sync : ggml

* add magika inference example

* ggml : fix unaligned accesses in custom ops

* ggml : fix FP32 GELU for values that exceed the FP16 range

* use ggml_pool_1d

* add README

* Update README.md

* pad inputs if the files are too small

* cleanup

ggml-ci

.devops/full-cuda.Dockerfile

@@ -14,7 +14,8 @@ 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.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/full-rocm.Dockerfile

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

.devops/full.Dockerfile

@@ -5,7 +5,8 @@ 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.txt   requirements.txt
+COPY requirements       requirements
 
 RUN pip install --upgrade pip setuptools wheel \
     && pip install -r requirements.txt

.devops/main-intel.Dockerfile

@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN mkdir build && \
+    cd build && \
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target main
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/main /main
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

.devops/main-rocm.Dockerfile

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

.devops/main-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target main
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/main /main && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/main" ]

.devops/nix/apps.nix

@@ -0,0 +1,22 @@
+{
+  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

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

.devops/nix/docker.nix

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

.devops/nix/jetson-support.nix

@@ -0,0 +1,39 @@
+{ 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

@@ -0,0 +1,47 @@
+{ 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

@@ -0,0 +1,290 @@
+{
+  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,
+  vulkan-headers,
+  vulkan-loader,
+  clblast,
+  useBlas ? builtins.all (x: !x) [
+    useCuda
+    useMetalKit
+    useOpenCL
+    useRocm
+    useVulkan
+  ],
+  useCuda ? config.cudaSupport,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
+  useMpi ? false, # Increases the runtime closure size by ~700M
+  useOpenCL ? false,
+  useRocm ? config.rocmSupport,
+  useVulkan ? false,
+  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" ]
+    ++ lib.optionals useVulkan [ "Vulkan" ];
+
+  pnameSuffix =
+    strings.optionalString (suffices != [ ])
+      "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
+  descriptionSuffix =
+    strings.optionalString (suffices != [ ])
+      ", accelerated with ${strings.concatStringsSep ", " suffices}";
+
+  # TODO: package the Python in this repository in a Nix-like way.
+  # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
+  # is PEP 517-compatible, and ensure the correct .dist-info is generated.
+  # https://peps.python.org/pep-0517/
+  llama-python = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+    ]
+  );
+
+  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+  llama-python-extra = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+      ps.tiktoken
+      ps.torchWithoutCuda
+      ps.transformers
+    ]
+  );
+
+  # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
+  # separately
+  darwinBuildInputs =
+    with darwin.apple_sdk.frameworks;
+    [
+      Accelerate
+      CoreVideo
+      CoreGraphics
+    ]
+    ++ optionals useMetalKit [ MetalKit ];
+
+  cudaBuildInputs = with cudaPackages; [
+    cuda_cccl.dev # <nv/target>
+
+    # A temporary hack for reducing the closure size, remove once cudaPackages
+    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
+    cuda_cudart.dev
+    cuda_cudart.lib
+    cuda_cudart.static
+    libcublas.dev
+    libcublas.lib
+    libcublas.static
+  ];
+
+  rocmBuildInputs = with rocmPackages; [
+    clr
+    hipblas
+    rocblas
+  ];
+
+  vulkanBuildInputs = [
+    vulkan-headers
+    vulkan-loader
+  ];
+in
+
+effectiveStdenv.mkDerivation (
+  finalAttrs: {
+    pname = "llama-cpp${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
+      ++ optionals useVulkan vulkanBuildInputs;
+
+    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)
+        (cmakeBool "LLAMA_VULKAN" useVulkan)
+      ]
+      ++ 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 $out/include/
+    '';
+
+    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
+    passthru = {
+      inherit
+        useBlas
+        useCuda
+        useMetalKit
+        useMpi
+        useOpenCL
+        useRocm
+        useVulkan
+        ;
+
+      shell = mkShell {
+        name = "shell-${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

@@ -0,0 +1,19 @@
+{
+  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 { };
+    docker = self.callPackage ./docker.nix { };
+    docker-min = self.callPackage ./docker.nix { interactive = false; };
+    sif = self.callPackage ./sif.nix { };
+  }
+)

.devops/nix/sif.nix

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

.devops/server-cuda.Dockerfile

@@ -0,0 +1,32 @@
+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

@@ -0,0 +1,28 @@
+ARG ONEAPI_VERSION=2024.0.1-devel-ubuntu22.04
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
+
+ARG LLAMA_SYCL_F16=OFF
+RUN apt-get update && \
+    apt-get install -y git
+
+WORKDIR /app
+
+COPY . .
+
+RUN mkdir build && \
+    cd build && \
+    if [ "${LLAMA_SYCL_F16}" = "ON" ]; then \
+        echo "LLAMA_SYCL_F16 is set" && \
+        export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
+    fi && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake --build . --config Release --target server
+
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
+
+COPY --from=build /app/build/bin/server /server
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server-rocm.Dockerfile

@@ -0,0 +1,45 @@
+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-vulkan.Dockerfile

@@ -0,0 +1,29 @@
+ARG UBUNTU_VERSION=jammy
+
+FROM ubuntu:$UBUNTU_VERSION as build
+
+# Install build tools
+RUN apt update && apt install -y git build-essential cmake wget
+
+# Install Vulkan SDK
+RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add - && \
+    wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list && \
+    apt update -y && \
+    apt-get install -y vulkan-sdk
+
+# Build it
+WORKDIR /app
+COPY . .
+RUN mkdir build && \
+    cd build && \
+    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake --build . --config Release --target server
+
+# Clean up
+WORKDIR /
+RUN cp /app/build/bin/server /server && \
+    rm -rf /app
+
+ENV LC_ALL=C.utf8
+
+ENTRYPOINT [ "/server" ]

.devops/server.Dockerfile

@@ -0,0 +1,20 @@
+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

@@ -13,6 +13,8 @@ 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
@@ -34,6 +36,8 @@ 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"

.ecrc

@@ -1,4 +1,5 @@
 {
+  "Exclude": ["^\\.gitmodules$"],
   "Disable": {
     "IndentSize": true
   }

.editorconfig

@@ -15,8 +15,14 @@ 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

.flake8

@@ -1,2 +1,3 @@
 [flake8]
 max-line-length = 125
+ignore = W503

.github/ISSUE_TEMPLATE/bug.md

@@ -6,179 +6,6 @@ assignees: ''
 
 ---
 
-# Prerequisites
+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.
 
-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
-```
+If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/workflows/build.yml

@@ -37,6 +37,8 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -65,14 +67,14 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake ..
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-sanitizer:
     runs-on: ubuntu-latest
@@ -100,14 +102,14 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -141,8 +143,93 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest --verbose
+          ctest -L main --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)
+
+  ubuntu-22-cmake-sycl-fp16:
+    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 -DLLAMA_SYCL_F16=ON ..
+          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
 
@@ -159,15 +246,21 @@ jobs:
 
       - name: Build
         id: make_build
+        env:
+            LLAMA_FATAL_WARNINGS: 1
         run: |
-          make -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: make_test
         run: |
-          make tests -j $(sysctl -n hw.logicalcpu)
-          make test -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 make tests -j $(sysctl -n hw.logicalcpu)
+          LLAMA_NO_METAL=1 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
 
@@ -188,14 +281,14 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   macOS-latest-cmake-ios:
     runs-on: macos-latest
@@ -288,6 +381,8 @@ jobs:
       OPENBLAS_VERSION: 0.3.23
       OPENCL_VERSION: 2023.04.17
       CLBLAST_VERSION: 1.6.0
+      SDE_VERSION: 9.33.0-2024-01-07
+      VULKAN_VERSION: 1.3.261.1
 
     strategy:
       matrix:
@@ -304,6 +399,10 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CLBLAST=ON -DBUILD_SHARED_LIBS=ON -DCMAKE_PREFIX_PATH="$env:RUNNER_TEMP/clblast"'
           - build: 'openblas'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_BLAS=ON -DBUILD_SHARED_LIBS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+          - build: 'kompute'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'vulkan'
+            defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -312,6 +411,12 @@ jobs:
         with:
           fetch-depth: 0
 
+      - name: Clone Kompute submodule
+        id: clone_kompute
+        if: ${{ matrix.build == 'kompute' }}
+        run: |
+          git submodule update --init kompute
+
       - name: Download OpenCL SDK
         id: get_opencl
         if: ${{ matrix.build == 'clblast' }}
@@ -346,6 +451,15 @@ jobs:
           $lib =  $(join-path $msvc 'bin\Hostx64\x64\lib.exe')
           & $lib /machine:x64 "/def:${env:RUNNER_TEMP}/openblas/lib/libopenblas.def" "/out:${env:RUNNER_TEMP}/openblas/lib/openblas.lib" /name:openblas.dll
 
+      - name: Install Vulkan SDK
+        id: get_vulkan
+        if: ${{ matrix.build == 'kompute' || matrix.build == 'vulkan' }}
+        run: |
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+
       - name: Build
         id: cmake_build
         run: |
@@ -383,10 +497,23 @@ jobs:
 
       - name: Test
         id: cmake_test
-        if: ${{ matrix.build != 'clblast' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }} # Test AVX-512 only when possible
+        # not all machines have native AVX-512
+        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
-          ctest -C Release --verbose --timeout 900
+          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
 
       - name: Determine tag name
         id: tag
@@ -485,6 +612,65 @@ jobs:
           path: |
             cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
+  windows-latest-cmake-sycl:
+    runs-on: windows-latest
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
+      WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
+
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
+      - name: Install
+        run:  scripts/install-oneapi.bat $WINDOWS_BASEKIT_URL $WINDOWS_DPCPP_MKL
+
+      - name: Build
+        id: cmake_build
+        run:  examples/sycl/win-build-sycl.bat
+
+  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
+
+          ./gradlew build --no-daemon
+
 #  freeBSD-latest:
 #    runs-on: macos-12
 #    steps:

.github/workflows/docker.yml

@@ -28,13 +28,18 @@ 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
@@ -52,6 +57,36 @@ 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
@@ -59,7 +94,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
           file: ${{ matrix.config.dockerfile }}
 
       - name: Build and push Docker image (tagged)
@@ -68,5 +103,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/ggerganov/llama.cpp:${{ matrix.config.tag }}"
+          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 }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/editorconfig.yml

@@ -1,6 +1,12 @@
 name: EditorConfig Checker
 
 on:
+  workflow_dispatch: # allows manual triggering
+    inputs:
+      create_release:
+        description: 'Create new release'
+        required: true
+        type: boolean
   push:
     branches:
       - master

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

@@ -0,0 +1,61 @@
+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:
+    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: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-platforms = aarch64-linux
+          extra-system-features = nixos-test kvm
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: llama-cpp
+    - 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

@@ -0,0 +1,68 @@
+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: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - 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:
+    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: ${{ secrets.GITHUB_TOKEN }}
+        extra-conf: |
+          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+    - uses: DeterminateSystems/magic-nix-cache-action@v2
+      with:
+        upstream-cache: https://${{ matrix.cachixName }}.cachix.org
+    - name: Set-up cachix to push the results to
+      uses: cachix/cachix-action@v13
+      with:
+        authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
+        name: llama-cpp
+    - name: Build
+      run: >
+          nix run github:Mic92/nix-fast-build
+          -- --skip-cached --no-nom
+          --flake
+          ".#checks.$(nix eval --raw --impure --expr builtins.currentSystem)"

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

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

@@ -0,0 +1,36 @@
+# 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

@@ -0,0 +1,29 @@
+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

@@ -0,0 +1,20 @@
+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,W503"
+            exclude: "examples/*,examples/*/**,*/**/__init__.py"

.github/workflows/server.yml

@@ -0,0 +1,83 @@
+# Server build and tests
+name: Server
+
+on:
+  workflow_dispatch: # allows manual triggering
+  push:
+    branches:
+      - master
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+  pull_request:
+    types: [opened, synchronize, reopened]
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+
+jobs:
+  server:
+    runs-on: ubuntu-latest
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
+        include:
+          - build_type: Release
+            sanitizer: ""
+        exclude:
+          - build_type: Release
+            sanitizer: ADDRESS
+          - build_type: Release
+            sanitizer: THREAD
+          - build_type: Release
+            sanitizer: UNDEFINED
+
+    container:
+      image: ubuntu:latest
+      ports:
+        - 8888
+      options: --cpus 4
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          apt-get update
+          apt-get -y install \
+            build-essential \
+            git \
+            cmake \
+            python3-pip \
+            wget \
+            psmisc
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
+              -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
+
+      - name: Tests dependencies
+        id: test_dependencies
+        run: |
+          pip install -r examples/server/tests/requirements.txt
+
+      - name: Download models
+        id: download_models
+        run: |
+          cd examples/server/tests
+          ../../../scripts/hf.sh --repo ggml-org/models --file tinyllamas/stories260K.gguf
+
+      - name: Tests
+        id: server_integration_test
+        run: |
+          cd examples/server/tests
+          PORT=8888 ./tests.sh

.gitignore

@@ -15,6 +15,7 @@
 .DS_Store
 .build/
 .cache/
+.ccls-cache/
 .direnv/
 .envrc
 .swiftpm
@@ -22,11 +23,13 @@
 .clang-tidy
 .vs/
 .vscode/
+.idea/
 
 lcov-report/
 gcovr-report/
 
-build*/
+build*
+cmake-build-*
 out/
 tmp/
 
@@ -42,12 +45,16 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
+/imatrix
 /infill
 /libllama.so
 /llama-bench
-/llava
+/llava-cli
+/lookahead
+/lookup
 /main
 /metal
+/passkey
 /perplexity
 /q8dot
 /quantize
@@ -63,8 +70,9 @@ models-mnt
 /speculative
 /parallel
 /train-text-from-scratch
+/tokenize
 /vdot
-build-info.h
+/common/build-info.cpp
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
@@ -83,17 +91,4 @@ 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
+nppBackup

