGML files older than GGJTv2.'
         elif (self.file_format == GGMLFormat.GGJT and self.format_version == 2):
-            if ftype in (GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
-                         GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
+            if ftype in ( GGMLFType.MOSTLY_Q4_0, GGMLFType.MOSTLY_Q4_1,
+                          GGMLFType.MOSTLY_Q4_1_SOME_F16, GGMLFType.MOSTLY_Q8_0):
                 err = 'Q4 and Q8 quantizations changed in GGJTv3.'
         if len(err) > 0:
             raise ValueError(f'{err} Sorry, your {self.file_format.name}v{self.format_version} file of type {ftype.name} is not eligible for conversion.')
@@ -193,7 +208,6 @@ class GGMLModel:
         hp.set_n_ff(self)
         return offset
 
-
 class GGMLToGGUF:
     def __init__(self, ggml_model, data, cfg, params_override = None, vocab_override = None, special_vocab = None):
         hp = ggml_model.hyperparameters
@@ -224,7 +238,7 @@ class GGMLToGGUF:
         gguf_writer = gguf.GGUFWriter(
             self.cfg.output,
             gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA],
-            use_temp_file = False)
+            use_temp_file = False )
         self.add_params(gguf_writer)
         self.add_vocab(gguf_writer)
         if self.special_vocab is not None:
@@ -348,8 +362,7 @@ class GGMLToGGUF:
                 mapped_name,
                 data[tensor.start_offset:tensor.start_offset + tensor.len_bytes],
                 raw_shape = tempdims,
-                raw_dtype = tensor.dtype)
-
+                raw_dtype = tensor.dtype )
 
 def handle_metadata(cfg, hp):
     import convert
@@ -371,38 +384,40 @@ def handle_metadata(cfg, hp):
         params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
     else:
         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 = convert.load_vocab(
+        cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir,
+        cfg.vocabtype )
+    # FIXME: Respect cfg.vocab_dir?
+    svocab = gguf.SpecialVocab(cfg.model_metadata_dir,
+        load_merges = cfg.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
     convert.check_vocab_size(params, vocab)
-    return params, vocab, special_vocab
-
+    return (params, vocab, svocab)
 
 def handle_args():
     parser = argparse.ArgumentParser(description = 'Convert GGML models to GGUF')
     parser.add_argument('--input', '-i', type = Path, required = True,
-                        help = 'Input GGMLv3 filename')
+        help = 'Input GGMLv3 filename')
     parser.add_argument('--output', '-o', type = Path, required = True,
-                        help ='Output GGUF filename')
+        help ='Output GGUF filename')
     parser.add_argument('--name',
-                        help = 'Set model name')
+        help = 'Set model name')
     parser.add_argument('--desc',
-                        help = 'Set model description')
+        help = 'Set model description')
     parser.add_argument('--gqa', type = int, default = 1,
-                        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
+        help = 'grouped-query attention factor (use 8 for LLaMA2 70B)')
     parser.add_argument('--eps', default = '5.0e-06',
-                        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
+        help = 'RMS norm eps: Use 1e-6 for LLaMA1 and OpenLLaMA, use 1e-5 for LLaMA2')
     parser.add_argument('--context-length', '-c', type=int, default = 2048,
-                        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
+        help = 'Default max context length: LLaMA1 is typically 2048, LLaMA2 is typically 4096')
     parser.add_argument('--model-metadata-dir', '-m', type = Path,
-                        help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
+        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")
+        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)")
+        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
-
 def main():
     cfg = handle_args()
     print(f'* Using config: {cfg}')
@@ -412,7 +427,7 @@ def main():
     data = np.memmap(cfg.input, mode = 'r')
     model = GGMLModel()
     print('* Scanning GGML input file')
-    offset = model.load(data, 0)  # noqa
+    offset = model.load(data, 0)
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
@@ -427,15 +442,12 @@ def main():
         print('\n=== WARNING === Special tokens may not be converted correctly. Use --model-metadata-dir if possible === WARNING ===\n')
         if model.file_format == GGMLFormat.GGML:
             print('! This is a very old GGML file that does not contain vocab scores. Strongly recommend using model metadata!')
-    converter = GGMLToGGUF(
-        model, data, cfg,
+    converter = GGMLToGGUF(model, data, cfg,
         params_override = params_override,
         vocab_override = vocab_override,
-        special_vocab = special_vocab
-    )
+        special_vocab = special_vocab )
     converter.save()
     print(f'* Successful completion. Output saved to: {cfg.output}')
 
-
 if __name__ == '__main__':
     main()

convert-lora-to-ggml.py

@@ -3,21 +3,51 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import struct
 import sys
-from pathlib import Path
 from typing import Any, BinaryIO, Sequence
 
 import numpy as np
 import torch
 
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
-import gguf
-
 NUMPY_TYPE_TO_FTYPE: dict[str, int] = {"float32": 0, "float16": 1}
 
 
+HF_SUBLAYER_TO_GGML = {
+    "self_attn.q_proj": "attn_q",
+    "self_attn.k_proj": "attn_k",
+    "self_attn.v_proj": "attn_v",
+    "self_attn.o_proj": "attn_output",
+    "mlp.gate_proj": "ffn_gate",
+    "mlp.down_proj": "ffn_down",
+    "mlp.up_proj": "ffn_up",
+    "input_layernorm": "attn_norm",
+    "post_attention_layernorm": "ffn_norm",
+}
+
+
+def translate_tensor_name(t: str) -> str:
+    match = re.match(r".*layers\.(\d+)\.(\w+\.\w+)\.lora_(A|B)\.weight", t)
+    if match:
+        nn = match.group(1)
+        sub_layer = match.group(2)
+        lora_type = match.group(3)
+
+        sub_layer_renamed = HF_SUBLAYER_TO_GGML.get(sub_layer)
+        if sub_layer_renamed is None:
+            print(f"Error: unrecognized sub-layer {sub_layer} in tensor {t}")
+            sys.exit(1)
+
+        output_string = (
+            f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
+        )
+        return output_string
+    else:
+        print(f"Error: unrecognized tensor {t}")
+        sys.exit(1)
+
+
 def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(b"ggla"[::-1])  # magic (ggml lora)
     fout.write(struct.pack("i", 1))  # file version
@@ -31,7 +61,9 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
-def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
+def write_tensor_header(
+    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
+) -> None:
     sname = name.encode("utf-8")
     fout.write(
         struct.pack(
@@ -46,103 +78,60 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty
     fout.seek((fout.tell() + 31) & -32)
 
 
-if __name__ == '__main__':
-    if len(sys.argv) < 2:
-        print(f"Usage: python {sys.argv[0]} <path> [arch]")
-        print(
-            "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
-        )
-        print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-        sys.exit(1)
+if len(sys.argv) != 2:
+    print(f"Usage: python {sys.argv[0]} <path>")
+    print(
+        "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
+    )
+    sys.exit(1)
 
-    input_json = os.path.join(sys.argv[1], "adapter_config.json")
-    input_model = os.path.join(sys.argv[1], "adapter_model.bin")
-    output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
+input_json = os.path.join(sys.argv[1], "adapter_config.json")
+input_model = os.path.join(sys.argv[1], "adapter_model.bin")
+output_path = os.path.join(sys.argv[1], "ggml-adapter-model.bin")
 
-    if os.path.exists(input_model):
-        model = torch.load(input_model, map_location="cpu")
-    else:
-        input_model = os.path.join(sys.argv[1], "adapter_model.safetensors")
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-        model = load_file(input_model, device="cpu")
+model = torch.load(input_model, map_location="cpu")
 
-    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+with open(input_json, "r") as f:
+    params = json.load(f)
 
-    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-        print(f"Error: unsupported architecture {arch_name}")
-        sys.exit(1)
+if params["peft_type"] != "LORA":
+    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+    sys.exit(1)
 
-    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
+if params["fan_in_fan_out"] is True:
+    print("Error: param fan_in_fan_out is not supported")
+    sys.exit(1)
 
-    with open(input_json, "r") as f:
-        params = json.load(f)
+if params["bias"] is not None and params["bias"] != "none":
+    print("Error: param bias is not supported")
+    sys.exit(1)
 
-    if params["peft_type"] != "LORA":
-        print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-        sys.exit(1)
+# TODO: these seem to be layers that have been trained but without lora.
+# doesn't seem widely used but eventually should be supported
+if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
+    print("Error: param modules_to_save is not supported")
+    sys.exit(1)
 
-    if params["fan_in_fan_out"] is True:
-        print("Error: param fan_in_fan_out is not supported")
-        sys.exit(1)
+with open(output_path, "wb") as fout:
+    fout.truncate()
 
-    if params["bias"] is not None and params["bias"] != "none":
-        print("Error: param bias is not supported")
-        sys.exit(1)
-
-    # TODO: these seem to be layers that have been trained but without lora.
-    # doesn't seem widely used but eventually should be supported
-    if params["modules_to_save"] is not None and len(params["modules_to_save"]) > 0:
-        print("Error: param modules_to_save is not supported")
-        sys.exit(1)
-
-    with open(output_path, "wb") as fout:
-        fout.truncate()
-
-        write_file_header(fout, params)
-        for k, v in model.items():
-            orig_k = k
-            if k.endswith(".default.weight"):
-                k = k.replace(".default.weight", ".weight")
-            if k in ["llama_proj.weight", "llama_proj.bias"]:
-                continue
-            if k.endswith("lora_A.weight"):
-                if v.dtype != torch.float16 and v.dtype != torch.float32:
-                    v = v.float()
-                v = v.T
-            else:
+    write_file_header(fout, params)
+    for k, v in model.items():
+        if k.endswith(".default.weight"):
+            k = k.replace(".default.weight", ".weight")
+        if k in ["llama_proj.weight", "llama_proj.bias"]:
+            continue
+        if k.endswith("lora_A.weight"):
+            if v.dtype != torch.float16 and v.dtype != torch.float32:
                 v = v.float()
+            v = v.T
+        else:
+            v = v.float()
 
-            t = v.detach().numpy()
+        t = v.detach().numpy()
+        tname = translate_tensor_name(k)
+        print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+        write_tensor_header(fout, tname, t.shape, t.dtype)
+        t.tofile(fout)
 
-            prefix = "base_model.model."
-            if k.startswith(prefix):
-                k = k[len(prefix) :]
-
-            lora_suffixes = (".lora_A.weight", ".lora_B.weight")
-            if k.endswith(lora_suffixes):
-                suffix = k[-len(lora_suffixes[0]):]
-                k = k[: -len(lora_suffixes[0])]
-            else:
-                print(f"Error: unrecognized tensor name {orig_k}")
-                sys.exit(1)
-
-            tname = name_map.get_name(k)
-            if tname is None:
-                print(f"Error: could not map tensor name {orig_k}")
-                print(" Note: the arch parameter must be specified if the model is not llama")
-                sys.exit(1)
-
-            if suffix == ".lora_A.weight":
-                tname += ".weight.loraA"
-            elif suffix == ".lora_B.weight":
-                tname += ".weight.loraB"
-            else:
-                assert False
-
-            print(f"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
-            write_tensor_header(fout, tname, t.shape, t.dtype)
-            t.tofile(fout)
-
-    print(f"Converted {input_json} and {input_model} to {output_path}")
+print(f"Converted {input_json} and {input_model} to {output_path}")

convert-mpt-hf-to-gguf.py

@@ -0,0 +1,227 @@
+#!/usr/bin/env python3
+# HF mpt--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert an MPT model to a GGML compatible file")
+    parser.add_argument(
+        "--vocab-only", action="store_true",
+        help="extract only the vocab",
+    )
+    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 (*.bin)",
+    )
+    parser.add_argument(
+        "ftype", type=int, choices=[0, 1], default=1, nargs='?',
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "MPTForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit()
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.MPT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layers"]
+
+gguf_writer.add_name(dir_model.name)
+gguf_writer.add_context_length(hparams["max_seq_len"])
+gguf_writer.add_embedding_length(hparams["d_model"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_feed_forward_length(4 * hparams["d_model"])
+gguf_writer.add_head_count(hparams["n_heads"])
+if kv_n_heads := hparams["attn_config"].get("kv_n_heads"):
+    gguf_writer.add_head_count_kv(kv_n_heads)
+gguf_writer.add_layer_norm_eps(1e-05)
+if hparams["attn_config"]["clip_qkv"] is not None:
+    gguf_writer.add_clamp_kqv(hparams["attn_config"]["clip_qkv"])
+gguf_writer.add_max_alibi_bias(hparams["attn_config"]["alibi_bias_max"])
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# MPT token embedding tensors have dimension 50432 (hparams["vocab_size"]), but
+# there are only 50254 (len(tokenizer.vocab)) tokens in the vocab, presumably to
+# accomodate some "reserved" tokens; this is causing problems down the line in
+# llama.cpp, so we pad the vocab with dummy tokens:
+
+vocab_size = hparams["vocab_size"]
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+added_vocab = tokenizer.get_added_vocab()
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(f"{dir_model}/{part_name}", map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Cannot map tensor '" + name + "'")
+            continue # for the sake of compatibility with some old published models, don't quit
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(new_name + ", n_dims = " + str(n_dims) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        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":
+            gguf_writer.add_tensor("output.weight", data)
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-persimmon-to-gguf.py

@@ -1,18 +1,14 @@
-#!/usr/bin/env python3
-import argparse
-import os
-import sys
-from pathlib import Path
-from pprint import pprint
-
 import torch
+import os
+from pprint import pprint
+import sys
+import argparse
+from pathlib import Path
 from sentencepiece import SentencePieceProcessor
-
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
-
 def _flatten_dict(dct, tensors, prefix=None):
     assert isinstance(dct, dict)
     for key in dct.keys():
@@ -25,7 +21,6 @@ def _flatten_dict(dct, tensors, prefix=None):
             raise ValueError(type(dct[key]))
     return None
 
-
 def _get_sentencepiece_tokenizer_info(dir_model: Path):
     tokenizer_path = dir_model / 'adept_vocab.model'
     print('gguf: getting sentencepiece tokenizer from', tokenizer_path)
@@ -59,7 +54,6 @@ def _get_sentencepiece_tokenizer_info(dir_model: Path):
         pass
     return tokens, scores, toktypes
 
-
 def main():
     parser = argparse.ArgumentParser(description="Convert a Persimmon model from Adept (e.g. Persimmon 8b chat) to a GGML compatible file")
     parser.add_argument("--outfile",             type=Path, help="path to write to; default: based on input")
@@ -71,7 +65,7 @@ def main():
     persimmon_model = torch.load(args.ckpt_path)
     hparams = persimmon_model['args']
     pprint(hparams)
-    tensors: dict[str, torch.Tensor] = {}
+    tensors = {}
     _flatten_dict(persimmon_model['model'], tensors, None)
 
     arch = gguf.MODEL_ARCH.PERSIMMON
@@ -131,5 +125,6 @@ def main():
     print("")
 
 
+
 if __name__ == '__main__':
     main()

convert-refact-hf-to-gguf.py

@@ -0,0 +1,272 @@
+#!/usr/bin/env python3
+# HF refact--> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import sys
+from pathlib import Path
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if "NO_LOCAL_GGUF" not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / "gguf-py" / "gguf"))
+import gguf
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Convert a Refact model to a GGML compatible file"
+    )
+    parser.add_argument(
+        "--vocab-only",
+        action="store_true",
+        help="extract only the vocab",
+    )
+    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 (*.bin)",
+    )
+    parser.add_argument(
+        "ftype",
+        type=int,
+        choices=[0, 1],
+        default=1,
+        nargs="?",
+        help="output format - use 0 for float32, 1 for float16",
+    )
+    return parser.parse_args()
+
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f"Error: {args.model} is not a directory", file=sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f"ggml-model-{ftype_str[ftype]}.gguf"
+
+print("gguf: loading model " + dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTRefactForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH = gguf.MODEL_ARCH.REFACT
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+# Get refact feed forward dimension
+hidden_dim = hparams["n_embd"]
+inner_dim = 4 * hidden_dim
+hidden_dim = int(2 * inner_dim / 3)
+multiple_of = 256
+ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("Refact")
+# refact uses Alibi. So this is from config.json which might be used by training.
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+
+gguf_writer.add_feed_forward_length(ff_dim)
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_rms_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+added_vocab = tokenizer.get_added_vocab()
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_types(toktypes)
+
+special_vocab = gguf.SpecialVocab(dir_model, load_merges=True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH, block_count)
+
+# params for qkv transform
+n_head = hparams["n_head"]
+n_head_kv = 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for i in range(block_count):
+        if f"transformer.h.{i}.attn.kv.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.attn.kv.weight"]
+            model_part[f"model.layers.{i}.self_attn.k_proj.weight"] = data[
+                : n_head_kv * head_dim
+            ]
+            model_part[f"model.layers.{i}.self_attn.v_proj.weight"] = data[
+                n_head_kv * head_dim :
+            ]
+            del model_part[f"transformer.h.{i}.attn.kv.weight"]
+        if f"transformer.h.{i}.attn.q.weight" in model_part:
+            model_part[f"model.layers.{i}.self_attn.q_proj.weight"] = model_part[
+                f"transformer.h.{i}.attn.q.weight"
+            ]
+            del model_part[f"transformer.h.{i}.attn.q.weight"]
+        if f"transformer.h.{i}.mlp.gate_up_proj.weight" in model_part:
+            data = model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+            model_part[f"model.layers.{i}.mlp.gate_proj.weight"] = data[:ff_dim]
+            model_part[f"model.layers.{i}.mlp.up_proj.weight"] = data[ff_dim:]
+            del model_part[f"transformer.h.{i}.mlp.gate_up_proj.weight"]
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes=(".weight",))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if (
+            ftype == 1
+            and data_dtype == np.float32
+            and name.endswith(".weight")
+            and n_dims == 2
+        ):
+            data = data.astype(np.float16)
+
+        print(
+            new_name
+            + ", n_dims = "
+            + str(n_dims)
+            + ", "
+            + str(old_dtype)
+            + " --> "
+            + str(data.dtype)
+        )
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert-starcoder-hf-to-gguf.py

