ggml : remove K_QUANTS_PER_ITERATION (#8306)

ggml-ci

.devops/full-cuda.Dockerfile

@@ -1,16 +1,18 @@
 ARG UBUNTU_VERSION=22.04
+
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.6.0
+ARG CUDA_VERSION=11.7.1
+
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
 
-# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=default
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -22,12 +24,13 @@ WORKDIR /app
 
 COPY . .
 
-# Use the default CUDA archs if not specified
-RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release -j$(nproc) && \
-    cp build/bin/* .
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV GGML_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
+
+RUN make -j$(nproc)
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full-rocm.Dockerfile

@@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
 
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878

.devops/full.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1

.devops/llama-cli-cann.Dockerfile

@@ -1,44 +0,0 @@
-ARG ASCEND_VERSION=8.0.rc2.alpha003-910b-openeuler22.03-py3.8
-
-FROM cosdt/cann:$ASCEND_VERSION AS build
-
-WORKDIR /app
-
-COPY . .
-
-RUN yum install -y gcc g++ cmake make
-ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
-ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
-ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
-ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
-ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
-ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
-ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
-ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
-ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
-
-# find libascend_hal.so, because the drive hasn`t been mounted.
-ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/runtime/lib64/stub:$LD_LIBRARY_PATH
-
-RUN echo "Building with static libs" && \
-    source /usr/local/Ascend/ascend-toolkit/set_env.sh --force && \
-    cmake -B build -DGGML_CANN=ON -DBUILD_SHARED_LIBS=OFF  && \
-    cmake --build build --config Release --target llama-cli
-
-# TODO: use image with NNRT
-FROM cosdt/cann:$ASCEND_VERSION AS runtime
-COPY --from=build /app/build/bin/llama-cli /llama-cli
-
-ENV LC_ALL=C.utf8
-
-ENV ASCEND_TOOLKIT_HOME=/usr/local/Ascend/ascend-toolkit/latest
-ENV LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:$LIBRARY_PATH
-ENV LD_LIBRARY_PATH=${ASCEND_TOOLKIT_HOME}/lib64:${ASCEND_TOOLKIT_HOME}/lib64/plugin/opskernel:${ASCEND_TOOLKIT_HOME}/lib64/plugin/nnengine:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe/op_tiling:${LD_LIBRARY_PATH}
-ENV PYTHONPATH=${ASCEND_TOOLKIT_HOME}/python/site-packages:${ASCEND_TOOLKIT_HOME}/opp/built-in/op_impl/ai_core/tbe:${PYTHONPATH}
-ENV PATH=${ASCEND_TOOLKIT_HOME}/bin:${ASCEND_TOOLKIT_HOME}/compiler/ccec_compiler/bin:${PATH}
-ENV ASCEND_AICPU_PATH=${ASCEND_TOOLKIT_HOME}
-ENV ASCEND_OPP_PATH=${ASCEND_TOOLKIT_HOME}/opp
-ENV TOOLCHAIN_HOME=${ASCEND_TOOLKIT_HOME}/toolkit
-ENV ASCEND_HOME_PATH=${ASCEND_TOOLKIT_HOME}
-
-ENTRYPOINT ["/llama-cli" ]

.devops/llama-cli-cuda.Dockerfile

@@ -1,37 +1,35 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.6.0
+ARG CUDA_VERSION=11.7.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
 
-# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=default
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake
+    apt-get install -y build-essential git
 
 WORKDIR /app
 
 COPY . .
 
-# Use the default CUDA archs if not specified
-RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_CUDA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release --target llama-cli -j$(nproc)
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV GGML_CUDA=1
 
-FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+RUN make -j$(nproc) llama-cli
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1
 
-COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
-COPY --from=build /app/build/src/libllama.so /libllama.so
-COPY --from=build /app/build/bin/llama-cli /llama-cli
+COPY --from=build /app/llama-cli /llama-cli
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-intel.Dockerfile

@@ -1,6 +1,6 @@
 ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
@@ -14,12 +14,10 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         echo "GGML_SYCL_F16 is set" && \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
-    echo "Building with static libs" && \
-    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx \
-    ${OPT_SYCL_F16} -DBUILD_SHARED_LIBS=OFF && \
+    cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
     cmake --build build --config Release --target llama-cli
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 COPY --from=build /app/build/bin/llama-cli /llama-cli
 

.devops/llama-cli-rocm.Dockerfile

@@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
 
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878

.devops/llama-cli-vulkan.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=jammy
 
-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build
 
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget libgomp1

.devops/llama-cli.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
     apt-get install -y build-essential git
@@ -11,7 +11,7 @@ COPY . .
 