.gitmodules

@@ -0,0 +1,3 @@
+[submodule "kompute"]
+	path = kompute
+	url = https://github.com/nomic-ai/kompute.git

Makefile

@@ -1,14 +1,15 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	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
+	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
 
 # 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-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
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -25,20 +26,6 @@ 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 ($(UNAME_S),Darwin)
@@ -56,10 +43,6 @@ ifeq ($(UNAME_S),Darwin)
 	endif
 endif
 
-ifneq '' '$(or $(filter clean,$(MAKECMDGOALS)),$(LLAMA_METAL))'
-BUILD_TARGETS += metal
-endif
-
 default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
@@ -78,7 +61,7 @@ test: $(TEST_TARGETS)
 			./$$test_target; \
 		fi; \
 		if [ $$? -ne 0 ]; then \
-			printf 'Test $$test_target FAILED!\n\n' $$test_target; \
+			printf 'Test %s FAILED!\n\n' $$test_target; \
 			failures=$$(( failures + 1 )); \
 		else \
 			printf 'Test %s passed.\n\n' $$test_target; \
@@ -114,20 +97,34 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = -std=c11   -fPIC
-MK_CXXFLAGS = -std=c++11 -fPIC
+MK_CPPFLAGS  = -I. -Icommon
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
-MK_CFLAGS        += -Ofast
-MK_HOST_CXXFLAGS += -Ofast
-MK_CUDA_CXXFLAGS += -O3
+MK_CFLAGS     += -Ofast
+HOST_CXXFLAGS += -Ofast
+MK_NVCCFLAGS  += -O3
 else
-MK_CFLAGS        += -O3
-MK_CXXFLAGS      += -O3
+MK_CFLAGS     += -O3
+MK_CXXFLAGS   += -O3
+MK_NVCCFLAGS  += -O3
 endif
 
+ifndef LLAMA_NO_CCACHE
+CCACHE := $(shell which ccache)
+ifdef CCACHE
+export CCACHE_SLOPPINESS = time_macros
+$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.)
+CC    := $(CCACHE) $(CC)
+CXX   := $(CCACHE) $(CXX)
+else
+$(info I ccache not found. Consider installing it for faster compilation.)
+endif # CCACHE
+endif # LLAMA_NO_CCACHE
+
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
 # posix_memalign came in POSIX.1-2001 / SUSv3
@@ -174,6 +171,10 @@ ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
+
+	ifeq ($(UNAME_S),Linux)
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
 endif
@@ -215,28 +216,14 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
-ifeq ($(CC_IS_CLANG), 1)
-	# clang options
-	MK_CFLAGS        += -Wunreachable-code-break -Wunreachable-code-return
-	MK_HOST_CXXFLAGS += -Wunreachable-code-break -Wunreachable-code-return -Wmissing-prototypes -Wextra-semi
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
 
-	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 $(CC_VER) \>= 080100), 1)
-		MK_HOST_CXXFLAGS += -Wextra-semi
-	endif
+# this version of Apple ld64 is buggy
+ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
+	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
 endif
 
 # OS specific
@@ -284,8 +271,8 @@ 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
-	MK_HOST_CXXFLAGS += -march=native -mtune=native
+	MK_CFLAGS     += -march=native -mtune=native
+	HOST_CXXFLAGS += -march=native -mtune=native
 
 	# Usage AVX-only
 	#MK_CFLAGS   += -mfma -mf16c -mavx
@@ -296,19 +283,31 @@ 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)),)
@@ -337,18 +336,20 @@ 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.
@@ -365,7 +366,7 @@ 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
@@ -380,62 +381,80 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include
-	MK_LDFLAGS   += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib
-	OBJS      += ggml-cuda.o
-	NVCCFLAGS = --forward-unknown-to-host-compiler -use_fast_math
+	ifneq ('', '$(wildcard /opt/cuda)')
+		CUDA_PATH ?= /opt/cuda
+	else
+		CUDA_PATH ?= /usr/local/cuda
+	endif
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	OBJS         += ggml-cuda.o
+	MK_NVCCFLAGS += -use_fast_math
+ifdef LLAMA_FATAL_WARNINGS
+	MK_NVCCFLAGS += -Werror all-warnings
+endif # LLAMA_FATAL_WARNINGS
+ifndef JETSON_EOL_MODULE_DETECT
+	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
+endif # JETSON_EOL_MODULE_DETECT
+ifdef LLAMA_DEBUG
+	MK_NVCCFLAGS += -lineinfo
+endif # LLAMA_DEBUG
 ifdef LLAMA_CUDA_NVCC
-	NVCC = $(LLAMA_CUDA_NVCC)
+	NVCC = $(CCACHE) $(LLAMA_CUDA_NVCC)
 else
-	NVCC = nvcc
+	NVCC = $(CCACHE) nvcc
 endif #LLAMA_CUDA_NVCC
 ifdef CUDA_DOCKER_ARCH
-	NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
-else
-	NVCCFLAGS += -arch=native
+	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
+else ifndef CUDA_POWER_ARCH
+	MK_NVCCFLAGS += -arch=native
 endif # CUDA_DOCKER_ARCH
 ifdef LLAMA_CUDA_FORCE_DMMV
-	NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
 ifdef LLAMA_CUDA_FORCE_MMQ
-	NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
+	MK_NVCCFLAGS += -DGGML_CUDA_FORCE_MMQ
 endif # LLAMA_CUDA_FORCE_MMQ
 ifdef LLAMA_CUDA_DMMV_X
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=$(LLAMA_CUDA_DMMV_X)
 else
-	NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
+	MK_NVCCFLAGS += -DGGML_CUDA_DMMV_X=32
 endif # LLAMA_CUDA_DMMV_X
 ifdef LLAMA_CUDA_MMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_MMV_Y)
 else ifdef LLAMA_CUDA_DMMV_Y
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=$(LLAMA_CUDA_DMMV_Y) # for backwards compatibility
 else
-	NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
+	MK_NVCCFLAGS += -DGGML_CUDA_MMV_Y=1
 endif # LLAMA_CUDA_MMV_Y
 ifdef LLAMA_CUDA_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_F16
 ifdef LLAMA_CUDA_DMMV_F16
-	NVCCFLAGS += -DGGML_CUDA_F16
+	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # LLAMA_CUDA_DMMV_F16
 ifdef LLAMA_CUDA_KQUANTS_ITER
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(LLAMA_CUDA_KQUANTS_ITER)
 else
-	NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
+	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
 endif
 ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(LLAMA_CUDA_PEER_MAX_BATCH_SIZE)
 else
-	NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
+	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 #ifdef LLAMA_CUDA_CUBLAS
-#	NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
 #endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
-	NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
+	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
-	$(NVCC) $(NVCCFLAGS) -c $< -o $@
+ifdef JETSON_EOL_MODULE_DETECT
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+else
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
@@ -456,14 +475,48 @@ 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
+
+ifdef LLAMA_VULKAN_DEBUG
+	MK_CPPFLAGS  += -DGGML_VULKAN_DEBUG
+endif
+
+ifdef LLAMA_VULKAN_VALIDATE
+	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
+endif
+
+ifdef LLAMA_VULKAN_RUN_TESTS
+	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
+endif
+
+ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif # LLAMA_VULKAN
+
 ifdef LLAMA_HIPBLAS
-	ROCM_PATH	?= /opt/rocm
-	HIPCC	    ?= $(ROCM_PATH)/bin/hipcc
-	GPU_TARGETS ?= $(shell $(ROCM_PATH)/llvm/bin/amdgpu-arch)
+
+	ifeq ($(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                   ?= $(CCACHE) $(ROCM_PATH)/bin/hipcc
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+ifdef LLAMA_HIP_UMA
+	MK_CPPFLAGS += -DGGML_HIP_UMA
+endif # LLAMA_HIP_UMA
 	MK_LDFLAGS  += -L$(ROCM_PATH)/lib -Wl,-rpath=$(ROCM_PATH)/lib
 	MK_LDFLAGS	+= -lhipblas -lamdhip64 -lrocblas
 	HIPFLAGS    += $(addprefix --offload-arch=,$(GPU_TARGETS))
@@ -485,11 +538,29 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
+ifdef LLAMA_METAL_EMBED_LIBRARY
+	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
+	OBJS        += ggml-metal-embed.o
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
+
+ifdef LLAMA_METAL_EMBED_LIBRARY
+ggml-metal-embed.o: ggml-metal.metal
+	@echo "Embedding Metal library"
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -497,21 +568,23 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
-ifndef LLAMA_NO_K_QUANTS
-k_quants.o: k_quants.c k_quants.h
-	$(CC) $(CFLAGS) -c $< -o $@
-endif # LLAMA_NO_K_QUANTS
+GF_CC := $(CC)
+include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-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)
+override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
+override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
-# save CXXFLAGS before we add host-only options
-NVCCFLAGS := $(NVCCFLAGS) $(CXXFLAGS) $(CUDA_CXXFLAGS) -Wno-pedantic -Xcompiler "$(HOST_CXXFLAGS)"
-override CXXFLAGS += $(HOST_CXXFLAGS)
+# identify CUDA host compiler
+ifdef LLAMA_CUBLAS
+GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
+include scripts/get-flags.mk
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
+endif
 
 #
 # Print build information
@@ -525,8 +598,19 @@ $(info I CFLAGS:    $(CFLAGS))
 $(info I CXXFLAGS:  $(CXXFLAGS))
 $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
-$(info I CC:        $(shell $(CC) --version | head -n 1))
-$(info I CXX:       $(shell $(CXX) --version | head -n 1))
+$(info I CC:        $(shell $(CC)   --version | head -n 1))
+$(info I CXX:       $(shell $(CXX)  --version | head -n 1))
+ifdef LLAMA_CUBLAS
+$(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
+endif # LLAMA_CUBLAS
 $(info )
 
 #
@@ -542,13 +626,16 @@ 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 $@
 
-OBJS += ggml-alloc.o ggml-backend.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
 
 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 build-info.h common/log.h
-COMMON_DEPS   = $(COMMON_H_DEPS) common.o sampling.o grammar-parser.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.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -568,108 +655,166 @@ train.o: common/train.cpp common/train.h
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
+libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
+	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
+
 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	find examples pocs -type f -name "*.o" -delete
 
 #
 # Examples
 #
 
-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)
+# $< is the first prerequisite, i.e. the source file.
+# Explicitly compile this to an object file so that it can be cached with ccache.
+# The source file is then filtered out from $^ (the list of all prerequisites) and the object file is added instead.
+
+# Helper function that replaces .c, .cpp, and .cu file endings with .o:
+GET_OBJ_FILE = $(patsubst %.c,%.o,$(patsubst %.cpp,%.o,$(patsubst %.cu,%.o,$(1))))
+
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo
 
-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)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-batched: examples/batched/batched.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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-quantize-stats: examples/quantize-stats/quantize-stats.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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-embedding: examples/embedding/embedding.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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-save-load-state: examples/save-load-state/save-load-state.cpp 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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/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
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-gguf: examples/gguf/gguf.cpp ggml.o llama.o $(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) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
+
+gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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)
+llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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
+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)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
 
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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)
+export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-ifdef LLAMA_METAL
-metal: examples/metal/metal.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
-endif
+lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif
 
-build-info.h: $(wildcard .git/index) scripts/build-info.sh
-	@sh scripts/build-info.sh $(CC) > $@.tmp
+common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
+	@sh scripts/build-info.sh "$(CC)" > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
 	else \
 		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.h ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 run-benchmark-matmult: benchmark-matmult
 	./$@
@@ -677,46 +822,80 @@ run-benchmark-matmult: benchmark-matmult
 .PHONY: run-benchmark-matmult swift
 
 vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-llama-grammar: tests/test-llama-grammar.cpp ggml.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-opt: tests/test-opt.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-fns: tests/test-quantize-fns.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-quantize-perf: tests/test-quantize-perf.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-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-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-rope: tests/test-rope.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

Package.swift

@@ -2,33 +2,14 @@
 
 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: platforms,
+    platforms: [
+        .macOS(.v12),
+        .iOS(.v14),
+        .watchOS(.v4),
+        .tvOS(.v14)
+    ],
     products: [
         .library(name: "llama", targets: ["llama"]),
     ],
@@ -36,26 +17,40 @@ let package = Package(
         .target(
             name: "llama",
             path: ".",
-            exclude: exclude,
+            exclude: [
+               "cmake",
+               "examples",
+               "scripts",
+               "models",
+               "tests",
+               "CMakeLists.txt",
+               "ggml-cuda.cu",
+               "ggml-cuda.h",
+               "Makefile"
+            ],
             sources: [
                 "ggml.c",
                 "llama.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
-                "k_quants.c",
-            ] + additionalSources,
-            resources: resources,
+                "ggml-quants.c",
+                "ggml-metal.m",
+            ],
+            resources: [
+                .process("ggml-metal.metal")
+            ],
             publicHeadersPath: "spm-headers",
             cSettings: [
                 .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_K_QUANTS"),
-                .define("GGML_USE_ACCELERATE")
+                .define("GGML_USE_ACCELERATE"),
+                .unsafeFlags(["-fno-objc-arc"]),
+                .define("GGML_USE_METAL"),
                 // 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-sycl.md