@@ -0,0 +1,210 @@
+#!/usr/bin/env python3
+# HF starcoder --> gguf conversion
+
+from __future__ import annotations
+
+import argparse
+import json
+import os
+import struct
+import sys
+from pathlib import Path
+from typing import Any
+
+import numpy as np
+import torch
+from transformers import AutoTokenizer  # type: ignore[import]
+
+if 'NO_LOCAL_GGUF' not in os.environ:
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+import gguf
+
+
+def count_model_parts(dir_model: Path) -> int:
+    num_parts = 0
+    for filename in os.listdir(dir_model):
+        if filename.startswith("pytorch_model-"):
+            num_parts += 1
+
+    if num_parts > 0:
+        print("gguf: found " + str(num_parts) + " model parts")
+    return num_parts
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(description="Convert a StarCoder model to a GGML compatible file")
+    parser.add_argument("--vocab-only", action="store_true", help="extract only the vocab")
+    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 (*.bin)")
+    parser.add_argument("ftype",        type=int,            help="output format - use 0 for float32, 1 for float16", choices=[0, 1], default = 1)
+    return parser.parse_args()
+
+args = parse_args()
+
+dir_model = args.model
+ftype = args.ftype
+if not dir_model.is_dir():
+    print(f'Error: {args.model} is not a directory', file = sys.stderr)
+    sys.exit(1)
+
+# possible tensor data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+if args.outfile is not None:
+    fname_out = args.outfile
+else:
+    # output in the same directory as the model by default
+    fname_out = dir_model / f'ggml-model-{ftype_str[ftype]}.gguf'
+
+print("gguf: loading model "+dir_model.name)
+
+with open(dir_model / "config.json", "r", encoding="utf-8") as f:
+    hparams = json.load(f)
+
+if hparams["architectures"][0] != "GPTBigCodeForCausalLM":
+    print("Model architecture not supported: " + hparams["architectures"][0])
+
+    sys.exit(1)
+
+# get number of model parts
+num_parts = count_model_parts(dir_model)
+
+ARCH=gguf.MODEL_ARCH.STARCODER
+gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH])
+
+print("gguf: get model metadata")
+
+block_count = hparams["n_layer"]
+
+gguf_writer.add_name("StarCoder")
+gguf_writer.add_context_length(hparams["n_positions"])
+gguf_writer.add_embedding_length(hparams["n_embd"])
+gguf_writer.add_feed_forward_length(4 * hparams["n_embd"])
+gguf_writer.add_block_count(block_count)
+gguf_writer.add_head_count(hparams["n_head"])
+gguf_writer.add_head_count_kv(1)
+gguf_writer.add_layer_norm_eps(hparams["layer_norm_epsilon"])
+gguf_writer.add_file_type(ftype)
+
+# TOKENIZATION
+
+print("gguf: get tokenizer metadata")
+
+tokens: list[bytearray] = []
+scores: list[float] = []
+toktypes: list[int] = []
+
+# gpt2 tokenizer
+gguf_writer.add_tokenizer_model("gpt2")
+
+print("gguf: get gpt2 tokenizer vocab")
+
+# ref: https://github.com/cmp-nct/ggllm.cpp/blob/master/falcon_convert.py
+tokenizer = AutoTokenizer.from_pretrained(dir_model)
+
+# The number of tokens in tokenizer.json can differ from the expected vocab size.
+# This causes downstream issues with mismatched tensor sizes when running the inference
+vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+assert max(tokenizer.vocab.values()) < vocab_size
+
+added_vocab = tokenizer.get_added_vocab()
+reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.vocab.items()}
+
+for i in range(vocab_size):
+    if i not in reverse_vocab:
+        tokens.append(f"[PAD{i}]")
+        toktypes.append(gguf.TokenType.USER_DEFINED)
+    elif reverse_vocab[i] in added_vocab:
+        tokens.append(reverse_vocab[i])
+        if tokenizer.added_tokens_decoder[i].special:
+            toktypes.append(gguf.TokenType.CONTROL)
+        else:
+            toktypes.append(gguf.TokenType.USER_DEFINED)
+    else:
+        tokens.append(reverse_vocab[i])
+        toktypes.append(gguf.TokenType.NORMAL)
+
+gguf_writer.add_token_list(tokens)
+gguf_writer.add_token_types(toktypes)
+special_vocab = gguf.SpecialVocab(dir_model, load_merges = True, n_vocab = len(tokens))
+special_vocab.add_to_gguf(gguf_writer)
+
+# TENSORS
+
+tensor_map = gguf.get_tensor_name_map(ARCH,block_count)
+
+# params for qkv transform
+n_head    = hparams["n_head"]
+n_head_kv = hparams["n_head_kv"] if "n_head_kv" in hparams else 1
+
+head_dim = hparams["n_embd"] // n_head
+
+# tensor info
+print("gguf: get tensor metadata")
+
+if num_parts == 0:
+    part_names = iter(("pytorch_model.bin",))
+else:
+    part_names = (
+        f"pytorch_model-{n:05}-of-{num_parts:05}.bin" for n in range(1, num_parts + 1)
+    )
+
+for part_name in part_names:
+    if args.vocab_only:
+        break
+    print("gguf: loading model part '" + part_name + "'")
+    model_part = torch.load(dir_model / part_name, map_location="cpu")
+
+    for name in model_part.keys():
+        data = model_part[name]
+
+        old_dtype = data.dtype
+
+        # convert any unsupported data types to float32
+        if data.dtype != torch.float16 and data.dtype != torch.float32:
+            data = data.to(torch.float32)
+
+        data = data.squeeze().numpy()
+
+        # map tensor names
+        new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
+        if new_name is None:
+            print("Can not map tensor '" + name + "'")
+            sys.exit()
+
+        n_dims = len(data.shape)
+        data_dtype = data.dtype
+
+        # if f32 desired, convert any float16 to float32
+        if 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 ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            data = data.astype(np.float32)
+
+        # if f16 desired, convert any float32 2-dim weight tensors to float16
+        if ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+            data = data.astype(np.float16)
+
+        print(name, "=>", new_name + ", shape = " + str(data.shape) + ", " + str(old_dtype) + " --> " + str(data.dtype))
+
+        gguf_writer.add_tensor(new_name, data)
+
+
+print("gguf: write header")
+gguf_writer.write_header_to_file()
+print("gguf: write metadata")
+gguf_writer.write_kv_data_to_file()
+if not args.vocab_only:
+    print("gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+
+gguf_writer.close()
+
+print(f"gguf: model successfully exported to '{fname_out}'")
+print("")

convert.py

@@ -3,14 +3,15 @@ from __future__ import annotations
 
 import argparse
 import concurrent.futures
+import copy
 import enum
 import faulthandler
 import functools
+import io
 import itertools
 import json
 import math
 import mmap
-import os
 import pickle
 import re
 import signal
@@ -22,13 +23,14 @@ from abc import ABCMeta, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
+from typing import IO, TYPE_CHECKING, Any, Callable, Generator, Iterable, Literal, Sequence, TypeVar
 
 import numpy as np
-from sentencepiece import SentencePieceProcessor
+from sentencepiece import SentencePieceProcessor  # type: ignore[import]
 
+import os
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
 import gguf
 
 if TYPE_CHECKING:
@@ -42,12 +44,10 @@ NDArray: TypeAlias = 'np.ndarray[Any, Any]'
 ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
-
 #
 # data types
 #
 
-
 @dataclass(frozen=True)
 class DataType:
     name: str
@@ -57,17 +57,14 @@ class DataType:
     def elements_to_bytes(self, n_elements: int) -> int:
         return n_elements * self.dtype.itemsize
 
-
 @dataclass(frozen=True)
 class UnquantizedDataType(DataType):
     pass
 
-
-DT_F16  = UnquantizedDataType('F16',  dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
-DT_F32  = UnquantizedDataType('F32',  dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
-DT_I32  = UnquantizedDataType('I32',  dtype = np.dtype(np.int16),   valid_conversions = [])
-DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16),  valid_conversions = ['F32', 'F16', 'Q8_0'])
-
+DT_F16  = UnquantizedDataType('F16', dtype = np.dtype(np.float16), valid_conversions = ['F32', 'Q8_0'])
+DT_F32  = UnquantizedDataType('F32', dtype = np.dtype(np.float32), valid_conversions = ['F16', 'Q8_0'])
+DT_I32  = UnquantizedDataType('I32', dtype = np.dtype(np.int16), valid_conversions = [])
+DT_BF16 = UnquantizedDataType('BF16', dtype = np.dtype(np.uint16), valid_conversions = ['F32', 'F16', 'Q8_0'])
 
 @dataclass(frozen=True)
 class QuantizedDataType(DataType):
@@ -82,7 +79,6 @@ class QuantizedDataType(DataType):
         assert n_elements % self.block_size == 0, f'Invalid number of elements {n_elements} for {self.name} with block size {self.block_size}'
         return self.quantized_dtype.itemsize * (n_elements // self.block_size)
 
-
 @dataclass(frozen=True)
 class Q8_0QuantizedDataType(QuantizedDataType):
     # Mini Q8_0 quantization in Python!
@@ -92,7 +88,6 @@ class Q8_0QuantizedDataType(QuantizedDataType):
         n_blocks = arr.size // self.block_size
         blocks = arr.reshape((n_blocks, self.block_size))
         # Much faster implementation of block quantization contributed by @Cebtenzzre
-
         def quantize_blocks_q8_0(blocks: NDArray) -> Iterable[tuple[Any, Any]]:
             d = abs(blocks).max(axis = 1) / np.float32(127)
             with np.errstate(divide = 'ignore'):
@@ -101,11 +96,10 @@ class Q8_0QuantizedDataType(QuantizedDataType):
             yield from zip(d, qs)
         return np.fromiter(quantize_blocks_q8_0(blocks), count = n_blocks, dtype = self.quantized_dtype)
 
-
 DT_Q8_0 = Q8_0QuantizedDataType('Q8_0',
-                                dtype = np.dtype(np.float32), valid_conversions = [],
-                                ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
-                                quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
+    dtype = np.dtype(np.float32), valid_conversions = [],
+    ggml_type = gguf.GGMLQuantizationType.Q8_0, block_size = 32,
+    quantized_dtype = np.dtype([('d', '<f2'), ('qs', 'i1', (32,))]))
 
 # Quantized types skipped here because they may also map to np.float32
 NUMPY_TYPE_TO_DATA_TYPE: dict[np.dtype[Any], DataType] = {}
@@ -124,8 +118,6 @@ SAFETENSORS_DATA_TYPES: dict[str, DataType] = {
 # TODO: match this with `llama_ftype`
 # TODO: rename to LLAMAFileType
 # TODO: move to `gguf.py`
-
-
 class GGMLFileType(enum.IntEnum):
     AllF32     = 0
     MostlyF16  = 1  # except 1d tensors
@@ -138,7 +130,6 @@ class GGMLFileType(enum.IntEnum):
         # 1D tensors are always F32.
         return dt if len(tensor.shape) > 1 else DT_F32
 
-
 GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
     GGMLFileType.AllF32    : DT_F32,
     GGMLFileType.MostlyF16 : DT_F16,
@@ -149,25 +140,19 @@ GGML_FILE_TYPE_TO_DATA_TYPE: dict[GGMLFileType, DataType] = {
 # hparams loading
 #
 
-
 @dataclass
 class Params:
-    n_vocab:        int
-    n_embd:         int
-    n_layer:        int
-    n_ctx:          int
-    n_ff:           int
-    n_head:         int
-    n_head_kv:      int
-    n_experts:      int | None = None
-    n_experts_used: int | None = None
-    f_norm_eps:     float | None = None
+    n_vocab:    int
+    n_embd:     int
+    n_layer:    int
+    n_ctx:      int
+    n_ff:       int
+    n_head:     int
+    n_head_kv:  int
+    f_norm_eps: float
 
-    rope_scaling_type: gguf.RopeScalingType | None = None
     f_rope_freq_base: float | None = None
     f_rope_scale: float | None = None
-    n_orig_ctx: int | None = None
-    rope_finetuned: bool | None = None
 
     ftype: GGMLFileType | None = None
 
@@ -181,11 +166,11 @@ class Params:
 
         # try transformer naming first
         if "model.layers.0.self_attn.q_proj.weight" in model:
-            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
+            n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.q_proj.weight" not in model)
         elif "model.layers.0.self_attn.W_pack.weight" in model:   # next: try baichuan naming
-            n_layer = next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
+            n_layer=next(i for i in itertools.count() if f"model.layers.{i}.self_attn.W_pack.weight" not in model)
         else:
-            n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
+            n_layer=next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
 
         if n_layer < 1:
             raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
@@ -213,20 +198,20 @@ class Params:
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
-        rope_scaling = config.get("rope_scaling")
+        n_vocab          = config["vocab_size"]
+        n_embd           = config["hidden_size"]
+        n_layer          = config["num_hidden_layers"]
+        n_ff             = config["intermediate_size"]
+        n_head           = config["num_attention_heads"]
+        n_head_kv        = config["num_key_value_heads"] if "num_key_value_heads" in config else n_head
+        f_norm_eps       = config["rms_norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
 
-        if rope_scaling is not None and (typ := rope_scaling.get("type")):
-            rope_factor = rope_scaling.get("factor")
-            f_rope_scale = rope_factor
-            if typ == "linear":
-                rope_scaling_type = gguf.RopeScalingType.LINEAR
-            elif typ == "yarn":
-                rope_scaling_type = gguf.RopeScalingType.YARN
-                n_orig_ctx = rope_scaling['original_max_position_embeddings']
-                rope_finetuned = rope_scaling['finetuned']
-            else:
-                raise NotImplementedError(f'Unknown rope scaling type: {typ}')
+        rope_scaling = config.get("rope_scaling")
+        if isinstance(rope_scaling, dict) and rope_scaling.get("type") == "linear":
+            f_rope_scale = config["rope_scaling"].get("factor")
+        else:
+            f_rope_scale = None
 
         if "max_sequence_length" in config:
             n_ctx = config["max_sequence_length"]
@@ -236,29 +221,17 @@ class Params:
             raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
                             "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
 
-        n_experts      = None
-        n_experts_used = None
-
-        if "num_local_experts" in config:
-            n_experts = config["num_local_experts"]
-            n_experts_used = config["num_experts_per_tok"]
-
         return Params(
-            n_vocab           = config["vocab_size"],
-            n_embd            = config["hidden_size"],
-            n_layer           = config["num_hidden_layers"],
-            n_ctx             = n_ctx,
-            n_ff              = config["intermediate_size"],
-            n_head            = (n_head := config["num_attention_heads"]),
-            n_head_kv         = config.get("num_key_value_heads", n_head),
-            n_experts         = n_experts,
-            n_experts_used    = n_experts_used,
-            f_norm_eps        = config["rms_norm_eps"],
-            f_rope_freq_base  = config.get("rope_theta"),
-            rope_scaling_type = rope_scaling_type,
-            f_rope_scale      = f_rope_scale,
-            n_orig_ctx        = n_orig_ctx,
-            rope_finetuned    = rope_finetuned,
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
+            n_ctx            = n_ctx,
+            n_ff             = n_ff,
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
+            f_rope_scale     = f_rope_scale,
         )
 
     # LLaMA v2 70B params.json
@@ -267,15 +240,17 @@ class Params:
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
         config = json.load(open(config_path))
 
-        n_experts      = None
-        n_experts_used = None
-        f_rope_freq_base = None
+        n_vocab          = config["vocab_size"] if "vocab_size" in config else -1
+        n_embd           = config["dim"]
+        n_layer          = config["n_layers"]
+        n_ff             = -1
+        n_head           = config["n_heads"]
+        n_head_kv        = config["n_kv_heads"] if "n_kv_heads" in config else n_head
+        f_norm_eps       = config["norm_eps"]
+        f_rope_freq_base = config["rope_theta"] if "rope_theta" in config else None
 
         # hack to determine LLaMA v1 vs v2 vs CodeLlama
-        if config.get("moe"):
-            # Mixtral
-            n_ctx = 32768
-        elif config.get("rope_theta") == 1000000:
+        if f_rope_freq_base == 1000000:
             # CodeLlama
             n_ctx = 16384
         elif config["norm_eps"] == 1e-05:
@@ -285,27 +260,22 @@ class Params:
             # LLaMA v1
             n_ctx = 2048
 
-        if "layers.0.feed_forward.w1.weight" in model:
+        if n_vocab == -1:
+            n_vocab = model["tok_embeddings.weight"].shape[0]
+
+        if n_ff == -1:
             n_ff = model["layers.0.feed_forward.w1.weight"].shape[0]
 
-        if config.get("moe"):
-            n_ff = model["layers.0.feed_forward.experts.0.w1.weight"].shape[0]
-            n_experts      = config["moe"]["num_experts"]
-            n_experts_used = config["moe"]["num_experts_per_tok"]
-            f_rope_freq_base = 1e6
-
         return Params(
-            n_vocab          = model["tok_embeddings.weight"].shape[0],
-            n_embd           = config["dim"],
-            n_layer          = config["n_layers"],
+            n_vocab          = n_vocab,
+            n_embd           = n_embd,
+            n_layer          = n_layer,
             n_ctx            = n_ctx,
             n_ff             = n_ff,
-            n_head           = (n_head := config["n_heads"]),
-            n_head_kv        = config.get("n_kv_heads", n_head),
-            n_experts        = n_experts,
-            n_experts_used   = n_experts_used,
-            f_norm_eps       = config["norm_eps"],
-            f_rope_freq_base = config.get("rope_theta", f_rope_freq_base),
+            n_head           = n_head,
+            n_head_kv        = n_head_kv,
+            f_norm_eps       = f_norm_eps,
+            f_rope_freq_base = f_rope_freq_base,
         )
 
