clip : refactor + bug fixes (#4696)

* clip : refactor + bug fixes

ggml-ci

* server : add log message

.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-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/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/jetson-support.nix

@@ -0,0 +1,32 @@
+{ inputs, ... }:
+{
+  perSystem =
+    {
+      config,
+      system,
+      lib,
+      pkgsCuda,
+      ...
+    }:
+    lib.optionalAttrs (system == "aarch64-linux") {
+      packages =
+        let
+          caps.jetson-xavier = "7.2";
+          caps.jetson-orin = "8.7";
+          caps.jetson-nano = "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 { }).llama-cpp) caps;
+    };
+}

.devops/nix/nixpkgs-instances.nix

@@ -0,0 +1,35 @@
+{ 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 = {
+        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 +1,265 @@
+{
+  lib,
+  config,
+  stdenv,
+  mkShell,
+  cmake,
+  ninja,
+  pkg-config,
+  git,
+  python3,
+  mpi,
+  openblas, # TODO: Use the generic `blas` so users could switch betwen alternative implementations
+  cudaPackages,
+  darwin,
+  rocmPackages,
+  clblast,
+  useBlas ? builtins.all (x: !x) [
+    useCuda
+    useMetalKit
+    useOpenCL
+    useRocm
+  ],
+  useCuda ? config.cudaSupport,
+  useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
+  useMpi ? false, # Increases the runtime closure size by ~700M
+  useOpenCL ? false,
+  useRocm ? config.rocmSupport,
+  llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
+}@inputs:
+
+let
+  inherit (lib)
+    cmakeBool
+    cmakeFeature
+    optionals
+    strings
+    versionOlder
+    ;
+
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
+  stdenv = throw "Use effectiveStdenv instead";
+  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
+
+  suffices =
+    lib.optionals useBlas [ "BLAS" ]
+    ++ lib.optionals useCuda [ "CUDA" ]
+    ++ lib.optionals useMetalKit [ "MetalKit" ]
+    ++ lib.optionals useMpi [ "MPI" ]
+    ++ lib.optionals useOpenCL [ "OpenCL" ]
+    ++ lib.optionals useRocm [ "ROCm" ];
+
+  pnameSuffix =
+    strings.optionalString (suffices != [ ])
+      "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
+  descriptionSuffix =
+    strings.optionalString (suffices != [ ])
+      ", accelerated with ${strings.concatStringsSep ", " suffices}";
+
+  # TODO: package the Python in this repository in a Nix-like way.
+  # It'd be nice to migrate to buildPythonPackage, as well as ensure this repo
+  # is PEP 517-compatible, and ensure the correct .dist-info is generated.
+  # https://peps.python.org/pep-0517/
+  llama-python = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+    ]
+  );
+
+  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+  llama-python-extra = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+      ps.torchWithoutCuda
+      ps.transformers
+    ]
+  );
+
+  # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
+  # separately
+  darwinBuildInputs =
+    with darwin.apple_sdk.frameworks;
+    [
+      Accelerate
+      CoreVideo
+      CoreGraphics
+    ]
+    ++ optionals useMetalKit [ MetalKit ];
+
+  cudaBuildInputs = with cudaPackages; [
+    cuda_cccl.dev # <nv/target>
+
+    # A temporary hack for reducing the closure size, remove once cudaPackages
+    # have stopped using lndir: https://github.com/NixOS/nixpkgs/issues/271792
+    cuda_cudart.dev
+    cuda_cudart.lib
+    cuda_cudart.static
+    libcublas.dev
+    libcublas.lib
+    libcublas.static
+  ];
+
+  rocmBuildInputs = with rocmPackages; [
+    clr
+    hipblas
+    rocblas
+  ];
+in
+
+effectiveStdenv.mkDerivation (
+  finalAttrs: {
+    pname = "llama-cpp${pnameSuffix}";
+    version = llamaVersion;
+
+    src = lib.cleanSourceWith {
+      filter =
+        name: type:
+        !(builtins.any (_: _) [
+          (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
+          (name == "README.md") # Ignore *.md changes whe computing outPaths
+          (lib.hasPrefix "." name) # Skip hidden files and directories
+        ]);
+      src = lib.cleanSource ../../.;
+    };
+
+    postPatch = ''
+      substituteInPlace ./ggml-metal.m \
+        --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+
+      # TODO: Package up each Python script or service appropriately.
+      # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
+      # we could make those *.py into setuptools' entrypoints
+      substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
+    '';
+
+    nativeBuildInputs =
+      [
+        cmake
+        ninja
+        pkg-config
+        git
+      ]
+      ++ optionals useCuda [
+        cudaPackages.cuda_nvcc
+
+        # TODO: Replace with autoAddDriverRunpath
+        # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
+        cudaPackages.autoAddOpenGLRunpathHook
+      ];
+
+    buildInputs =
+      optionals effectiveStdenv.isDarwin darwinBuildInputs
+      ++ optionals useCuda cudaBuildInputs
+      ++ optionals useMpi [ mpi ]
+      ++ optionals useOpenCL [ clblast ]
+      ++ optionals useRocm rocmBuildInputs;
+
+    cmakeFlags =
+      [
+        (cmakeBool "LLAMA_NATIVE" true)
+        (cmakeBool "LLAMA_BUILD_SERVER" true)
+        (cmakeBool "BUILD_SHARED_LIBS" true)
+        (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+        (cmakeBool "LLAMA_BLAS" useBlas)
+        (cmakeBool "LLAMA_CLBLAST" useOpenCL)
+        (cmakeBool "LLAMA_CUBLAS" useCuda)
+        (cmakeBool "LLAMA_HIPBLAS" useRocm)
+        (cmakeBool "LLAMA_METAL" useMetalKit)
+        (cmakeBool "LLAMA_MPI" useMpi)
+      ]
+      ++ optionals useCuda [
+        (
+          with cudaPackages.flags;
+          cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
+            builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
+          )
+        )
+      ]
+      ++ optionals useRocm [
+        (cmakeFeature "CMAKE_C_COMPILER" "hipcc")
+        (cmakeFeature "CMAKE_CXX_COMPILER" "hipcc")
+
+        # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
+        # in https://github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
+        # and select the line that matches the current nixpkgs version of rocBLAS.
+        # Should likely use `rocmPackages.clr.gpuTargets`.
+        "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
+      ]
+      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
+      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
+
+    # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
+    # if they haven't been added yet.
+    postInstall = ''
+      mv $out/bin/main $out/bin/llama
+      mv $out/bin/server $out/bin/llama-server
+      mkdir -p $out/include
+      cp $src/llama.h $out/include/
+    '';
+
+    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
+    passthru = {
+      inherit
+        useBlas
+        useCuda
+        useMetalKit
+        useMpi
+        useOpenCL
+        useRocm
+        ;
+
+      shell = mkShell {
+        name = "shell-${finalAttrs.finalPackage.name}";
+        description = "contains numpy and sentencepiece";
+        buildInputs = [ llama-python ];
+        inputsFrom = [ finalAttrs.finalPackage ];
+      };
+
+      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 +1,12 @@
+{
+  lib,
+  newScope,
+  llamaVersion ? "0.0.0",
+}:
+
+lib.makeScope newScope (
+  self: {
+    inherit llamaVersion;
+    llama-cpp = self.callPackage ./package.nix { };
+  }
+)

.github/workflows/nix-flakestry.yml

@@ -0,0 +1,23 @@
+# Make the flake discoverable on https://flakestry.dev
+name: "Publish a flake to flakestry"
+on:
+    push:
+        tags:
+        - "v?[0-9]+.[0-9]+.[0-9]+"
+        - "v?[0-9]+.[0-9]+"
+    workflow_dispatch:
+        inputs:
+            tag:
+                description: "The existing tag to publish"
+                type: "string"
+                required: true
+jobs:
+    publish-flake:
+        runs-on: ubuntu-latest
+        permissions:
+            id-token: "write"
+            contents: "read"
+        steps:
+            - uses: flakestry/flakestry-publish@main
+              with:
+                version: "${{ inputs.tag || github.ref_name }}"

.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

README.md

@@ -385,16 +385,30 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [BLIS.md](docs/BLIS.md) for more information.
 
-- #### Intel MKL
+- #### Intel oneMKL
+  - Using manual oneAPI installation:
+    By default, `LLAMA_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DLLAMA_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
+      ```bash
+      mkdir build
+      cd build
+      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-runtime docker image, only required for manual installation
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
+      cmake --build . --config Release
+      ```
 
-  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:
+  - Using oneAPI docker image:
+    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-runtime](https://hub.docker.com/r/intel/oneapi-runtime)
 
-  ```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
-  ```
+      ```bash
+      mkdir build
+      cd build
+      cmake .. -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_NATIVE=ON
+      cmake --build . --config Release
+      ```
+
+  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni.
+
+  Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
 
 - #### cuBLAS
 

common/CMakeLists.txt

@@ -65,4 +65,4 @@ endif()
 
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama build_info)
+target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)

common/common.cpp

@@ -1394,6 +1394,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
     fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
     fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
+    fprintf(stream, "cpu_has_avx_vnni: %s\n",    ggml_cpu_has_avx_vnni()    ? "true" : "false");
     fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
     fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");

convert-hf-to-gguf.py

@@ -242,7 +242,7 @@ class Model:
         tokens: list[bytearray] = []
         toktypes: list[int] = []
 
-        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model)
         vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
         assert max(tokenizer.vocab.values()) < vocab_size
@@ -856,7 +856,7 @@ class StableLMModel(Model):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name(dir_model.name)
+        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -902,7 +902,7 @@ class QwenModel(Model):
         tokens: list[bytearray] = []
         toktypes: list[int] = []
 
-        from transformers import AutoTokenizer  # type: ignore[attr-defined]
+        from transformers import AutoTokenizer
         tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True)
         vocab_size = hparams["vocab_size"]
         assert max(tokenizer.get_vocab().values()) < vocab_size
@@ -1185,57 +1185,62 @@ def parse_args() -> argparse.Namespace:
     return parser.parse_args()
 
 
-args = parse_args()
+def main() -> None:
+    args = parse_args()
 
-dir_model = args.model
+    dir_model = args.model
 
-if args.awq_path:
-    sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
-    from awq.apply_awq import add_scale_weights
-    tmp_model_path = args.model / "weighted_model"
-    dir_model = tmp_model_path
-    if tmp_model_path.is_dir():
-        print(f"{tmp_model_path} exists as a weighted model.")
+    if args.awq_path:
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
+        from awq.apply_awq import add_scale_weights
+        tmp_model_path = args.model / "weighted_model"
+        dir_model = tmp_model_path
+        if tmp_model_path.is_dir():
+            print(f"{tmp_model_path} exists as a weighted model.")
+        else:
+            tmp_model_path.mkdir(parents=True, exist_ok=True)
+            print("Saving new weighted model ...")
+            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
+            print(f"Saved weighted model at {tmp_model_path}.")
+
+    if not dir_model.is_dir():
+        print(f'Error: {args.model} is not a directory', file=sys.stderr)
+        sys.exit(1)
+
+    ftype_map = {
+        "f32": gguf.GGMLQuantizationType.F32,
+        "f16": gguf.GGMLQuantizationType.F16,
+    }
+
+    if args.outfile is not None:
+        fname_out = args.outfile
     else:
-        tmp_model_path.mkdir(parents=True, exist_ok=True)
-        print("Saving new weighted model ...")
-        add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
-        print(f"Saved weighted model at {tmp_model_path}.")
+        # output in the same directory as the model by default
+        fname_out = dir_model / f'ggml-model-{args.outtype}.gguf'
 