@@ -0,0 +1,494 @@
+# llama.cpp for SYCL
+
+- [Background](#background)
+- [OS](#os)
+- [Intel GPU](#intel-gpu)
+- [Docker](#docker)
+- [Linux](#linux)
+- [Windows](#windows)
+- [Environment Variable](#environment-variable)
+- [Known Issue](#known-issue)
+- [Q&A](#q&a)
+- [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, Fedora Silverblue 39|
+|Windows|Support|Windows 11|
+
+
+## Intel GPU
+
+### Verified
+
+|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, 730M|
+|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
+|Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
+
+Note: If the EUs (Execution Unit) in iGPU is less than 80, the inference speed will be too slow to use.
+
+### Memory
+
+The memory is a limitation to run LLM on GPUs.
+
+When run llama.cpp, there is print log to show the applied memory on GPU. You could know how much memory to be used in your case. Like `llm_load_tensors:            buffer size =  3577.56 MiB`.
+
+For iGPU, please make sure the shared memory from host memory is enough. For llama-2-7b.Q4_0, recommend the host memory is 8GB+.
+
+For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
+
+## Docker
+
+Note:
+- Only docker on Linux is tested. Docker on WSL may not work.
+- You may need to install Intel GPU driver on the host machine (See the [Linux](#linux) section to know how to do that)
+
+### Build the image
+
+You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
+
+
+```sh
+# For F16:
+#docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
+
+# Or, for F32:
+docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
+
+# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example
+```
+
+### Run
+
+```sh
+# Firstly, find all the DRI cards:
+ls -la /dev/dri
+# Then, pick the card that you want to use.
+
+# For example with "/dev/dri/card1"
+docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+```
+
+## 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.
+
+```sh
+sudo usermod -aG render username
+sudo usermod -aG video username
+```
+
+Note: re-login to enable it.
+
+c. Check
+
+```sh
+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
+
+```sh
+source /opt/intel/oneapi/setvars.sh
+
+sycl-ls
+```
+
+There should be one or more level-zero devices. Please confirm that at least one GPU is present, 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:
+
+Note:
+- You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
+- 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.
+
+```sh
+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
+
+# Or, for FP32:
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Build example/main only
+#cmake --build . --config Release --target main
+
+# Or, build all binary
+cmake --build . --config Release -v
+
+cd ..
+```
+
+or
+
+```sh
+./examples/sycl/build.sh
+```
+
+### Run
+
+1. Put model file to folder **models**
+
+You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+
+2. Enable oneAPI running environment
+
+```
+source /opt/intel/oneapi/setvars.sh
+```
+
+3. List device ID
+
+Run without parameter:
+
+```sh
+./build/bin/ls-sycl-device
+
+# or running the "main" executable and look at the output log:
+
+./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**
+
+```sh
+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:
+
+```sh
+./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
+```
+
+## Windows
+
+### Setup Environment
+
+1. Install Intel GPU driver.
+
+Please install Intel GPU driver by official guide: [Install GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+
+Note: **The driver is mandatory for compute function**.
+
+2. Install Visual Studio.
+
+Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact oneAPI environment enabling in Windows.
+
+3. 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: **C:\Program Files (x86)\Intel\oneAPI**.
+
+Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
+
+b. Enable oneAPI running environment:
+
+- In Search, input 'oneAPI'.
+
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+c. Check GPU
+
+In oneAPI command line:
+
+```
+sycl-ls
+```
+
+There should be one or more level-zero devices. Please confirm that at least one GPU is present, 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, 11th Gen Intel(R) Core(TM) i7-1185G7 @ 3.00GHz OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Iris(R) Xe Graphics OpenCL 3.0 NEO  [31.0.101.5186]
+[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
+```
+
+4. Install cmake & make
+
+a. Download & install cmake for Windows: https://cmake.org/download/
+
+b. Download & install mingw-w64 make for Windows provided by w64devkit
+
+- Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+
+- Extract `w64devkit` on your pc.
+
+- Add the **bin** folder path in the Windows system PATH environment, like `C:\xxx\w64devkit\bin\`.
+
+### Build locally:
+
+In oneAPI command line window:
+
+```
+mkdir -p build
+cd build
+@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
+
+::  for FP16
+::  faster for long-prompt inference
+::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+
+::  for FP32
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+
+::  build example/main only
+::  make main
+
+::  build all binary
+make -j
+cd ..
+```
+
+or
+
+```
+.\examples\sycl\win-build-sycl.bat
+```
+
+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**
+
+You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+
+2. Enable oneAPI running environment
+
+- In Search, input 'oneAPI'.
+
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
+```
+"C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
+```
+
+3. List device ID
+
+Run without parameter:
+
+```
+build\bin\ls-sycl-device.exe
+
+or
+
+build\bin\main.exe
+```
+
+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 **set GGML_SYCL_DEVICE=0**
+
+```
+set GGML_SYCL_DEVICE=0
+build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0
+```
+or run by script:
+
+```
+.\examples\sycl\win-run-llama2.bat
+```
+
+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 (Linux), icx (Windows)|use icpx/icx 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
+
+- 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** or **--mmap 0**.
+
+## Q&A
+
+- 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`.
+
+- In Windows, no result, not error.
+
+  Miss to enable oneAPI running environment.
+
+- Meet compile error.
+
+  Remove folder **build** and try again.
+
+- I can **not** see **[ext_oneapi_level_zero:gpu:0]** afer install GPU driver in Linux.
+
+  Please run **sudo sycl-ls**.
+
+  If you see it in result, please add video/render group to your ID:
+
+  ```
+  sudo usermod -aG render username
+  sudo usermod -aG video username
+  ```
+
+  Then **relogin**.
+
+  If you do not see it, please check the installation GPU steps again.
+
+## Todo
+
+- Support multiple cards.

README.md

@@ -2,17 +2,18 @@
 
 ![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)
 
-Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
+Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
 ### Hot topics
 
-- 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)
+- Support for chat templates: [Wiki (contributions welcome)](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+- Support for Gemma models: https://github.com/ggerganov/llama.cpp/pull/5631
+- Non-linear quantization IQ4_NL: https://github.com/ggerganov/llama.cpp/pull/5590
+- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
 
@@ -28,17 +29,14 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
         <li><a href="#get-the-code">Get the Code</a></li>
         <li><a href="#build">Build</a></li>
         <li><a href="#blas-build">BLAS Build</a></li>
-        <li><a href="#prepare-data--run">Prepare Data & Run</a></li>
+        <li><a href="#prepare-and-quantize">Prepare and Quantize</a></li>
+        <li><a href="#run-the-quantized-model">Run the quantized model</a></li>
         <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
         <li><a href="#quantization">Quantization</a></li>
         <li><a href="#interactive-mode">Interactive mode</a></li>
         <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
-        <li><a href="#instruction-mode-with-alpaca">Instruction mode with Alpaca</a></li>
-        <li><a href="#using-openllama">Using OpenLLaMA</a></li>
-        <li><a href="#using-gpt4all">Using GPT4All</a></li>
-        <li><a href="#using-pygmalion-7b--metharme-7b">Using Pygmalion 7B & Metharme 7B</a></li>
-        <li><a href="#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data">Obtaining the Facebook LLaMA original model and Stanford Alpaca model data</a></li>
-        <li><a href="#verifying-the-model-files">Verifying the model files</a></li>
+        <li><a href="#instruct-mode">Instruct mode</a></li>
+        <li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
         <li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
         <li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
         <li><a href="#android">Android</a></li>
@@ -53,18 +51,20 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ## Description
 
-The main goal of `llama.cpp` is to run the LLaMA model using 4-bit integer quantization on a MacBook
+The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
+variety of hardware - locally and in the cloud.
 
-- Plain C/C++ implementation without dependencies
-- Apple silicon first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
+- Plain C/C++ implementation without any dependencies
+- Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- Mixed F16 / F32 precision
-- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit and 8-bit integer quantization support
-- CUDA, Metal and OpenCL GPU backend support
+- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
+- Vulkan, SYCL, and (partial) OpenCL backend support
+- CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
-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
-as the main playground for developing new features for the [ggml](https://github.com/ggerganov/ggml) library.
+Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomment-1465108022), the project has
+improved significantly thanks to many contributions. It is the main playground for developing new features for the
+[ggml](https://github.com/ggerganov/ggml) library.
 
 **Supported platforms:**
 
@@ -72,55 +72,100 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] Linux
 - [X] Windows (via CMake)
 - [X] Docker
+- [X] FreeBSD
 
 **Supported models:**
 
+Typically finetunes of the base models below are supported as well.
+
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
+- [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
 - [X] Falcon
-- [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
-- [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
-- [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
-- [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
-- [X] [Pygmalion/Metharme](#using-pygmalion-7b--metharme-7b)
-- [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 - [X] [Baichuan 1 & 2](https://huggingface.co/models?search=baichuan-inc/Baichuan) + [derivations](https://huggingface.co/hiyouga/baichuan-7b-sft)
 - [X] [Aquila 1 & 2](https://huggingface.co/models?search=BAAI/Aquila)
 - [X] [Starcoder models](https://github.com/ggerganov/llama.cpp/pull/3187)
-- [X] [Mistral AI v0.1](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [X] [Refact](https://huggingface.co/smallcloudai/Refact-1_6B-fim)
 - [X] [Persimmon 8B](https://github.com/ggerganov/llama.cpp/pull/3410)
 - [X] [MPT](https://github.com/ggerganov/llama.cpp/pull/3417)
 - [X] [Bloom](https://github.com/ggerganov/llama.cpp/pull/3553)
+- [x] [Yi models](https://huggingface.co/models?search=01-ai/Yi)
+- [X] [StableLM models](https://huggingface.co/stabilityai)
+- [x] [Deepseek models](https://huggingface.co/models?search=deepseek-ai/deepseek)
+- [x] [Qwen models](https://huggingface.co/models?search=Qwen/Qwen)
+- [x] [PLaMo-13B](https://github.com/ggerganov/llama.cpp/pull/3557)
+- [x] [Phi models](https://huggingface.co/models?search=microsoft/phi)
+- [x] [GPT-2](https://huggingface.co/gpt2)
+- [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
+- [x] [InternLM2](https://huggingface.co/models?search=internlm2)
+- [x] [CodeShell](https://github.com/WisdomShell/codeshell)
+- [x] [Gemma](https://ai.google.dev/gemma)
 
+**Multimodal models:**
+
+- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
+- [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
+- [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
+- [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
+- [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
+- [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
+
+**HTTP server**
+
+[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
 - 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)
+- JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
-- Rust: [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
+- 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)
 - 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:**
 
+Unless otherwise noted these projects are open-source with permissive licensing:
+
+- [iohub/collama](https://github.com/iohub/coLLaMA)
+- [janhq/jan](https://github.com/janhq/jan) (AGPL)
 - [nat/openplayground](https://github.com/nat/openplayground)
-- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
+- [Faraday](https://faraday.dev/) (proprietary)
+- [LMStudio](https://lmstudio.ai/) (proprietary)
+- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
+- [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
+- [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
+- [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
+- [ollama/ollama](https://github.com/ollama/ollama)
+- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui) (AGPL)
+- [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
+- [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
+- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
+- [Msty](https://msty.app) (proprietary)
+- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
 
 ---
 
 Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
-```java
+```
 $ make -j && ./main -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
 I llama.cpp build info:
 I UNAME_S:  Darwin
@@ -204,7 +249,7 @@ https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8
 
 ## Usage
 
-Here are the end-to-end binary build and model conversion steps for the LLaMA-7B model.
+Here are the end-to-end binary build and model conversion steps for most supported models.
 
 ### Get the Code
 
@@ -265,7 +310,7 @@ In order to build llama.cpp you have three different options.
         sudo pkg install gmake automake autoconf pkgconf llvm15 clinfo clover \
             opencl clblast openblas
 
-            gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
         ```
 
     **Notes:** With this packages you can build llama.cpp with OPENBLAS and
@@ -321,7 +366,7 @@ 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). BLAS doesn't affect the normal generation performance. There are currently three different implementations of it:
+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:
 
 - #### Accelerate Framework:
 
@@ -365,20 +410,37 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [BLIS.md](docs/BLIS.md) for more information.
 
-- #### Intel MKL
+- #### SYCL
+  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
 
-  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:
+  llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
 
-  ```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
-  ```
+  For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md).
+
+- #### Intel oneMKL
+  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md).
+
+  - 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-basekit 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
+      ```
+
+  - 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-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
+
+  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.
 
 - #### 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
@@ -411,22 +473,39 @@ 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`:
+  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
     ```bash
-    mkdir build
-    cd build
-    CC=/opt/rocm/llvm/bin/clang CXX=/opt/rocm/llvm/bin/clang++ cmake .. -DLLAMA_HIPBLAS=ON
-    cmake --build .
+    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++ ..
+    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 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.
+  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.
   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 |
@@ -539,34 +618,87 @@ 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.
 
-### Prepare Data & Run
+- #### Vulkan
+
+  **With docker**:
+
+  You don't need to install Vulkan SDK. It will be installed inside the container.
+
+  ```sh
+  # Build the image
+  docker build -t llama-cpp-vulkan -f .devops/main-vulkan.Dockerfile .
+
+  # Then, use it:
+  docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+  ```
+
+  **Without docker**:
+
+  Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+
+  For example, on Ubuntu 22.04 (jammy), use the command below:
+
+  ```bash
+  wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
+  wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+  apt update -y
+  apt-get install -y vulkan-sdk
+  # To verify the installation, use the command below:
+  vulkaninfo
+  ```
+
+  Alternatively your package manager might be able to provide the appropiate libraries. For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
+
+  Then, build llama.cpp using the cmake command below:
+
+  ```bash
+  mkdir -p build
+  cd build
+  cmake .. -DLLAMA_VULKAN=1
+  cmake --build . --config Release
+  # Test the output binary (with "-ngl 33" to offload all layers to GPU)
+  ./bin/main -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+
+  # You should see in the output, ggml_vulkan detected your GPU. For example:
+  # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
+  ```
+
+### Prepare and Quantize
+
+To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
 
 ```bash
-# obtain the original LLaMA model weights and place them in ./models
+# obtain the official LLaMA model weights and place them in ./models
 ls ./models
-65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
-  # [Optional] for models using BPE tokenizers
-  ls ./models
-  65B 30B 13B 7B vocab.json
+llama-2-7b tokenizer_checklist.chk tokenizer.model
+# [Optional] for models using BPE tokenizers
+ls ./models
+<folder containing weights and tokenizer json> vocab.json
+# [Optional] for PyTorch .bin models like Mistral-7B
+ls ./models
+<folder containing weights and tokenizer json>
 
 # install Python dependencies
 python3 -m pip install -r requirements.txt
 
-# convert the 7B model to ggml FP16 format
-python3 convert.py models/7B/
+# convert the model to ggml FP16 format
+python3 convert.py models/mymodel/
 
-  # [Optional] for models using BPE tokenizers
-  python convert.py models/7B/ --vocabtype bpe
+# [Optional] for models using BPE tokenizers
+python convert.py models/mymodel/ --vocab-type bpe
 
-# quantize the model to 4-bits (using q4_0 method)
-./quantize ./models/7B/ggml-model-f16.gguf ./models/7B/ggml-model-q4_0.gguf q4_0
+# quantize the model to 4-bits (using Q4_K_M method)
+./quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
 
-# update the gguf filetype to current if older version is unsupported by another application
-./quantize ./models/7B/ggml-model-q4_0.gguf ./models/7B/ggml-model-q4_0-v2.gguf COPY
+# update the gguf filetype to current version if older version is now unsupported
+./quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
+```
 
+### Run the quantized model
 
-# run the inference
-./main -m ./models/7B/ggml-model-q4_0.gguf -n 128
+```bash
+# start inference on a gguf model
+./main -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
 ```
 