     @staticmethod
@@ -334,10 +304,6 @@ class Params:
 class BpeVocab:
     def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        try:
-            self.vocab = self.bpe_tokenizer["model"]["vocab"]
-        except:
-            self.vocab = self.bpe_tokenizer
         added_tokens: dict[str, int]
         if fname_added_tokens is not None:
             # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
@@ -353,9 +319,9 @@ class BpeVocab:
                     (item['content'], item['id'])
                     for item in tokenizer_json.get('added_tokens', [])
                     # Added tokens here can be duplicates of the main vocabulary.
-                    if item['content'] not in self.bpe_tokenizer)
+                    if item['content'] not in self.bpe_tokenizer )
 
-        vocab_size: int = len(self.vocab)
+        vocab_size: int = len(self.bpe_tokenizer)
         expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
         actual_ids      = sorted(added_tokens.values())
         if expected_ids != actual_ids:
@@ -363,7 +329,6 @@ class BpeVocab:
             raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
-        self.added_tokens_dict    = added_tokens
         self.added_tokens_list    = [text for (text, idx) in items]
         self.vocab_size_base: int = vocab_size
         self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
@@ -371,9 +336,11 @@ class BpeVocab:
         self.fname_added_tokens   = fname_added_tokens
 
     def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
+        tokenizer = self.bpe_tokenizer
+        from transformers.models.gpt2 import tokenization_gpt2  # type: ignore[import]
+        reverse_vocab = {id: encoded_tok for encoded_tok, id in tokenizer.items()}
 
-        for i, _ in enumerate(self.vocab):
+        for i, _ in enumerate(tokenizer):
             yield reverse_vocab[i], 0.0, gguf.TokenType.NORMAL
 
     def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
@@ -408,7 +375,6 @@ class SentencePieceVocab:
             raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
 
         # Token pieces that were added to the base vocabulary.
-        self.added_tokens_dict = added_tokens
         self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
         self.vocab_size_base    = vocab_size
         self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
@@ -451,120 +417,20 @@ class SentencePieceVocab:
     def __repr__(self) -> str:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
-
-class HfVocab:
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
-        try:
-            from transformers import AutoTokenizer
-        except ImportError as e:
-            raise ImportError(
-                "To use HfVocab, please install the `transformers` package. "
-                "You can install it with `pip install transformers`."
-            ) from e
-
-        print("fname_tokenizer:", fname_tokenizer)
-        # Allow the tokenizer to default to slow or fast versions.
-        # Explicitly set tokenizer to use local paths.
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            fname_tokenizer,
-            cache_dir=fname_tokenizer,
-            local_files_only=True,
-        )
-
-        # Initialize lists and dictionaries for added tokens
-        self.added_tokens_list = []
-        self.added_tokens_dict = dict()
-        self.added_tokens_ids  = set()
-
-        # Process added tokens
-        for tok, tokidx in sorted(
-            self.tokenizer.get_added_vocab().items(), key=lambda x: x[1]
-        ):
-            # Only consider added tokens that are not in the base vocabulary
-            if tokidx >= self.tokenizer.vocab_size:
-                self.added_tokens_list.append(tok)
-                self.added_tokens_dict[tok] = tokidx
-                self.added_tokens_ids.add(tokidx)
-
-        # Store special tokens and their IDs
-        self.specials = {
-            tok: self.tokenizer.get_vocab()[tok]
-            for tok in self.tokenizer.all_special_tokens
-        }
-        self.special_ids = set(self.tokenizer.all_special_ids)
-
-        # Set vocabulary sizes
-        self.vocab_size_base = self.tokenizer.vocab_size
-        self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
-
-        self.fname_tokenizer    = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
-
-    def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        reverse_vocab = {
-            id: encoded_tok for encoded_tok, id in self.tokenizer.get_vocab().items()
-        }
-
-        for token_id in range(self.vocab_size_base):
-            # Skip processing added tokens here
-            if token_id in self.added_tokens_ids:
-                continue
-
-            # Convert token text to bytes
-            token_text = reverse_vocab[token_id].encode("utf-8")
-
-            # Yield token text, score, and type
-            yield token_text, self.get_token_score(token_id), self.get_token_type(
-                token_id, self.special_ids  # Reuse already stored special IDs
-            )
-
-    def get_token_type(self, token_id: int, special_ids: set[int]) -> gguf.TokenType:
-        # Determine token type based on whether it's a special token
-        return gguf.TokenType.CONTROL if token_id in special_ids else gguf.TokenType.NORMAL
-
-    def get_token_score(self, token_id: int) -> float:
-        # Placeholder for actual logic to determine the token's score
-        # This needs to be implemented based on specific requirements
-        return -1000.0  # Default score
-
-    def added_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        for text in self.added_tokens_list:
-            if text in self.specials:
-                toktype = self.get_token_type(self.specials[text], self.special_ids)
-                score = self.get_token_score(self.specials[text])
-            else:
-                toktype = gguf.TokenType.USER_DEFINED
-                score = -1000.0
-
-            yield text.encode("utf-8"), score, toktype
-
-    def has_newline_token(self):
-        return "<0x0A>" in self.tokenizer.vocab or "\n" in self.tokenizer.vocab
-
-    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
-        yield from self.hf_tokens()
-        yield from self.added_tokens()
-
-    def __repr__(self) -> str:
-        return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
-
+Vocab: TypeAlias = 'BpeVocab | SentencePieceVocab'
 
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)
 #
 
-
 def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
-    # print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
+    #print( "permute debug " + str(weights.shape[0]) + " x " + str(weights.shape[1]) + " nhead " + str(n_head) + " nheadkv " + str(n_kv_head) )
     if n_head_kv is not None and n_head != n_head_kv:
         n_head = n_head_kv
     return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
-            .swapaxes(1, 2)
-            .reshape(weights.shape))
+                .swapaxes(1, 2)
+                .reshape(weights.shape))
 
 
 class Tensor(metaclass=ABCMeta):
@@ -645,7 +511,7 @@ class LazyTensor:
         ret = self._load()
         # Should be okay if it maps to the same numpy type?
         assert ret.data_type == self.data_type or (self.data_type.dtype == ret.data_type.dtype), \
-            (self.data_type, ret.data_type, self.description)
+                (self.data_type, ret.data_type, self.description)
         return ret
 
     def astype(self, data_type: DataType) -> LazyTensor:
@@ -718,14 +584,14 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
 
     if any("model.embed_tokens.weight" in mp.model for mp in models_plus):
         # Transformers models put different tensors in different files, but
-        # don't split individual tensors between files.
+        # don't split indivdual tensors between files.
         model: LazyModel = {}
         for mp in models_plus:
             model.update(mp.model)
     else:
         model = merge_sharded([mp.model for mp in models_plus])
 
-    return ModelPlus(model, paths, format, vocab)  # pytype: disable=wrong-arg-types
+    return ModelPlus(model, paths, format, vocab)
 
 
 def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
@@ -733,7 +599,6 @@ def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTe
         return lazy_tensor.load().permute(n_head, n_head_kv)
     return LazyTensor(load, lazy_tensor.shape, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-
 def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_kv: int) -> LazyTensor:
     def load() -> Tensor:
         return lazy_tensor.load().permute_part(n_part, n_head, n_head_kv)
@@ -741,7 +606,6 @@ def permute_part_lazy(lazy_tensor: LazyTensor, n_part: int, n_head: int, n_head_
     s[0] = s[0] // 3
     return LazyTensor(load, s, lazy_tensor.data_type, f'permute({n_head}, {n_head_kv}) ' + lazy_tensor.description)
 
-
 def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     def load() -> Tensor:
         return lazy_tensor.load().part(n_part)
@@ -811,7 +675,7 @@ class LazyUnpickler(pickle.Unpickler):
         return func(*args)
 
     CLASSES: dict[tuple[str, str], Any] = {
-        # getattr used here as a workaround for mypy not being smart enough to determine
+        # getattr used here as a workaround for mypy not being smart enough to detrmine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
         ('torch._utils', '_rebuild_tensor_v2'): getattr(lazy_rebuild_tensor_v2, '__func__'),
@@ -837,7 +701,6 @@ def lazy_load_torch_file(outer_fp: IO[bytes], path: Path) -> ModelPlus:
                               data_base_path=pickle_paths[0][:-4],
                               zip_file=zf)
     model = unpickler.load()
-    if 'model' in model: model = model['model']
     as_dict = dict(model.items())
     return ModelPlus(model=as_dict, paths=[path], format='torch', vocab=None)
 
@@ -891,7 +754,6 @@ def lazy_load_file(path: Path) -> ModelPlus:
 In = TypeVar('In')
 Out = TypeVar('Out')
 
-
 def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], concurrency: int, max_workers: int | None = None, use_processpool_executor: bool = False) -> Iterable[Out]:
     '''Parallel map, but with backpressure.  If the caller doesn't call `next`
     fast enough, this will stop calling `func` at some point rather than
@@ -906,7 +768,7 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
         executor_class = ProcessPoolExecutor
     else:
         executor_class = ThreadPoolExecutor
-    with executor_class(max_workers=max_workers) as executor:
+    with executor_class(max_workers = max_workers) as executor:
         futures: list[concurrent.futures.Future[Out]] = []
         done = False
         for _ in range(concurrency):
@@ -926,41 +788,25 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
                     break
             yield result
 
-
-def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
-    # Handle special case where the model's vocab size is not set
-    if params.n_vocab == -1:
-        raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
-        )
-
-    # Check for a vocab size mismatch
-    if params.n_vocab == vocab.vocab_size:
-        print("Ignoring added_tokens.json since model matches vocab size without it.")
-        return
-
-    if pad_vocab and params.n_vocab > vocab.vocab_size:
-        pad_count = params.n_vocab - vocab.vocab_size
-        print(
-            f"Padding vocab with {pad_count} token(s) - <dummy00001> through <dummy{pad_count:05}>"
-        )
-        for i in range(1, pad_count + 1):
-            vocab.added_tokens_dict[f"<dummy{i:05}>"] = -1
-            vocab.added_tokens_list.append(f"<dummy{i:05}>")
-        vocab.vocab_size = params.n_vocab
-        return
-
-    msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer} has {vocab.vocab_size})."
-    if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20:
-        msg += f"  Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
-    if vocab.vocab_size < params.n_vocab:
-        msg += " Add the --pad-vocab option and try again."
-
-    raise Exception(msg)
+def check_vocab_size(params: Params, vocab: Vocab) -> None:
+    if params.n_vocab != vocab.vocab_size:
+        assert isinstance(vocab, BpeVocab) or isinstance(vocab, SentencePieceVocab)
+        if params.n_vocab == vocab.vocab_size_base:
+            print("Ignoring added_tokens.json since model matches vocab size without it.")
+            vocab.added_tokens_list = []
+            vocab.vocab_size = vocab.vocab_size_base
+            return
+        msg = f"Vocab size mismatch (model has {params.n_vocab}, but {vocab.fname_tokenizer}"
+        if vocab.fname_added_tokens is not None:
+            msg += f" combined with {vocab.fname_added_tokens}"
+        msg += f" has {vocab.vocab_size})."
+        if vocab.vocab_size < params.n_vocab < vocab.vocab_size + 20 and vocab.fname_added_tokens is None:
+            msg += f"  Most likely you are missing added_tokens.json (should be in {vocab.fname_tokenizer.parent})."
+        raise Exception(msg)
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
@@ -980,75 +826,33 @@ class OutputFile:
         self.gguf.add_rope_dimension_count(params.n_embd // params.n_head)
         self.gguf.add_head_count          (params.n_head)
         self.gguf.add_head_count_kv       (params.n_head_kv)
-
-        if params.n_experts:
-            self.gguf.add_expert_count(params.n_experts)
-
-        if params.n_experts_used:
-            self.gguf.add_expert_used_count(params.n_experts_used)
-
-        if params.f_norm_eps:
-            self.gguf.add_layer_norm_rms_eps(params.f_norm_eps)
-        else:
-            raise ValueError('f_norm_eps is None')
+        self.gguf.add_layer_norm_rms_eps  (params.f_norm_eps)
 
         if params.f_rope_freq_base is not None:
             self.gguf.add_rope_freq_base(params.f_rope_freq_base)
 
-        if params.rope_scaling_type:
-            assert params.f_rope_scale is not None
-            self.gguf.add_rope_scaling_type(params.rope_scaling_type)
-            self.gguf.add_rope_scaling_factor(params.f_rope_scale)
-
-        if params.n_orig_ctx is not None:
-            self.gguf.add_rope_scaling_orig_ctx_len(params.n_orig_ctx)
-
-        if params.rope_finetuned is not None:
-            self.gguf.add_rope_scaling_finetuned(params.rope_finetuned)
+        if params.f_rope_scale is not None:
+            self.gguf.add_rope_scale_linear(params.f_rope_scale)
 
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
-    def handle_tokenizer_model(self, vocab: Vocab) -> str:
-        # Map the vocab types to the supported tokenizer models
-        tokenizer_model = {
-            SentencePieceVocab: "llama",
-            HfVocab: "llama",
-            BpeVocab: "gpt2",
-        }.get(type(vocab))
-
-        # Block if vocab type is not predefined
-        if tokenizer_model is None:
-            raise ValueError("Unknown vocab type: Not supported")
-
-        return tokenizer_model
-
-    def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
+    def add_meta_vocab(self, vocab: Vocab) -> None:
         tokens = []
         scores = []
         toktypes = []
-
         # NOTE: `all_tokens` returns the base vocabulary and added tokens
         for text, score, toktype in vocab.all_tokens():
             tokens.append(text)
             scores.append(score)
             toktypes.append(toktype)
 
-        assert len(tokens) == vocab.vocab_size
-
-        return tokens, scores, toktypes
-
-    def add_meta_vocab(self, vocab: Vocab) -> None:
-        # Handle the tokenizer model
-        tokenizer_model = self.handle_tokenizer_model(vocab)
-
-        # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(tokenizer_model)
-
-        # Extract model vocabulary for model conversion
-        tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
-
-        # Add extracted token information for model conversion
+        if isinstance(vocab, SentencePieceVocab):
+            self.gguf.add_tokenizer_model("llama")
+        elif isinstance(vocab, BpeVocab):
+            self.gguf.add_tokenizer_model("gpt2")
+        else:
+            raise ValueError(f'Unknown vocab type: Not BpeVocab or SentencePieceVocab')
         self.gguf.add_token_list(tokens)
         self.gguf.add_token_scores(scores)
         self.gguf.add_token_types(toktypes)
@@ -1061,7 +865,7 @@ class OutputFile:
         raw_dtype = getattr(tensor.data_type, 'ggml_type', None)
         data_type = getattr(tensor.data_type, 'quantized_type', None) or tensor.data_type.dtype
         data_nbytes = tensor.data_type.elements_to_bytes(n_elements)
-        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype=raw_dtype)
+        self.gguf.add_tensor_info(name, tensor.shape, data_type, data_nbytes, raw_dtype = raw_dtype)
 
     def write_meta(self) -> None:
         self.gguf.write_header_to_file()
@@ -1074,11 +878,8 @@ class OutputFile:
         self.gguf.close()
 
     @staticmethod
-    def write_vocab_only(
-        fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
-    ) -> None:
-        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
+    def write_vocab_only(fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab, endianess:gguf.GGUFEndian=gguf.GGUFEndian.LITTLE) -> None:
+        check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1105,12 +906,8 @@ class OutputFile:
         return dt.quantize(arr)
 
     @staticmethod
-    def write_all(
-        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
-        concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
-        pad_vocab: bool = False,
-    ) -> None:
-        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
+    def write_all(fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab, concurrency: int = DEFAULT_CONCURRENCY, endianess=gguf.GGUFEndian.LITTLE) -> None:
+        check_vocab_size(params, vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1129,10 +926,7 @@ class OutputFile:
         # tensor data
         ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
         if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
+            ndarrays = bounded_parallel_map(OutputFile.maybe_do_quantize, ndarrays_inner, concurrency = concurrency, max_workers = concurrency, use_processpool_executor = True)
         else:
             ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
 
@@ -1141,16 +935,13 @@ class OutputFile:
             elapsed = time.time() - start
             size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
             padi = len(str(len(model)))
-            print(
-                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
+            print(f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}")
             of.gguf.write_tensor_data(ndarray)
 
         of.close()
 
-
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
-    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
+    wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0)+".weight"].data_type
 
     if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
@@ -1163,12 +954,10 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
 
     raise Exception(f"Unexpected combination of types: {name_to_type}")
 
-
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
     return {name: tensor.astype(output_type.type_for_tensor(name, tensor))
             for (name, tensor) in model.items()}
 