-if not dir_model.is_dir():
-    print(f'Error: {args.model} is not a directory', file=sys.stderr)
-    sys.exit(1)
+    print(f"Loading model: {dir_model.name}")
 
-ftype_map = {
-    "f32": gguf.GGMLQuantizationType.F32,
-    "f16": gguf.GGMLQuantizationType.F16,
-}
+    hparams = Model.load_hparams(dir_model)
 
-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-{args.outtype}.gguf'
+    with torch.inference_mode():
+        model_class = Model.from_model_architecture(hparams["architectures"][0])
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
 
-print(f"Loading model: {dir_model.name}")
+        print("Set model parameters")
+        model_instance.set_gguf_parameters()
 
-hparams = Model.load_hparams(dir_model)
+        print("Set model tokenizer")
+        model_instance.set_vocab()
 
-with torch.inference_mode():
-    model_class = Model.from_model_architecture(hparams["architectures"][0])
-    model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+        if args.vocab_only:
+            print(f"Exporting model vocab to '{fname_out}'")
+            model_instance.write_vocab()
+        else:
+            print(f"Exporting model to '{fname_out}'")
+            model_instance.write()
 
-    print("Set model parameters")
-    model_instance.set_gguf_parameters()
+        print(f"Model successfully exported to '{fname_out}'")
 
-    print("Set model tokenizer")
-    model_instance.set_vocab()
 
-    if args.vocab_only:
-        print(f"Exporting model vocab to '{fname_out}'")
-        model_instance.write_vocab()
-    else:
-        print(f"Exporting model to '{fname_out}'")
-        model_instance.write()
-
-    print(f"Model successfully exported to '{fname_out}'")
+if __name__ == '__main__':
+    main()

convert-lora-to-ggml.py

@@ -47,95 +47,96 @@ def write_tensor_header(fout: BinaryIO, name: str, shape: Sequence[int], data_ty
     fout.seek((fout.tell() + 31) & -32)
 
 
-if len(sys.argv) < 2:
-    print(f"Usage: python {sys.argv[0]} <path> [arch]")
-    print(
-        "Path must contain HuggingFace PEFT LoRA files 'adapter_config.json' and 'adapter_model.bin'"
-    )
-    print(f"Arch must be one of {list(gguf.MODEL_ARCH_NAMES.values())} (default: llama)")
-    sys.exit(1)
+if __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")
-arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
+    model = torch.load(input_model, map_location="cpu")
+    arch_name = sys.argv[2] if len(sys.argv) == 3 else "llama"
 
-if arch_name not in gguf.MODEL_ARCH_NAMES.values():
-    print(f"Error: unsupported architecture {arch_name}")
-    sys.exit(1)
+    if arch_name not in gguf.MODEL_ARCH_NAMES.values():
+        print(f"Error: unsupported architecture {arch_name}")
+        sys.exit(1)
 
-arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
-name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
+    arch = list(gguf.MODEL_ARCH_NAMES.keys())[list(gguf.MODEL_ARCH_NAMES.values()).index(arch_name)]
+    name_map = gguf.TensorNameMap(arch, 200) # 200 layers ought to be enough for anyone
 
-with open(input_json, "r") as f:
-    params = json.load(f)
+    with open(input_json, "r") as f:
+        params = json.load(f)
 
-if params["peft_type"] != "LORA":
-    print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
-    sys.exit(1)
+    if params["peft_type"] != "LORA":
+        print(f"Error: unsupported adapter type {params['peft_type']}, expected LORA")
+        sys.exit(1)
 
-if params["fan_in_fan_out"] is True:
-    print("Error: param fan_in_fan_out is not supported")
-    sys.exit(1)
+    if params["fan_in_fan_out"] is True:
+        print("Error: param fan_in_fan_out is not supported")
+        sys.exit(1)
 
-if params["bias"] is not None and params["bias"] != "none":
-    print("Error: param bias is not supported")
-    sys.exit(1)
+    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)
+    # 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()
+    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:
+        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()
+            t = v.detach().numpy()
 
-        prefix = "base_model.model."
-        if k.startswith(prefix):
-            k = k[len(prefix) :]
+            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)
+            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)
+            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
+            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"{k} => {tname} {t.shape} {t.dtype} {t.nbytes/1024/1024:.2f}MB")
+            write_tensor_header(fout, tname, t.shape, t.dtype)
+            t.tofile(fout)
 
-print(f"Converted {input_json} and {input_model} to {output_path}")
+    print(f"Converted {input_json} and {input_model} to {output_path}")

convert-persimmon-to-gguf.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import torch
 import os
 from pprint import pprint

examples/llava/CMakeLists.txt

@@ -24,7 +24,8 @@ endif()
 
 if (NOT MSVC)
     target_compile_options(llava PRIVATE -Wno-cast-qual) # stb_image.h
-    endif()
+endif()
+
 if(TARGET BUILD_INFO)
     add_dependencies(llava BUILD_INFO)
 endif()
@@ -32,5 +33,5 @@ endif()
 set(TARGET llava-cli)
 add_executable(llava-cli llava-cli.cpp)
 install(TARGETS llava-cli RUNTIME)
-target_link_libraries(llava-cli PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(llava-cli PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(llava PRIVATE cxx_std_11)

examples/llava/clip.cpp

@@ -16,12 +16,19 @@
 #include "clip.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
+#include "ggml-backend.h"
+
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
+
+#ifdef GGML_USE_METAL
+#include "ggml-metal.h"
+#endif
 
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
-#define CLIP_DEBUG
-
 static std::string format(const char * fmt, ...) {
     va_list ap;
     va_list ap2;
@@ -139,6 +146,27 @@ static std::string get_ftype(int ftype) {
     }
 }
 
+//
+// image data
+//
+
+// RGB uint8 image
+struct clip_image_u8 {
+    int nx;
+    int ny;
+
+    std::vector<uint8_t> buf;
+};
+
+// RGB float32 image (NHWC)
+// Memory layout: RGBRGBRGB...
+struct clip_image_f32 {
+    int nx;
+    int ny;
+
+    std::vector<float> buf;
+};
+
 //
 // clip layers
 //
@@ -196,39 +224,31 @@ struct clip_vision_model {
     struct ggml_tensor * mm_2_b;
 };
 
-// Replacement for std::vector<uint8_t> that doesn't require zero-initialization.
-struct clip_buffer {
-    uint8_t * data = NULL;
-    size_t size = 0;
-
-    void resize(size_t size) {
-        delete[] data;
-        data = new uint8_t[size];
-        this->size = size;
-    }
-
-    ~clip_buffer() { delete[] data; }
-};
-
 struct clip_ctx {
-    bool has_text_encoder = false;
-    bool has_vision_encoder = false;
+    bool has_text_encoder    = false;
+    bool has_vision_encoder  = false;
     bool has_llava_projector = false;
+
     struct clip_vision_model vision_model;
+
     float image_mean[3];
     float image_std[3];
     bool use_gelu = false;
     int32_t ftype = 1;
-    struct ggml_context * ctx;
+
     struct gguf_context * ctx_gguf;
+    struct ggml_context * ctx_data;
+
+    std::vector<uint8_t> buf_compute_meta;
 
     // memory buffers to evaluate the model
-    clip_buffer buf_compute;
-    clip_buffer buf_alloc;
-    ggml_allocr * alloc = NULL;
+    ggml_backend_buffer_t params_buffer = NULL;
+    ggml_backend_buffer_t compute_buffer = NULL;
+    ggml_backend_t backend = NULL;
+    ggml_allocr * compute_alloc = NULL;
 };
 
-static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_image_f32_batch * imgs) {
+static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return nullptr;
@@ -249,28 +269,24 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
     //const int projection_dim = hparams.projection_dim;
     const float eps = hparams.eps;
     int batch_size = imgs->size;
-    if(ctx->has_llava_projector) {
+    if (ctx->has_llava_projector) {
         GGML_ASSERT(batch_size == 1);
     }
 
-    const auto & buf_compute = ctx->buf_compute;
-
     struct ggml_init_params params = {
-        /*.mem_size =*/ buf_compute.size,
-        /*.mem_buffer =*/ buf_compute.data,
-        /*.no_alloc =*/ false,
+        /*.mem_size   =*/ ctx->buf_compute_meta.size(),
+        /*.mem_buffer =*/ ctx->buf_compute_meta.data(),
+        /*.no_alloc   =*/ true,
     };
 
-    params.no_alloc = true;
-
     struct ggml_context * ctx0 = ggml_init(params);
     struct ggml_cgraph * gf = ggml_new_graph(ctx0);
 
     struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
-    ggml_allocr_alloc(ctx->alloc, inp_raw);
+    ggml_allocr_alloc(ctx->compute_alloc, inp_raw);
 
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        float * data = (float *)ggml_get_data(inp_raw);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        float * data = (float *)malloc(ggml_nbytes(inp_raw));
 
         for (size_t i = 0; i < imgs->size; i++) {
             const int nx = imgs->data[i].nx;
@@ -283,12 +299,14 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
                 for (int k = 0; k < 3; k++) {
                     for (int y = 0; y < ny; y++) {
                         for (int x = 0; x < nx; x++) {
-                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].data[3 * (y * nx + x) + k];
+                            data[(b * 3 * n) + k * n + y * nx + x] = imgs->data[b].buf[3 * (y * nx + x) + k];
                         }
                     }
                 }
             }
         }
+        ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
+        free(data);
     }
 
     struct ggml_tensor * inp = ggml_conv_2d(ctx0, model.patch_embeddings, inp_raw, patch_size, patch_size, 0, 0, 1, 1);
@@ -298,36 +316,39 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
 
     // concat class_embeddings and patch_embeddings
     struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
-    ggml_allocr_alloc(ctx->alloc, embeddings);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
-        ggml_set_zero(embeddings);
+    ggml_allocr_alloc(ctx->compute_alloc, embeddings);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        void* zero_mem = malloc(ggml_nbytes(embeddings));
+        memset(zero_mem, 0, ggml_nbytes(embeddings));
+        ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
+        free(zero_mem);
     }
 
-    struct ggml_tensor * temp = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, 1, batch_size);
-    ggml_allocr_alloc(ctx->alloc, temp);
+    embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
 
-    embeddings = ggml_acc(ctx0, embeddings, ggml_repeat(ctx0, model.class_embedding, temp), embeddings->nb[1],
-                          embeddings->nb[2], embeddings->nb[3], 0);
-    embeddings =
-        ggml_acc(ctx0, embeddings, inp, embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    embeddings = ggml_acc(ctx0, embeddings, inp,
+            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
-    ggml_allocr_alloc(ctx->alloc, positions);
-    if (!ggml_allocr_is_measure(ctx->alloc)) {
+    ggml_allocr_alloc(ctx->compute_alloc, positions);
+    if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+        int* positions_data = (int*)malloc(ggml_nbytes(positions));
         for (int i = 0; i < num_positions; i++) {
-            ggml_set_i32_1d(positions, i, i);
+            positions_data[i] = i;
         }
+        ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
+        free(positions_data);
     }
 
     embeddings =
-        ggml_add(ctx0, embeddings, ggml_repeat(ctx0, ggml_get_rows(ctx0, model.position_embeddings, positions), embeddings));
+        ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
     // pre-layernorm
     {
         embeddings = ggml_norm(ctx0, embeddings, eps);
 