 When running the larger models, make sure you have enough disk space to store all the intermediate files.
@@ -587,7 +719,7 @@ From the unzipped folder, open a terminal/cmd window here and place a pre-conver
 
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
-| Model | Original size | Quantized size (4-bit) |
+| Model | Original size | Quantized size (Q4_0) |
 |------:|--------------:|-----------------------:|
 |    7B |         13 GB |                 3.9 GB |
 |   13B |         24 GB |                 7.8 GB |
@@ -614,9 +746,21 @@ Several quantization methods are supported. They differ in the resulting model d
 |   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
 - [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
-- recent k-quants improvements
+- recent k-quants improvements and new i-quants
   - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
   - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
+  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
 
 ### Perplexity (measuring model quality)
 
@@ -628,7 +772,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
+1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```
@@ -691,9 +835,9 @@ The `grammars/` folder contains a handful of sample grammars. To write your own,
 
 For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
 
-### Instruction mode with Alpaca
+### Instruct mode
 
-1. First, download the `ggml` Alpaca model into the `./models` folder
+1. First, download and place the `ggml` model into the `./models` folder
 2. Run the `main` tool like this:
 
 ```
@@ -719,50 +863,6 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 >
 ```
 
-### Using [OpenLLaMA](https://github.com/openlm-research/open_llama)
-
-OpenLLaMA is an openly licensed reproduction of Meta's original LLaMA model. It uses the same architecture and is a drop-in replacement for the original LLaMA weights.
-
-- Download the [3B](https://huggingface.co/openlm-research/open_llama_3b), [7B](https://huggingface.co/openlm-research/open_llama_7b), or [13B](https://huggingface.co/openlm-research/open_llama_13b) model from Hugging Face.
-- Convert the model to ggml FP16 format using `python convert.py <path to OpenLLaMA directory>`
-
-### Using [GPT4All](https://github.com/nomic-ai/gpt4all)
-
-*Note: these instructions are likely obsoleted by the GGUF update*
-
-- Obtain the `tokenizer.model` file from LLaMA model and put it to `models`
-- Obtain the `added_tokens.json` file from Alpaca model and put it to `models`
-- Obtain the `gpt4all-lora-quantized.bin` file from GPT4All model and put it to `models/gpt4all-7B`
-- It is distributed in the old `ggml` format which is now obsoleted
-- You have to convert it to the new format using `convert.py`:
-
-```bash
-python3 convert.py models/gpt4all-7B/gpt4all-lora-quantized.bin
-```
-
-- You can now use the newly generated `models/gpt4all-7B/ggml-model-q4_0.bin` model in exactly the same way as all other models
-
-- The newer GPT4All-J model is not yet supported!
-
-### Using Pygmalion 7B & Metharme 7B
-
-- Obtain the [LLaMA weights](#obtaining-the-facebook-llama-original-model-and-stanford-alpaca-model-data)
-- Obtain the [Pygmalion 7B](https://huggingface.co/PygmalionAI/pygmalion-7b/) or [Metharme 7B](https://huggingface.co/PygmalionAI/metharme-7b) XOR encoded weights
-- Convert the LLaMA model with [the latest HF convert script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py)
-- Merge the XOR files with the converted LLaMA weights by running the [xor_codec](https://huggingface.co/PygmalionAI/pygmalion-7b/blob/main/xor_codec.py) script
-- Convert to `ggml` format using the `convert.py` script in this repo:
-```bash
-python3 convert.py pygmalion-7b/ --outtype q4_1
-```
-> The Pygmalion 7B & Metharme 7B weights are saved in [bfloat16](https://en.wikipedia.org/wiki/Bfloat16_floating-point_format) precision. If you wish to convert to `ggml` without quantizating, please specify the `--outtype` as `f32` instead of `f16`.
-
-
-### Obtaining the Facebook LLaMA original model and Stanford Alpaca model data
-
-- **Under no circumstances should IPFS, magnet links, or any other links to model downloads be shared anywhere in this repository, including in issues, discussions, or pull requests. They will be immediately deleted.**
-- The LLaMA models are officially distributed by Facebook and will **never** be provided through this repository.
-- Refer to [Facebook's LLaMA repository](https://github.com/facebookresearch/llama/pull/73/files) if you need to request access to the model data.
-
 ### Obtaining and using the Facebook LLaMA 2 model
 
 - Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
@@ -774,20 +874,6 @@ python3 convert.py pygmalion-7b/ --outtype q4_1
   - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
   - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
 
-### Verifying the model files
-
-Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files to confirm that you have the correct model data files before creating an issue relating to your model files.
-- The following python script will verify if you have all possible latest files in your self-installed `./models` subdirectory:
-
-```bash
-# run the verification script
-./scripts/verify-checksum-models.py
-```
-
-- On linux or macOS it is also possible to run the following commands to verify if you have all possible latest files in your self-installed `./models` subdirectory:
-    - On Linux: `sha256sum --ignore-missing -c SHA256SUMS`
-    - on macOS: `shasum -a 256 --ignore-missing -c SHA256SUMS`
-
 ### Seminal papers and background on the models
 
 If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
@@ -872,19 +958,22 @@ 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 two Docker images available for this project:
+We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable 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 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).
+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).
 
 #### Usage
 
@@ -908,6 +997,12 @@ 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.
@@ -917,6 +1012,7 @@ 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.
@@ -930,6 +1026,7 @@ 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
 
@@ -938,6 +1035,7 @@ 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
@@ -957,6 +1055,8 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-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
 

SHA256SUMS

@@ -1,40 +0,0 @@
-700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
-666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
-ec2f2d1f0dfb73b72a4cbac7fa121abbe04c37ab327125a38248f930c0f09ddf  models/7B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/7B/ggml-model-q5_1.bin
-7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265  models/7B/params.json
-745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
-d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
-2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
-fad169e6f0f575402cf75945961cb4a8ecd824ba4da6be2af831f320c4348fa5  models/13B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/13B/ggml-model-q5_1.bin
-4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f  models/13B/params.json
-e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/consolidated.00.pth
-4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff  models/30B/consolidated.01.pth
-24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
-1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
-7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
-d2a441403944819492ec8c2002cc36fa38468149bfb4b7b4c52afc7bd9a7166d  models/30B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/30B/ggml-model-q5_1.bin
-2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb  models/30B/params.json
-135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe  models/65B/consolidated.00.pth
-9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde  models/65B/consolidated.01.pth
-e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770  models/65B/consolidated.02.pth
-73176ffb426b40482f2aa67ae1217ef79fbbd1fff5482bae5060cdc5a24ab70e  models/65B/consolidated.03.pth
-882e6431d0b08a8bc66261a0d3607da21cbaeafa96a24e7e59777632dbdac225  models/65B/consolidated.04.pth
-a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/consolidated.05.pth
-72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
-d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
-60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
-cde053439fa4910ae454407e2717cc46cc2c2b4995c00c93297a2b52e790fa92  models/65B/ggml-model-q4_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q4_1.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_0.bin
-ffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffffff  models/65B/ggml-model-q5_1.bin
-999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b  models/65B/params.json
-9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347  models/tokenizer.model

awq-py/README.md

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

awq-py/awq/apply_awq.py

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

awq-py/requirements.txt

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

build.zig

@@ -10,7 +10,6 @@ const Maker = struct {
     builder: *std.build.Builder,
     target: CrossTarget,
     optimize: Mode,
-    config_header: *ConfigHeader,
     enable_lto: bool,
 
     include_dirs: ArrayList([]const u8),
@@ -41,26 +40,24 @@ const Maker = struct {
         const commit_hash = try std.ChildProcess.exec(
             .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
         );
-        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),
-            },
-        );
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
+            \\int LLAMA_BUILD_NUMBER = {};
+            \\char const *LLAMA_COMMIT = "{s}";
+            \\char const *LLAMA_COMPILER = "Zig {s}";
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
+            \\
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
         var m = Maker{
             .builder = builder,
             .target = target,
             .optimize = builder.standardOptimizeOption(.{}),
-            .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(&.{});
@@ -72,7 +69,7 @@ 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();
@@ -85,7 +82,6 @@ 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;
@@ -105,7 +101,6 @@ 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;
@@ -116,31 +111,28 @@ 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");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
+    const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, 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 });
+    _ = 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 });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, llama, common, sampling, grammar_parser, clip });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/README.md

@@ -22,4 +22,8 @@ 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

@@ -10,6 +10,9 @@
 # # 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>"
@@ -22,14 +25,32 @@ mkdir -p "$2"
 OUT=$(realpath "$1")
 MNT=$(realpath "$2")
 
-rm -v $OUT/*.log
-rm -v $OUT/*.exit
-rm -v $OUT/*.md
+rm -f "$OUT/*.log"
+rm -f "$OUT/*.exit"
+rm -f "$OUT/*.md"
 
 sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
+CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
+
+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
@@ -81,10 +102,10 @@ function gg_run_ctest_debug {
 
     set -e
 
-    (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 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 ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+    (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
     set +e
 }
@@ -109,13 +130,13 @@ function gg_run_ctest_release {
 
     set -e
 
-    (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
+    (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
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
-        (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     else
-        (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
+        (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
     fi
 
     set +e
@@ -131,6 +152,61 @@ 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 {
@@ -143,7 +219,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -154,8 +230,8 @@ function gg_run_open_llama_3b_v2 {
 
     set -e
 
-    (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
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_QKK_64=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                             ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -208,6 +284,8 @@ 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 {
@@ -235,6 +313,8 @@ 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"
@@ -276,7 +356,6 @@ 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
 }
 
@@ -286,6 +365,7 @@ 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)"
@@ -321,7 +401,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"
@@ -331,8 +411,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (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
+    (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
 
     python3 ../convert.py ${path_models}
 
@@ -385,6 +465,8 @@ 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 {
@@ -412,6 +494,8 @@ 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="$1"
@@ -463,6 +547,7 @@ function gg_sum_open_llama_7b_v2 {
     gg_printf 'OpenLLaMA 7B-v2:\n'
     gg_printf '- status: %s\n' "$(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)"
@@ -483,17 +568,69 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
+# bge-small
+
+function gg_run_embd_bge_small {
+    cd ${SRC}
+
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
+
+    gg_wget models-mnt/bge-small/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/bge-small"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert-hf-to-gguf.py ${path_models}
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
+
+    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
+
+    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+
+    set +e
+}
+
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    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)"
+}
+
 ## 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
 
-    python3 -m pip install -r ${SRC}/requirements.txt
-    python3 -m pip install --editable gguf-py
+    # Create a fresh python3 venv and enter it
+    python3 -m venv "$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
 fi
 
 ret=0
@@ -502,12 +639,16 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
+    test $ret -eq 0 && gg_run embd_bge_small
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2
         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

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

@@ -1,8 +1,52 @@
 # 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(${TARGET} OBJECT
+add_library(${TARGET} STATIC
+    base64.hpp
     common.h
     common.cpp
     sampling.h
@@ -21,4 +65,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
+target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)

common/base64.hpp

@@ -0,0 +1,392 @@
+/*
+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

@@ -0,0 +1,4 @@
+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.h

@@ -9,6 +9,7 @@
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
+#include <cmath>
 #include <string>
 #include <vector>
 #include <random>
@@ -25,35 +26,58 @@
 #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__, BUILD_NUMBER, BUILD_COMMIT);         \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+#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);    \
 } 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
-    int32_t n_threads                       = get_num_physical_cores();
-    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
+    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_MODE_LAYER; // how to split the model across GPUs
+    int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
+    float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
+    int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
+    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
+    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
+    ggml_numa_strategy numa       = GGML_NUMA_STRATEGY_DISABLED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -68,6 +92,9 @@ 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
@@ -77,14 +104,22 @@ 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
 
@@ -101,15 +136,22 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
-    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
+    bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
+    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);
@@ -120,6 +162,15 @@ std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
+//
+// String utils
+//
+
+std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
+std::vector<std::string> string_split(std::string input, char separator);
+std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
+
 //
 // Model utils
 //
@@ -181,6 +232,10 @@ 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
 //
@@ -194,3 +249,13 @@ 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 preceeding item to */+/? at ") + pos);
+                    throw std::runtime_error(std::string("expecting preceding 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 diffrent function
+//  The log target can also be redirected to a different function
 //  like so:
 //
-//  #define LOG_TARGET log_handler_diffrent()
+//  #define LOG_TARGET log_handler_different()
 //  #include "log.h"
 //
-//  FILE* log_handler_diffrent()
+//  FILE* log_handler_different()
 //  {
 //      return stderr;
 //  }
@@ -97,38 +97,56 @@
     #define LOG_TEE_TARGET stderr
 #endif
 