 RUN make -j$(nproc) llama-cli
 
-FROM ubuntu:$UBUNTU_VERSION AS runtime
+FROM ubuntu:$UBUNTU_VERSION as runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1

.devops/llama-server-cuda.Dockerfile

@@ -1,41 +1,38 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=12.6.0
+ARG CUDA_VERSION=11.7.1
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
 ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
 
-FROM ${BASE_CUDA_DEV_CONTAINER} AS build
+FROM ${BASE_CUDA_DEV_CONTAINER} as build
 
-# CUDA architecture to build for (defaults to all supported archs)
-ARG CUDA_DOCKER_ARCH=default
+# Unless otherwise specified, we make a fat build.
+ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-# Use the default CUDA archs if not specified
-RUN if [ "${CUDA_DOCKER_ARCH}" != "default" ]; then \
-        export CMAKE_ARGS="-DCMAKE_CUDA_ARCHITECTURES=${CUDA_DOCKER_ARCH}"; \
-    fi && \
-    cmake -B build -DGGML_CUDA=ON -DLLAMA_CURL=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
-    cmake --build build --config Release --target llama-server -j$(nproc)
+# Set nvcc architecture
+ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
+# Enable CUDA
+ENV GGML_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
 
-FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
+RUN make -j$(nproc) llama-server
+
+FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
-COPY --from=build /app/build/src/libllama.so /libllama.so
-COPY --from=build /app/build/bin/llama-server /llama-server
-
-# Must be set to 0.0.0.0 so it can listen to requests from host machine
-ENV LLAMA_ARG_HOST=0.0.0.0
+COPY --from=build /app/llama-server /llama-server
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-intel.Dockerfile

@@ -1,6 +1,6 @@
 ARG ONEAPI_VERSION=2024.1.1-devel-ubuntu22.04
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS build
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
 ARG GGML_SYCL_F16=OFF
 RUN apt-get update && \
@@ -14,11 +14,10 @@ RUN if [ "${GGML_SYCL_F16}" = "ON" ]; then \
         echo "GGML_SYCL_F16 is set" && \
         export OPT_SYCL_F16="-DGGML_SYCL_F16=ON"; \
     fi && \
-    echo "Building with dynamic libs" && \
     cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
     cmake --build build --config Release --target llama-server
 
-FROM intel/oneapi-basekit:$ONEAPI_VERSION AS runtime
+FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev curl
@@ -26,8 +25,6 @@ RUN apt-get update && \
 COPY --from=build /app/build/bin/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
-# Must be set to 0.0.0.0 so it can listen to requests from host machine
-ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-rocm.Dockerfile

@@ -6,7 +6,7 @@ ARG ROCM_VERSION=5.6
 # Target the CUDA build image
 ARG BASE_ROCM_DEV_CONTAINER=rocm/dev-ubuntu-${UBUNTU_VERSION}:${ROCM_VERSION}-complete
 
-FROM ${BASE_ROCM_DEV_CONTAINER} AS build
+FROM ${BASE_ROCM_DEV_CONTAINER} as build
 
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
@@ -39,8 +39,6 @@ ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
-# Must be set to 0.0.0.0 so it can listen to requests from host machine
-ENV LLAMA_ARG_HOST=0.0.0.0
 
 # Enable cURL
 ENV LLAMA_CURL=1

.devops/llama-server-vulkan.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=jammy
 
-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build
 
 # Install build tools
 RUN apt update && apt install -y git build-essential cmake wget
@@ -23,8 +23,6 @@ RUN cp /app/build/bin/llama-server /llama-server && \
     rm -rf /app
 
 ENV LC_ALL=C.utf8
-# Must be set to 0.0.0.0 so it can listen to requests from host machine
-ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server.Dockerfile