-
 def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
     should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
@@ -1181,7 +970,7 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
             print(f"Permuting layer {i}")
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.q_proj.weight"], params.n_head, params.n_head)
             tmp[f"model.layers.{i}.self_attn.k_proj.weight"] = permute_lazy(model[f"model.layers.{i}.self_attn.k_proj.weight"], params.n_head, params.n_head_kv)
-            # tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
+           #tmp[f"model.layers.{i}.self_attn.v_proj.weight"] =              model[f"model.layers.{i}.self_attn.v_proj.weight"]
         elif f"model.layers.{i}.self_attn.W_pack.weight" in model:
             print(f"Unpacking and permuting layer {i}")
             tmp[f"model.layers.{i}.self_attn.q_proj.weight"] = permute_part_lazy(model[f"model.layers.{i}.self_attn.W_pack.weight"], 0, params.n_head, params.n_head)
@@ -1206,7 +995,6 @@ def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
 
     return out
 
-
 def nth_multifile_path(path: Path, n: int) -> Path | None:
     '''Given any path belonging to a multi-file model (e.g. foo.bin.1), return
     the nth path in the model.
@@ -1251,8 +1039,7 @@ def load_some_model(path: Path) -> ModelPlus:
     # Be extra-friendly and accept either a file or a directory:
     if path.is_dir():
         # Check if it's a set of safetensors files first
-        globs = ["model-00001-of-*.safetensors", "model.safetensors"]
-        files = [file for glob in globs for file in path.glob(glob)]
+        files = list(path.glob("model-00001-of-*.safetensors"))
         if not files:
             # Try the PyTorch patterns too, with lower priority
             globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
@@ -1273,74 +1060,35 @@ def load_some_model(path: Path) -> ModelPlus:
     return model_plus
 
 
-class VocabFactory:
-    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()
-
-    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 _select_file(self, vocabtype: str | None) -> Path:
-        if vocabtype in ["spm", "bpe"]:
-            for file_key in self.files.keys():
-                if (file := self.files[file_key]) is not None:
-                    return file
-            raise FileNotFoundError(f"{vocabtype} vocab not found.")
-        if vocabtype == "hfft":
-            # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
-            return self.path
-        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
-
-    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocabtype == "bpe"
-        n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
-        return gguf.SpecialVocab(
-            model_parent_path,
-            load_merges=load_merges,
-            special_token_types=None,  # Predetermined or passed as a parameter
-            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}'")
-
-        added_tokens_path = path.parent / "added_tokens.json"
-        vocab: Vocab
-        if vocabtype == "bpe":
-            vocab = BpeVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        elif vocabtype == "spm":
-            vocab = SentencePieceVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        elif vocabtype == "hfft":
-            vocab = HfVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
+def load_vocab(path: Path, vocabtype: str | None) -> Vocab:
+    # Be extra-friendly and accept either a file or a directory.  Also, if it's
+    # a directory, it might be the model directory, and tokenizer.model might
+    # be in the parent of that.
+    if path.is_dir():
+        vocab_file = "tokenizer.model"
+        if vocabtype == 'bpe':
+            vocab_file = "vocab.json"
+        path2 = path / vocab_file
+        # Use `.parent` instead of /.. to handle the symlink case better.
+        path3 = path.parent / vocab_file
+        if path2.exists():
+            path = path2
+        elif path3.exists():
+            path = path3
         else:
-            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
-        # FIXME: Respect --vocab-dir?
-        special_vocab = self._create_special_vocab(
-            vocab,
-            vocabtype,
-            model_parent_path,
-        )
-        return vocab, special_vocab
+            raise FileNotFoundError(
+                f"Could not find {vocab_file} in {path} or its parent; "
+                "if it's in another directory, pass the directory as --vocab-dir")
+
+    print(f"Loading vocab file '{path}', type '{vocabtype}'")
+
+    added_tokens_path = path.parent / "added_tokens.json"
+    if vocabtype == "bpe":
+        return BpeVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    elif vocabtype == "spm":
+        return SentencePieceVocab(path, added_tokens_path if added_tokens_path.exists() else None)
+    else:
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
 
 
 def default_outfile(model_paths: list[Path], file_type: GGMLFileType) -> Path:
@@ -1367,40 +1115,20 @@ def do_dump_model(model_plus: ModelPlus) -> None:
 
 
 def main(args_in: list[str] | None = None) -> None:
-    output_choices = ["f32", "f16"]
-    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("--outtype",     choices=["f32", "f16", "q8_0"], 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("--vocabtype",   choices=["spm", "bpe"], help="vocab format (default: spm)", default="spm")
     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("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default = DEFAULT_CONCURRENCY)
+    parser.add_argument("--bigendian",   action="store_true",    help="model is executed on big endian machine")
 
     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)
         do_dump_model(model_plus)
@@ -1415,7 +1143,7 @@ def main(args_in: list[str] | None = None) -> None:
         do_dump_model(model_plus)
         return
     endianess = gguf.GGUFEndian.LITTLE
-    if args.big_endian:
+    if args.bigendian:
         endianess = gguf.GGUFEndian.BIG
 
     params = Params.load(model_plus)
@@ -1436,25 +1164,29 @@ def main(args_in: list[str] | None = None) -> None:
 
     print(f"params = {params}")
 
-    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: Vocab
     if args.vocab_only:
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
+        # FIXME: Try to respect vocab_dir somehow?
+        vocab = load_vocab(args.vocab_dir or args.model, args.vocabtype)
+        special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+            load_merges = args.vocabtype == 'bpe',
+            n_vocab = vocab.vocab_size)
         outfile = args.outfile
-        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab,
-                                    endianess=endianess, pad_vocab=args.pad_vocab)
+        OutputFile.write_vocab_only(outfile, params, vocab, special_vocab)
         print(f"Wrote {outfile}")
         return
 
     if model_plus.vocab is not None and args.vocab_dir is None:
         vocab = model_plus.vocab
-
-    print(f"Vocab info: {vocab}")
-    print(f"Special vocab info: {special_vocab}")
+    else:
+        vocab_dir = args.vocab_dir if args.vocab_dir else model_plus.paths[0].parent
+        vocab = load_vocab(vocab_dir, args.vocabtype)
+    # FIXME: Try to respect vocab_dir somehow?
+    special_vocab = gguf.SpecialVocab(model_plus.paths[0].parent,
+        load_merges = args.vocabtype == 'bpe',
+        n_vocab = vocab.vocab_size)
 
     model   = model_plus.model
     model   = convert_model_names(model, params)
@@ -1465,8 +1197,7 @@ def main(args_in: list[str] | None = None) -> None:
     params.ftype = ftype
     print(f"Writing {outfile}, format {ftype}")
 
-    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab,
-                         concurrency=args.concurrency, endianess=endianess, pad_vocab=args.pad_vocab)
+    OutputFile.write_all(outfile, ftype, params, model, vocab, special_vocab, concurrency = args.concurrency, endianess=endianess)
     print(f"Wrote {outfile}")
 
 

docs/token_generation_performance_tips.md

@@ -17,7 +17,7 @@ llama_model_load_internal: [cublas] total VRAM used: 17223 MB
 If you see these lines, then the GPU is being used.
 
 ## Verifying that the CPU is not oversaturated
-llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physical CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down.
+llama accepts a `-t N` (or `--threads N`) parameter. It's extremely important that this parameter is not too large. If your token generation is extremely slow, try setting this number to 1. If this significantly improves your token generation speed, then your CPU is being oversaturated and you need to explicitly set this parameter to the number of the physicial CPU cores on your machine (even if you utilize a GPU). If in doubt, start with 1 and double the amount until you hit a performance bottleneck, then scale the number down.
 
 # Example of runtime flags effect on inference speed benchmark
 These runs were tested on the following machine:

examples/CMakeLists.txt

@@ -23,23 +23,18 @@ else()
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
-    if (LLAMA_SYCL)
-        add_subdirectory(sycl)
-    endif()
     add_subdirectory(main)
-    add_subdirectory(tokenize)
     add_subdirectory(parallel)
     add_subdirectory(perplexity)
     add_subdirectory(quantize)
     add_subdirectory(quantize-stats)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
-    add_subdirectory(passkey)
     add_subdirectory(speculative)
-    add_subdirectory(lookahead)
-    add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
-    add_subdirectory(imatrix)
+    if (LLAMA_METAL)
+        add_subdirectory(metal)
+    endif()
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -575,7 +575,10 @@ static struct ggml_tensor * forward(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -841,7 +844,10 @@ static struct ggml_tensor * forward_batch(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
             assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
 
             // KQ_masked = mask_past(KQ_scaled)
@@ -1125,7 +1131,10 @@ static struct ggml_tensor * forward_lora(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -1249,9 +1258,9 @@ static struct ggml_tensor * forward_lora(
 }
 
 static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
-    assert(ggml_is_matrix(logits));
-    assert(ggml_is_matrix(probs));
-    assert(ggml_is_vector(best_samples));
+    assert(logits->n_dims == 2);
+    assert(probs->n_dims == 2);
+    assert(best_samples->n_dims == 1);
     assert(logits->ne[1] == best_samples->ne[0]);
     assert(logits->ne[0] == probs->ne[0]);
     assert(logits->ne[1] == probs->ne[1]);
@@ -1283,9 +1292,9 @@ static void sample_softmax_batch(
     struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
     struct ggml_tensor * best_samples
 ) {
-    GGML_ASSERT(ggml_is_matrix(best_samples));
-    GGML_ASSERT(ggml_is_3d(logits));
-    GGML_ASSERT(ggml_is_3d(probs));
+    GGML_ASSERT(best_samples->n_dims == 2);
+    GGML_ASSERT(logits->n_dims == 3);
+    GGML_ASSERT(probs->n_dims == 3);
     int n_tokens = best_samples->ne[0];
     int n_batch  = best_samples->ne[1];
     int n_vocab  = logits->ne[0];
@@ -1325,7 +1334,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(ggml_is_matrix(probs));
+    assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = ggml_get_f32_1d(probs, i*probs->ne[0] + k);
@@ -1377,8 +1386,8 @@ static void get_example_targets(int example_id, struct ggml_tensor * tokens_inpu
 static void get_example_targets_batch(
     struct ggml_context * ctx, int example_id, struct ggml_tensor * tokens_input, struct ggml_tensor * targets
 ) {
-    GGML_ASSERT(ggml_is_matrix(tokens_input));
-    GGML_ASSERT(ggml_is_3d(targets));
+    GGML_ASSERT(tokens_input->n_dims == 2);
+    GGML_ASSERT(     targets->n_dims == 3);
     int n_tokens = tokens_input->ne[0];
     int n_batch  = tokens_input->ne[1];
     GGML_ASSERT(n_tokens == targets->ne[1]);

examples/base-translate.sh

@@ -1,61 +0,0 @@
-#!/bin/bash
-#
-# Few-shot translation example.
-# Requires a base model (i.e. no fine-tuned or instruct models).
-#
-# Usage:
-#
-#   cd llama.cpp
-#   make -j
-#
-#   ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
-#
-
-if [ $# -lt 2 ]; then
-  echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
-  exit 1
-fi
-
-eargs=""
-if [ $# -gt 2 ]; then
-  eargs="${@:3}"
-fi
-
-ftmp="__llama.cpp_example_tmp__.txt"
-trap "rm -f $ftmp" EXIT
-
-echo "Translate from English to French:
-
-===
-
-sea otter, peppermint, plush girafe:
-
-sea otter => loutre de mer
-peppermint => menthe poivrée
-plush girafe => girafe peluche
-
-===
-
-violin
-
-violin => violon
-
-===
-
-phone, computer, mouse, keyboard:
-
-phone => téléphone
-computer => ordinateur
-mouse => souris
-keyboard => clavier
-
-===
-" > $ftmp
-
-echo "$2
-" >> $ftmp
-
-model=$1
-
-# generate the most likely continuation until the string "===" is found
-./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs

examples/batched-bench/batched-bench.cpp

@@ -88,10 +88,7 @@ int main(int argc, char ** argv) {
 
     llama_model_params model_params = llama_model_default_params();
 
-    const std::vector<float> t_split (LLAMA_MAX_DEVICES, 0.0f);
-
     model_params.n_gpu_layers = n_gpu_layers;
-    model_params.tensor_split = t_split.data();
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -158,7 +155,7 @@ 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, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
     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");
@@ -188,7 +185,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_clear(ctx);
+                llama_kv_cache_tokens_rm(ctx, -1, -1);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
                     LOG_TEE("%s: llama_decode() failed\n", __func__);

examples/batched.swift/README.md

@@ -1,4 +1,4 @@
 This is a swift clone of `examples/batched`.
 
 $ `make`
-$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -153,7 +153,7 @@ while n_cur <= n_len {
         // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
-        if new_token_id == llama_token_eos(model) || n_cur == n_len {
+        if new_token_id == llama_token_eos(context) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {
@@ -215,10 +215,9 @@ print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end
 llama_print_timings(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
-    let utf8Count = text.utf8.count
-    let n_tokens = utf8Count + (add_bos ? 1 : 0)
+    let n_tokens = text.count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
+    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])
@@ -231,15 +230,18 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
     var result = [CChar](repeating: 0, count: 8)
     let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
     if nTokens < 0 {
-        let actualTokensCount = -Int(nTokens)
-        result = .init(repeating: 0, count: actualTokensCount)
+        if result.count >= -Int(nTokens) {
+            result.removeLast(-Int(nTokens))
+        } else {
+            result.removeAll()
+        }
         let check = llama_token_to_piece(
             model,
             token,
             &result,
             Int32(result.count)
         )
-        assert(check == actualTokensCount)
+        assert(check == nTokens)
     } else {
         result.removeLast(result.count - Int(nTokens))
     }
@@ -257,4 +259,5 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
         buffer = []
         return bufferString
     }
+    return nil
 }

examples/batched/batched.cpp

@@ -69,7 +69,6 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
-
     const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
 
     // initialize the context

examples/benchmark/CMakeLists.txt

@@ -1,6 +1,9 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/benchmark/benchmark-matmult.cpp

@@ -1,3 +1,4 @@
+#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 
@@ -129,13 +130,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(qtype);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
     ctx_size += 1024*1024*16;
 
     printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
@@ -171,8 +172,7 @@ int main(int argc, char ** argv)  {
     struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
 
     // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf, m11xm2);
+    struct ggml_cgraph gf = ggml_build_forward(m11xm2);
 
     printf("n_threads=%i\n", benchmark_params.n_threads);
 
@@ -181,9 +181,9 @@ int main(int argc, char ** argv)  {
 
     std::vector<uint8_t> work_buffer;
 
-    ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
+    ggml_graph_compute_helper(work_buffer, &gf, benchmark_params.n_threads);
 
-    TENSOR_DUMP(gf->nodes[0]);
+    TENSOR_DUMP(gf.nodes[0]);
 
     printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
@@ -194,27 +194,25 @@ int main(int argc, char ** argv)  {
     // 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, hist_cur.data());
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
     struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
 
     // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf31, q31);
+    struct ggml_cgraph gf31 = ggml_build_forward(q31);
 
     // 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, hist_cur.data());
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
 
     //printf("Creating compute graph\n");
-    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf32, q32);
+    struct ggml_cgraph gf32 = ggml_build_forward(q32);
     printf("n_threads=%i\n", benchmark_params.n_threads);
 
     const int dimx = sizex;
@@ -226,7 +224,7 @@ int main(int argc, char ** argv)  {
 
 
     // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
+    float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
     printf("=====================================================================================\n");
@@ -236,7 +234,7 @@ int main(int argc, char ** argv)  {
 
         long long int start = ggml_time_us();
         //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
+        ggml_graph_compute_helper(work_buffer, &gf31, benchmark_params.n_threads);
 
         long long int stop = ggml_time_us();
         long long int usec = stop-start;
@@ -254,7 +252,7 @@ int main(int argc, char ** argv)  {
 
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
+        float sum_of_Q4_result = tensor_sum_elements(gf31.nodes[0]);
         float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 
@@ -269,7 +267,7 @@ int main(int argc, char ** argv)  {
         }
 
         // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
+        ggml_graph_compute_helper(work_buffer, &gf32, benchmark_params.n_threads);
     }
     printf("\n");
     printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));

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

@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(ggml_is_matrix(probs));
+    assert(probs->n_dims == 2);
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
@@ -639,7 +639,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
     int ct;
-    switch (ggml_n_dims(gg_weights)) {
+    switch (gg_weights->n_dims){
         case 1:
             ct = 0;
             for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

examples/embedding/CMakeLists.txt

@@ -3,3 +3,6 @@ add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/embedding/embedding.cpp

@@ -1,3 +1,4 @@
+#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 

examples/export-lora/export-lora.cpp

@@ -240,13 +240,14 @@ static struct lora_data * load_lora(struct lora_info * info) {
     }
 
     struct ggml_init_params params_ggml;
-    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
     params_ggml.mem_buffer = NULL;
     params_ggml.no_alloc   = true;
     result->ctx = ggml_init(params_ggml);
 
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
     uint32_t magic   = file.read_u32();
-    if (magic != LLAMA_FILE_MAGIC_GGLA) {
+    if (magic != LLAMA_FILE_MAGIC_LORA) {
         die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
     }
     uint32_t version = file.read_u32();
@@ -308,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
 ) {
     struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
     if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, scaling);
+        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
 
@@ -333,7 +334,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
     float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
 
     struct ggml_init_params params;
-    params.mem_size   = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_size   = GGML_OBJECT_SIZE + GGML_GRAPH_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
     params.mem_buffer = NULL;
     params.no_alloc   = true;
     struct ggml_context * ctx = NULL;

examples/finetune/README.md

@@ -21,7 +21,7 @@ wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/s
 ./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
 ```
 
-**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
+Finetune output files will be saved every N iterations (config with `--save-every N`).
 The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
 So in above example after 10 iterations these files will be written:
 - chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
@@ -61,7 +61,7 @@ For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' L
   --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
 ```
 
-The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values too big will sometimes result in worse output. Play around to find good values.
+The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values to big will sometimes result in worse output. Play around to find good values.
 
 Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
 If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.

examples/finetune/convert-finetune-checkpoint-to-gguf.py

@@ -3,7 +3,9 @@
 
 import argparse
 import gguf
+import os
 import struct
+import sys
 import numpy as np
 from pathlib import Path
 

examples/finetune/finetune.cpp

@@ -3,9 +3,15 @@
 #include "llama.h"
 #include "common.h"
 #include "train.h"
+#include <unordered_map>
 #include <vector>
+#include <cassert>
+#include <climits>
 #include <cstring>
+#include <cstdarg>
 #include <ctime>
+#include <random>
+#include <stdexcept>
 #include <algorithm>
 #include <string>
 
@@ -190,13 +196,13 @@ 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               : %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_ff                  : %u\n", __func__, params->n_ff);
-    printf("%s: n_head                : %u\n", __func__, params->n_head);
-    printf("%s: n_head_kv             : %u\n", __func__, params->n_head_kv);
-    printf("%s: n_layer               : %u\n", __func__, params->n_layer);
+    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_ff:      %u\n", __func__, params->n_ff);
+    printf("%s: n_head:    %u\n", __func__, params->n_head);
+    printf("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    printf("%s: n_layer:   %u\n", __func__, params->n_layer);
     printf("%s: norm_rms_eps          : %f\n", __func__, params->f_norm_rms_eps);
     printf("%s: rope_freq_base        : %f\n", __func__, params->rope_freq_base);
     printf("%s: rope_freq_scale       : %f\n", __func__, params->rope_freq_scale);
@@ -263,7 +269,7 @@ static void load_model_hparams_gguf(struct gguf_context * ctx, struct my_llama_h
     float rope_freq_scale = 1.0f;
     GGUF_GET_KEY(ctx, hparams->f_norm_rms_eps, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS));
     GGUF_GET_KEY(ctx, hparams->rope_freq_base, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_FREQ_BASE));
-    GGUF_GET_KEY(ctx, rope_freq_scale,         gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
+    GGUF_GET_KEY(ctx, rope_freq_scale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, kv(LLM_KV_ROPE_SCALE_LINEAR));
     if (rope_freq_scale != 1.0f) {
         hparams->rope_freq_scale = 1.0f / rope_freq_scale;
     }
@@ -542,35 +548,35 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
     struct random_normal_distribution * rnd = init_random_normal_distribution(seed, mean, std, min, max);
 
     randomize_tensor_normal(lora->tok_embeddings_a, rnd);
-    ggml_set_zero(lora->tok_embeddings_b);
+    randomize_tensor_normal(lora->tok_embeddings_b, rnd);
     randomize_tensor_normal(lora->norm_a,           rnd);
-    ggml_set_zero(lora->norm_b);
+    randomize_tensor_normal(lora->norm_b,           rnd);
     randomize_tensor_normal(lora->output_a,         rnd);
-    ggml_set_zero(lora->output_b);
+    randomize_tensor_normal(lora->output_b,         rnd);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = lora->layers[i];
         randomize_tensor_normal(layer.attention_norm_a, rnd);
-        ggml_set_zero(layer.attention_norm_b);
+        randomize_tensor_normal(layer.attention_norm_b, rnd);
 
         randomize_tensor_normal(layer.wq_a, rnd);
-        ggml_set_zero(layer.wq_b);
+        randomize_tensor_normal(layer.wq_b, rnd);
         randomize_tensor_normal(layer.wk_a, rnd);
-        ggml_set_zero(layer.wk_b);
+        randomize_tensor_normal(layer.wk_b, rnd);
         randomize_tensor_normal(layer.wv_a, rnd);
-        ggml_set_zero(layer.wv_b);
+        randomize_tensor_normal(layer.wv_b, rnd);
         randomize_tensor_normal(layer.wo_a, rnd);
-        ggml_set_zero(layer.wo_b);
+        randomize_tensor_normal(layer.wo_b, rnd);
 
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
-        ggml_set_zero(layer.ffn_norm_b);
+        randomize_tensor_normal(layer.ffn_norm_b, rnd);
 
         randomize_tensor_normal(layer.w1_a, rnd);
-        ggml_set_zero(layer.w1_b);
+        randomize_tensor_normal(layer.w1_b, rnd);
         randomize_tensor_normal(layer.w2_a, rnd);
-        ggml_set_zero(layer.w2_b);
+        randomize_tensor_normal(layer.w2_b, rnd);
         randomize_tensor_normal(layer.w3_a, rnd);
-        ggml_set_zero(layer.w3_b);
+        randomize_tensor_normal(layer.w3_b, rnd);
     }
 
     free_random_normal_distribution(rnd);
@@ -606,7 +612,6 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     const int n_rot       = hparams.n_embd_head();
     const int n_embd_head = hparams.n_embd_head();
     const int n_embd_gqa  = hparams.n_embd_gqa();
-
     const float rms_norm_eps    = hparams.f_norm_rms_eps;
     const float rope_freq_base  = hparams.rope_freq_base;
     const float rope_freq_scale = hparams.rope_freq_scale;
@@ -637,9 +642,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const int rope_mode = 0;
 
         return ggml_rope_custom(ctx,
-            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
-            rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
-        );
+            t, KQ_pos, n_rot, rope_mode, n_ctx,
+            rope_freq_base, rope_freq_scale);
     };
 
     set_name(tokens_input, "tokens_input");
@@ -648,7 +652,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     GGML_ASSERT(tokens_input->type == GGML_TYPE_I32);
 
     auto add_to_f32 = [] (struct ggml_context * ctx, struct ggml_tensor * a, struct ggml_tensor * b) {
-        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
+        if (ggml_is_quantized(a->type)) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);
@@ -675,7 +679,10 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         checkpoints.push_back(t01);
     }
 
-    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
+    struct ggml_tensor * kv_scale = NULL;
+    if (!enable_flash_attn) {
+        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
+    }
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -764,7 +771,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     if (enable_checkpointing) {
         ggml_build_backward_gradient_checkpointing(ctx, gf, gb, gb_tmp, checkpoints.data(), (int) checkpoints.size());
     } else {
-        ggml_graph_cpy(gf, gb);
+        *gb = *gf;
         ggml_build_backward_expand(ctx, gf, gb, true);
     }
 
@@ -773,32 +780,32 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
     // make sure some tensors are not reallocated by inserting new temporary nodes depending on them
     int n_leafs_before = gb->n_leafs;
     int n_nodes_before = gb->n_nodes;
-
+    struct ggml_tensor * one = ggml_new_f32(ctx, 1.0f);
     // output tensors
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
     // input gradient
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
     ggml_allocr_alloc(alloc, t36->grad);
     // KQ_pos
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
 
     // make sure base model tensors data cannot be used in viewable operations
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, 1.0f));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->tok_embeddings, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->norm, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, model->output, one));
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.attention_norm, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, 1.0f));
-        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.attention_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_norm, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wq, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, one));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, one));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -1102,7 +1109,7 @@ static void write_tensor(struct llama_file * file, struct ggml_tensor * tensor,
         name = ggml_get_name(tensor);
     }
     uint32_t name_len = strlen(name);
-    uint32_t nd = ggml_n_dims(tensor);
+    uint32_t nd = tensor->n_dims;
     uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                        (uint32_t)tensor->ne[1],
                        (uint32_t)tensor->ne[2],
@@ -1138,8 +1145,9 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         return tn_buf.data();
     };
 
+    uint32_t LLAMA_FILE_MAGIC_LORA = 0x67676C61; // 'ggla'
     // write_magic
-    file.write_u32(LLAMA_FILE_MAGIC_GGLA);   // magic
+    file.write_u32(LLAMA_FILE_MAGIC_LORA);   // magic
     file.write_u32(1); // version
     // write_hparams
     file.write_u32(lora->hparams.lora_r);
@@ -1537,7 +1545,6 @@ int main(int argc, char ** argv) {
     srand(params.common.seed);
 
     struct llama_model_params llama_mparams = llama_model_default_params();
-    llama_mparams.n_gpu_layers = params.common.n_gpu_layers;
     llama_mparams.vocab_only = false;
 
     printf("%s: model base = '%s'\n", __func__, params.fn_model_base);
@@ -1595,7 +1602,6 @@ int main(int argc, char ** argv) {
     opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
-    opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;
     opt->params.n_threads               = params.common.n_threads;
     opt->params.past                    = params.common.opt_past;
     opt->params.delta                   = params.common.opt_delta;
@@ -1611,6 +1617,8 @@ int main(int argc, char ** argv) {
     opt->params.adam.gclip              = params.common.adam_gclip;
     opt->params.adam.eps_f              = params.common.adam_eps_f;
 
+    ggml_allocr * alloc = NULL;
+
     printf("%s: init model\n", __func__);
     bool existed = load_checkpoint_lora_file(params.common.fn_checkpoint_in, &model, &lora, train);
 
@@ -1714,14 +1722,17 @@ int main(int argc, char ** argv) {
 
     // allocate input tensors
     mem_input_data.resize(max_input_size);
-    ggml_allocr_t alloc_inps = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
-    ggml_allocr_alloc(alloc_inps, tokens_input);
-    ggml_allocr_alloc(alloc_inps, target_probs);
+    alloc = ggml_allocr_new(mem_input_data.data(), mem_input_data.size(), tensor_alignment);
+    ggml_allocr_alloc(alloc, tokens_input);
+    ggml_allocr_alloc(alloc, target_probs);
+    ggml_allocr_free(alloc);
 
     // context for compute tensors without their data
-    const size_t estimated_compute_size_wo_data = (
-            2*LLAMA_TRAIN_MAX_NODES*ggml_tensor_overhead() +
-            (params.common.use_checkpointing ? 3 : 2)*(GGML_OBJECT_SIZE+ggml_graph_overhead_custom(LLAMA_TRAIN_MAX_NODES, true))
+    size_t estimated_compute_size_wo_data = (
+        ggml_tensor_overhead()*GGML_MAX_NODES*2
+      + (GGML_OBJECT_SIZE+GGML_GRAPH_SIZE)*(
+            params.common.use_checkpointing ? 3 : 2
+        )
     );
     struct ggml_init_params ctx_compute_params = {
         estimated_compute_size_wo_data, // mem_size
@@ -1743,12 +1754,12 @@ int main(int argc, char ** argv) {
     // find best evaluation order
     for (unsigned order = 0; order < (unsigned) GGML_CGRAPH_EVAL_ORDER_COUNT; ++order) {
         ctx_compute = ggml_init(ctx_compute_params);
-        ggml_allocr_t alloc = ggml_allocr_new_measure(tensor_alignment);
-        gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
+        alloc = ggml_allocr_new_measure(tensor_alignment);
+        gf = ggml_new_graph(ctx_compute);
         gf->order = (enum ggml_cgraph_eval_order) order;
-        gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
+        gb = ggml_new_graph(ctx_compute);
         gb_tmp = params.common.use_checkpointing
-            ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
+            ? ggml_new_graph(ctx_compute)
             : NULL;
         loss = llama_build_lora_finetune_graphs(
             &model, &lora, alloc, ctx_compute,
@@ -1776,12 +1787,12 @@ int main(int argc, char ** argv) {
     // allocate compute tensors
     mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    ggml_allocr_t alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
-    gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
+    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
+    gf = ggml_new_graph(ctx_compute);
     gf->order = best_order;
-    gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
+    gb = ggml_new_graph(ctx_compute);
     gb_tmp = params.common.use_checkpointing
-        ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
+        ? ggml_new_graph(ctx_compute)
         : NULL;
     loss = llama_build_lora_finetune_graphs(
         &model, &lora, alloc, ctx_compute,
@@ -1792,16 +1803,12 @@ int main(int argc, char ** argv) {
         params.common.use_checkpointing
     );
     ggml_allocr_free(alloc);
-    ggml_allocr_free(alloc_inps);
-
 
     // tokenize data
     std::vector<llama_token> train_tokens;
     std::vector<size_t> train_samples_begin;
     std::vector<size_t> train_samples_size;
-    printf("%s: tokenize training data from %s\n", __func__, params.common.fn_train_data);
-    printf("%s: sample-start: %s\n", __func__, params.common.sample_start.c_str());
-    printf("%s: include-sample-start: %s\n", __func__, params.common.include_sample_start ? "true" : "false");
+    printf("%s: tokenize training data\n", __func__);
     tokenize_file(lctx,
             params.common.fn_train_data,
             params.common.sample_start,

examples/finetune/finetune.sh

@@ -1,34 +0,0 @@
-#!/bin/bash
-cd `dirname $0`
-cd ../..
-
-EXE="./finetune"
-
-if [[ ! $LLAMA_MODEL_DIR ]]; then LLAMA_MODEL_DIR="./models"; fi
-if [[ ! $LLAMA_TRAINING_DIR ]]; then LLAMA_TRAINING_DIR="."; fi
-
-# MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2-q8_0.gguf" # This is the model the readme uses.
-MODEL="$LLAMA_MODEL_DIR/openllama-3b-v2.gguf" # An f16 model. Note in this case with "-g", you get an f32-format .BIN file that isn't yet supported if you use it with "main --lora" with GPU inferencing.
-
-while getopts "dg" opt; do
-  case $opt in
-    d)
-      DEBUGGER="gdb --args"
-      ;;
-    g)
-      EXE="./build/bin/Release/finetune"
-      GPUARG="--gpu-layers 25"
-      ;;
-  esac
-done
-
-$DEBUGGER $EXE \
-        --model-base $MODEL \
-        $GPUARG \
-        --checkpoint-in  chk-ol3b-shakespeare-LATEST.gguf \
-        --checkpoint-out chk-ol3b-shakespeare-ITERATION.gguf \
-        --lora-out lora-ol3b-shakespeare-ITERATION.bin \
-        --train-data "$LLAMA_TRAINING_DIR\shakespeare.txt" \
-        --save-every 10 \
-        --threads 10 --adam-iter 30 --batch 4 --ctx 64 \
-        --use-checkpointing

examples/gguf/CMakeLists.txt

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

examples/gguf/gguf.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "llama.h"
 
 #include <cstdio>
 #include <cinttypes>
@@ -194,7 +195,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
             struct ggml_tensor * cur = ggml_get_tensor(ctx_data, name);
 
-            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, ggml_n_dims(cur), cur->name, cur->data);
+            printf("%s: tensor[%d]: n_dims = %d, name = %s, data = %p\n", __func__, i, cur->n_dims, cur->name, cur->data);
 
             // print first 10 elements
             const float * data = (const float *) cur->data;

examples/imatrix/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET imatrix)
-add_executable(${TARGET} imatrix.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/imatrix/README.md

@@ -1,32 +0,0 @@
-# llama.cpp/examples/imatrix
-
-Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models.
-More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
-
-## Usage
-
-```
-./imatrix -m <some_fp_model> -f <some_training_data> [-o <output_file>] [--verbosity <verbosity_level>]
-        [-ofreq num_chunks] [-ow <0 or 1>] [other common params]
-```
-
-Here `-m` with a model name and `-f` with a file containing training data (such as e.g. `wiki.train.raw`) are mandatory.
-The parameters in square brackets are optional and have the following meaning:
-* `-o` (or `--output-file`) specifies the name of the file where the computed data will be stored. If missing `imatrix.dat` is used.
-* `--verbosity` specifies the verbosity level. If set to `0`, no output other than the perplexity of the processed chunks will be generated. If set to `1`, each time the results are saved a message is written to `stderr`. If `>=2`, a message is output each time data is collected for any tensor. Default verbosity level is `1`.
-* `-ofreq` (or `--output-frequency`) specifies how often the so far computed result is saved to disk. Default is 10 (i.e., every 10 chunks)
-* `-ow` (or `--output-weight`) specifies if data will be collected for the `output.weight` tensor. My experience is that it is better to not utilize the importance matrix when quantizing `output.weight`, so this is set to `false` by default.
-
-For faster computation, make sure to use GPU offloading via the `-ngl` argument
-
-## Example
-
-```bash
-LLAMA_CUBLAS=1 make -j
-
-# generate importance matrix (imatrix.dat)
-./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99
-
-# use the imatrix to perform a Q4_K_M quantization
-./quantize --imatrix imatrix.dat ggml-model-f16.gguf ./ggml-model-q4_k_m.gguf q4_k_m
-```