-        embeddings = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.pre_ln_w, embeddings), embeddings),
-                              ggml_repeat(ctx0, model.pre_ln_b, embeddings));
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.pre_ln_w), model.pre_ln_b);
     }
 
     // loop over layers
@@ -340,15 +361,15 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         {
             cur = ggml_norm(ctx0, cur, eps);
 
-            cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].ln_1_w, cur), cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_1_w),
+                           model.layers[il].ln_1_b);
         }
 
         // self-attention
         {
 
             struct ggml_tensor * Q =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].q_b, cur), ggml_mul_mat(ctx0, model.layers[il].q_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].q_w, cur), model.layers[il].q_b);
 
             Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
             Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_positions, batch_size);
@@ -356,14 +377,14 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
             Q = ggml_reshape_3d(ctx0, Q, d_head, num_positions, n_head * batch_size);
 
             struct ggml_tensor * K =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].k_b, cur), ggml_mul_mat(ctx0, model.layers[il].k_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].k_w, cur), model.layers[il].k_b);
 
             K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
             K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
             K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
 
             struct ggml_tensor * V =
-                ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].v_b, cur), ggml_mul_mat(ctx0, model.layers[il].v_w, cur));
+                ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].v_w, cur), model.layers[il].v_b);
 
             V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
             V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
@@ -379,7 +400,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         }
 
         // attention output
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].o_b, cur), ggml_mul_mat(ctx0, model.layers[il].o_w, cur));
+        cur = ggml_add(ctx0, ggml_mul_mat(ctx0, model.layers[il].o_w, cur), model.layers[il].o_b);
 
         // re-add the layer input, e.g., residual
         cur = ggml_add(ctx0, cur, embeddings);
@@ -390,12 +411,11 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         {
             cur = ggml_norm(ctx0, cur, eps);
 
-            cur = ggml_add(ctx0, ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].ln_2_w, cur), cur),
-                           ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
+            cur = ggml_add(ctx0, ggml_mul(ctx0, cur, model.layers[il].ln_2_w), model.layers[il].ln_2_b);
         }
 
         cur = ggml_mul_mat(ctx0, model.layers[il].ff_i_w, cur);
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].ff_i_b, cur), cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_i_b);
 
         if (ctx->use_gelu) {
             cur = ggml_gelu_inplace(ctx0, cur);
@@ -404,7 +424,7 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         }
 
         cur = ggml_mul_mat(ctx0, model.layers[il].ff_o_w, cur);
-        cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].ff_o_b, cur), cur);
+        cur = ggml_add(ctx0, cur, model.layers[il].ff_o_b);
 
         // residual 2
         cur = ggml_add(ctx0, embeddings, cur);
@@ -417,23 +437,26 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
 
         struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
-        ggml_allocr_alloc(ctx->alloc, patches);
-        if (!ggml_allocr_is_measure(ctx->alloc)) {
-            for (int i = 0; i < num_patches; ++i) {
-                ggml_set_i32_1d(patches, i, i+1);
+        ggml_allocr_alloc(ctx->compute_alloc, patches);
+        if (!ggml_allocr_is_measure(ctx->compute_alloc)) {
+            int* patches_data = (int*)malloc(ggml_nbytes(patches));
+            for (int i = 0; i < num_patches; i++) {
+                patches_data[i] = i + 1;
             }
+            ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
+            free(patches_data);
         }
 
         embeddings = ggml_get_rows(ctx0, embeddings, patches);
 
         // mm projection 0
         embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
-        embeddings = ggml_add(ctx0, ggml_repeat(ctx0, model.mm_0_b, embeddings), embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
 
         embeddings = ggml_gelu(ctx0, embeddings);
 
         embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
-        embeddings = ggml_add(ctx0, ggml_repeat(ctx0, model.mm_2_b, embeddings), embeddings);
+        embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
     }
 
     // build the graph
@@ -446,7 +469,6 @@ static ggml_cgraph * clip_image_build_graph(const clip_ctx * ctx, const clip_ima
 
 // read and create ggml_context containing the tensors and their data
 struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
-
     struct ggml_context * meta = NULL;
 
     struct gguf_init_params params = {
@@ -479,7 +501,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         printf("%s: ftype:        %s\n", __func__, ftype_str.c_str());
         printf("\n");
     }
-
+    const int n_tensors = gguf_get_n_tensors(ctx);
     // kv
     if (verbosity >= 3) {
         const int n_kv = gguf_get_n_kv(ctx);
@@ -493,27 +515,38 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     }
 
     // data
-    size_t ctx_size = 0;
+    size_t buffer_size = 0;
     {
-        const int n_tensors = gguf_get_n_tensors(ctx);
-
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             const size_t offset = gguf_get_tensor_offset(ctx, i);
-
             struct ggml_tensor * cur = ggml_get_tensor(meta, name);
-            ctx_size += sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE;
             size_t tensor_size = ggml_nbytes(cur);
-            size_t padded_size = ggml_nbytes_pad(cur);
-            ctx_size += padded_size;
+            buffer_size += tensor_size;
             if (verbosity >= 3) {
-                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, padded_size=%zu, offset=%zu\n", __func__, i,
-                       ggml_n_dims(cur), cur->name, tensor_size, padded_size, offset);
+                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu\n", __func__, i,
+                       ggml_n_dims(cur), cur->name, tensor_size, offset);
             }
         }
     }
 
+    buffer_size += n_tensors * 128 /* CLIP PADDING */;
+
     clip_ctx * new_clip = new clip_ctx;
+#ifdef GGML_USE_CUBLAS
+    new_clip->backend = ggml_backend_cuda_init(0);
+    printf("%s: CLIP using CUDA backend\n", __func__);
+#endif
+
+#ifdef GGML_USE_METAL
+    new_clip->backend = ggml_backend_metal_init();
+    printf("%s: CLIP using Metal backend\n", __func__);
+#endif
+
+    if (!new_clip->backend) {
+        new_clip->backend = ggml_backend_cpu_init();
+        printf("%s: CLIP using CPU backend\n", __func__);
+    }
 
     // model size and capabilities
     {
@@ -539,21 +572,24 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             printf("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
             printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
             printf("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            printf("%s: model size:     %.2f MB\n", __func__, (ctx_size / 1024.0 / 1024.0));
+            printf("%s: model size:     %.2f MB\n", __func__, buffer_size / 1024.0 / 1024.0);
             printf("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
+    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, buffer_size / (1024.0 * 1024.0), n_tensors);
+
     // load tensors
     {
+        std::vector<uint8_t> read_buf;
         struct ggml_init_params params = {
-            /*.mem_size =*/ ctx_size,
+            /*.mem_size =*/ (n_tensors + 1) * ggml_tensor_overhead(),
             /*.mem_buffer =*/ NULL,
-            /*.no_alloc =*/ false,
+            /*.no_alloc =*/ true,
         };
 
-        new_clip->ctx = ggml_init(params);
-        if (!new_clip->ctx) {
+        new_clip->ctx_data = ggml_init(params);
+        if (!new_clip->ctx_data) {
             fprintf(stderr, "%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
             return nullptr;
@@ -566,13 +602,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             return nullptr;
         }
 
-        const int n_tensors = gguf_get_n_tensors(ctx);
+        // add tensors to context
         for (int i = 0; i < n_tensors; ++i) {
             const char * name = gguf_get_tensor_name(ctx, i);
             struct ggml_tensor * t = ggml_get_tensor(meta, name);
-            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx, t);
+            struct ggml_tensor * cur = ggml_dup_tensor(new_clip->ctx_data, t);
             ggml_set_name(cur, name);
+        }
 
+        // alloc memory and offload data
+        new_clip->params_buffer = ggml_backend_alloc_buffer(new_clip->backend, buffer_size);
+        ggml_allocr* alloc = ggml_allocr_new_from_buffer(new_clip->params_buffer);
+        for (int i = 0; i < n_tensors; ++i) {
+            const char * name = gguf_get_tensor_name(ctx, i);
+            struct ggml_tensor * cur = ggml_get_tensor(new_clip->ctx_data, name);
+            ggml_allocr_alloc(alloc, cur);
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
@@ -580,10 +624,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 clip_free(new_clip);
                 return nullptr;
             }
-
-            fin.read(reinterpret_cast<char *>(cur->data), ggml_nbytes(t));
+            int num_bytes = ggml_nbytes(cur);
+            if (ggml_backend_buffer_is_host(new_clip->params_buffer)) {
+                // for the CPU and Metal backend, we can read directly into the tensor
+                fin.read(reinterpret_cast<char *>(cur->data), num_bytes);
+            } else {
+                // read into a temporary buffer first, then copy to device memory
+                read_buf.resize(num_bytes);
+                fin.read(reinterpret_cast<char *>(read_buf.data()), num_bytes);
+                ggml_backend_tensor_set(cur, read_buf.data(), 0, num_bytes);
+            }
         }
-
+        ggml_allocr_free(alloc);
         fin.close();
     }
 
@@ -592,20 +644,20 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         // load vision model
         auto & vision_model = new_clip->vision_model;
         auto & hparams = vision_model.hparams;
-        hparams.hidden_size = get_u32(ctx, format(KEY_N_EMBD, "vision"));
-        hparams.n_head = get_u32(ctx, format(KEY_N_HEAD, "vision"));
+        hparams.hidden_size    = get_u32(ctx, format(KEY_N_EMBD, "vision"));
+        hparams.n_head         = get_u32(ctx, format(KEY_N_HEAD, "vision"));
         hparams.n_intermediate = get_u32(ctx, format(KEY_N_FF, "vision"));
-        hparams.n_layer = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
-        hparams.image_size = get_u32(ctx, KEY_IMAGE_SIZE);
-        hparams.patch_size = get_u32(ctx, KEY_PATCH_SIZE);
+        hparams.n_layer        = get_u32(ctx, format(KEY_N_BLOCK, "vision"));
+        hparams.image_size     = get_u32(ctx, KEY_IMAGE_SIZE);
+        hparams.patch_size     = get_u32(ctx, KEY_PATCH_SIZE);
         hparams.projection_dim = get_u32(ctx, format(KEY_PROJ_DIM, "vision"));
-        hparams.eps = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
+        hparams.eps            = get_f32(ctx, format(KEY_LAYER_NORM_EPS, "vision"));
 
         int idx_mean = get_key_idx(ctx, KEY_IMAGE_MEAN);
-        int idx_std = get_key_idx(ctx, KEY_IMAGE_STD);
+        int idx_std  = get_key_idx(ctx, KEY_IMAGE_STD);
         for (int i = 0; i < 3; ++i) {
             new_clip->image_mean[i] = *((const float *)gguf_get_arr_data(ctx, idx_mean));
-            new_clip->image_std[i] = *((const float *)gguf_get_arr_data(ctx, idx_std));
+            new_clip->image_std[i]  = *((const float *)gguf_get_arr_data(ctx, idx_std));
         }
 
         if (verbosity >= 2) {
@@ -619,35 +671,35 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             printf("v_n_layer          %d\n", hparams.n_layer);
         }
 