-// NOTE: currently disabled as it produces too many log files
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
+
 // 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(log_file_basename, log_file_extension)
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
 
 // INTERNAL, DO NOT USE
-inline std::string log_filename_generator_impl(const std::string & log_file_basename, const std::string & log_file_extension)
+inline std::string log_filename_generator_impl(LogTriState multilog, 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;
-    //buf << ".";
-    //buf << log_get_pid();
+    if (_multilog)
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
     buf << ".";
     buf << log_file_extension;
 
@@ -213,15 +231,6 @@ inline std::string log_filename_generator_impl(const std::string & log_file_base
     #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
 //
@@ -315,16 +324,23 @@ enum LogTriState
 #endif
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+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)
 {
-    static bool _initialized{false};
-    static bool _disabled{(filename.empty() && target == nullptr)};
+    static bool _initialized = false;
+    static bool _append = 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
@@ -377,7 +393,7 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
             }
         }
 
-        logfile = fopen(filename.c_str(), "w");
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
     }
 
     if (!logfile)
@@ -398,20 +414,20 @@ inline FILE *log_handler1_impl(bool change = false, LogTriState disable = LogTri
 }
 
 // INTERNAL, DO NOT USE
-inline FILE *log_handler2_impl(bool change = false, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+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)
 {
-    return log_handler1_impl(change, disable, filename, target);
+    return log_handler1_impl(change, append, disable, filename, target);
 }
 
 // Disables logs entirely at runtime.
 //  Makes LOG() and LOG_TEE() produce no output,
-//  untill enabled back.
+//  until enabled back.
 #define log_disable() log_disable_impl()
 
 // INTERNAL, DO NOT USE
 inline FILE *log_disable_impl()
 {
-    return log_handler1_impl(true, LogTriStateTrue);
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
 }
 
 // Enables logs at runtime.
@@ -420,19 +436,31 @@ inline FILE *log_disable_impl()
 // INTERNAL, DO NOT USE
 inline FILE *log_enable_impl()
 {
-    return log_handler1_impl(true, LogTriStateFalse);
+    return log_handler1_impl(true, LogTriStateSame, 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, filename); }
-inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, target); }
+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); }
 
 // 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();
@@ -494,6 +522,18 @@ 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;
 }
 
@@ -523,7 +563,9 @@ 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 file will be tagged with unique ID and written as \"<name>.<ID>.log\"\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");
 }
 
 #define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)

common/sampling.cpp

@@ -13,6 +13,7 @@ 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;
         }
 
@@ -39,6 +40,7 @@ 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()) {
@@ -89,29 +91,79 @@ 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, typical_p = %.3f, temp = %.3f\n"
+            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_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.typical_p, params.temp,
+            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
             params.mirostat, params.mirostat_eta, params.mirostat_tau);
 
     return std::string(result);
 }
 
-llama_token llama_sampling_sample(
+std::string llama_sampling_order_print(const llama_sampling_params & params) {
+    std::string result = "CFG -> Penalties ";
+    if (params.mirostat == 0) {
+        for (auto sampler_type : params.samplers_sequence) {
+            const auto sampler_type_name = sampler_type_to_name_string(sampler_type);
+            if (!sampler_type_name.empty()) {
+                result += "-> " + sampler_type_name + " ";
+            }
+        }
+    } 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 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;
+    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::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
+
+    for (auto sampler_type : samplers_sequence) {
+        switch (sampler_type) {
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case llama_sampler_type::TEMPERATURE:
+                if (dynatemp_range > 0) {
+                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
+                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
+                    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(
                   struct llama_sampling_context * ctx_sampling,
                   struct llama_context * ctx_main,
                   struct llama_context * ctx_cfg,
-                  const int idx) {
+                  const int idx,
+                  bool is_resampling) {  // Add a parameter to indicate if we are resampling
     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;
@@ -126,13 +178,27 @@ llama_token llama_sampling_sample(
 
     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++) {
@@ -141,17 +207,15 @@ llama_token llama_sampling_sample(
 
     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
-    if (!prev.empty()) {
+    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
+    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
+    if (penalty_tokens_used_size) {
         const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
 
         llama_sample_repetition_penalties(ctx_main, &cur_p,
-                prev.data() + prev.size() - penalty_last_n,
-                penalty_last_n, penalty_repeat, penalty_freq, penalty_present);
+                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
+                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
 
         if (!penalize_nl) {
             for (size_t idx = 0; idx < cur_p.size; idx++) {
@@ -163,7 +227,8 @@ llama_token llama_sampling_sample(
         }
     }
 
-    if (ctx_sampling->grammar != NULL) {
+    // If we are in the resampling phase, apply grammar checks before sampling logic
+    if (is_resampling && ctx_sampling->grammar != NULL) {
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
@@ -184,13 +249,9 @@ llama_token llama_sampling_sample(
             id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
             // temperature sampling
-            size_t min_keep = std::max(1, params.n_probs);
+            size_t min_keep = std::max(1, params.min_keep);
 
-            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);
+            sampler_queue(ctx_main, params, cur_p, min_keep);
 
             id = llama_sample_token(ctx_main, &cur_p);
 
@@ -205,13 +266,44 @@ llama_token llama_sampling_sample(
             //    }
             //}
 
-            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+        }
+    }
+
+    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

@@ -8,23 +8,46 @@
 #include <vector>
 #include <unordered_map>
 
+// sampler types
+enum class llama_sampler_type : char {
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
+};
+
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t 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
+    int32_t     n_prev                = 64;       // number of previous tokens to remember
+    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t     top_k                 = 40;       // <= 0 to use vocab size
+    float       top_p                 = 0.95f;    // 1.0 = disabled
+    float       min_p                 = 0.05f;    // 0.0 = disabled
+    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::vector<llama_sampler_type> samplers_sequence = {
+        llama_sampler_type::TOP_K,
+        llama_sampler_type::TFS_Z,
+        llama_sampler_type::TYPICAL_P,
+        llama_sampler_type::TOP_P,
+        llama_sampler_type::MIN_P,
+        llama_sampler_type::TEMPERATURE
+    };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
 
@@ -34,6 +57,9 @@ 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
@@ -79,6 +105,9 @@ 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

@@ -31,7 +31,8 @@ struct train_state  * init_train_state() {
 
     state->opt = new struct ggml_opt_context;
     state->opt->ctx = NULL;
-    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
     return state;
@@ -70,7 +71,7 @@ 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 (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             scale /= sqrtf((float) tensor->ne[0]);
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
@@ -118,7 +119,7 @@ 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 (tensor->n_dims) {
+    switch (ggml_n_dims(tensor)) {
         case 1:
             for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
                 float * dst = (float *) ((char *) tensor->data + i0*tensor->nb[0]);
@@ -182,25 +183,27 @@ 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);
@@ -224,8 +227,8 @@ int64_t get_example_targets_batch(
     bool                   sample_random_offsets
 ) {
     GGML_ASSERT(samples_count > 0);
-    GGML_ASSERT(tokens_input->n_dims  == 2);
-    GGML_ASSERT(target_probs->n_dims  == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(target_probs));
     int64_t n_vocab  = target_probs->ne[0];
     int64_t n_tokens = tokens_input->ne[0];
     int64_t n_batch  = tokens_input->ne[1];
@@ -553,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
     std::string opt_type;
     GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
     if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
-        opt->params.type = GGML_OPT_ADAM;
+        opt->params.type = GGML_OPT_TYPE_ADAM;
 
         GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
         GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@@ -565,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
         copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
     } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
-        opt->params.type = GGML_OPT_LBFGS;
+        opt->params.type = GGML_OPT_TYPE_LBFGS;
 
         GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
         GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@@ -600,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
     gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
 
     switch (opt->params.type) {
-        case GGML_OPT_ADAM:
+        case GGML_OPT_TYPE_ADAM:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
                 gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
@@ -619,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
                     gguf_add_tensor(fctx, opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_LBFGS:
+        case GGML_OPT_TYPE_LBFGS:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
                 gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);
@@ -1045,6 +1048,7 @@ 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;
 
@@ -1080,6 +1084,7 @@ 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;
 }
 
@@ -1102,7 +1107,7 @@ 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 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, "  --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, "  --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)" : "");
@@ -1133,6 +1138,7 @@ 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");
 }
 
@@ -1352,6 +1358,17 @@ 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;
+            }
+            if (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");
+            }
     } else if (arg == "-h" || arg == "--help") {
         params->print_usage = true;
         return true;

common/train.h

@@ -9,6 +9,8 @@
 #include "ggml.h"
 #include "llama.h"
 
+#define LLAMA_TRAIN_MAX_NODES 16384
+
 typedef std::string mt19937_state;
 
 struct train_state {
@@ -44,6 +46,7 @@ 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

@@ -1,316 +0,0 @@
-#!/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

@@ -1,247 +0,0 @@
-#!/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

@@ -1,253 +0,0 @@
-#!/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

@@ -1,221 +0,0 @@
-#!/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 math
+import os
 import struct
 import sys
 from enum import IntEnum
@@ -10,36 +10,17 @@ 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' / 'gguf'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 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
@@ -59,6 +40,7 @@ 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
@@ -90,6 +72,7 @@ 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 = []
@@ -111,6 +94,7 @@ class Vocab:
             self.items.append((item_text, item_score))
         return offset - orig_offset
 
+
 class Tensor:
     def __init__(self, use_padding = True):
         self.name = None
@@ -125,7 +109,7 @@ 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 = GGML_QUANT_SIZES.get(dtype)
+        quant = gguf.GGML_QUANT_SIZES.get(dtype)
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
@@ -144,6 +128,7 @@ 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
@@ -180,8 +165,8 @@ 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.')
@@ -208,6 +193,7 @@ 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
@@ -238,7 +224,7 @@ 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:
@@ -362,7 +348,8 @@ 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
@@ -384,40 +371,38 @@ def handle_metadata(cfg, hp):
         params = convert.Params.loadOriginalParamsJson(fakemodel, orig_config_path)
     else:
         raise ValueError('Unable to load metadata')
-    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)
+    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)
     convert.check_vocab_size(params, vocab)
-    return (params, vocab, svocab)
+    return params, vocab, special_vocab
+
 
 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}')
@@ -427,7 +412,7 @@ def main():
     data = np.memmap(cfg.input, mode = 'r')
     model = GGMLModel()
     print('* Scanning GGML input file')
-    offset = model.load(data, 0)
+    offset = model.load(data, 0)  # noqa
     print(f'* GGML model hyperparameters: {model.hyperparameters}')
     vocab_override = None
     params_override = None
@@ -442,12 +427,15 @@ 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

@@ -3,51 +3,21 @@ 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
@@ -61,9 +31,7 @@ def write_file_header(fout: BinaryIO, params: dict[str, Any]) -> None:
     fout.write(struct.pack("i", int(params["lora_alpha"])))
 
 
-def write_tensor_header(
-    self, name: str, shape: Sequence[int], data_type: np.dtype[Any]
-) -> None:
+def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_type: np.dtype[Any]) -> None:
     sname = name.encode("utf-8")
     fout.write(
         struct.pack(
@@ -78,60 +46,103 @@ def write_tensor_header(
     fout.seek((fout.tell() + 31) & -32)
 
 
-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)
+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)
 
-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")
 
-model = torch.load(input_model, map_location="cpu")
+    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")
 
-with open(input_json, "r") as f:
-    params = json.load(f)
+    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
 
-if params["peft_type"] != "LORA":
-    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-    sys.exit(1)
+    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
+        print(f"Error: unsupported architecture {arch_name}")
+        sys.exit(1)
 
-if params["fan_in_fan_out"] is True:
-    print("Error: param fan_in_fan_out is not supported")
-    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["bias"] is not None and params["bias"] != "none":
-    print("Error: param bias is not supported")
-    sys.exit(1)
+    with open(input_json, "r") as f:
+        params = json.load(f)
 
-# 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["peft_type"] != "LORA":
+        print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+        sys.exit(1)
 
-with open(output_path, "wb") as fout:
-    fout.truncate()
+    if params["fan_in_fan_out"] is True:
+        print("Error: param fan_in_fan_out is not supported")
+        sys.exit(1)
 
-    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:
+    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:
                 v = v.float()
-            v = v.T
-        else:
-            v = v.float()
 
-        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)
+            t = v.detach().numpy()
 
-print(f"Converted {input_json} and {input_model} to {output_path}")
+            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}")

convert-mpt-hf-to-gguf.py

@@ -1,227 +0,0 @@
-#!/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

@@ -1,14 +1,18 @@
-import torch
-import os
-from pprint import pprint
-import sys
+#!/usr/bin/env python3
 import argparse
+import os
+import sys
 from pathlib import Path
+from pprint import pprint
+
+import torch
 from sentencepiece import SentencePieceProcessor
+
 if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py' / 'gguf'))
+    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
+
 def _flatten_dict(dct, tensors, prefix=None):
     assert isinstance(dct, dict)
     for key in dct.keys():
@@ -21,6 +25,7 @@ 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)
@@ -54,6 +59,7 @@ 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")
@@ -65,7 +71,7 @@ def main():
     persimmon_model = torch.load(args.ckpt_path)
     hparams = persimmon_model['args']
     pprint(hparams)
-    tensors = {}
+    tensors: dict[str, torch.Tensor] = {}
     _flatten_dict(persimmon_model['model'], tensors, None)
 
     arch = gguf.MODEL_ARCH.PERSIMMON
@@ -82,7 +88,8 @@ def main():
     gguf_writer.add_embedding_length(hidden_size)
     gguf_writer.add_block_count(block_count)
     gguf_writer.add_feed_forward_length(hparams.ffn_hidden_size)
-    gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+    # ref: https://github.com/ggerganov/llama.cpp/pull/4889/commits/eea19039fc52ea2dbd1aab45b59ab4e3e29a3443
+    gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
     gguf_writer.add_head_count(head_count)
     gguf_writer.add_head_count_kv(head_count_kv)
     gguf_writer.add_rope_freq_base(hparams.rotary_emb_base)
@@ -125,6 +132,5 @@ def main():
     print("")
 
 
-
 if __name__ == '__main__':
     main()

convert-refact-hf-to-gguf.py

@@ -1,272 +0,0 @@
-#!/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

@@ -1,210 +0,0 @@
-#!/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("")

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 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.
+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.
 