examples/imatrix/imatrix.cpp

@@ -1,513 +0,0 @@
-#include "common.h"
-#include "llama.h"
-
-#include <cmath>
-#include <cstdio>
-#include <cstring>
-#include <ctime>
-#include <sstream>
-#include <thread>
-#include <mutex>
-#include <vector>
-#include <fstream>
-#include <unordered_map>
-#include <algorithm>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#endif
-
-struct Stats {
-    std::vector<float> values;
-    int ncall = 0;
-};
-
-struct StatParams {
-    std::string ofile = "imatrix.dat";
-    int         n_output_frequency = 10;
-    int         verbosity = 1;
-    int         keep_every = 0;
-    bool        collect_output_weight = false;
-};
-
-class IMatrixCollector {
-public:
-    IMatrixCollector() = default;
-    void set_parameters(StatParams&& params) { m_params = std::move(params); }
-    bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
-    void save_imatrix() const;
-private:
-    std::unordered_map<std::string, Stats> m_stats;
-    StatParams                             m_params;
-    std::mutex                             m_mutex;
-    int                                    m_last_call = 0;
-    std::vector<float>                     m_src1_data;
-    std::vector<int>                       m_ids; // the expert ids from ggml_mul_mat_id
-                                                  //
-    void save_imatrix(const char * file_name) const;
-    void keep_imatrix(int ncall) const;
-};
-
-bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
-    GGML_UNUSED(user_data);
-
-    const struct ggml_tensor * src0 = t->src[0];
-    const struct ggml_tensor * src1 = t->src[1];
-
-    // when ask is true, the scheduler wants to know if we are interested in data from this tensor
-    // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
-    if (ask) {
-        if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
-        if (t->op != GGML_OP_MUL_MAT) return false;
-        if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
-        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
-        return true;
-    }
-
-    std::lock_guard<std::mutex> lock(m_mutex);
-
-    // copy the data from the GPU memory if needed
-    const bool is_host = ggml_backend_buffer_is_host(src1->buffer);
-
-    if (!is_host) {
-        m_src1_data.resize(ggml_nelements(src1));
-        ggml_backend_tensor_get(src1, m_src1_data.data(), 0, ggml_nbytes(src1));
-    }
-
-    const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
-
-    if (t->op == GGML_OP_MUL_MAT_ID) {
-        const int idx  = ((int32_t *) t->op_params)[0];
-        const int n_as = ((int32_t *) t->op_params)[1];
-
-        // the top-k selected expert ids are stored in the src0 tensor
-        // for simplicity, always copy src0 to host, because it is small
-        // take into account that src0 is not contiguous!
-        GGML_ASSERT(src0->ne[1] == src1->ne[1]);
-        GGML_ASSERT(n_as*ggml_nrows(src0)*sizeof(int) == GGML_PAD(ggml_nbytes(src0), n_as*sizeof(int)));
-        m_ids.resize(ggml_nbytes(src0)/sizeof(int));
-        ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0));
-
-        // loop over all possible experts, regardless if they are used or not in the batch
-        // this is necessary to guarantee equal number of "ncall" for each tensor
-        for (int ex = 0; ex < n_as; ++ex) {
-            src0 = t->src[2 + ex];
-            auto& e = m_stats[src0->name];
-            if (e.values.empty()) {
-                e.values.resize(src1->ne[0], 0);
-            }
-            else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
-                exit(1); //GGML_ASSERT(false);
-            }
-            // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
-            //       using the following line, we can correct for that if needed
-            //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
-            ++e.ncall;
-            if (m_params.verbosity > 1) {
-                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
-            }
-            for (int row = 0; row < (int)src1->ne[1]; ++row) {
-                const int excur = m_ids[row*n_as + idx];
-                GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
-                if (excur != ex) continue;
-                const float * x = data + row * src1->ne[0];
-                for (int j = 0; j < (int)src1->ne[0]; ++j) {
-                    e.values[j] += x[j]*x[j];
-                }
-            }
-            if (e.ncall > m_last_call) {
-                m_last_call = e.ncall;
-                if (m_last_call % m_params.n_output_frequency == 0) {
-                    save_imatrix();
-                }
-                if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
-                    keep_imatrix(m_last_call);
-                }
-            }
-        }
-    } else {
-        auto& e = m_stats[src0->name];
-        if (e.values.empty()) {
-            e.values.resize(src1->ne[0], 0);
-        }
-        else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
-            exit(1); //GGML_ASSERT(false);
-        }
-        ++e.ncall;
-        if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
-        }
-        for (int row = 0; row < (int)src1->ne[1]; ++row) {
-            const float * x = data + row * src1->ne[0];
-            for (int j = 0; j < (int)src1->ne[0]; ++j) {
-                e.values[j] += x[j]*x[j];
-            }
-        }
-        if (e.ncall > m_last_call) {
-            m_last_call = e.ncall;
-            if (m_last_call % m_params.n_output_frequency == 0) {
-                save_imatrix();
-            }
-            if (m_params.keep_every > 0 && m_last_call%m_params.keep_every == 0) {
-                keep_imatrix(m_last_call);
-            }
-        }
-    }
-
-    return true;
-}
-
-void IMatrixCollector::save_imatrix() const {
-    save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str());
-}
-
-void IMatrixCollector::keep_imatrix(int ncall) const {
-    auto file_name = m_params.ofile;
-    if (file_name.empty()) file_name = "imatrix.dat";
-    file_name += ".at_";
-    file_name += std::to_string(ncall);
-    save_imatrix(file_name.c_str());
-}
-
-void IMatrixCollector::save_imatrix(const char * fname) const {
-    std::ofstream out(fname, std::ios::binary);
-    int n_entries = m_stats.size();
-    out.write((const char*)&n_entries, sizeof(n_entries));
-    for (auto& p : m_stats) {
-        int len = p.first.size();
-        out.write((const char*)&len, sizeof(len));
-        out.write(p.first.c_str(), len);
-        out.write((const char*)&p.second.ncall, sizeof(p.second.ncall));
-        int nval = p.second.values.size();
-        out.write((const char*)&nval, sizeof(nval));
-        if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float));
-    }
-    if (m_params.verbosity > 0) {
-        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname);
-    }
-}
-
-static IMatrixCollector g_collector;
-
-static bool ik_collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
-    return g_collector.collect_imatrix(t, ask, user_data);
-}
-
-
-struct results_log_softmax {
-    double log_softmax;
-    float  logit;
-    float  prob;
-};
-
-static std::vector<float> softmax(const std::vector<float>& logits) {
-    std::vector<float> probs(logits.size());
-    float max_logit = logits[0];
-    for (float v : logits) {
-        max_logit = std::max(max_logit, v);
-    }
-    double sum_exp = 0.0;
-    for (size_t i = 0; i < logits.size(); i++) {
-        // Subtract the maximum logit value from the current logit value for numerical stability
-        const float logit = logits[i] - max_logit;
-        const float exp_logit = expf(logit);
-        sum_exp += exp_logit;
-        probs[i] = exp_logit;
-    }
-    for (size_t i = 0; i < probs.size(); i++) {
-        probs[i] /= sum_exp;
-    }
-    return probs;
-}
-
-static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
-    float max_logit = logits[0];
-    for (int i = 1; i < n_vocab; ++i) {
-        max_logit = std::max(max_logit, logits[i]);
-    }
-    double sum_exp = 0.0;
-    for (int i = 0; i < n_vocab; ++i) {
-        sum_exp += expf(logits[i] - max_logit);
-    }
-    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
-}
-
-static void process_logits(
-    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
-    double & nll, double & nll2, float * logit_history, float * prob_history
-) {
-    std::mutex mutex;
-    int counter = 0;
-    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
-        double local_nll  = 0;
-        double local_nll2 = 0;
-        while (true) {
-            std::unique_lock<std::mutex> lock(mutex);
-            int i = counter++;
-            if (i >= n_token) {
-                nll += local_nll; nll2 += local_nll2;
-                break;
-            }
-            lock.unlock();
-            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
-            const double v = -results.log_softmax;
-            local_nll += v;
-            local_nll2 += v*v;
-
-            logit_history[i] = results.logit;
-            prob_history[i]  = results.prob;
-        }
-    };
-    for (auto & w : workers) {
-        w = std::thread(compute);
-    }
-    compute();
-    for (auto & w : workers) {
-        w.join();
-    }
-}
-
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl) {
-
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    const int n_ctx = llama_n_ctx(ctx);
-
-    auto tim1 = std::chrono::high_resolution_clock::now();
-    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
-
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
-
-    auto tim2 = std::chrono::high_resolution_clock::now();
-    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
-
-    if (int(tokens.size()) < 2*n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
-                n_ctx);
-        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
-        return false;
-    }
-
-    std::vector<float> logit_history;
-    std::vector<float> prob_history;
-
-    if (compute_ppl) {
-        logit_history.resize(tokens.size());
-        prob_history.resize(tokens.size());
-    }
-
-    const int n_chunk_max = tokens.size() / n_ctx;
-
-    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-    const int n_batch = params.n_batch;
-
-    int count = 0;
-    double nll = 0.0;
-    double nll2 = 0.0;
-
-    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
-
-    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
-
-    const int num_batches = (n_ctx + n_batch - 1) / n_batch;
-
-    std::vector<float> logits;
-    if (compute_ppl && num_batches > 1) {
-        logits.reserve((size_t)n_ctx * n_vocab);
-    }
-
-    for (int i = 0; i < n_chunk; ++i) {
-        const int start =     i * n_ctx;
-        const int end   = start + n_ctx;
-
-        std::vector<float> logits;
-
-        const auto t_start = std::chrono::high_resolution_clock::now();
-
-        // clear the KV cache
-        llama_kv_cache_clear(ctx);
-
-        for (int j = 0; j < num_batches; ++j) {
-            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];
-
-            // add BOS token for the first batch of each chunk
-            if (add_bos && j == 0) {
-                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
-            }
-
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
-                fprintf(stderr, "%s : failed to eval\n", __func__);
-                return false;
-            }
-
-            // restore the original token in case it was set to BOS
-            tokens[batch_start] = token_org;
-
-            if (compute_ppl && num_batches > 1) {
-                const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
-            }
-        }
-
-        const auto t_end = std::chrono::high_resolution_clock::now();
-
-        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);
-            if (total_seconds >= 60*60) {
-                fprintf(stderr, "%d hours ", total_seconds / (60*60));
-                total_seconds = total_seconds % (60*60);
-            }
-            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
-        }
-
-        if (compute_ppl) {
-            const int first = n_ctx/2;
-            const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
-            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;
-
-            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
-            fflush(stdout);
-
-            logits.clear();
-        }
-    }
-    printf("\n");
-
-    if (compute_ppl) {
-        nll2 /= count;
-        nll /= count;
-        const double ppl = exp(nll);
-        nll2 -= nll * nll;
-        if (nll2 > 0) {
-            nll2 = sqrt(nll2/(count-1));
-            printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
-        } else {
-            printf("Unexpected negative standard deviation of log(prob)\n");
-        }
-    }
-
-    return true;
-}
-
-int main(int argc, char ** argv) {
-
-    StatParams sparams;
-    bool compute_ppl = true;
-    std::vector<char*> args;
-    args.push_back(argv[0]);
-    int iarg = 1;
-    for (; iarg < argc-1; ++iarg) {
-        std::string arg{argv[iarg]};
-        if (arg == "-o" || arg == "--output-file") {
-            sparams.ofile = argv[++iarg];
-        }
-        else if (arg == "-ofreq" || arg == "--output-frequency") {
-            sparams.n_output_frequency = std::stoi(argv[++iarg]);
-        }
-        else if (arg == "-ow" || arg == "--output-weight") {
-            sparams.collect_output_weight = std::stoi(argv[++iarg]);
-        }
-        else if (arg == "--verbosity") {
-            sparams.verbosity = std::stoi(argv[++iarg]);
-        } else if (arg == "--no-ppl") {
-            compute_ppl = false;
-        } else if (arg == "--keep-imatrix") {
-            sparams.keep_every = std::stoi(argv[++iarg]);
-        } else {
-            args.push_back(argv[iarg]);
-        }
-    }
-    if (iarg < argc) {
-        std::string arg{argv[iarg]};
-        if (arg == "--no-ppl") {
-            compute_ppl = false;
-        } else {
-            args.push_back(argv[iarg]);
-        }
-    }
-
-    gpt_params params;
-    params.n_batch = 512;
-    if (!gpt_params_parse(args.size(), args.data(), params)) {
-        return 1;
-    }
-
-    g_collector.set_parameters(std::move(sparams));
-
-    params.logits_all = true;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
-
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
-    }
-
-    llama_backend_init(params.numa);
-
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        return 1;
-    }
-
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
-    // pass the callback to the backend scheduler
-    // it will be executed for each node during the graph computation
-    cparams.cb_eval = ik_collect_imatrix;
-    cparams.cb_eval_user_data = NULL;
-
-    llama_context * ctx = llama_new_context_with_model(model, cparams);
-    if (ctx == NULL) {
-        fprintf(stderr, "%s: error: unable to create context\n", __func__);
-        return 1;
-    }
-
-    const int n_ctx_train = llama_n_ctx_train(model);
-    if (params.n_ctx > n_ctx_train) {
-        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, params.n_ctx);
-    }
-
-    // print system information
-    {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", get_system_info(params).c_str());
-    }
-
-    bool OK = compute_imatrix(ctx, params, compute_ppl);
-    if (!OK) {
-        return 1;
-    }
-
-    g_collector.save_imatrix();
-
-    llama_print_timings(ctx);
-
-    llama_free(ctx);
-    llama_free_model(model);
-
-    llama_backend_free();
-
-    return 0;
-}

examples/infill/CMakeLists.txt

@@ -3,3 +3,6 @@ add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+    add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/infill/infill.cpp

@@ -2,6 +2,7 @@
 
 #include "console.h"
 #include "llama.h"
+#include "build-info.h"
 #include "grammar-parser.h"
 
 #include <cassert>
@@ -146,13 +147,6 @@ int main(int argc, char ** argv) {
 
         return 0;
     }
-    if (params.chatml) {
-        printf("\n************\n");
-        printf("%s: please use the 'main' tool for chatml mode\n", __func__);
-        printf("************\n\n");
-
-        return 0;
-    }
     if (!params.antiprompt.empty()) {
         printf("\n************\n");
         printf("%s: please use the 'main' tool for antiprompt mode\n", __func__);
@@ -190,8 +184,8 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -237,11 +231,11 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", get_system_info(params).c_str());
     }
-    const bool add_bos = llama_should_add_bos_token(model);
+    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
     LOG("add_bos: %d\n", add_bos);
 
     bool suff_rm_leading_spc = params.escape;
-    if (suff_rm_leading_spc && params.input_suffix.find_first_of(' ') == 0 && params.input_suffix.size() > 1) {
+    if (suff_rm_leading_spc && params.input_suffix.find_first_of(" ") == 0 && params.input_suffix.size() > 1) {
         params.input_suffix.erase(0, 1);
         suff_rm_leading_spc = false;
     }

examples/llama-bench/CMakeLists.txt

@@ -3,3 +3,6 @@ add_executable(${TARGET} llama-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+if(TARGET BUILD_INFO)
+  add_dependencies(${TARGET} BUILD_INFO)
+endif()

examples/llama-bench/llama-bench.cpp

@@ -19,6 +19,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "build-info.h"
 #include "ggml-cuda.h"
 
 // utils
@@ -53,13 +54,6 @@ static std::vector<T> split(const std::string & str, char delim) {
     return values;
 }
 
-template<typename T, typename F>
-static std::vector<std::string> transform_to_str(const std::vector<T> & values, F f) {
-    std::vector<std::string> str_values;
-    std::transform(values.begin(), values.end(), std::back_inserter(str_values), f);
-    return str_values;
-}
-
 template<typename T>
 static T avg(const std::vector<T> & v) {
     if (v.empty()) {
@@ -128,37 +122,15 @@ static std::string get_gpu_info() {
 // command line params
 enum output_formats {CSV, JSON, MARKDOWN, SQL};
 
-static const char * output_format_str(output_formats format) {
-    switch (format) {
-        case CSV:      return "csv";
-        case JSON:     return "json";
-        case MARKDOWN: return "md";
-        case SQL:      return "sql";
-        default: GGML_ASSERT(!"invalid output 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";
-        default: GGML_ASSERT(!"invalid split mode");
-    }
-}
-
 struct cmd_params {
     std::vector<std::string> model;
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<ggml_type> type_k;
-    std::vector<ggml_type> type_v;
+    std::vector<bool> f32_kv;
     std::vector<int> n_threads;
     std::vector<int> n_gpu_layers;
-    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::array<float, LLAMA_MAX_DEVICES>> tensor_split;
     int reps;
@@ -171,13 +143,10 @@ static const cmd_params cmd_params_defaults = {
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
     /* n_batch       */ {512},
-    /* type_k        */ {GGML_TYPE_F16},
-    /* type_v        */ {GGML_TYPE_F16},
+    /* f32_kv        */ {false},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
-    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
-    /* no_kv_offload */ {false},
     /* mul_mat_q     */ {true},
     /* tensor_split  */ {{}},
     /* reps          */ 5,
@@ -190,50 +159,24 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help\n");
-    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
-    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("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
-    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
-    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
-    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
-    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
+    printf("  -m, --model <filename>            (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                   (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -b, --batch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  --memory-f32 <0|1>                (default: %s)\n", join(cmd_params_defaults.f32_kv, ",").c_str());
+    printf("  -t, --threads <n>                 (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>          (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -mg, --main-gpu <i>               (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -mmq, --mul-mat-q <0|1>           (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>               \n");
+    printf("  -r, --repetitions <n>             (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>    (default: %s)\n", cmd_params_defaults.output_format == CSV ? "csv" : cmd_params_defaults.output_format == JSON ? "json" : cmd_params_defaults.output_format == MARKDOWN ? "md" : "sql");
+    printf("  -v, --verbose                     (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
+
 }
 