-        vision_model.patch_embeddings = get_tensor(new_clip->ctx, TN_PATCH_EMBD);
-        vision_model.class_embedding = get_tensor(new_clip->ctx, TN_CLASS_EMBD);
-        vision_model.position_embeddings = get_tensor(new_clip->ctx, format(TN_POS_EMBD, "v"));
-        vision_model.pre_ln_w = get_tensor(new_clip->ctx, format(TN_LN_PRE, "v", "weight"));
-        vision_model.pre_ln_b = get_tensor(new_clip->ctx, format(TN_LN_PRE, "v", "bias"));
-        vision_model.mm_0_w = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 0, "weight"));
-        vision_model.mm_0_b = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 0, "bias"));
-        vision_model.mm_2_w = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 2, "weight"));
-        vision_model.mm_2_b = get_tensor(new_clip->ctx, format(TN_LLAVA_PROJ, 2, "bias"));
+        vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
+        vision_model.class_embedding     = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+        vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
+        vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+        vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+        vision_model.mm_0_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "weight"));
+        vision_model.mm_0_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 0, "bias"));
+        vision_model.mm_2_w              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "weight"));
+        vision_model.mm_2_b              = get_tensor(new_clip->ctx_data, format(TN_LLAVA_PROJ, 2, "bias"));
 
         vision_model.layers.resize(hparams.n_layer);
         for (int il = 0; il < hparams.n_layer; ++il) {
             auto & layer = vision_model.layers[il];
-            layer.k_w = get_tensor(new_clip->ctx, format(TN_ATTN_K, "v", il, "weight"));
-            layer.q_w = get_tensor(new_clip->ctx, format(TN_ATTN_Q, "v", il, "weight"));
-            layer.v_w = get_tensor(new_clip->ctx, format(TN_ATTN_V, "v", il, "weight"));
-            layer.o_w = get_tensor(new_clip->ctx, format(TN_ATTN_OUTPUT, "v", il, "weight"));
-            layer.ln_1_w = get_tensor(new_clip->ctx, format(TN_LN_1, "v", il, "weight"));
-            layer.ln_2_w = get_tensor(new_clip->ctx, format(TN_LN_2, "v", il, "weight"));
-            layer.ff_i_w = get_tensor(new_clip->ctx, format(TN_FFN_DOWN, "v", il, "weight"));
-            layer.ff_o_w = get_tensor(new_clip->ctx, format(TN_FFN_UP, "v", il, "weight"));
-            layer.k_b = get_tensor(new_clip->ctx, format(TN_ATTN_K, "v", il, "bias"));
-            layer.q_b = get_tensor(new_clip->ctx, format(TN_ATTN_Q, "v", il, "bias"));
-            layer.v_b = get_tensor(new_clip->ctx, format(TN_ATTN_V, "v", il, "bias"));
-            layer.o_b = get_tensor(new_clip->ctx, format(TN_ATTN_OUTPUT, "v", il, "bias"));
-            layer.ln_1_b = get_tensor(new_clip->ctx, format(TN_LN_1, "v", il, "bias"));
-            layer.ln_2_b = get_tensor(new_clip->ctx, format(TN_LN_2, "v", il, "bias"));
-            layer.ff_i_b = get_tensor(new_clip->ctx, format(TN_FFN_DOWN, "v", il, "bias"));
-            layer.ff_o_b = get_tensor(new_clip->ctx, format(TN_FFN_UP, "v", il, "bias"));
+            layer.k_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "weight"));
+            layer.q_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "weight"));
+            layer.v_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "weight"));
+            layer.o_w    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "weight"));
+            layer.ln_1_w = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "weight"));
+            layer.ln_2_w = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "weight"));
+            layer.ff_i_w = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "weight"));
+            layer.ff_o_w = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "weight"));
+            layer.k_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_K,      "v", il, "bias"));
+            layer.q_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_Q,      "v", il, "bias"));
+            layer.v_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_V,      "v", il, "bias"));
+            layer.o_b    = get_tensor(new_clip->ctx_data, format(TN_ATTN_OUTPUT, "v", il, "bias"));
+            layer.ln_1_b = get_tensor(new_clip->ctx_data, format(TN_LN_1,        "v", il, "bias"));
+            layer.ln_2_b = get_tensor(new_clip->ctx_data, format(TN_LN_2,        "v", il, "bias"));
+            layer.ff_i_b = get_tensor(new_clip->ctx_data, format(TN_FFN_DOWN,    "v", il, "bias"));
+            layer.ff_o_b = get_tensor(new_clip->ctx_data, format(TN_FFN_UP,      "v", il, "bias"));
         }
     }
 
@@ -655,45 +707,45 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
     new_clip->ctx_gguf = ctx;
 
-// measure mem requirement and allocate
+    // measure mem requirement and allocate
     {
-        static const size_t tensor_alignment = 32;
-        new_clip->buf_compute.resize(ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead());
-        new_clip->alloc = ggml_allocr_new_measure(tensor_alignment);
+        new_clip->buf_compute_meta.resize(GGML_DEFAULT_GRAPH_SIZE * ggml_tensor_overhead() + ggml_graph_overhead());
+        new_clip->compute_alloc = ggml_allocr_new_measure_from_backend(new_clip->backend);
         clip_image_f32_batch batch;
         batch.size = 1;
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
-        size_t alloc_size = ggml_allocr_alloc_graph(new_clip->alloc, gf) + tensor_alignment;
-        ggml_allocr_free(new_clip->alloc);
-        new_clip->buf_alloc.resize(alloc_size);
-        new_clip->alloc = ggml_allocr_new(new_clip->buf_alloc.data, new_clip->buf_alloc.size, tensor_alignment);
+        size_t compute_memory_buffer_size = ggml_allocr_alloc_graph(new_clip->compute_alloc, gf);
+        ggml_allocr_free(new_clip->compute_alloc);
+        new_clip->compute_buffer = ggml_backend_alloc_buffer(new_clip->backend, compute_memory_buffer_size);
+        new_clip->compute_alloc = ggml_allocr_new_from_buffer(new_clip->compute_buffer);
 
-        printf("%s: total allocated memory: %.2f MB\n", __func__, (new_clip->buf_compute.size + alloc_size)/1024.0/1024.0);
+        printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
     return new_clip;
 }
 
-clip_image_u8 * make_clip_image_u8() {
-    auto img = new clip_image_u8();
-    return img;
+struct clip_image_u8 * clip_image_u8_init() {
+    return new clip_image_u8();
 }
-clip_image_f32 * make_clip_image_f32() { return new clip_image_f32(); }
 
-void clip_image_u8_free(clip_image_u8 * img) { if (img->data) { delete[] img->data; } delete img; }
-void clip_image_f32_free(clip_image_f32 * img) { if (img->data) { delete[] img->data; } delete img; }
+struct clip_image_f32 * clip_image_f32_init() {
+    return new clip_image_f32();
+}
+
+void clip_image_u8_free (struct clip_image_u8  * img) { delete img; }
+void clip_image_f32_free(struct clip_image_f32 * img) { delete img; }
 
 static void build_clip_img_from_data(const stbi_uc * data, int nx, int ny, clip_image_u8 * img) {
     img->nx = nx;
     img->ny = ny;
-    img->size = nx * ny * 3;
-    img->data = new uint8_t[img->size]();
-    memcpy(img->data, data, img->size);
+    img->buf.resize(3 * nx * ny);
+    memcpy(img->buf.data(), data, img->buf.size());
 }
 
 bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
     int nx, ny, nc;
-    auto data = stbi_load(fname, &nx, &ny, &nc, 3);
+    auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
     if (!data) {
         fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
         return false;
@@ -705,7 +757,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
 
 bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img) {
     int nx, ny, nc;
-    auto data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
+    auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
     if (!data) {
         fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
         return false;
@@ -717,7 +769,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
 
 // normalize: x = (x - mean) / std
 // TODO: implement bicubic interpolation instead of linear.
-bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) {
+bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32 * res, const bool pad2square) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -726,18 +778,17 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
     // the logic below is to pad the shorter side to the longer side with a background color: rgb(122, 116, 104)
     // see https://github.com/haotian-liu/LLaVA/blob/e854a2bf85118c504f6f16bf5c3c7c92f8fa8c6b/llava/conversation.py#L113-L156
 
-    clip_image_u8 * temp = make_clip_image_u8(); // we will keep the input image data here temporarily
+    clip_image_u8 * temp = clip_image_u8_init(); // we will keep the input image data here temporarily
     if (pad2square && img->nx != img->ny) {
         int longer_side = std::max(img->nx, img->ny);
         temp->nx = longer_side;
         temp->ny = longer_side;
-        temp->size = 3 * longer_side * longer_side;
-        temp->data = new uint8_t[temp->size]();
-        uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
+        temp->buf.resize(3 * longer_side * longer_side);
+        const uint8_t bc[3] = {122, 116, 104}; // background color in RGB from LLaVA
 
         // fill with background color
-        for (size_t i = 0; i < temp->size; i++) {
-            temp->data[i] = bc[i % 3];
+        for (size_t i = 0; i < temp->buf.size(); i++) {
+            temp->buf[i] = bc[i % 3];
         }
 
         // copy from the input image
@@ -745,17 +796,16 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
             for (int x = 0; x < img->nx; x++) {
                 const int i = 3 * (y * img->nx + x);
                 const int j = 3 * (y * temp->nx + x);
-                temp->data[j] = img->data[i];
-                temp->data[j+1] = img->data[i+1];
-                temp->data[j+2] = img->data[i+2];
+                temp->buf[j]   = img->buf[i];
+                temp->buf[j+1] = img->buf[i+1];
+                temp->buf[j+2] = img->buf[i+2];
             }
         }
     } else {
-        temp->nx   = img->nx;
-        temp->ny   = img->ny;
-        temp->size = img->size;
-        temp->data = new uint8_t[temp->size]();
-        memcpy(&temp->data[0], &img->data[0], temp->size); // copy
+        temp->nx = img->nx;
+        temp->ny = img->ny;
+        temp->buf.resize(img->buf.size());
+        memcpy(temp->buf.data(), img->buf.data(), temp->buf.size());
     }
 
     const int nx = temp->nx;
@@ -766,8 +816,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 
     res->nx = nx2;
     res->ny = ny2;
-    res->size = 3 * nx2 * ny2;
-    res->data = new float[res->size]();
+    res->buf.resize(3 * nx2 * ny2);
 
     const float scale = std::max(nx, ny) / (float)ctx->vision_model.hparams.image_size;
 
@@ -798,10 +847,10 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
                 const int j10 = 3 * (y1 * nx + x0) + c;
                 const int j11 = 3 * (y1 * nx + x1) + c;
 