@@ -1,9 +1,9 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION AS build
+FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git libcurl4-openssl-dev curl
 
 WORKDIR /app
 
@@ -13,16 +13,14 @@ ENV LLAMA_CURL=1
 
 RUN make -j$(nproc) llama-server
 
-FROM ubuntu:$UBUNTU_VERSION AS runtime
+FROM ubuntu:$UBUNTU_VERSION as runtime
 
 RUN apt-get update && \
-    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+    apt-get install -y libcurl4-openssl-dev libgomp1
 
 COPY --from=build /app/llama-server /llama-server
 
 ENV LC_ALL=C.utf8
-# Must be set to 0.0.0.0 so it can listen to requests from host machine
-ENV LLAMA_ARG_HOST=0.0.0.0
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/nix/apps.nix

@@ -10,6 +10,7 @@
             "llama-embedding"
             "llama-server"
             "llama-quantize"
+            "llama-train-text-from-scratch"
           ];
           mkApp = name: {
             type = "app";

.devops/nix/devshells.nix

@@ -1,52 +1,13 @@
-{ inputs, ... }:
-
 {
   perSystem =
-    {
-      config,
-      lib,
-      system,
-      ...
-    }:
+    { config, lib, ... }:
     {
       devShells =
-        let
-          pkgs = import inputs.nixpkgs { inherit system; };
-          stdenv = pkgs.stdenv;
-          scripts = config.packages.python-scripts;
-        in
-        lib.pipe (config.packages) [
-          (lib.concatMapAttrs (
-            name: package: {
-              ${name} = pkgs.mkShell {
-                name = "${name}";
-                inputsFrom = [ package ];
-                shellHook = ''
-                  echo "Entering ${name} devShell"
-                '';
-              };
-              "${name}-extra" =
-                if (name == "python-scripts") then
-                  null
-                else
-                  pkgs.mkShell {
-                    name = "${name}-extra";
-                    inputsFrom = [
-                      package
-                      scripts
-                    ];
-                    # Extra packages that *may* be used by some scripts
-                    packages = [
-                        pkgs.python3Packages.tiktoken
-                    ];
-                    shellHook = ''
-                      echo "Entering ${name} devShell"
-                      addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib stdenv.cc.cc}/lib"
-                    '';
-                  };
-            }
-          ))
-          (lib.filterAttrs (name: value: value != null))
-        ];
+        lib.concatMapAttrs
+          (name: package: {
+            ${name} = package.passthru.shell;
+            ${name + "-extra"} = package.passthru.shell-extra;
+          })
+          config.packages;
     };
 }

.devops/nix/nixpkgs-instances.nix

@@ -26,14 +26,16 @@
           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 ]);
+            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 {

.devops/nix/package-gguf-py.nix

@@ -1,36 +0,0 @@
-{
-  lib,
-  llamaVersion,
-  numpy,
-  tqdm,
-  sentencepiece,
-  pyyaml,
-  poetry-core,
-  buildPythonPackage,
-  pytestCheckHook,
-}:
-
-buildPythonPackage {
-  pname = "gguf";
-  version = llamaVersion;
-  pyproject = true;
-  nativeBuildInputs = [ poetry-core ];
-  propagatedBuildInputs = [
-    numpy
-    tqdm
-    sentencepiece
-    pyyaml
-  ];
-  src = lib.cleanSource ../../gguf-py;
-  pythonImportsCheck = [
-    "numpy"
-    "gguf"
-  ];
-  nativeCheckInputs = [ pytestCheckHook ];
-  doCheck = true;
-  meta = with lib; {
-    description = "Python package for writing binary files in the GGUF format";
-    license = licenses.mit;
-    maintainers = [ maintainers.ditsuke ];
-  };
-}