-static ggml_type ggml_type_from_name(const std::string & s) {
-    if (s == "f16") {
-        return GGML_TYPE_F16;
-    }
-    if (s == "q8_0") {
-        return GGML_TYPE_Q8_0;
-    }
-    if (s == "q4_0") {
-        return GGML_TYPE_Q4_0;
-    }
-    if (s == "q4_1") {
-        return GGML_TYPE_Q4_1;
-    }
-    if (s == "q5_0") {
-        return GGML_TYPE_Q5_0;
-    }
-    if (s == "q5_1") {
-        return GGML_TYPE_Q5_1;
-    }
-
-    return GGML_TYPE_COUNT;
-}
-
-
 static cmd_params parse_cmd_params(int argc, char ** argv) {
     cmd_params params;
     std::string arg;
@@ -282,38 +225,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "-ctk" || arg == "--cache-type-k") {
+        } else if (arg == "--memory-f32") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            auto p = split<std::string>(argv[i], split_delim);
-            std::vector<ggml_type> types;
-            for (const auto & t : p) {
-                ggml_type gt = ggml_type_from_name(t);
-                if (gt == GGML_TYPE_COUNT) {
-                    invalid_param = true;
-                    break;
-                }
-                types.push_back(gt);
-            }
-            params.type_k.insert(params.type_k.end(), types.begin(), types.end());
-        } else if (arg == "-ctv" || arg == "--cache-type-v") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<std::string>(argv[i], split_delim);
-            std::vector<ggml_type> types;
-            for (const auto & t : p) {
-                ggml_type gt = ggml_type_from_name(t);
-                if (gt == GGML_TYPE_COUNT) {
-                    invalid_param = true;
-                    break;
-                }
-                types.push_back(gt);
-            }
-            params.type_v.insert(params.type_v.end(), types.begin(), types.end());
+            auto p = split<int>(argv[i], split_delim);
+            params.f32_kv.insert(params.f32_kv.end(), p.begin(), p.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -328,41 +246,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
-        } else if (arg == "-sm" || arg == "--split-mode") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<std::string>(argv[i], split_delim);
-            std::vector<llama_split_mode> modes;
-            for (const auto & m : p) {
-                llama_split_mode mode;
-                if (m == "none") {
-                    mode = LLAMA_SPLIT_NONE;
-                } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_LAYER;
-                } else if (m == "row") {
-                    mode = LLAMA_SPLIT_ROW;
-                } else {
-                    invalid_param = true;
-                    break;
-                }
-                modes.push_back(mode);
-            }
-            params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.main_gpu = split<int>(argv[i], split_delim);
-        } else if (arg == "-nkvo" || arg == "--no-kv-offload") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            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;
@@ -433,12 +322,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
-    if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
+    if (params.f32_kv.empty())       { params.f32_kv = cmd_params_defaults.f32_kv; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
-    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.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
@@ -451,13 +337,10 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    ggml_type type_k;
-    ggml_type type_v;
+    bool f32_kv;
     int n_threads;
     int n_gpu_layers;
-    llama_split_mode split_mode;
     int main_gpu;
-    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
 
@@ -465,7 +348,6 @@ struct cmd_params_instance {
         llama_model_params mparams = llama_model_default_params();
 
         mparams.n_gpu_layers = n_gpu_layers;
-        mparams.split_mode = split_mode;
         mparams.main_gpu = main_gpu;
         mparams.tensor_split = tensor_split.data();
 
@@ -475,7 +357,6 @@ struct cmd_params_instance {
     bool equal_mparams(const cmd_params_instance & other) const {
         return model == other.model &&
                n_gpu_layers == other.n_gpu_layers &&
-               split_mode == other.split_mode &&
                main_gpu == other.main_gpu &&
                tensor_split == other.tensor_split;
     }
@@ -485,29 +366,53 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.type_k = type_k;
-        cparams.type_v = type_v;
+        cparams.f16_kv = !f32_kv;
         cparams.mul_mat_q = mul_mat_q;
-        cparams.offload_kqv = !no_kv_offload;
 
         return cparams;
     }
 };
 
-static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
+static std::vector<cmd_params_instance> get_cmd_params_instances_int(const cmd_params & params, int n_gen, int n_prompt) {
     std::vector<cmd_params_instance> instances;
 
-    // this ordering minimizes the number of times that each model needs to be reloaded
     for (const auto & m : params.model)
     for (const auto & nl : params.n_gpu_layers)
-    for (const auto & sm : params.split_mode)
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & nb : params.n_batch)
-    for (const auto & tk : params.type_k)
-    for (const auto & tv : params.type_v)
+    for (const auto & fk : params.f32_kv)
+    for (const auto & mmq : params.mul_mat_q)
+    for (const auto & nt : params.n_threads) {
+        cmd_params_instance instance = {
+            /* .model        = */ m,
+            /* .n_prompt     = */ n_prompt,
+            /* .n_gen        = */ n_gen,
+            /* .n_batch      = */ nb,
+            /* .f32_kv       = */ fk,
+            /* .n_threads    = */ nt,
+            /* .n_gpu_layers = */ nl,
+            /* .main_gpu     = */ mg,
+            /* .mul_mat_q    = */ mmq,
+            /* .tensor_split = */ ts,
+        };
+        instances.push_back(instance);
+    }
+    return instances;
+}
+
+static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_params & params) {
+    std::vector<cmd_params_instance> instances;
+
+#if 1
+    // this ordering minimizes the number of times that each model needs to be reloaded
+    for (const auto & m : params.model)
+    for (const auto & nl : params.n_gpu_layers)
+    for (const auto & mg : params.main_gpu)
+    for (const auto & ts : params.tensor_split)
+    for (const auto & nb : params.n_batch)
+    for (const auto & fk : params.f32_kv)
     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) {
             if (n_prompt == 0) {
@@ -518,13 +423,10 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .type_k       = */ tk,
-                /* .type_v       = */ tv,
+                /* .f32_kv       = */ fk,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
-                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
-                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
@@ -540,19 +442,34 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .type_k       = */ tk,
-                /* .type_v       = */ tv,
+                /* .f32_kv       = */ fk,
                 /* .n_threads    = */ nt,
                 /* .n_gpu_layers = */ nl,
-                /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
-                /* .no_kv_offload= */ nkvo,
                 /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
             };
             instances.push_back(instance);
         }
     }
+#else
+    // this ordering separates the prompt and generation tests
+    for (const auto & n_prompt : params.n_prompt) {
+        if (n_prompt == 0) {
+            continue;
+        }
+        auto instances_prompt = get_cmd_params_instances_int(params, 0, n_prompt);
+        instances.insert(instances.end(), instances_prompt.begin(), instances_prompt.end());
+    }
+
+    for (const auto & n_gen : params.n_gen) {
+        if (n_gen == 0) {
+            continue;
+        }
+        auto instances_gen = get_cmd_params_instances_int(params, n_gen, 0);
+        instances.insert(instances.end(), instances_gen.begin(), instances_gen.end());
+    }
+#endif
 
     return instances;
 }
@@ -562,7 +479,6 @@ struct test {
     static const int build_number;
     static const bool cuda;
     static const bool opencl;
-    static const bool vulkan;
     static const bool metal;
     static const bool gpu_blas;
     static const bool blas;
@@ -574,12 +490,9 @@ struct test {
     uint64_t model_n_params;
     int n_batch;
     int n_threads;
-    ggml_type type_k;
-    ggml_type type_v;
+    bool f32_kv;
     int n_gpu_layers;
-    llama_split_mode split_mode;
     int main_gpu;
-    bool no_kv_offload;
     bool mul_mat_q;
     std::array<float, LLAMA_MAX_DEVICES> tensor_split;
     int n_prompt;
@@ -596,12 +509,9 @@ struct test {
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
         n_threads = inst.n_threads;
-        type_k = inst.type_k;
-        type_v = inst.type_v;
+        f32_kv = inst.f32_kv;
         n_gpu_layers = inst.n_gpu_layers;
-        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;
         n_prompt = inst.n_prompt;
@@ -644,9 +554,6 @@ struct test {
         if (opencl) {
             return "OpenCL";
         }
-        if (vulkan) {
-            return "Vulkan";
-        }
         if (metal) {
             return "Metal";
         }
@@ -662,13 +569,11 @@ struct test {
     static const std::vector<std::string> & get_fields() {
         static const std::vector<std::string> fields = {
             "build_commit", "build_number",
-            "cuda", "opencl", "vulkan", "metal", "gpu_blas", "blas",
+            "cuda", "opencl", "metal", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_threads", "type_k", "type_v",
-            "n_gpu_layers", "split_mode",
-            "main_gpu", "no_kv_offload",
-            "mul_mat_q", "tensor_split",
+            "n_batch", "n_threads", "f16_kv",
+            "n_gpu_layers", "main_gpu", "mul_mat_q", "tensor_split",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -686,8 +591,8 @@ struct test {
             field == "avg_ns" || field == "stddev_ns") {
             return INT;
         }
-        if (field == "cuda" || field == "opencl"  || field == "vulkan"|| field == "metal" || field == "gpu_blas" || field == "blas" ||
-            field == "f16_kv" || field == "no_kv_offload" || field == "mul_mat_q") {
+        if (field == "cuda" || field == "opencl" || field == "metal" || field == "gpu_blas" || field == "blas" ||
+            field == "f16_kv" || field == "mul_mat_q") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -714,13 +619,11 @@ struct test {
         }
         std::vector<std::string> values = {
             build_commit, std::to_string(build_number),
-            std::to_string(cuda), std::to_string(opencl), std::to_string(vulkan), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
+            std::to_string(cuda), std::to_string(opencl), std::to_string(metal), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            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(n_batch), std::to_string(n_threads), std::to_string(!f32_kv),
+            std::to_string(n_gpu_layers), std::to_string(main_gpu), std::to_string(mul_mat_q), tensor_split_str,
             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())
@@ -738,11 +641,10 @@ struct test {
     }
 };
 
-const std::string test::build_commit = LLAMA_COMMIT;
-const int         test::build_number = LLAMA_BUILD_NUMBER;
+const std::string test::build_commit = BUILD_COMMIT;
+const int         test::build_number = BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
-const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::metal        = !!ggml_cpu_has_metal();
 const bool        test::gpu_blas     = !!ggml_cpu_has_gpublas();
 const bool        test::blas         = !!ggml_cpu_has_blas();
@@ -876,18 +778,12 @@ struct markdown_printer : public printer {
         if (field == "n_gpu_layers") {
             return "ngl";
         }
-        if (field == "split_mode") {
-            return "sm";
-        }
         if (field == "n_threads") {
             return "threads";
         }
         if (field == "mul_mat_q") {
             return "mmq";
         }
-        if (field == "no_kv_offload") {
-            return "nkvo";
-        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -910,24 +806,15 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.push_back("n_batch");
         }
-        if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
-            fields.push_back("type_k");
-        }
-        if (params.type_v.size() > 1 || params.type_v != cmd_params_defaults.type_v) {
-            fields.push_back("type_v");
+        if (params.f32_kv.size() > 1 || params.f32_kv != cmd_params_defaults.f32_kv) {
+            fields.push_back("f16_kv");
         }
         if (params.main_gpu.size() > 1 || params.main_gpu != cmd_params_defaults.main_gpu) {
             fields.push_back("main_gpu");
         }
-        if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
-            fields.push_back("split_mode");
-        }
         if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
             fields.push_back("mul_mat_q");
         }
-        if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
-            fields.push_back("no_kv_offload");
-        }
         if (params.tensor_split.size() > 1 || params.tensor_split != cmd_params_defaults.tensor_split) {
             fields.push_back("tensor_split");
         }
@@ -1150,7 +1037,7 @@ int main(int argc, char ** argv) {
 
         test t(inst, lmodel, ctx);
 
-        llama_kv_cache_clear(ctx);
+        llama_kv_cache_tokens_rm(ctx, -1, -1);
 
         // warmup run
         if (t.n_prompt > 0) {
@@ -1161,7 +1048,7 @@ int main(int argc, char ** argv) {
         }
 
         for (int i = 0; i < params.reps; i++) {
-            llama_kv_cache_clear(ctx);
+            llama_kv_cache_tokens_rm(ctx, -1, -1);
 
             uint64_t t_start = get_time_ns();
             if (t.n_prompt > 0) {

examples/llama.android/.gitignore

@@ -1,33 +0,0 @@
-# Gradle files
-.gradle/
-build/
-
-# Local configuration file (sdk path, etc)
-local.properties
-
-# Log/OS Files
-*.log
-
-# Android Studio generated files and folders
-captures/
-.externalNativeBuild/
-.cxx/
-*.apk
-output.json
-
-# IntelliJ
-*.iml
-.idea/
-misc.xml
-deploymentTargetDropDown.xml
-render.experimental.xml
-
-# Keystore files
-*.jks
-*.keystore
-
-# Google Services (e.g. APIs or Firebase)
-google-services.json
-
-# Android Profiling
-*.hprof

examples/llama.android/app/.gitignore

@@ -1 +0,0 @@
-/build

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

@@ -1,92 +0,0 @@
-plugins {
-    id("com.android.application")
-    id("org.jetbrains.kotlin.android")
-}
-
-android {
-    namespace = "com.example.llama"
-    compileSdk = 34
-
-    ndkVersion = "26.1.10909125"
-
-    defaultConfig {
-        applicationId = "com.example.llama"
-        minSdk = 33
-        targetSdk = 34
-        versionCode = 1
-        versionName = "1.0"
-
-        testInstrumentationRunner = "androidx.test.runner.AndroidJUnitRunner"
-        vectorDrawables {
-            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")
-            }
-        }
-        externalNativeBuild {
-            cmake {
-                arguments += "-DCMAKE_BUILD_TYPE=Release"
-                cppFlags += listOf()
-                arguments += listOf()
-            }
-        }
-    }
-
-    buildTypes {
-        release {
-            isMinifyEnabled = false
-            proguardFiles(
-                getDefaultProguardFile("proguard-android-optimize.txt"),
-                "proguard-rules.pro"
-            )
-        }
-    }
-    compileOptions {
-        sourceCompatibility = JavaVersion.VERSION_1_8
-        targetCompatibility = JavaVersion.VERSION_1_8
-    }
-    kotlinOptions {
-        jvmTarget = "1.8"
-    }
-    buildFeatures {
-        compose = true
-    }
-    composeOptions {
-        kotlinCompilerExtensionVersion = "1.5.1"
-    }
-    packaging {
-        resources {
-            excludes += "/META-INF/{AL2.0,LGPL2.1}"
-        }
-    }
-    externalNativeBuild {
-        cmake {
-            path = file("src/main/cpp/CMakeLists.txt")
-            version = "3.22.1"
-        }
-    }
-}
-
-dependencies {
-
-    implementation("androidx.core:core-ktx:1.12.0")
-    implementation("androidx.lifecycle:lifecycle-runtime-ktx:2.6.2")
-    implementation("androidx.activity:activity-compose:1.8.2")
-    implementation(platform("androidx.compose:compose-bom:2023.08.00"))
-    implementation("androidx.compose.ui:ui")
-    implementation("androidx.compose.ui:ui-graphics")
-    implementation("androidx.compose.ui:ui-tooling-preview")
-    implementation("androidx.compose.material3:material3")
-    testImplementation("junit:junit:4.13.2")
-    androidTestImplementation("androidx.test.ext:junit:1.1.5")
-    androidTestImplementation("androidx.test.espresso:espresso-core:3.5.1")
-    androidTestImplementation(platform("androidx.compose:compose-bom:2023.08.00"))
-    androidTestImplementation("androidx.compose.ui:ui-test-junit4")
-    debugImplementation("androidx.compose.ui:ui-tooling")
-    debugImplementation("androidx.compose.ui:ui-test-manifest")
-}

examples/llama.android/app/proguard-rules.pro

@@ -1,21 +0,0 @@
-# Add project specific ProGuard rules here.
-# You can control the set of applied configuration files using the
-# proguardFiles setting in build.gradle.
-#
-# For more details, see
-#   http://developer.android.com/guide/developing/tools/proguard.html
-
-# If your project uses WebView with JS, uncomment the following
-# and specify the fully qualified class name to the JavaScript interface
-# class:
-#-keepclassmembers class fqcn.of.javascript.interface.for.webview {
-#   public *;
-#}
-
-# Uncomment this to preserve the line number information for
-# debugging stack traces.
-#-keepattributes SourceFile,LineNumberTable
-
-# If you keep the line number information, uncomment this to
-# hide the original source file name.
-#-renamesourcefileattribute SourceFile

examples/llama.android/app/src/main/AndroidManifest.xml

@@ -1,30 +0,0 @@
-<?xml version="1.0" encoding="utf-8"?>
-<manifest xmlns:android="http://schemas.android.com/apk/res/android"
-    xmlns:tools="http://schemas.android.com/tools">
-
-    <uses-permission android:name="android.permission.INTERNET" />
-
-    <application
-        android:allowBackup="true"
-        android:dataExtractionRules="@xml/data_extraction_rules"
-        android:fullBackupContent="@xml/backup_rules"
-        android:icon="@mipmap/ic_launcher"
-        android:label="@string/app_name"
-        android:roundIcon="@mipmap/ic_launcher_round"
-        android:supportsRtl="true"
-        android:theme="@style/Theme.LlamaAndroid"
-        >
-
-        <activity
-            android:name=".MainActivity"
-            android:exported="true"
-            android:theme="@style/Theme.LlamaAndroid">
-            <intent-filter>
-                <action android:name="android.intent.action.MAIN" />
-
-                <category android:name="android.intent.category.LAUNCHER" />
-            </intent-filter>
-        </activity>
-    </application>
-
-</manifest>

examples/llama.android/app/src/main/cpp/CMakeLists.txt

@@ -1,50 +0,0 @@
-
-# For more information about using CMake with Android Studio, read the
-# documentation: https://d.android.com/studio/projects/add-native-code.html.
-# For more examples on how to use CMake, see https://github.com/android/ndk-samples.
-
-# Sets the minimum CMake version required for this project.
-cmake_minimum_required(VERSION 3.22.1)
-
-# Declares the project name. The project name can be accessed via ${ PROJECT_NAME},
-# Since this is the top level CMakeLists.txt, the project name is also accessible
-# with ${CMAKE_PROJECT_NAME} (both CMake variables are in-sync within the top level
-# build script scope).
-project("llama-android")
-
-include(FetchContent)
-FetchContent_Declare(
-        llama
-        GIT_REPOSITORY https://github.com/ggerganov/llama.cpp
-        GIT_TAG        master
-)
-
-# Also provides "common"
-FetchContent_MakeAvailable(llama)
-
-# Creates and names a library, sets it as either STATIC
-# or SHARED, and provides the relative paths to its source code.
-# You can define multiple libraries, and CMake builds them for you.
-# Gradle automatically packages shared libraries with your APK.
-#
-# In this top level CMakeLists.txt, ${CMAKE_PROJECT_NAME} is used to define
-# the target library name; in the sub-module's CMakeLists.txt, ${PROJECT_NAME}
-# is preferred for the same purpose.
-#
-# In order to load a library into your app from Java/Kotlin, you must call
-# System.loadLibrary() and pass the name of the library defined here;
-# for GameActivity/NativeActivity derived applications, the same library name must be
-# used in the AndroidManifest.xml file.
-add_library(${CMAKE_PROJECT_NAME} SHARED
-    # List C/C++ source files with relative paths to this CMakeLists.txt.
-    llama-android.cpp)
-
-# Specifies libraries CMake should link to your target library. You
-# can link libraries from various origins, such as libraries defined in this
-# build script, prebuilt third-party libraries, or Android system libraries.
-target_link_libraries(${CMAKE_PROJECT_NAME}
-    # List libraries link to the target library
-    llama
-    common
-    android
-    log)