-                const float v00 = temp->data[j00];
-                const float v01 = temp->data[j01];
-                const float v10 = temp->data[j10];
-                const float v11 = temp->data[j11];
+                const float v00 = temp->buf[j00];
+                const float v01 = temp->buf[j01];
+                const float v10 = temp->buf[j10];
+                const float v11 = temp->buf[j11];
 
                 const float v0 = v00 * (1.0f - dx) + v01 * dx;
                 const float v1 = v10 * (1.0f - dx) + v11 * dx;
@@ -812,7 +861,7 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 
                 const int i = 3 * (y * nx3 + x) + c;
 
-                res->data[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
+                res->buf[i] = ((float(v2) / 255.0f) - m3[c]) / s3[c];
             }
         }
     }
@@ -822,12 +871,13 @@ bool clip_image_preprocess(const clip_ctx * ctx, const clip_image_u8 * img, clip
 }
 
 void clip_free(clip_ctx * ctx) {
-    ggml_free(ctx->ctx);
+    ggml_free(ctx->ctx_data);
     gguf_free(ctx->ctx_gguf);
+
     delete ctx;
 }
 
-bool clip_image_encode(const clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
+bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -839,8 +889,7 @@ bool clip_image_encode(const clip_ctx * ctx, const int n_threads, clip_image_f32
     return clip_image_batch_encode(ctx, n_threads, &imgs, vec);
 }
 
-bool clip_image_batch_encode(const clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
-
+bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
     if (!ctx->has_vision_encoder) {
         printf("This gguf file seems to have no vision encoder\n");
         return false;
@@ -852,29 +901,29 @@ bool clip_image_batch_encode(const clip_ctx * ctx, const int n_threads, const cl
     }
 
     // reset alloc buffer to clean the memory from previous invocations
-    ggml_allocr_reset(ctx->alloc);
+    ggml_allocr_reset(ctx->compute_alloc);
 
     // build the inference graph
     ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
-    ggml_allocr_alloc_graph(ctx->alloc, gf);
+    ggml_allocr_alloc_graph(ctx->compute_alloc, gf);
 
-    struct ggml_cplan plan = ggml_graph_plan(gf, n_threads);
-    if (plan.work_size > 0) {
-        plan.work_data = (uint8_t *)malloc(plan.work_size);
+    if (ggml_backend_is_cpu(ctx->backend)) {
+        ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
     }
 
-    ggml_graph_compute(gf, &plan);
+#ifdef GGML_USE_METAL
+    if (ggml_backend_is_metal(ctx->backend)) {
+        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
+    }
+#endif
+
+    ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor
-struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
 
     // copy the embeddings to the location passed by the user
-    memcpy(vec, ggml_get_data_f32(embeddings), ggml_nbytes(embeddings));
-
-    if (plan.work_size > 0) {
-        free(plan.work_data);
-    }
-
+    ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));
     return true;
 }
 
@@ -883,31 +932,32 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     ggml_type type = GGML_TYPE_Q4_1;
 
     switch (itype) {
-    case 2:
-        type = GGML_TYPE_Q4_0;
-        break;
-    case 3:
-        type = GGML_TYPE_Q4_1;
-        break;
-    case 6:
-        type = GGML_TYPE_Q5_0;
-        break;
-    case 7:
-        type = GGML_TYPE_Q5_1;
-        break;
-    case 8:
-        type = GGML_TYPE_Q8_0;
-        break;
-    default:
-        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype);
-        return false;
+        case 2:
+            type = GGML_TYPE_Q4_0;
+            break;
+        case 3:
+            type = GGML_TYPE_Q4_1;
+            break;
+        case 6:
+            type = GGML_TYPE_Q5_0;
+            break;
+        case 7:
+            type = GGML_TYPE_Q5_1;
+            break;
+        case 8:
+            type = GGML_TYPE_Q8_0;
+            break;
+        default:
+            fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype);
+            return false;
     };
 
-    auto ctx_clip = clip_model_load(fname_inp, 2);
-    const auto & ctx_src = ctx_clip->ctx_gguf;
-    const auto & ctx_data = ctx_clip->ctx;
+    auto * ctx_clip = clip_model_load(fname_inp, 2);
 
-    auto ctx_out = gguf_init_empty();
+    const auto & ctx_src = ctx_clip->ctx_gguf;
+    const auto & ctx_data = ctx_clip->ctx_data;
+
+    auto * ctx_out = gguf_init_empty();
     gguf_set_kv(ctx_out, ctx_src);
     gguf_set_val_u32(ctx_out, "general.quantization_version", GGML_QNT_VERSION);
     gguf_set_val_u32(ctx_out, "general.file_type", itype);
@@ -1045,8 +1095,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     gguf_free(ctx_out);
 
     {
-        printf("%s: original size  = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
-        printf("%s: quantized size  = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+        printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
 
         int64_t sum_all = 0;
         for (size_t i = 0; i < hist_all.size(); ++i) {

examples/llava/clip.h

@@ -35,31 +35,14 @@ struct clip_vision_hparams {
     float eps;
 };
 
-/** load mmproj model */
-CLIP_API struct clip_ctx * clip_model_load(const char * fname, const int verbosity);
-/** free mmproj model */
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
 CLIP_API void clip_free(struct clip_ctx * ctx);
 
-size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-int clip_n_patches(const struct clip_ctx * ctx);
-int clip_n_mmproj_embd(const struct clip_ctx * ctx);
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
 
-// RGB uint8 image
-struct clip_image_u8 {
-    int nx;
-    int ny;
-    uint8_t * data = NULL;
-    size_t size;
-};
-
-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
-struct clip_image_f32 {
-    int nx;
-    int ny;
-    float * data = NULL;
-    size_t size;
-};
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -71,21 +54,22 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
-struct clip_image_u8 * make_clip_image_u8();
-struct clip_image_f32 * make_clip_image_f32();
-CLIP_API void clip_image_u8_free(clip_image_u8 * img);
-CLIP_API void clip_image_f32_free(clip_image_f32 * img);
+CLIP_API struct clip_image_u8  * clip_image_u8_init ();
+CLIP_API struct clip_image_f32 * clip_image_f32_init();
+
+CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
+CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
+
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
+
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-bool clip_image_preprocess(const struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, const bool pad2square);
-bool clip_image_encode(const struct clip_ctx * ctx, const int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
+CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
+CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 
-bool clip_image_batch_encode(const struct clip_ctx * ctx, const int n_threads, const struct clip_image_f32_batch * imgs,
-                             float * vec);
-
-bool clip_model_quantize(const char * fname_inp, const char * fname_out, const int itype);
+CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
 #ifdef __cplusplus
 }

examples/llava/llava-cli.cpp

@@ -39,73 +39,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
     return true;
 }
 
-// TODO: use common/sampling.h
-static llama_token sample_id(llama_context * ctx_llama, gpt_params & params) {
-    auto & sparams = params.sparams;
-
-    // out of user input, sample next token
-    const float   temp      = sparams.temp;
-    const int32_t top_k     = sparams.top_k <= 0 ? llama_n_vocab(llama_get_model(ctx_llama)) : sparams.top_k;
-    const float   top_p     = sparams.top_p;
-    const float   tfs_z     = sparams.tfs_z;
-    const float   typical_p = sparams.typical_p;
-    // const int32_t repeat_last_n   = sparams.repeat_last_n < 0 ? n_ctx : sparams.repeat_last_n;
-    // const float   repeat_penalty  = sparams.repeat_penalty;
-    // const float   alpha_presence  = sparams.presence_penalty;
-    // const float   alpha_frequency = sparams.frequency_penalty;
-    const int     mirostat     = sparams.mirostat;
-    const float   mirostat_tau = sparams.mirostat_tau;
-    const float   mirostat_eta = sparams.mirostat_eta;
-    // const bool    penalize_nl     = sparams.penalize_nl;
-
-    llama_token id = 0;
-    {
-        auto logits  = llama_get_logits(ctx_llama);
-        auto n_vocab = llama_n_vocab(llama_get_model(ctx_llama));
-
-        // Apply params.logit_bias map
-        for (auto it = sparams.logit_bias.begin(); it != sparams.logit_bias.end(); it++) {
-            logits[it->first] += it->second;
-        }
-
-        std::vector<llama_token_data> candidates;
-        candidates.reserve(n_vocab);
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-        }
-
-        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
-        if (temp <= 0) {
-              // Greedy sampling
-            id = llama_sample_token_greedy(ctx_llama, &candidates_p);
-        } else {
-            if (mirostat == 1) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                const  int mirostat_m    = 100;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, mirostat_m, &mirostat_mu);
-            } else if (mirostat == 2) {
-                static float mirostat_mu = 2.0f * mirostat_tau;
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token_mirostat_v2(ctx_llama, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
-            } else {
-                  // Temperature sampling
-                llama_sample_top_k(ctx_llama, &candidates_p, top_k, 1);
-                llama_sample_tail_free(ctx_llama, &candidates_p, tfs_z, 1);
-                llama_sample_typical(ctx_llama, &candidates_p, typical_p, 1);
-                llama_sample_top_p(ctx_llama, &candidates_p, top_p, 1);
-                llama_sample_temp(ctx_llama, &candidates_p, temp);
-                id = llama_sample_token(ctx_llama, &candidates_p);
-            }
-        }
-    }
-
-    return id;
-}
-
-static const char * sample(struct llama_context * ctx_llama, gpt_params & params, int * n_past) {
-    int id = sample_id(ctx_llama, params);
+static const char * sample(struct llama_sampling_context * ctx_sampling,
+                           struct llama_context * ctx_llama,
+                           int * n_past) {
+    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
+    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
     static std::string ret;
     if (id == llama_token_eos(llama_get_model(ctx_llama))) {
         ret = "</s>";
@@ -174,8 +112,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    printf("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    printf("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    fprintf(stderr, "\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    fprintf(stderr, "  note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
@@ -185,7 +123,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     auto prompt = params->prompt;
     if (prompt_contains_image(prompt)) {
         if (!params->image.empty()) {
-            printf("using base64 encoded image instead of command line image path\n");
+            fprintf(stderr, "using base64 encoded image instead of command line image path\n");
         }
         embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
         if (!embed) {
@@ -217,16 +155,19 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     // generate the response
 
-    printf("\n");
+    fprintf(stderr, "\n");
+
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
 
     for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_llava->ctx_llama, *params, &n_past);
+        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
         if (strcmp(tmp, "</s>") == 0) break;
 
         printf("%s", tmp);
         fflush(stdout);
     }
 
+    llama_sampling_free(ctx_sampling);
     printf("\n");
 }
 

examples/llava/llava.cpp

@@ -10,7 +10,7 @@
 #include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    clip_image_f32 * img_res = make_clip_image_f32();
+    clip_image_f32 * img_res = clip_image_f32_init();
     if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
         clip_image_f32_free(img_res);
@@ -86,7 +86,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
 }
 
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
-    clip_image_u8 * img = make_clip_image_u8();
+    clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
         fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);

examples/server/CMakeLists.txt

@@ -6,7 +6,7 @@ install(TARGETS ${TARGET} RUNTIME)
 target_compile_definitions(${TARGET} PRIVATE
     SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
 )
-target_link_libraries(${TARGET} PRIVATE common llama llava ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
 if (WIN32)
     TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()

examples/server/server.cpp

@@ -25,6 +25,7 @@
 #include <thread>
 #include <mutex>
 #include <chrono>
+#include <condition_variable>
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -81,7 +82,7 @@ static inline bool is_base64(uint8_t c)
     return (isalnum(c) || (c == '+') || (c == '/'));
 }
 