 # Example of runtime flags effect on inference speed benchmark
 These runs were tested on the following machine:

examples/CMakeLists.txt

@@ -23,18 +23,24 @@ 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(gguf)
     add_subdirectory(train-text-from-scratch)
-    if (LLAMA_METAL)
-        add_subdirectory(metal)
-    endif()
+    add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)
         add_subdirectory(server)
     endif()

examples/baby-llama/baby-llama.cpp

@@ -575,10 +575,7 @@ 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,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -844,10 +841,7 @@ 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,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 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)
@@ -1131,10 +1125,7 @@ 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,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -1258,9 +1249,9 @@ static struct ggml_tensor * forward_lora(
 }
 
 static void sample_softmax(struct ggml_tensor * logits, struct ggml_tensor * probs, struct ggml_tensor * best_samples) {
-    assert(logits->n_dims == 2);
-    assert(probs->n_dims == 2);
-    assert(best_samples->n_dims == 1);
+    assert(ggml_is_matrix(logits));
+    assert(ggml_is_matrix(probs));
+    assert(ggml_is_vector(best_samples));
     assert(logits->ne[1] == best_samples->ne[0]);
     assert(logits->ne[0] == probs->ne[0]);
     assert(logits->ne[1] == probs->ne[1]);
@@ -1292,9 +1283,9 @@ static void sample_softmax_batch(
     struct ggml_context * ctx, struct ggml_tensor * logits, struct ggml_tensor * probs,
     struct ggml_tensor * best_samples
 ) {
-    GGML_ASSERT(best_samples->n_dims == 2);
-    GGML_ASSERT(logits->n_dims == 3);
-    GGML_ASSERT(probs->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(best_samples));
+    GGML_ASSERT(ggml_is_3d(logits));
+    GGML_ASSERT(ggml_is_3d(probs));
     int n_tokens = best_samples->ne[0];
     int n_batch  = best_samples->ne[1];
     int n_vocab  = logits->ne[0];
@@ -1334,7 +1325,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     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);
@@ -1386,8 +1377,8 @@ 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(tokens_input->n_dims == 2);
-    GGML_ASSERT(     targets->n_dims == 3);
+    GGML_ASSERT(ggml_is_matrix(tokens_input));
+    GGML_ASSERT(ggml_is_3d(targets));
     int n_tokens = tokens_input->ne[0];
     int n_batch  = tokens_input->ne[1];
     GGML_ASSERT(n_tokens == targets->ne[1]);
@@ -1542,27 +1533,28 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        ggml_cgraph gf = {};
+        struct ggml_cgraph * gf = NULL;
+        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(&gf, e);
-        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(gf, e);
+        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1609,13 +1601,14 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            ggml_cgraph gf = {};
+            struct ggml_cgraph * gf = NULL;
+            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(&gf, logits);
-            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(gf, logits);
+            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/base-translate.sh

@@ -0,0 +1,61 @@
+#!/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

@@ -82,13 +82,17 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
     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);
 
@@ -155,7 +159,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq);
+    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, ctx_params.n_threads, ctx_params.n_threads_batch);
     LOG_TEE("\n");
 
     LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
@@ -185,7 +189,7 @@ int main(int argc, char ** argv) {
 
                 const auto t_pp_start = ggml_time_us();
 
-                llama_kv_cache_tokens_rm(ctx, -1, -1);
+                llama_kv_cache_clear(ctx);
 
                 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`
-$ `./swift MODEL_PATH [PROMPT] [PARALLEL]`
+$ `./batched_swift MODEL_PATH [PROMPT] [PARALLEL]`

examples/batched.swift/Sources/main.swift

@@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
 let n_len: Int = 32
 
 // init LLM
-llama_backend_init(false)
+llama_backend_init()
 defer {
     llama_backend_free()
 }
@@ -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(context) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {
@@ -215,9 +215,10 @@ 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 n_tokens = text.count + (add_bos ? 1 : 0)
+    let utf8Count = text.utf8.count
+    let n_tokens = utf8Count + (add_bos ? 1 : 0)
     let tokens = UnsafeMutablePointer<llama_token>.allocate(capacity: n_tokens)
-    let tokenCount = llama_tokenize(model, text, Int32(text.count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
+    let tokenCount = llama_tokenize(model, text, Int32(utf8Count), tokens, Int32(n_tokens), add_bos, /*special tokens*/ false)
     var swiftTokens: [llama_token] = []
     for i in 0 ..< tokenCount {
         swiftTokens.append(tokens[Int(i)])
@@ -230,18 +231,15 @@ 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 {
-        if result.count >= -Int(nTokens) {
-            result.removeLast(-Int(nTokens))
-        } else {
-            result.removeAll()
-        }
+        let actualTokensCount = -Int(nTokens)
+        result = .init(repeating: 0, count: actualTokensCount)
         let check = llama_token_to_piece(
             model,
             token,
             &result,
             Int32(result.count)
         )
-        assert(check == nTokens)
+        assert(check == actualTokensCount)
     } else {
         result.removeLast(result.count - Int(nTokens))
     }
@@ -259,5 +257,4 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
         buffer = []
         return bufferString
     }
-    return nil
 }

examples/batched/batched.cpp

@@ -50,7 +50,8 @@ int main(int argc, char ** argv) {
 
     // init LLM
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     // initialize the model
 
@@ -69,6 +70,7 @@ 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
@@ -90,7 +92,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {

examples/beam-search/beam-search.cpp

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

examples/benchmark/CMakeLists.txt

@@ -1,9 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${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

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 
@@ -130,13 +129,13 @@ int main(int argc, char ** argv)  {
     const ggml_type qtype = GGML_TYPE_Q4_1;
 
     size_t ctx_size = 0;
-    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 += 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 += 1024*1024*16;
 
     printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
@@ -172,7 +171,8 @@ 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_build_forward(m11xm2);
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf, 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));
 
@@ -194,25 +194,27 @@ 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, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
     struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
 
     // printf("Creating compute graph\n");
-    struct ggml_cgraph gf31 = ggml_build_forward(q31);
+    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf31, 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, hist_cur.data());
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
 
     //printf("Creating compute graph\n");
-    struct ggml_cgraph gf32 = ggml_build_forward(q32);
+    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
+    ggml_build_forward_expand(gf32, q32);
     printf("n_threads=%i\n", benchmark_params.n_threads);
 
     const int dimx = sizex;
@@ -224,7 +226,7 @@ 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");
@@ -234,7 +236,7 @@ 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;
@@ -252,7 +254,7 @@ 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
 
@@ -267,7 +269,7 @@ 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

@@ -325,14 +325,14 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %d\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %d\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %d\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %d\n", __func__, params->n_mult);
-    printf("%s: n_head:  %d\n", __func__, params->n_head);
-    printf("%s: n_ff:    %d\n", __func__, params->n_ff);
-    printf("%s: n_layer: %d\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %d\n", __func__, params->n_rot);
+    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
+    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
+    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
+    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
+    printf("%s: n_head:  %u\n", __func__, params->n_head);
+    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
+    printf("%s: n_layer: %u\n", __func__, params->n_layer);
+    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -350,25 +350,25 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x [%d] = [%d] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
 
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%d] float space for model->norm\n",__func__,n_embd);
+    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
 
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
 
     // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wq for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wk for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wv for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.wo for [%d] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%d] float space for layer.ffn_norm for [%d] layers\n",__func__,n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
 
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w1 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w2 for [%d] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%d] x[%d] = [%d] float space for layer.w3 for [%d] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
+    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");
@@ -427,7 +427,7 @@ static void print_row(struct ggml_tensor * probs, int i) {
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
-    assert(probs->n_dims == 2);
+    assert(ggml_is_matrix(probs));
     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 (gg_weights->n_dims){
+    switch (ggml_n_dims(gg_weights)) {
         case 1:
             ct = 0;
             for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){

examples/embedding/CMakeLists.txt

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

@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"
 
@@ -8,6 +7,51 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
+static std::vector<std::string> split_lines(const std::string & s) {
+    std::string line;
+    std::vector<std::string> lines;
+    std::stringstream ss(s);
+    while (std::getline(ss, line)) {
+        lines.push_back(line);
+    }
+    return lines;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, false);
+    }
+}
+
+static void normalize(float * vec, float * out, int n) {
+    float norm = 0;
+    for (int i = 0; i < n; i++) {
+        norm += vec[i] * vec[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n; i++) {
+        out[i] = vec[i] / norm;
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // clear previous kv_cache values (irrelevant for embeddings)
+    llama_kv_cache_clear(ctx);
+
+    // run model
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
+    }
+
+    // normalize on copy
+    for (int k = 0; k < n_seq; k++) {
+        float * emb = llama_get_embeddings_ith(ctx, k);
+        float * out = output + k * n_embd;
+        normalize(emb, out, n_embd);
+    }
+}
+
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -30,7 +74,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -56,49 +101,84 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    int n_past = 0;
+    // split the prompt into lines
+    std::vector<std::string> prompts = split_lines(params.prompt);
 
-    // tokenize the prompt
-    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch == params.n_ctx);
 
+    // tokenize the prompts and trim
+    std::vector<std::vector<int32_t>> inputs;
+    for (const auto & prompt : prompts) {
+        auto inp = ::llama_tokenize(ctx, prompt, true);
+        if (inp.size() > n_batch) {
+            inp.resize(n_batch);
+        }
+        inputs.push_back(inp);
+    }
+
+    // tokenization stats
     if (params.verbose_prompt) {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-        for (int i = 0; i < (int) embd_inp.size(); i++) {
-            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        for (int i = 0; i < (int) inputs.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            for (int j = 0; j < (int) inputs[i].size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
         }
-        fprintf(stderr, "\n");
     }
 
-    if (embd_inp.size() > (size_t)n_ctx) {
-        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
-                __func__, embd_inp.size(), n_ctx);
-        return 1;
-    }
-
-    while (!embd_inp.empty()) {
-        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
-        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
-            return 1;
-        }
-        n_past += n_tokens;
-        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
-    }
+    // initialize batch
+    const int n_prompts = prompts.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, n_prompts);
 
+    // allocate output
     const int n_embd = llama_n_embd(model);
-    const auto * embeddings = llama_get_embeddings(ctx);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    float * emb = embeddings.data();
 
-    for (int i = 0; i < n_embd; i++) {
-        printf("%f ", embeddings[i]);
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_prompts; k++) {
+        // clamp to n_batch tokens
+        auto & inp = inputs[k];
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
     }
-    printf("\n");
 
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // print first 3 embeddings
+    for (int j = 0; j < std::min(3, n_prompts); j++) {
+        fprintf(stderr, "embedding %d: ", j);
+        for (int i = 0; i < n_embd; i++) {
+            fprintf(stderr, "%f ", emb[j * n_embd + i]);
+        }
+        fprintf(stderr, "\n\n");
+    }
+    fprintf(stderr, "\n");
+
+    // clean up
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
-
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

@@ -7,8 +7,6 @@
 #include <string>
 #include <thread>
 
-static const size_t tensor_alignment = 32;
-
 struct lora_info {
     std::string filename;
     float scale;
@@ -240,14 +238,13 @@ static struct lora_data * load_lora(struct lora_info * info) {
     }
 
     struct ggml_init_params params_ggml;
-    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_MAX_NODES;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE;
     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_LORA) {
+    if (magic != LLAMA_FILE_MAGIC_GGLA) {
         die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
     }
     uint32_t version = file.read_u32();
@@ -309,7 +306,7 @@ 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, ggml_new_f32(ctx, scaling));
+        ab = ggml_scale(ctx, ab, scaling);
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
 
@@ -334,28 +331,18 @@ 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_SIZE + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_size   = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
     params.mem_buffer = NULL;
     params.no_alloc   = true;
     struct ggml_context * ctx = NULL;
-    struct ggml_allocr * alloc = NULL;
-    struct ggml_cgraph * gf = NULL;
+    struct ggml_gallocr * alloc = NULL;
+    struct ggml_cgraph  * gf = NULL;
 
     ctx   = ggml_init(params);
-    alloc = ggml_allocr_new_measure(tensor_alignment);
+    alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
     gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    size_t alloc_size = ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
-    ggml_free(ctx);
 
-    static std::vector<uint8_t> data_compute;
-    data_compute.resize(alloc_size + tensor_alignment);
-
-    ctx   = ggml_init(params);
-    alloc = ggml_allocr_new(data_compute.data(), data_compute.size(), tensor_alignment);
-    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
-    ggml_allocr_alloc_graph(alloc, gf);
-    ggml_allocr_free(alloc);
+    ggml_gallocr_alloc_graph(alloc, gf);
 
     struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
     static std::vector<uint8_t> data_work;
@@ -364,6 +351,7 @@ static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int
 
     ggml_graph_compute(gf, &cplan);
 