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

@@ -1,394 +0,0 @@
-#include <android/log.h>
-#include <jni.h>
-#include <iomanip>
-#include <math.h>
-#include <string>
-#include <unistd.h>
-#include "llama.h"
-#include "common/common.h"
-
-// Write C++ code here.
-//
-// Do not forget to dynamically load the C++ library into your application.
-//
-// For instance,
-//
-// In MainActivity.java:
-//    static {
-//       System.loadLibrary("llama-android");
-//    }
-//
-// Or, in MainActivity.kt:
-//    companion object {
-//      init {
-//         System.loadLibrary("llama-android")
-//      }
-//    }
-
-#define TAG "llama-android.cpp"
-#define LOGi(...) __android_log_print(ANDROID_LOG_INFO, TAG, __VA_ARGS__)
-#define LOGe(...) __android_log_print(ANDROID_LOG_ERROR, TAG, __VA_ARGS__)
-
-jclass la_int_var;
-jmethodID la_int_var_value;
-jmethodID la_int_var_inc;
-
-static void log_callback(ggml_log_level level, const char * fmt, void * data) {
-    if (level == GGML_LOG_LEVEL_ERROR)     __android_log_print(ANDROID_LOG_ERROR, TAG, fmt, data);
-    else if (level == GGML_LOG_LEVEL_INFO) __android_log_print(ANDROID_LOG_INFO, TAG, fmt, data);
-    else if (level == GGML_LOG_LEVEL_WARN) __android_log_print(ANDROID_LOG_WARN, TAG, fmt, data);
-    else __android_log_print(ANDROID_LOG_DEFAULT, TAG, fmt, data);
-}
-
-extern "C"
-JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_load_1model(JNIEnv *env, jobject, jstring filename) {
-    llama_model_params model_params = llama_model_default_params();
-
-    auto path_to_model = env->GetStringUTFChars(filename, 0);
-    LOGi("Loading model from %s", path_to_model);
-
-    auto model = llama_load_model_from_file(path_to_model, model_params);
-    env->ReleaseStringUTFChars(filename, path_to_model);
-
-    if (!model) {
-        LOGe("load_model() failed");
-        env->ThrowNew(env->FindClass("java/lang/IllegalStateException"), "load_model() failed");
-        return 0;
-    }
-
-    return reinterpret_cast<jlong>(model);
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1model(JNIEnv *, jobject, jlong model) {
-    llama_free_model(reinterpret_cast<llama_model *>(model));
-}
-
-extern "C"
-JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1context(JNIEnv *env, jobject, jlong jmodel) {
-    auto model = reinterpret_cast<llama_model *>(jmodel);
-
-    if (!model) {
-        LOGe("new_context(): model cannot be null");
-        env->ThrowNew(env->FindClass("java/lang/IllegalArgumentException"), "Model cannot be null");
-        return 0;
-    }
-
-    int n_threads = std::max(1, std::min(8, (int) sysconf(_SC_NPROCESSORS_ONLN) - 2));
-    LOGi("Using %d threads", n_threads);
-
-    llama_context_params ctx_params = llama_context_default_params();
-    ctx_params.seed  = 1234;
-    ctx_params.n_ctx = 2048;
-    ctx_params.n_threads       = n_threads;
-    ctx_params.n_threads_batch = n_threads;
-
-    llama_context * context = llama_new_context_with_model(model, ctx_params);
-
-    if (!context) {
-        LOGe("llama_new_context_with_model() returned null)");
-        env->ThrowNew(env->FindClass("java/lang/IllegalStateException"),
-                      "llama_new_context_with_model() returned null)");
-        return 0;
-    }
-
-    return reinterpret_cast<jlong>(context);
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1context(JNIEnv *, jobject, jlong context) {
-    llama_free(reinterpret_cast<llama_context *>(context));
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1free(JNIEnv *, jobject) {
-    llama_backend_free();
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_log_1to_1android(JNIEnv *, jobject) {
-    llama_log_set(log_callback, NULL);
-}
-
-extern "C"
-JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_bench_1model(
-        JNIEnv *env,
-        jobject,
-        jlong context_pointer,
-        jlong model_pointer,
-        jlong batch_pointer,
-        jint pp,
-        jint tg,
-        jint pl,
-        jint nr
-        ) {
-    auto pp_avg = 0.0;
-    auto tg_avg = 0.0;
-    auto pp_std = 0.0;
-    auto tg_std = 0.0;
-
-    const auto context = reinterpret_cast<llama_context *>(context_pointer);
-    const auto model = reinterpret_cast<llama_model *>(model_pointer);
-    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
-
-    const int n_ctx = llama_n_ctx(context);
-
-    LOGi("n_ctx = %d", n_ctx);
-
-    int i, j;
-    int nri;
-    for (nri = 0; nri < nr; nri++) {
-        LOGi("Benchmark prompt processing (pp)");
-
-        llama_batch_clear(*batch);
-
-        const int n_tokens = pp;
-        for (i = 0; i < n_tokens; i++) {
-            llama_batch_add(*batch, 0, i, { 0 }, false);
-        }
-
-        batch->logits[batch->n_tokens - 1] = true;
-        llama_kv_cache_clear(context);
-
-        const auto t_pp_start = ggml_time_us();
-        if (llama_decode(context, *batch) != 0) {
-            LOGi("llama_decode() failed during prompt processing");
-        }
-        const auto t_pp_end = ggml_time_us();
-
-        // bench text generation
-
-        LOGi("Benchmark text generation (tg)");
-
-        llama_kv_cache_clear(context);
-        const auto t_tg_start = ggml_time_us();
-        for (i = 0; i < tg; i++) {
-
-            llama_batch_clear(*batch);
-            for (j = 0; j < pl; j++) {
-                llama_batch_add(*batch, 0, i, { j }, true);
-            }
-
-            LOGi("llama_decode() text generation: %d", i);
-            if (llama_decode(context, *batch) != 0) {
-                LOGi("llama_decode() failed during text generation");
-            }
-        }
-
-        const auto t_tg_end = ggml_time_us();
-
-        llama_kv_cache_clear(context);
-
-        const auto t_pp = double(t_pp_end - t_pp_start) / 1000000.0;
-        const auto t_tg = double(t_tg_end - t_tg_start) / 1000000.0;
-
-        const auto speed_pp = double(pp) / t_pp;
-        const auto speed_tg = double(pl * tg) / t_tg;
-
-        pp_avg += speed_pp;
-        tg_avg += speed_tg;
-
-        pp_std += speed_pp * speed_pp;
-        tg_std += speed_tg * speed_tg;
-
-        LOGi("pp %f t/s, tg %f t/s", speed_pp, speed_tg);
-    }
-
-    pp_avg /= double(nr);
-    tg_avg /= double(nr);
-
-    if (nr > 1) {
-        pp_std = sqrt(pp_std / double(nr - 1) - pp_avg * pp_avg * double(nr) / double(nr - 1));
-        tg_std = sqrt(tg_std / double(nr - 1) - tg_avg * tg_avg * double(nr) / double(nr - 1));
-    } else {
-        pp_std = 0;
-        tg_std = 0;
-    }
-
-    char model_desc[128];
-    llama_model_desc(model, model_desc, sizeof(model_desc));
-
-    const auto model_size     = double(llama_model_size(model)) / 1024.0 / 1024.0 / 1024.0;
-    const auto model_n_params = double(llama_model_n_params(model)) / 1e9;
-
-    const auto backend    = "(Android)"; // TODO: What should this be?
-
-    std::stringstream result;
-    result << std::setprecision(2);
-    result << "| model | size | params | backend | test | t/s |\n";
-    result << "| --- | --- | --- | --- | --- | --- |\n";
-    result << "| " << model_desc << " | " << model_size << "GiB | " << model_n_params << "B | " << backend << " | pp " << pp << " | " << pp_avg << " ± " << pp_std << " |\n";
-    result << "| " << model_desc << " | " << model_size << "GiB | " << model_n_params << "B | " << backend << " | tg " << tg << " | " << tg_avg << " ± " << tg_std << " |\n";
-
-    return env->NewStringUTF(result.str().c_str());
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
-    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
-}
-
-extern "C"
-JNIEXPORT jlong JNICALL
-Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
-
-    // Source: Copy of llama.cpp:llama_batch_init but heap-allocated.
-
-    llama_batch *batch = new llama_batch {
-        0,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        nullptr,
-        0,
-        0,
-        0,
-    };
-
-    if (embd) {
-        batch->embd = (float *) malloc(sizeof(float) * n_tokens * embd);
-    } else {
-        batch->token = (llama_token *) malloc(sizeof(llama_token) * n_tokens);
-    }
-
-    batch->pos      = (llama_pos *)     malloc(sizeof(llama_pos)      * n_tokens);
-    batch->n_seq_id = (int32_t *)       malloc(sizeof(int32_t)        * n_tokens);
-    batch->seq_id   = (llama_seq_id **) malloc(sizeof(llama_seq_id *) * n_tokens);
-    for (int i = 0; i < n_tokens; ++i) {
-        batch->seq_id[i] = (llama_seq_id *) malloc(sizeof(llama_seq_id) * n_seq_max);
-    }
-    batch->logits   = (int8_t *)        malloc(sizeof(int8_t)         * n_tokens);
-
-    return reinterpret_cast<jlong>(batch);
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
-    llama_backend_init(numa);
-}
-
-extern "C"
-JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_system_1info(JNIEnv *env, jobject) {
-    return env->NewStringUTF(llama_print_system_info());
-}
-
-extern "C"
-JNIEXPORT jint JNICALL
-Java_com_example_llama_Llm_completion_1init(
-        JNIEnv *env,
-        jobject,
-        jlong context_pointer,
-        jlong batch_pointer,
-        jstring jtext,
-        jint n_len
-    ) {
-
-    const auto text = env->GetStringUTFChars(jtext, 0);
-    const auto context = reinterpret_cast<llama_context *>(context_pointer);
-    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
-
-    const auto tokens_list = llama_tokenize(context, text, 1);
-
-    auto n_ctx = llama_n_ctx(context);
-    auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
-
-    LOGi("n_len = %d, n_ctx = %d, n_kv_req = %d", n_len, n_ctx, n_kv_req);
-
-    if (n_kv_req > n_ctx) {
-        LOGe("error: n_kv_req > n_ctx, the required KV cache size is not big enough");
-    }
-
-    for (auto id : tokens_list) {
-        LOGi("%s", llama_token_to_piece(context, id).c_str());
-    }
-
-    llama_batch_clear(*batch);
-
-    // evaluate the initial prompt
-    for (auto i = 0; i < tokens_list.size(); i++) {
-        llama_batch_add(*batch, tokens_list[i], i, { 0 }, false);
-    }
-
-    // llama_decode will output logits only for the last token of the prompt
-    batch->logits[batch->n_tokens - 1] = true;
-
-    if (llama_decode(context, *batch) != 0) {
-        LOGe("llama_decode() failed");
-    }
-
-    env->ReleaseStringUTFChars(jtext, text);
-
-    return batch->n_tokens;
-}
-
-extern "C"
-JNIEXPORT jstring JNICALL
-Java_com_example_llama_Llm_completion_1loop(
-        JNIEnv * env,
-        jobject,
-        jlong context_pointer,
-        jlong batch_pointer,
-        jint n_len,
-        jobject intvar_ncur
-) {
-    const auto context = reinterpret_cast<llama_context *>(context_pointer);
-    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
-    const auto model = llama_get_model(context);
-
-    if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
-    if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
-    if (!la_int_var_inc) la_int_var_inc = env->GetMethodID(la_int_var, "inc", "()V");
-
-    auto n_vocab = llama_n_vocab(model);
-    auto logits = llama_get_logits_ith(context, batch->n_tokens - 1);
-
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-    }
-
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-    // sample the most likely token
-    const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
-
-    const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
-    if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
-        return env->NewStringUTF("");
-    }
-
-    auto new_token_chars = llama_token_to_piece(context, new_token_id);
-    LOGi("new_token_chars: `%s`", new_token_chars.c_str());
-    auto new_token = env->NewStringUTF(new_token_chars.c_str());
-
-    llama_batch_clear(*batch);
-    llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
-
-    env->CallVoidMethod(intvar_ncur, la_int_var_inc);
-
-    if (llama_decode(context, *batch) != 0) {
-        LOGe("llama_decode() returned null");
-    }
-
-    return new_token;
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_kv_1cache_1clear(JNIEnv *, jobject, jlong context) {
-    llama_kv_cache_clear(reinterpret_cast<llama_context *>(context));
-}

examples/llama.android/app/src/main/java/com/example/llama/Downloadable.kt

@@ -1,119 +0,0 @@
-package com.example.llama
-
-import android.app.DownloadManager
-import android.net.Uri
-import android.util.Log
-import androidx.compose.material3.Button
-import androidx.compose.material3.Text
-import androidx.compose.runtime.Composable
-import androidx.compose.runtime.getValue
-import androidx.compose.runtime.mutableDoubleStateOf
-import androidx.compose.runtime.mutableStateOf
-import androidx.compose.runtime.remember
-import androidx.compose.runtime.rememberCoroutineScope
-import androidx.compose.runtime.setValue
-import androidx.core.database.getLongOrNull
-import androidx.core.net.toUri
-import kotlinx.coroutines.delay
-import kotlinx.coroutines.launch
-import java.io.File
-
-data class Downloadable(val name: String, val source: Uri, val destination: File) {
-    companion object {
-        @JvmStatic
-        private val tag: String? = this::class.qualifiedName
-
-        sealed interface State
-        data object Ready: State
-        data class Downloading(val id: Long): State
-        data class Downloaded(val downloadable: Downloadable): State
-        data class Error(val message: String): State
-
-        @JvmStatic
-        @Composable
-        fun Button(viewModel: MainViewModel, dm: DownloadManager, item: Downloadable) {
-            var status: State by remember {
-                mutableStateOf(
-                    if (item.destination.exists()) Downloaded(item)
-                    else Ready
-                )
-            }
-            var progress by remember { mutableDoubleStateOf(0.0) }
-
-            val coroutineScope = rememberCoroutineScope()
-
-            suspend fun waitForDownload(result: Downloading, item: Downloadable): State {
-                while (true) {
-                    val cursor = dm.query(DownloadManager.Query().setFilterById(result.id))
-
-                    if (cursor == null) {
-                        Log.e(tag, "dm.query() returned null")
-                        return Error("dm.query() returned null")
-                    }
-
-                    if (!cursor.moveToFirst() || cursor.count < 1) {
-                        cursor.close()
-                        Log.i(tag, "cursor.moveToFirst() returned false or cursor.count < 1, download canceled?")
-                        return Ready
-                    }
-
-                    val pix = cursor.getColumnIndex(DownloadManager.COLUMN_BYTES_DOWNLOADED_SO_FAR)
-                    val tix = cursor.getColumnIndex(DownloadManager.COLUMN_TOTAL_SIZE_BYTES)
-                    val sofar = cursor.getLongOrNull(pix) ?: 0
-                    val total = cursor.getLongOrNull(tix) ?: 1
-                    cursor.close()
-
-                    if (sofar == total) {
-                        return Downloaded(item)
-                    }
-
-                    progress = (sofar * 1.0) / total
-
-                    delay(1000L)
-                }
-            }
-
-            fun onClick() {
-                when (val s = status) {
-                    is Downloaded -> {
-                        viewModel.load(item.destination.path)
-                    }
-
-                    is Downloading -> {
-                        coroutineScope.launch {
-                            status = waitForDownload(s, item)
-                        }
-                    }
-
-                    else -> {
-                        item.destination.delete()
-
-                        val request = DownloadManager.Request(item.source).apply {
-                            setTitle("Downloading model")
-                            setDescription("Downloading model: ${item.name}")
-                            setAllowedNetworkTypes(DownloadManager.Request.NETWORK_WIFI)
-                            setDestinationUri(item.destination.toUri())
-                        }
-
-                        viewModel.log("Saving ${item.name} to ${item.destination.path}")
-                        Log.i(tag, "Saving ${item.name} to ${item.destination.path}")
-
-                        val id = dm.enqueue(request)
-                        status = Downloading(id)
-                        onClick()
-                    }
-                }
-            }
-
-            Button(onClick = { onClick() }, enabled = status !is Downloading) {
-                when (status) {
-                    is Downloading -> Text(text = "Downloading ${(progress * 100).toInt()}%")
-                    is Downloaded -> Text("Load ${item.name}")
-                    is Ready -> Text("Download ${item.name}")
-                    is Error -> Text("Download ${item.name}")
-                }
-            }
-        }
-
-    }
-}