-static std::vector<uint8_t> base64_decode(std::string const &encoded_string)
+static std::vector<uint8_t> base64_decode(const std::string & encoded_string)
 {
     int i = 0;
     int j = 0;
@@ -208,10 +209,10 @@ struct slot_image
     int32_t id;
 
     bool request_encode_image = false;
-    float* image_embedding = nullptr;
+    float * image_embedding = nullptr;
     int32_t image_tokens = 0;
 
-    clip_image_u8 img_data;
+    clip_image_u8 * img_data;
 
     std::string prefix_prompt; // before of this image
 };
@@ -433,10 +434,12 @@ struct llama_client_slot
 
         generated_token_probs.clear();
 
-        for (slot_image &img : images)
+        for (slot_image & img : images)
         {
             free(img.image_embedding);
-            delete[] img.img_data.data;
+            if (img.img_data) {
+                clip_image_u8_free(img.img_data);
+            }
             img.prefix_prompt = "";
         }
 
@@ -541,7 +544,9 @@ struct llama_server_context
     std::vector<task_result> queue_results;
     std::vector<task_multi>  queue_multitasks;
     std::mutex mutex_tasks; // also guards id_gen, and queue_multitasks
+    std::condition_variable condition_tasks;
     std::mutex mutex_results;
+    std::condition_variable condition_results;
 
     ~llama_server_context()
     {
@@ -848,24 +853,17 @@ struct llama_server_context
             {
                 for (const auto &img : *images_data)
                 {
-                    std::string data_b64 = img["data"].get<std::string>();
+                    const std::vector<uint8_t> image_buffer = base64_decode(img["data"].get<std::string>());
+
                     slot_image img_sl;
                     img_sl.id = img.count("id") != 0 ? img["id"].get<int>() : slot->images.size();
-                    int width, height, channels;
-                    std::vector<uint8_t> image_buffer = base64_decode(data_b64);
-                    data_b64.clear();
-                    auto data = stbi_load_from_memory(image_buffer.data(), image_buffer.size(), &width, &height, &channels, 3);
-                    if (!data) {
+                    img_sl.img_data = clip_image_u8_init();
+                    if (!clip_image_load_from_bytes(image_buffer.data(), image_buffer.size(), img_sl.img_data))
+                    {
                         LOG_TEE("slot %i - failed to load image [id: %i]\n", slot->id, img_sl.id);
                         return false;
                     }
-                    LOG_TEE("slot %i - image loaded [id: %i] resolution (%i x %i)\n", slot->id, img_sl.id, width, height);
-                    img_sl.img_data.nx = width;
-                    img_sl.img_data.ny = height;
-                    img_sl.img_data.size = width * height * 3;
-                    img_sl.img_data.data = new uint8_t[width * height * 3]();
-                    memcpy(img_sl.img_data.data, data, width * height * 3);
-                    stbi_image_free(data);
+                    LOG_TEE("slot %i - loaded image\n", slot->id);
                     img_sl.request_encode_image = true;
                     slot->images.push_back(img_sl);
                 }
@@ -1140,8 +1138,8 @@ struct llama_server_context
             {
                 continue;
             }
-            clip_image_f32 img_res;
-            if (!clip_image_preprocess(clp_ctx, &img.img_data, &img_res, /*pad2square =*/ true))
+            clip_image_f32 * img_res = clip_image_f32_init();
+            if (!clip_image_preprocess(clp_ctx, img.img_data, img_res, /*pad2square =*/ true))
             {
                 LOG_TEE("Error processing the given image");
                 clip_free(clp_ctx);
@@ -1156,11 +1154,12 @@ struct llama_server_context
                 return false;
             }
             LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
-            if (!clip_image_encode(clp_ctx, params.n_threads, &img_res, img.image_embedding))
+            if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
             {
                 LOG_TEE("Unable to encode image\n");
                 return false;
             }
+            clip_image_f32_free(img_res);
             img.request_encode_image = false;
         }
 
@@ -1169,7 +1168,7 @@ struct llama_server_context
 
     void send_error(task_server& task, std::string error)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = task.id;
         res.multitask_id = task.multitask_id;
@@ -1177,6 +1176,7 @@ struct llama_server_context
         res.error = true;
         res.result_json = { { "content", error } };
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void add_multi_task(int id, std::vector<int>& sub_ids)
@@ -1186,6 +1186,7 @@ struct llama_server_context
         multi.id = id;
         std::copy(sub_ids.begin(), sub_ids.end(), std::inserter(multi.subtasks_remaining, multi.subtasks_remaining.end()));
         queue_multitasks.push_back(multi);
+        condition_tasks.notify_one();
     }
 
     void update_multi_task(int multitask_id, int subtask_id, task_result& result)
@@ -1197,6 +1198,7 @@ struct llama_server_context
             {
                 multitask.subtasks_remaining.erase(subtask_id);
                 multitask.results.push_back(result);
+                condition_tasks.notify_one();
             }
         }
     }
@@ -1244,7 +1246,7 @@ struct llama_server_context
 
     void send_partial_response(llama_client_slot &slot, completion_token_output tkn)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1280,11 +1282,12 @@ struct llama_server_context
         }
 
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void send_final_response(llama_client_slot &slot)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1340,11 +1343,12 @@ struct llama_server_context
         }
 
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     void send_embedding(llama_client_slot &slot)
     {
-        std::lock_guard<std::mutex> lock(mutex_results);
+        std::unique_lock<std::mutex> lock(mutex_results);
         task_result res;
         res.id = slot.task_id;
         res.multitask_id = slot.multitask_id;
@@ -1372,6 +1376,7 @@ struct llama_server_context
             };
         }
         queue_results.push_back(res);
+        condition_results.notify_all();
     }
 
     int request_completion(json data, bool infill, bool embedding, int multitask_id)
@@ -1395,6 +1400,7 @@ struct llama_server_context
 
         // otherwise, it's a single-prompt task, we actually queue it
         queue_tasks.push_back(task);
+        condition_tasks.notify_one();
         return task.id;
     }
 
@@ -1402,13 +1408,10 @@ struct llama_server_context
     {
         while (true)
         {
-            std::this_thread::sleep_for(std::chrono::microseconds(5));
-            std::lock_guard<std::mutex> lock(mutex_results);
-
-            if (queue_results.empty())
-            {
-                continue;
-            }
+            std::unique_lock<std::mutex> lock(mutex_results);
+            condition_results.wait(lock, [&]{
+                return !queue_results.empty();
+            });
 
             for (int i = 0; i < (int) queue_results.size(); i++)
             {
@@ -1504,12 +1507,13 @@ struct llama_server_context
 
     void request_cancel(int task_id)
     {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
+        std::unique_lock<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
         task.type = CANCEL_TASK;
         task.target_id = task_id;
         queue_tasks.push_back(task);
+        condition_tasks.notify_one();
     }
 
     int split_multiprompt_task(task_server& multiprompt_task)
@@ -1535,7 +1539,7 @@ struct llama_server_context
 
     void process_tasks()
     {
-        std::lock_guard<std::mutex> lock(mutex_tasks);
+        std::unique_lock<std::mutex> lock(mutex_tasks);
         while (!queue_tasks.empty())
         {
             task_server task = queue_tasks.front();
@@ -1607,6 +1611,7 @@ struct llama_server_context
 
                 std::lock_guard<std::mutex> lock(mutex_results);
                 queue_results.push_back(aggregate_result);
+                condition_results.notify_all();
 
                 queue_iterator = queue_multitasks.erase(queue_iterator);
             }
@@ -1637,8 +1642,10 @@ struct llama_server_context
                 LOG_TEE("all slots are idle and system prompt is empty, clear the KV cache\n");
                 kv_cache_clear();
             }
-            // avoid 100% usage of cpu all time
-            std::this_thread::sleep_for(std::chrono::milliseconds(5));
+            std::unique_lock<std::mutex> lock(mutex_tasks);
+            condition_tasks.wait(lock, [&]{
+                return !queue_tasks.empty();
+            });
         }
 
         for (llama_client_slot &slot : slots)

flake.lock

@@ -1,30 +1,30 @@
 {
   "nodes": {
-    "flake-utils": {
+    "flake-parts": {
       "inputs": {
-        "systems": "systems"
+        "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1694529238,
-        "narHash": "sha256-zsNZZGTGnMOf9YpHKJqMSsa0dXbfmxeoJ7xHlrt+xmY=",
-        "owner": "numtide",
-        "repo": "flake-utils",
-        "rev": "ff7b65b44d01cf9ba6a71320833626af21126384",
+        "lastModified": 1701473968,
+        "narHash": "sha256-YcVE5emp1qQ8ieHUnxt1wCZCC3ZfAS+SRRWZ2TMda7E=",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
+        "rev": "34fed993f1674c8d06d58b37ce1e0fe5eebcb9f5",
         "type": "github"
       },
       "original": {
-        "owner": "numtide",
-        "repo": "flake-utils",
+        "owner": "hercules-ci",
+        "repo": "flake-parts",
         "type": "github"
       }
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1698318101,
-        "narHash": "sha256-gUihHt3yPD7bVqg+k/UVHgngyaJ3DMEBchbymBMvK1E=",
+        "lastModified": 1703559957,
+        "narHash": "sha256-x9PUuMEPGUOMB51zNxrDr2QoHbYWlCS2xhFedm9MC5Q=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "63678e9f3d3afecfeafa0acead6239cdb447574c",
+        "rev": "75dd68c36f458c6593c5bbb48abfd3e59bfed380",
         "type": "github"
       },
       "original": {
@@ -34,26 +34,29 @@
         "type": "github"
       }
     },
-    "root": {
-      "inputs": {
-        "flake-utils": "flake-utils",
-        "nixpkgs": "nixpkgs"
-      }
-    },
-    "systems": {
+    "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1681028828,
-        "narHash": "sha256-Vy1rq5AaRuLzOxct8nz4T6wlgyUR7zLU309k9mBC768=",
-        "owner": "nix-systems",
-        "repo": "default",
-        "rev": "da67096a3b9bf56a91d16901293e51ba5b49a27e",
+        "dir": "lib",
+        "lastModified": 1701253981,
+        "narHash": "sha256-ztaDIyZ7HrTAfEEUt9AtTDNoCYxUdSd6NrRHaYOIxtk=",
+        "owner": "NixOS",
+        "repo": "nixpkgs",
+        "rev": "e92039b55bcd58469325ded85d4f58dd5a4eaf58",
         "type": "github"
       },
       "original": {
-        "owner": "nix-systems",
-        "repo": "default",
+        "dir": "lib",
+        "owner": "NixOS",
+        "ref": "nixos-unstable",
+        "repo": "nixpkgs",
         "type": "github"
       }
+    },
+    "root": {
+      "inputs": {
+        "flake-parts": "flake-parts",
+        "nixpkgs": "nixpkgs"
+      }
     }
   },
   "root": "root",