+    ggml_gallocr_free(alloc);
     ggml_free(ctx);
     return true;
 }

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
 ```
 
-Finetune output files will be saved every N iterations (config with `--save-every N`).
+**Only llama based models are supported!** The 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 to 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 too 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`.
@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-w1 N                LORA rank for w1 tensor (default 4)
-  --rank-w2 N                LORA rank for w2 tensor (default 4)
-  --rank-w3 N                LORA rank for w3 tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

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

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

examples/finetune/finetune.cpp

@@ -1,17 +1,12 @@
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
 #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>
 
@@ -19,8 +14,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static const size_t tensor_alignment = 32;
-
 struct my_llama_hparams {
     uint32_t n_vocab    = 32000;
     uint32_t n_ctx      = 512;
@@ -67,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * w1;
-    struct ggml_tensor * w2;
-    struct ggml_tensor * w3;
+    struct ggml_tensor * ffn_gate; // w1
+    struct ggml_tensor * ffn_down; // w2
+    struct ggml_tensor * ffn_up;   // w3
 };
 
 struct my_llama_model {
@@ -92,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_w1 = 4;
-    uint32_t n_rank_w2 = 4;
-    uint32_t n_rank_w3 = 4;
+    uint32_t n_rank_ffn_gate = 4;
+    uint32_t n_rank_ffn_down = 4;
+    uint32_t n_rank_ffn_up = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -124,17 +117,17 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * w1_a;
-    struct ggml_tensor * w1_b;
-    struct ggml_tensor * w2_a;
-    struct ggml_tensor * w2_b;
-    struct ggml_tensor * w3_a;
-    struct ggml_tensor * w3_b;
+    struct ggml_tensor * ffn_gate_a;
+    struct ggml_tensor * ffn_gate_b;
+    struct ggml_tensor * ffn_down_a;
+    struct ggml_tensor * ffn_down_b;
+    struct ggml_tensor * ffn_up_a;
+    struct ggml_tensor * ffn_up_b;
 };
 
 struct my_llama_lora {
     struct ggml_context * ctx = NULL;
-    std::vector<uint8_t> data;
+    ggml_backend_buffer_t data;
 
     my_llama_lora_hparams hparams;
 
@@ -196,13 +189,13 @@ 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);
@@ -215,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
-    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
-    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
+    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
+    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
+    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -269,7 +262,7 @@ 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;
     }
@@ -326,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -336,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -369,69 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.w1_a);
-        ggml_set_param(ctx, layer.w1_b);
-        ggml_set_param(ctx, layer.w2_a);
-        ggml_set_param(ctx, layer.w2_b);
-        ggml_set_param(ctx, layer.w3_a);
-        ggml_set_param(ctx, layer.w3_b);
-    }
-}
-
-static void alloc_lora(struct ggml_allocr * alloc, struct my_llama_lora * lora) {
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b);
-    ggml_allocr_alloc(alloc, lora->norm_a);
-    ggml_allocr_alloc(alloc, lora->norm_b);
-    ggml_allocr_alloc(alloc, lora->output_a);
-    ggml_allocr_alloc(alloc, lora->output_b);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b);
-        ggml_allocr_alloc(alloc, layer.wq_a);
-        ggml_allocr_alloc(alloc, layer.wq_b);
-        ggml_allocr_alloc(alloc, layer.wk_a);
-        ggml_allocr_alloc(alloc, layer.wk_b);
-        ggml_allocr_alloc(alloc, layer.wv_a);
-        ggml_allocr_alloc(alloc, layer.wv_b);
-        ggml_allocr_alloc(alloc, layer.wo_a);
-        ggml_allocr_alloc(alloc, layer.wo_b);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b);
-        ggml_allocr_alloc(alloc, layer.w1_a);
-        ggml_allocr_alloc(alloc, layer.w1_b);
-        ggml_allocr_alloc(alloc, layer.w2_a);
-        ggml_allocr_alloc(alloc, layer.w2_b);
-        ggml_allocr_alloc(alloc, layer.w3_a);
-        ggml_allocr_alloc(alloc, layer.w3_b);
-    }
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_a->grad);
-    ggml_allocr_alloc(alloc, lora->tok_embeddings_b->grad);
-    ggml_allocr_alloc(alloc, lora->norm_a->grad);
-    ggml_allocr_alloc(alloc, lora->norm_b->grad);
-    ggml_allocr_alloc(alloc, lora->output_a->grad);
-    ggml_allocr_alloc(alloc, lora->output_b->grad);
-    for (uint32_t i = 0; i < lora->layers.size(); ++i) {
-        auto & layer = lora->layers[i];
-        ggml_allocr_alloc(alloc, layer.attention_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.attention_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.wq_a->grad);
-        ggml_allocr_alloc(alloc, layer.wq_b->grad);
-        ggml_allocr_alloc(alloc, layer.wk_a->grad);
-        ggml_allocr_alloc(alloc, layer.wk_b->grad);
-        ggml_allocr_alloc(alloc, layer.wv_a->grad);
-        ggml_allocr_alloc(alloc, layer.wv_b->grad);
-        ggml_allocr_alloc(alloc, layer.wo_a->grad);
-        ggml_allocr_alloc(alloc, layer.wo_b->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_a->grad);
-        ggml_allocr_alloc(alloc, layer.ffn_norm_b->grad);
-        ggml_allocr_alloc(alloc, layer.w1_a->grad);
-        ggml_allocr_alloc(alloc, layer.w1_b->grad);
-        ggml_allocr_alloc(alloc, layer.w2_a->grad);
-        ggml_allocr_alloc(alloc, layer.w2_b->grad);
-        ggml_allocr_alloc(alloc, layer.w3_a->grad);
-        ggml_allocr_alloc(alloc, layer.w3_b->grad);
+        ggml_set_param(ctx, layer.ffn_gate_a);
+        ggml_set_param(ctx, layer.ffn_gate_b);
+        ggml_set_param(ctx, layer.ffn_down_a);
+        ggml_set_param(ctx, layer.ffn_down_b);
+        ggml_set_param(ctx, layer.ffn_up_a);
+        ggml_set_param(ctx, layer.ffn_up_b);
     }
 }
 
@@ -499,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
-        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
-        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
-        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
-        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
-        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
+        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
+        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
+        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
+        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
+        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
+        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -518,28 +454,18 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
 
-    // measure data size
-    size_t size = 0;
-    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
-    }
-
-    // allocate data
-    struct ggml_allocr * alloc = NULL;
-    lora->data.resize(size + tensor_alignment);
-    alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
-    alloc_lora(alloc, lora);
-    ggml_allocr_free(alloc);
+    // allocate data for lora tensors
+    lora->data = ggml_backend_alloc_ctx_tensors_from_buft(ctx, ggml_backend_cpu_buffer_type());
 }
 
 static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, float std, float min, float max) {
@@ -548,35 +474,35 @@ 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);
-    randomize_tensor_normal(lora->tok_embeddings_b, rnd);
+    ggml_set_zero(lora->tok_embeddings_b);
     randomize_tensor_normal(lora->norm_a,           rnd);
-    randomize_tensor_normal(lora->norm_b,           rnd);
+    ggml_set_zero(lora->norm_b);
     randomize_tensor_normal(lora->output_a,         rnd);
-    randomize_tensor_normal(lora->output_b,         rnd);
+    ggml_set_zero(lora->output_b);
 
     for (uint32_t i = 0; i < n_layer; ++i) {
         auto & layer = lora->layers[i];
         randomize_tensor_normal(layer.attention_norm_a, rnd);
-        randomize_tensor_normal(layer.attention_norm_b, rnd);
+        ggml_set_zero(layer.attention_norm_b);
 
         randomize_tensor_normal(layer.wq_a, rnd);
-        randomize_tensor_normal(layer.wq_b, rnd);
+        ggml_set_zero(layer.wq_b);
         randomize_tensor_normal(layer.wk_a, rnd);
-        randomize_tensor_normal(layer.wk_b, rnd);
+        ggml_set_zero(layer.wk_b);
         randomize_tensor_normal(layer.wv_a, rnd);
-        randomize_tensor_normal(layer.wv_b, rnd);
+        ggml_set_zero(layer.wv_b);
         randomize_tensor_normal(layer.wo_a, rnd);
-        randomize_tensor_normal(layer.wo_b, rnd);
+        ggml_set_zero(layer.wo_b);
 
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
-        randomize_tensor_normal(layer.ffn_norm_b, rnd);
+        ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.w1_a, rnd);
-        randomize_tensor_normal(layer.w1_b, rnd);
-        randomize_tensor_normal(layer.w2_a, rnd);
-        randomize_tensor_normal(layer.w2_b, rnd);
-        randomize_tensor_normal(layer.w3_a, rnd);
-        randomize_tensor_normal(layer.w3_b, rnd);
+        randomize_tensor_normal(layer.ffn_gate_a, rnd);
+        ggml_set_zero(layer.ffn_gate_b);
+        randomize_tensor_normal(layer.ffn_down_a, rnd);
+        ggml_set_zero(layer.ffn_down_b);
+        randomize_tensor_normal(layer.ffn_up_a, rnd);
+        ggml_set_zero(layer.ffn_up_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -585,7 +511,7 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
 static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct my_llama_model * model,
         struct my_llama_lora  * lora,
-        struct ggml_allocr    * alloc,
+        ggml_gallocr_t          alloc,
         struct ggml_context   * ctx,
         struct ggml_cgraph    * gf,
         struct ggml_cgraph    * gb,
@@ -596,7 +522,8 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         const  int              n_tokens,
         const  int              n_batch,
         const  bool             enable_flash_attn,
-        const  bool             enable_checkpointing) {
+        const  bool             enable_checkpointing,
+        const  bool             measure_only) {
 
     ggml_set_scratch(ctx, { 0, 0, nullptr, });
     const int n_past = 0;
@@ -612,6 +539,7 @@ 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;
@@ -627,13 +555,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
     // KQ_pos - contains the positions
     struct ggml_tensor * KQ_pos = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, N);
-    ggml_allocr_alloc(alloc, KQ_pos);
-    if (!ggml_allocr_is_measure(alloc)) {
-        int * data = (int *) KQ_pos->data;
-        for (int i = 0; i < N; ++i) {
-            data[i] = n_past + i;
-        }
-    }
+    ggml_set_input(KQ_pos);
 
     // rope has so much parameters that we make a custom function for it
     auto rope = [ctx, KQ_pos, n_rot, n_ctx, rope_freq_base, rope_freq_scale]
@@ -642,8 +564,9 @@ 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,
-            rope_freq_base, rope_freq_scale);
+            t, KQ_pos, n_rot, rope_mode, n_ctx, 0,
+            rope_freq_base, rope_freq_scale, 0.0f, 1.0f, 0.0f, 0.0f
+        );
     };
 
     set_name(tokens_input, "tokens_input");
@@ -652,7 +575,7 @@ 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)) {
+        if (ggml_is_quantized(a->type) || a->type == GGML_TYPE_F16) {
             return ggml_add_cast(ctx, a, b, GGML_TYPE_F32);
         } else if (a->type == GGML_TYPE_F32) {
             return ggml_add(ctx, a, b);
@@ -679,10 +602,7 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         checkpoints.push_back(t01);
     }
 
-    struct ggml_tensor * kv_scale = NULL;
-    if (!enable_flash_attn) {
-        kv_scale = ggml_new_f32(ctx, 1.0f/sqrtf(float(n_embd)/n_head));
-    }
+    const float kv_scale = 1.0f/sqrtf(float(n_embd)/n_head);
 
     for (int il = 0; il < n_layer; ++il) {
         struct my_llama_layer & layer = model->layers[il];
@@ -690,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
-        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
-        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
+        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
+        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
+        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -739,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -771,7 +691,7 @@ 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 {
-        *gb = *gf;
+        ggml_graph_cpy(gf, gb);
         ggml_build_backward_expand(ctx, gf, gb, true);
     }
 
@@ -780,43 +700,55 @@ 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, one));
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t35, 1.0f));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36, 1.0f));
     // input gradient
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, t36->grad, 1.0f));
     GGML_ASSERT(t36->grad->data == NULL && t36->grad->view_src == NULL);
-    ggml_allocr_alloc(alloc, t36->grad);
+    ggml_set_input(t36->grad);
     // KQ_pos
-    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, one));
+    ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, KQ_pos, 1.0f));
 
     // make sure base model tensors data cannot be used in viewable operations
-    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));
+    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));
     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, 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));
+        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.ffn_gate, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
     // note: they will be freed in reverse order
     for (unsigned int i = 0; i < checkpoints.size(); ++i) {
         if (checkpoints[i]->data == NULL && checkpoints[i]->view_src == NULL) {
-            ggml_allocr_alloc(alloc, checkpoints[i]);
+            ggml_set_input(checkpoints[i]);
         }
     }
 
-    ggml_allocr_alloc_graph(alloc, gb);
+    if (measure_only) {
+        ggml_gallocr_reserve(alloc, gb);
+    } else {
+        ggml_gallocr_alloc_graph(alloc, gb);
+
+        // set KQ_pos
+        {
+            int * data = (int *) KQ_pos->data;
+            for (int i = 0; i < N; ++i) {
+                data[i] = n_past + i;
+            }
+        }
+    }
 
     // remove the additional nodes and leafs
     for (int i = n_leafs_before; i < gb->n_leafs; ++i) {
@@ -866,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -893,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
-        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
-        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
-        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
-        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
-        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
+        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
+        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
+        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
+        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
+        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
+        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
     }
 }
 
@@ -936,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -962,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.w1_a);
-        gguf_add_tensor(fctx, layer.w1_b);
-        gguf_add_tensor(fctx, layer.w2_a);
-        gguf_add_tensor(fctx, layer.w2_b);
-        gguf_add_tensor(fctx, layer.w3_a);
-        gguf_add_tensor(fctx, layer.w3_b);
+        gguf_add_tensor(fctx, layer.ffn_gate_a);
+        gguf_add_tensor(fctx, layer.ffn_gate_b);
+        gguf_add_tensor(fctx, layer.ffn_down_a);
+        gguf_add_tensor(fctx, layer.ffn_down_b);
+        gguf_add_tensor(fctx, layer.ffn_up_a);
+        gguf_add_tensor(fctx, layer.ffn_up_b);
     }
 }
 
@@ -1109,7 +1041,7 @@ 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 = tensor->n_dims;
+    uint32_t nd = ggml_n_dims(tensor);
     uint32_t ne[4] = { (uint32_t)tensor->ne[0],
                        (uint32_t)tensor->ne[1],
                        (uint32_t)tensor->ne[2],
@@ -1145,9 +1077,8 @@ 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_LORA);   // magic