flake.nix

@@ -1,139 +1,99 @@
 {
+  description = "Port of Facebook's LLaMA model in C/C++";
+
   inputs = {
     nixpkgs.url = "github:NixOS/nixpkgs/nixos-unstable";
-    flake-utils.url = "github:numtide/flake-utils";
+    flake-parts.url = "github:hercules-ci/flake-parts";
   };
-  outputs = { self, nixpkgs, flake-utils }:
-    flake-utils.lib.eachDefaultSystem (system:
-      let
-        name = "llama.cpp";
-        src = ./.;
-        meta.mainProgram = "llama";
-        inherit (pkgs.stdenv) isAarch32 isAarch64 isDarwin;
-        buildInputs = with pkgs; [ openmpi ];
-        osSpecific = with pkgs; buildInputs ++ (
-          if isAarch64 && isDarwin then
-            with pkgs.darwin.apple_sdk_11_0.frameworks; [
-              Accelerate
-              MetalKit
-            ]
-          else if isAarch32 && isDarwin then
-            with pkgs.darwin.apple_sdk.frameworks; [
-              Accelerate
-              CoreGraphics
-              CoreVideo
-            ]
-          else if isDarwin then
-            with pkgs.darwin.apple_sdk.frameworks; [
-              Accelerate
-              CoreGraphics
-              CoreVideo
-            ]
-          else
-            with pkgs; [ openblas ]
-        );
-        pkgs = import nixpkgs { inherit system; };
-        nativeBuildInputs = with pkgs; [ cmake ninja pkg-config ];
-        cudatoolkit_joined = with pkgs; symlinkJoin {
-          # HACK(Green-Sky): nix currently has issues with cmake findcudatoolkit
-          # see https://github.com/NixOS/nixpkgs/issues/224291
-          # copied from jaxlib
-          name = "${cudaPackages.cudatoolkit.name}-merged";
-          paths = [
-            cudaPackages.cudatoolkit.lib
-            cudaPackages.cudatoolkit.out
-          ] ++ lib.optionals (lib.versionOlder cudaPackages.cudatoolkit.version "11") [
-            # for some reason some of the required libs are in the targets/x86_64-linux
-            # directory; not sure why but this works around it
-            "${cudaPackages.cudatoolkit}/targets/${system}"
-          ];
-        };
-        llama-python =
-          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece ]);
-        # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
-        llama-python-extra =
-          pkgs.python3.withPackages (ps: with ps; [ numpy sentencepiece torchWithoutCuda transformers ]);
-        postPatch = ''
-          substituteInPlace ./ggml-metal.m \
-            --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-          substituteInPlace ./*.py --replace '/usr/bin/env python' '${llama-python}/bin/python'
-        '';
-        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/
-        '';
-        cmakeFlags = [ "-DLLAMA_NATIVE=OFF" "-DLLAMA_BUILD_SERVER=ON" "-DBUILD_SHARED_LIBS=ON" "-DCMAKE_SKIP_BUILD_RPATH=ON" ];
-      in
+
+  # For inspection, use `nix flake show github:ggerganov/llama.cpp` or the nix repl:
+  #
+  # ```bash
+  # ❯ nix repl
+  # nix-repl> :lf github:ggerganov/llama.cpp
+  # Added 13 variables.
+  # nix-repl> outputs.apps.x86_64-linux.quantize
+  # { program = "/nix/store/00000000000000000000000000000000-llama.cpp/bin/quantize"; type = "app"; }
+  # ```
+  outputs =
+    { self, flake-parts, ... }@inputs:
+    let
+      # We could include the git revisions in the package names but those would
+      # needlessly trigger rebuilds:
+      # llamaVersion = self.dirtyShortRev or self.shortRev;
+
+      # Nix already uses cryptographic hashes for versioning, so we'll just fix
+      # the fake semver for now:
+      llamaVersion = "0.0.0";
+    in
+    flake-parts.lib.mkFlake { inherit inputs; }
+
       {
-        packages.default = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = osSpecific;
-          cmakeFlags = cmakeFlags
-            ++ (if isAarch64 && isDarwin then [
-            "-DCMAKE_C_FLAGS=-D__ARM_FEATURE_DOTPROD=1"
-            "-DLLAMA_METAL=ON"
-          ] else [
-            "-DLLAMA_BLAS=ON"
-            "-DLLAMA_BLAS_VENDOR=OpenBLAS"
-          ]);
-        };
-        packages.opencl = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ clblast ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_CLBLAST=ON"
-          ];
-        };
-        packages.cuda = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs; buildInputs ++ [ cudatoolkit_joined ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_CUBLAS=ON"
-          ];
-        };
-        packages.rocm = pkgs.stdenv.mkDerivation {
-          inherit name src meta postPatch nativeBuildInputs postInstall;
-          buildInputs = with pkgs.rocmPackages; buildInputs ++ [ clr hipblas rocblas ];
-          cmakeFlags = cmakeFlags ++ [
-            "-DLLAMA_HIPBLAS=1"
-            "-DCMAKE_C_COMPILER=hipcc"
-            "-DCMAKE_CXX_COMPILER=hipcc"
-            # Build all targets supported by rocBLAS. When updating search for TARGET_LIST_ROCM
-            # in github.com/ROCmSoftwarePlatform/rocBLAS/blob/develop/CMakeLists.txt
-            # and select the line that matches the current nixpkgs version of rocBLAS.
-            "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
-          ];
-        };
-        apps.llama-server = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/llama-server";
-        };
-        apps.llama-embedding = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/embedding";
-        };
-        apps.llama = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/llama";
-        };
-        apps.quantize = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/quantize";
-        };
-        apps.train-text-from-scratch = {
-          type = "app";
-          program = "${self.packages.${system}.default}/bin/train-text-from-scratch";
-        };
-        apps.default = self.apps.${system}.llama;
-        devShells.default = pkgs.mkShell {
-          buildInputs = [ llama-python ];
-          packages = nativeBuildInputs ++ osSpecific;
-        };
-        devShells.extra = pkgs.mkShell {
-          buildInputs = [ llama-python-extra ];
-          packages = nativeBuildInputs ++ osSpecific;
-        };
-      });
+
+        imports = [
+          .devops/nix/nixpkgs-instances.nix
+          .devops/nix/apps.nix
+          .devops/nix/devshells.nix
+          .devops/nix/jetson-support.nix
+        ];
+
+        # An overlay can be used to have a more granular control over llama-cpp's
+        # dependencies and configuration, than that offered by the `.override`
+        # mechanism. Cf. https://nixos.org/manual/nixpkgs/stable/#chap-overlays.
+        #
+        # E.g. in a flake:
+        # ```
+        # { nixpkgs, llama-cpp, ... }:
+        # let pkgs = import nixpkgs {
+        #     overlays = [ (llama-cpp.overlays.default) ];
+        #     system = "aarch64-linux";
+        #     config.allowUnfree = true;
+        #     config.cudaSupport = true;
+        #     config.cudaCapabilities = [ "7.2" ];
+        #     config.cudaEnableForwardCompat = false;
+        # }; in {
+        #     packages.aarch64-linux.llamaJetsonXavier = pkgs.llamaPackages.llama-cpp;
+        # }
+        # ```
+        #
+        # Cf. https://nixos.org/manual/nix/unstable/command-ref/new-cli/nix3-flake.html?highlight=flake#flake-format
+        flake.overlays.default =
+          (final: prev: {
+            llamaPackages = final.callPackage .devops/nix/scope.nix { inherit llamaVersion; };
+            inherit (final.llamaPackages) llama-cpp;
+          });
+
+        systems = [
+          "aarch64-darwin"
+          "aarch64-linux"
+          "x86_64-darwin" # x86_64-darwin isn't tested (and likely isn't relevant)
+          "x86_64-linux"
+        ];
+
+        perSystem =
+          {
+            config,
+            lib,
+            pkgs,
+            pkgsCuda,
+            pkgsRocm,
+            ...
+          }:
+          {
+            # We don't use the overlay here so as to avoid making too many instances of nixpkgs,
+            # cf. https://zimbatm.com/notes/1000-instances-of-nixpkgs
+            packages =
+              {
+                default = (pkgs.callPackage .devops/nix/scope.nix { inherit llamaVersion; }).llama-cpp;
+              }
+              // lib.optionalAttrs pkgs.stdenv.isLinux {
+                opencl = config.packages.default.override { useOpenCL = true; };
+                cuda = (pkgsCuda.callPackage .devops/nix/scope.nix { inherit llamaVersion; }).llama-cpp;
+                rocm = (pkgsRocm.callPackage .devops/nix/scope.nix { inherit llamaVersion; }).llama-cpp;
+
+                mpi-cpu = config.packages.default.override { useMpi = true; };
+                mpi-cuda = config.packages.default.override { useMpi = true; };
+              };
+          };
+      };
 }

ggml-cuda.cu

@@ -119,7 +119,9 @@
 #define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
 #define CC_VOLTA      700
 #define CC_OFFSET_AMD 1000000
+#define CC_RDNA1      (CC_OFFSET_AMD + 1010)
 #define CC_RDNA2      (CC_OFFSET_AMD + 1030)
+#define CC_RDNA3      (CC_OFFSET_AMD + 1100)
 
 #define GGML_CUDA_MAX_NODES 8192
 
@@ -133,7 +135,6 @@
 
 // TODO: improve this to be correct for more hardware
 //       for example, currently fails for GeForce GTX 1660 which is TURING arch (> VOLTA) but does not have tensor cores
-//       probably other such cases, and not sure what happens on AMD hardware
 #if !defined(GGML_CUDA_FORCE_MMQ)
 #define CUDA_USE_TENSOR_CORES
 #endif
@@ -6662,7 +6663,7 @@ static void ggml_cuda_pool_free_leg(int device, void * ptr, size_t size) {
 // pool with virtual memory
 static CUdeviceptr g_cuda_pool_addr[GGML_CUDA_MAX_DEVICES] = {0};
 static size_t g_cuda_pool_used[GGML_CUDA_MAX_DEVICES] = {0};
-static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 36; // 64 GB
+static const size_t CUDA_POOL_VMM_MAX_SIZE = 1ull << 35; // 32 GB
 
 static void * ggml_cuda_pool_malloc_vmm(int device, size_t size, size_t * actual_size) {
     scoped_spin_lock lock(g_cuda_pool_lock);
@@ -8661,11 +8662,25 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
         }
     }
 
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+    const bool fp16_performance_good = min_compute_capability >= CC_RDNA1;
+    bool               use_mul_mat_q = ggml_is_quantized(src0->type);
 #ifdef CUDA_USE_TENSOR_CORES
-    const bool use_tensor_cores = true;
+    use_mul_mat_q = use_mul_mat_q && min_compute_capability < CC_RDNA3;
+#endif // CUDA_USE_TENSOR_CORES
+
 #else
-    const bool use_tensor_cores = false;
-#endif
+
+    const bool fp16_performance_good = min_compute_capability >= CC_VOLTA;
+    bool               use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
+#ifdef CUDA_USE_TENSOR_CORES
+    // when tensor cores are available, use them for large batch size
+    // ref: https://github.com/ggerganov/llama.cpp/pull/3776
+    use_mul_mat_q = use_mul_mat_q && !(fp16_performance_good && src1->ne[1] > MMQ_MAX_BATCH_SIZE);
+#endif // CUDA_USE_TENSOR_CORES
+
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
     // debug helpers
     //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
@@ -8675,13 +8690,13 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
     //printf("src0 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src0), ggml_is_transposed(src0), ggml_type_name(src0->type), src0->name);
     //printf("src1 is contiguous %d, transposed %d, type = %s, name = %s\n", ggml_is_contiguous(src1), ggml_is_transposed(src1), ggml_type_name(src1->type), src1->name);
 