+    file.write_u32(LLAMA_FILE_MAGIC_GGLA);   // magic
     file.write_u32(1); // version
     // write_hparams
     file.write_u32(lora->hparams.lora_r);
@@ -1173,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1208,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_w1;
-    uint32_t n_rank_w2;
-    uint32_t n_rank_w3;
+    uint32_t n_rank_ffn_gate;
+    uint32_t n_rank_ffn_down;
+    uint32_t n_rank_ffn_up;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1221,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_w1;
-    bool custom_n_rank_w2;
-    bool custom_n_rank_w3;
+    bool custom_n_rank_ffn_gate;
+    bool custom_n_rank_ffn_down;
+    bool custom_n_rank_ffn_up;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1255,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_w1             = 4;
-    params.n_rank_w2             = 4;
-    params.n_rank_w3             = 4;
+    params.n_rank_ffn_gate       = 4;
+    params.n_rank_ffn_down       = 4;
+    params.n_rank_ffn_up         = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1268,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_w1             = false;
-    params.custom_n_rank_w2             = false;
-    params.custom_n_rank_w3             = false;
+    params.custom_n_rank_ffn_gate       = false;
+    params.custom_n_rank_ffn_down       = false;
+    params.custom_n_rank_ffn_up         = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1301,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1438,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-w1") {
+        } else if (arg == "--rank-ffn_gate") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w1 = std::stoi(argv[i]);
-            params->custom_n_rank_w1 = true;
-        } else if (arg == "--rank-w2") {
+            params->n_rank_ffn_gate = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_gate = true;
+        } else if (arg == "--rank-ffn_down") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w2 = std::stoi(argv[i]);
-            params->custom_n_rank_w2 = true;
-        } else if (arg == "--rank-w3") {
+            params->n_rank_ffn_down = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_down = true;
+        } else if (arg == "--rank-ffn_up") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_w3 = std::stoi(argv[i]);
-            params->custom_n_rank_w3 = true;
+            params->n_rank_ffn_up = std::stoi(argv[i]);
+            params->custom_n_rank_ffn_up = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1521,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.w1_a);
-        nx += ggml_nelements(layer.w1_b);
-        nx += ggml_nelements(layer.w2_a);
-        nx += ggml_nelements(layer.w2_b);
-        nx += ggml_nelements(layer.w3_a);
-        nx += ggml_nelements(layer.w3_b);
+        nx += ggml_nelements(layer.ffn_gate_a);
+        nx += ggml_nelements(layer.ffn_gate_b);
+        nx += ggml_nelements(layer.ffn_down_a);
+        nx += ggml_nelements(layer.ffn_down_b);
+        nx += ggml_nelements(layer.ffn_up_a);
+        nx += ggml_nelements(layer.ffn_up_b);
     }
     return nx;
 }
@@ -1545,6 +1476,7 @@ 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);
@@ -1579,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
-    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
-    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
+    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
+    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
+    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1591,17 +1523,18 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_w1             = n_rank_w1;
-    lora.hparams.n_rank_w2             = n_rank_w2;
-    lora.hparams.n_rank_w3             = n_rank_w3;
+    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
+    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
+    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
+    opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;
     opt->params.n_threads               = params.common.n_threads;
     opt->params.past                    = params.common.opt_past;
     opt->params.delta                   = params.common.opt_delta;
@@ -1617,8 +1550,6 @@ 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);
 
@@ -1635,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_w1             != n_rank_w1)
-        || (lora.hparams.n_rank_w2             != n_rank_w2)
-        || (lora.hparams.n_rank_w3             != n_rank_w3)
+        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
+        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
+        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)
@@ -1671,7 +1602,7 @@ int main(int argc, char ** argv) {
     printf("%s: seen train_samples     %llu\n", __func__, (long long unsigned) train->train_samples);
     printf("%s: seen train_tokens      %llu\n", __func__, (long long unsigned) train->train_tokens);
     printf("%s: completed train_epochs %llu\n", __func__, (long long unsigned) train->train_epochs);
-    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + lora.data.size()), (float) (ggml_used_mem(lora.ctx) + lora.data.size()) / (1024.0f*1024.0f));
+    printf("%s: lora_size = %zu bytes (%.1f MB)\n", __func__, (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)), (float) (ggml_used_mem(lora.ctx) + ggml_backend_buffer_get_size(lora.data)) / (1024.0f*1024.0f));
 
     if (params.only_write_lora) {
         save_train_files_data save_data;
@@ -1698,10 +1629,6 @@ int main(int argc, char ** argv) {
     int n_vocab  = model.hparams.n_vocab;
     int n_batch  = params.common.n_batch;
 
-
-    std::vector<uint8_t> mem_input_data;
-    std::vector<uint8_t> mem_compute_data;
-
     // context for input tensors without their data
     struct ggml_init_params ctx_input_params = {
         ggml_tensor_overhead() * 2, // mem_size
@@ -1714,25 +1641,16 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * tokens_input  = ggml_new_tensor_2d(ctx_input, GGML_TYPE_I32, n_tokens, n_batch);
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
+    // allocate input tensors
     // measure required memory for input tensors
-    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
-                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
-                            tensor_alignment;
+    ggml_backend_buffer_t input_data = ggml_backend_alloc_ctx_tensors_from_buft(ctx_input, ggml_backend_cpu_buffer_type());
+    size_t max_input_size = ggml_backend_buffer_get_size(input_data);
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
-    // allocate input tensors
-    mem_input_data.resize(max_input_size);
-    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
-    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
-        )
+    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))
     );
     struct ggml_init_params ctx_compute_params = {
         estimated_compute_size_wo_data, // mem_size
@@ -1754,12 +1672,12 @@ 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);
-        alloc = ggml_allocr_new_measure(tensor_alignment);
-        gf = ggml_new_graph(ctx_compute);
+        ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
+        gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gf->order = (enum ggml_cgraph_eval_order) order;
-        gb = ggml_new_graph(ctx_compute);
+        gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
         gb_tmp = params.common.use_checkpointing
-            ? ggml_new_graph(ctx_compute)
+            ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
             : NULL;
         loss = llama_build_lora_finetune_graphs(
             &model, &lora, alloc, ctx_compute,
@@ -1767,14 +1685,15 @@ int main(int argc, char ** argv) {
             &logits, tokens_input, target_probs,
             n_tokens, n_batch,
             params.common.use_flash,
-            params.common.use_checkpointing
+            params.common.use_checkpointing,
+            true
         );
-        size_t max_compute_size = ggml_allocr_max_size(alloc) + tensor_alignment;
+        size_t max_compute_size = ggml_gallocr_get_buffer_size(alloc, 0); // FIXME: this will still allocate the buffer
         if (max_compute_size < best_compute_size) {
             best_compute_size = max_compute_size;
             best_order = gf->order;
         }
-        ggml_allocr_free(alloc);
+        ggml_gallocr_free(alloc);
         ggml_free(ctx_compute);
     }
     size_t max_compute_size = best_compute_size;
@@ -1785,14 +1704,13 @@ int main(int argc, char ** argv) {
         "invalid");
 
     // allocate compute tensors
-    mem_compute_data.resize(max_compute_size);
     ctx_compute = ggml_init(ctx_compute_params);
-    alloc = ggml_allocr_new(mem_compute_data.data(), mem_compute_data.size(), tensor_alignment);
-    gf = ggml_new_graph(ctx_compute);
+    ggml_gallocr_t alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
+    gf = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gf->order = best_order;
-    gb = ggml_new_graph(ctx_compute);
+    gb = ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true);
     gb_tmp = params.common.use_checkpointing
-        ? ggml_new_graph(ctx_compute)
+        ? ggml_new_graph_custom(ctx_compute, LLAMA_TRAIN_MAX_NODES, true)
         : NULL;
     loss = llama_build_lora_finetune_graphs(
         &model, &lora, alloc, ctx_compute,
@@ -1800,15 +1718,17 @@ int main(int argc, char ** argv) {
         &logits, tokens_input, target_probs,
         n_tokens, n_batch,
         params.common.use_flash,
-        params.common.use_checkpointing
+        params.common.use_checkpointing,
+        false
     );
-    ggml_allocr_free(alloc);
 
     // 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\n", __func__);
+    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");
     tokenize_file(lctx,
             params.common.fn_train_data,
             params.common.sample_start,
@@ -1915,6 +1835,8 @@ int main(int argc, char ** argv) {
     ggml_free(ctx_work);
     ggml_free(ctx_compute);
     ggml_free(ctx_input);
+    ggml_gallocr_free(alloc);
+
 
     int64_t t1 = ggml_time_ms();
     printf("%s: total training time: ", __func__);

examples/finetune/finetune.sh

@@ -0,0 +1,34 @@
+#!/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 llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gguf/gguf.cpp

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

examples/imatrix/CMakeLists.txt

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

@@ -0,0 +1,32 @@
+# 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

@@ -0,0 +1,622 @@
+#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;
+    bool load_imatrix(const char * file_name, bool add);
+    static bool load_imatrix(const char * file_name, std::unordered_map<std::string, Stats>& imatrix);
+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);
+    }
+}
+
+bool IMatrixCollector::load_imatrix(const char * imatrix_file, std::unordered_map<std::string, Stats>& imatrix_data) {
+    std::ifstream in(imatrix_file, std::ios::binary);
+    if (!in) {
+        printf("%s: failed to open %s\n",__func__,imatrix_file);
+        return false;
+    }
+    int n_entries;
+    in.read((char*)&n_entries, sizeof(n_entries));
+    if (in.fail() || n_entries < 1) {
+        printf("%s: no data in file %s\n", __func__, imatrix_file);
+        return false;
+    }
+    for (int i = 0; i < n_entries; ++i) {
+        int len; in.read((char *)&len, sizeof(len));
+        std::vector<char> name_as_vec(len+1);
+        in.read((char *)name_as_vec.data(), len);
+        if (in.fail()) {
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1,imatrix_file);
+            return false;
+        }
+        name_as_vec[len] = 0;
+        std::string name{name_as_vec.data()};
+        auto& e = imatrix_data[std::move(name)];
+        int ncall;
+        in.read((char*)&ncall, sizeof(ncall));
+        int nval;
+        in.read((char *)&nval, sizeof(nval));
+        if (in.fail() || nval < 1) {
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return false;
+        }
+        e.values.resize(nval);
+        in.read((char*)e.values.data(), nval*sizeof(float));
+        if (in.fail()) {
+            printf("%s: failed reading data for entry %d\n",__func__,i);
+            imatrix_data = {};
+            return false;
+        }
+        e.ncall = ncall;
+    }
+    return true;
+}
+
+bool IMatrixCollector::load_imatrix(const char * file_name, bool add) {
+    if (!add) {
+        m_stats.clear();
+    }
+    return load_imatrix(file_name, m_stats);
+}
+
+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, int from_chunk) {
+
+    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 (from_chunk > 0) {
+        if (size_t((from_chunk + 2)*n_ctx) >= tokens.size()) {
+            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, from_chunk);
+            return false;
+        }
+        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, from_chunk, from_chunk*n_ctx);
+        tokens.erase(tokens.begin(), tokens.begin() + from_chunk*n_ctx);
+    }
+
+    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;
+    std::string prev_result_file;
+    std::string combine_files;
+    bool compute_ppl = true;
+    int  from_chunk  = 0;
+    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 if (arg == "--continue-from") {
+            prev_result_file = argv[++iarg];
+        } else if (arg == "--combine") {
+            combine_files = argv[++iarg];
+        }
+        else if (arg == "--from-chunk") {
+            from_chunk = 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]);
+        }
+    }
+
+    g_collector.set_parameters(std::move(sparams));
+
+    if (!combine_files.empty()) {
+        std::vector<std::string> files;
+        size_t pos = 0;
+        while (true) {
+            auto new_pos = combine_files.find(',', pos);
+            if (new_pos != std::string::npos) {
+                files.emplace_back(combine_files.substr(pos, new_pos - pos));
+                pos = new_pos + 1;
+            } else {
+                files.emplace_back(combine_files.substr(pos));
+                break;
+            }
+        }
+        if (files.size() < 2) {
+            fprintf(stderr, "You must provide at least two comma separated files to use --combine\n");
+            return 1;
+        }
+        printf("Combining the following %d files\n", int(files.size()));
+        for (auto& file : files) {
+            printf("    %s\n", file.c_str());
+            if (!g_collector.load_imatrix(file.c_str(), true)) {
+                fprintf(stderr, "Failed to load %s\n", file.c_str());
+                return 1;
+            }
+        }
+        g_collector.save_imatrix();
+        return 0;
+    }
+
+    if (!prev_result_file.empty()) {
+        if (!g_collector.load_imatrix(prev_result_file.c_str(), false)) {
+            fprintf(stderr, "=============== Failed to load %s\n", prev_result_file.c_str());
+            return 1;
+        }
+    }
+
+    gpt_params params;
+    params.n_batch = 512;
+    if (!gpt_params_parse(args.size(), args.data(), params)) {
+        return 1;
+    }
+
+    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();
+    llama_numa_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, from_chunk);
+    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

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

@@ -2,7 +2,6 @@
 
 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"
 
 #include <cassert>
@@ -147,6 +146,13 @@ 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__);
@@ -184,8 +190,8 @@ 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__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    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);
 
     if (params.seed == LLAMA_DEFAULT_SEED) {
         params.seed = time(NULL);
@@ -196,7 +202,8 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init(params.numa);
+    llama_backend_init();
+    llama_numa_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -231,11 +238,11 @@ int main(int argc, char ** argv) {
         LOG_TEE("\n");
         LOG_TEE("%s\n", get_system_info(params).c_str());
     }
-    const bool add_bos = llama_vocab_type(model) == LLAMA_VOCAB_TYPE_SPM;
+    const bool add_bos = llama_should_add_bos_token(model);
     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;
     }
@@ -440,8 +447,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;