-    if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
+    if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && ggml_is_permuted(src0) && ggml_is_permuted(src1) && src1->ne[1] == 1) {
         // KQ single-batch
         ggml_cuda_mul_mat_vec_p021(src0, src1, dst);
-    } else if (!split && all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
+    } else if (!split && all_on_device && !fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
         // KQV single-batch
         ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    } else if (!split && all_on_device && use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
+    } else if (!split && all_on_device && fp16_performance_good && src0->type == GGML_TYPE_F16 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
         // KQ + KQV multi-batch
         ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
     } else if (src0->type == GGML_TYPE_F32) {
@@ -8701,14 +8716,6 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_dequantize_mul_mat_vec, false);
             }
         } else {
-            bool use_mul_mat_q = min_compute_capability >= MIN_CC_DP4A && ggml_is_quantized(src0->type);
-
-            // when tensor cores are available, use them for large batch size
-            // ref: https://github.com/ggerganov/llama.cpp/pull/3776
-            if (use_tensor_cores && min_compute_capability >= CC_VOLTA && src1->ne[1] > MMQ_MAX_BATCH_SIZE) {
-                use_mul_mat_q = false;
-            }
-
             if (use_mul_mat_q) {
                 ggml_cuda_op_mul_mat(src0, src1, dst, ggml_cuda_op_mul_mat_q, true);
             } else {

ggml.c

@@ -19638,6 +19638,14 @@ int ggml_cpu_has_avx(void) {
 #endif
 }
 
+int ggml_cpu_has_avx_vnni(void) {
+#if defined(__AVXVNNI__)
+    return 1;
+#else
+    return 0;
+#endif
+}
+
 int ggml_cpu_has_avx2(void) {
 #if defined(__AVX2__)
     return 1;

ggml.h

@@ -2198,6 +2198,7 @@ extern "C" {
     //
 
     GGML_API int ggml_cpu_has_avx        (void);
+    GGML_API int ggml_cpu_has_avx_vnni   (void);
     GGML_API int ggml_cpu_has_avx2       (void);
     GGML_API int ggml_cpu_has_avx512     (void);
     GGML_API int ggml_cpu_has_avx512_vbmi(void);

llama.cpp

@@ -10780,6 +10780,7 @@ const char * llama_print_system_info(void) {
 
     s  = "";
     s += "AVX = "         + std::to_string(ggml_cpu_has_avx())         + " | ";
+    s += "AVX_VNNI = "    + std::to_string(ggml_cpu_has_avx_vnni())    + " | ";
     s += "AVX2 = "        + std::to_string(ggml_cpu_has_avx2())        + " | ";
     s += "AVX512 = "      + std::to_string(ggml_cpu_has_avx512())      + " | ";
     s += "AVX512_VBMI = " + std::to_string(ggml_cpu_has_avx512_vbmi()) + " | ";

requirements-hf-to-gguf.txt

@@ -1,3 +0,0 @@
--r requirements.txt
-torch==2.1.1
-transformers==4.35.2

requirements.txt

@@ -1,5 +1,12 @@
-numpy==1.24.4
-sentencepiece==0.1.98
-transformers>=4.34.0
-gguf>=0.1.0
-protobuf>=4.21.0
+# These requirements include all dependencies for all top-level python scripts
+# for llama.cpp. Avoid adding packages here directly.
+#
+# Package versions must stay compatible across all top-level python scripts.
+#
+
+-r ./requirements/requirements-convert.txt
+
+-r ./requirements/requirements-convert-hf-to-gguf.txt
+-r ./requirements/requirements-convert-llama-ggml-to-gguf.txt
+-r ./requirements/requirements-convert-lora-to-ggml.txt
+-r ./requirements/requirements-convert-persimmon-to-gguf.txt

requirements/requirements-convert-hf-to-gguf.txt

@@ -0,0 +1,2 @@
+-r ./requirements-convert.txt
+torch~=2.1.1

requirements/requirements-convert-llama-ggml-to-gguf.txt

@@ -0,0 +1 @@
+-r ./requirements-convert.txt

requirements/requirements-convert-lora-to-ggml.txt

@@ -0,0 +1,2 @@
+-r ./requirements-convert.txt
+torch~=2.1.1

requirements/requirements-convert-persimmon-to-gguf.txt

@@ -0,0 +1,2 @@
+-r ./requirements-convert.txt
+torch~=2.1.1

requirements/requirements-convert.txt

@@ -0,0 +1,5 @@
+numpy~=1.24.4
+sentencepiece~=0.1.98
+transformers>=4.35.2,<5.0.0
+gguf>=0.1.0
+protobuf>=4.21.0,<5.0.0

scripts/check-requirements.sh

@@ -0,0 +1,174 @@
+#!/bin/bash
+set -euo pipefail
+
+#
+# check-requirements.sh checks all requirements files for each top-level
+# convert*.py script.
+#
+# WARNING: This is quite IO intensive, because a fresh venv is set up for every
+# python script. As of 2023-12-22, this writes ~2.7GB of data. An adequately
+# sized tmpfs /tmp or ramdisk is recommended if running this frequently.
+#
+# usage:    check-requirements.sh [<working_dir>]
+#           check-requirements.sh nocleanup [<working_dir>]
+#
+# where:
+#           - <working_dir> is a directory that can be used as the base for
+#               setting up the venvs. Defaults to `/tmp`.
+#           - 'nocleanup' as the first argument will disable automatic cleanup
+#               of the files created by this script.
+#
+# requires:
+#           - bash >= 3.2.57
+#           - shellcheck
+#
+# For each script, it creates a fresh venv, `pip install`s the requirements, and
+# finally imports the python script to check for `ImportError`.
+#
+
+log() {
+    local level=$1 msg=$2
+    printf >&2 '%s: %s\n' "$level" "$msg"
+}
+
+debug() {
+    log DEBUG "$@"
+}
+
+info() {
+    log INFO "$@"
+}
+
+fatal() {
+    log FATAL "$@"
+    exit 1
+}
+
+cleanup() {
+    if [[ -n ${workdir+x} && -d $workdir && -w $workdir ]]; then
+        info "Removing $workdir"
+        local count=0
+        rm -rfv -- "$workdir" | while read -r; do
+            if (( count++ > 750 )); then
+                printf .
+                count=0
+            fi
+        done
+        printf '\n'
+        info "Removed $workdir"
+    fi
+}
+
+do_cleanup=1
+if [[ ${1-} == nocleanup ]]; then
+    do_cleanup=0; shift
+fi
+
+if (( do_cleanup )); then
+    trap exit INT TERM
+    trap cleanup EXIT
+fi
+
+this=$(realpath -- "$0"); readonly this
+cd "$(dirname "$this")/.." # PWD should stay in llama.cpp project directory
+
+shellcheck "$this"
+
+readonly reqs_dir=requirements
+
+if [[ ${1+x} ]]; then
+    tmp_dir=$(realpath -- "$1")
+    if [[ ! ( -d $tmp_dir && -w $tmp_dir ) ]]; then
+        fatal "$tmp_dir is not a writable directory"
+    fi
+else
+    tmp_dir=/tmp
+fi
+
+workdir=$(mktemp -d "$tmp_dir/check-requirements.XXXX"); readonly workdir
+info "Working directory: $workdir"
+
+check_requirements() {
+    local reqs=$1
+
+    info "$reqs: beginning check"
+    pip --disable-pip-version-check install -qr "$reqs"
+    info "$reqs: OK"
+}
+
+check_convert_script() {
+    local py=$1             # e.g. ./convert-hf-to-gguf.py
+    local pyname=${py##*/}  # e.g. convert-hf-to-gguf.py
+    pyname=${pyname%.py}    # e.g. convert-hf-to-gguf
+
+    info "$py: beginning check"
+
+    local reqs="$reqs_dir/requirements-$pyname.txt"
+    if [[ ! -r $reqs ]]; then
+        fatal "$py missing requirements. Expected: $reqs"
+    fi
+
+    local venv="$workdir/$pyname-venv"
+    python3 -m venv "$venv"
+
+    (
+        # shellcheck source=/dev/null
+        source "$venv/bin/activate"
+
+        check_requirements "$reqs"
+
+        python - "$py" "$pyname" <<'EOF'
+import sys
+from importlib.machinery import SourceFileLoader
+py, pyname = sys.argv[1:]
+SourceFileLoader(pyname, py).load_module()
+EOF
+    )
+
+    if (( do_cleanup )); then
+        rm -rf -- "$venv"
+    fi
+
+    info "$py: imports OK"
+}
+
+readonly ignore_eq_eq='check_requirements: ignore "=="'
+
+for req in "$reqs_dir"/*; do
+    # Check that all sub-requirements are added to top-level requirements.txt
+    if ! grep -qF "$req" requirements.txt; then
+        fatal "$req needs to be added to requirements.txt"
+    fi
+
+    # Make sure exact release versions aren't being pinned in the requirements
+    # Filters out the ignore string
+    if grep -vF "$ignore_eq_eq" "$req" | grep -q '=='; then
+        tab=$'\t'
+        cat >&2 <<EOF
+FATAL: Avoid pinning exact package versions. Use '~=' instead.
+You can suppress this error by appending the following to the line:
+$tab# $ignore_eq_eq
+EOF
+        exit 1
+    fi
+done
+
+all_venv="$workdir/all-venv"
+python3 -m venv "$all_venv"
+
+(
+    # shellcheck source=/dev/null
+    source "$all_venv/bin/activate"
+    check_requirements requirements.txt
+)
+
+if (( do_cleanup )); then
+    rm -rf -- "$all_venv"
+fi
+
+check_convert_script convert.py
+for py in convert-*.py; do
+    check_convert_script "$py"
+done
+
+info 'Done! No issues found.'

tests/CMakeLists.txt

@@ -2,7 +2,7 @@ function(llama_build_executable source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_executable(${TEST_TARGET} ${source})
     install(TARGETS ${TEST_TARGET} RUNTIME)
-    target_link_libraries(${TEST_TARGET} PRIVATE llama common)
+    target_link_libraries(${TEST_TARGET} PRIVATE common)
 endfunction()
 
 function(llama_test_executable name source)
@@ -14,7 +14,7 @@ function(llama_build_and_test_executable source)
     get_filename_component(TEST_TARGET ${source} NAME_WE)
     add_executable(${TEST_TARGET} ${source})
     install(TARGETS ${TEST_TARGET} RUNTIME)
-    target_link_libraries(${TEST_TARGET} PRIVATE llama common)
+    target_link_libraries(${TEST_TARGET} PRIVATE common)
     add_test(NAME ${TEST_TARGET} COMMAND $<TARGET_FILE:${TEST_TARGET}> ${ARGN})
 endfunction()