.devops/nix/package.nix

@@ -3,33 +3,30 @@
   glibc,
   config,
   stdenv,
+  mkShell,
   runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
+  python3,
   mpi,
   blas,
   cudaPackages,
-  autoAddDriverRunpath,
   darwin,
   rocmPackages,
   vulkan-headers,
   vulkan-loader,
   curl,
-  shaderc,
-  useBlas ?
-    builtins.all (x: !x) [
-      useCuda
-      useMetalKit
-      useRocm
-      useVulkan
-    ]
-    && blas.meta.available,
+  useBlas ? builtins.all (x: !x) [
+    useCuda
+    useMetalKit
+    useRocm
+    useVulkan
+  ] && blas.meta.available,
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin,
-  # Increases the runtime closure size by ~700M
-  useMpi ? false,
+  useMpi ? false, # Increases the runtime closure size by ~700M
   useRocm ? config.rocmSupport,
   enableCurl ? true,
   useVulkan ? false,
@@ -39,8 +36,8 @@
   # otherwise we get libstdc++ errors downstream.
   effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
   enableStatic ? effectiveStdenv.hostPlatform.isStatic,
-  precompileMetalShaders ? false,
-}:
+  precompileMetalShaders ? false
+}@inputs:
 
 let
   inherit (lib)
@@ -48,6 +45,7 @@ let
     cmakeFeature
     optionals
     strings
+    versionOlder
     ;
 
   stdenv = throw "Use effectiveStdenv instead";
@@ -63,11 +61,38 @@ let
   pnameSuffix =
     strings.optionalString (suffices != [ ])
       "-${strings.concatMapStringsSep "-" strings.toLower suffices}";
-  descriptionSuffix = strings.optionalString (
-    suffices != [ ]
-  ) ", accelerated with ${strings.concatStringsSep ", " suffices}";
+  descriptionSuffix =
+    strings.optionalString (suffices != [ ])
+      ", accelerated with ${strings.concatStringsSep ", " suffices}";
 
-  xcrunHost = runCommand "xcrunHost" { } ''
+  executableSuffix = effectiveStdenv.hostPlatform.extensions.executable;
+
+  # 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/
+  #
+  # TODO: Package up each Python script or service appropriately, by making
+  # them into "entrypoints"
+  llama-python = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+    ]
+  );
+
+  # TODO(Green-Sky): find a better way to opt-into the heavy ml python runtime
+  llama-python-extra = python3.withPackages (
+    ps: [
+      ps.numpy
+      ps.sentencepiece
+      ps.tiktoken
+      ps.torchWithoutCuda
+      ps.transformers
+    ]
+  );
+
+  xcrunHost = runCommand "xcrunHost" {} ''
     mkdir -p $out/bin
     ln -s /usr/bin/xcrun $out/bin
   '';
@@ -84,9 +109,16 @@ let
     ++ optionals useMetalKit [ MetalKit ];
 
   cudaBuildInputs = with cudaPackages; [
-    cuda_cudart
-    cuda_cccl # <nv/target>
-    libcublas
+    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; [
@@ -98,149 +130,184 @@ let
   vulkanBuildInputs = [
     vulkan-headers
     vulkan-loader
-    shaderc
   ];
 in
 
-effectiveStdenv.mkDerivation (finalAttrs: {
-  pname = "llama-cpp${pnameSuffix}";
-  version = llamaVersion;
+effectiveStdenv.mkDerivation (
+  finalAttrs: {
+    pname = "llama-cpp${pnameSuffix}";
+    version = llamaVersion;
 
-  # Note: none of the files discarded here are visible in the sandbox or
-  # affect the output hash. This also means they can be modified without
-  # triggering a rebuild.
-  src = lib.cleanSourceWith {
-    filter =
-      name: type:
-      let
-        noneOf = builtins.all (x: !x);
-        baseName = baseNameOf name;
-      in
-      noneOf [
-        (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
-        (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
-        (lib.hasPrefix "." baseName) # Skip hidden files and directories
-        (baseName == "flake.lock")
+    # Note: none of the files discarded here are visible in the sandbox or
+    # affect the output hash. This also means they can be modified without
+    # triggering a rebuild.
+    src = lib.cleanSourceWith {
+      filter =
+        name: type:
+        let
+          noneOf = builtins.all (x: !x);
+          baseName = baseNameOf name;
+        in
+        noneOf [
+          (lib.hasSuffix ".nix" name) # Ignore *.nix files when computing outPaths
+          (lib.hasSuffix ".md" name) # Ignore *.md changes whe computing outPaths
+          (lib.hasPrefix "." baseName) # Skip hidden files and directories
+          (baseName == "flake.lock")
+        ];
+      src = lib.cleanSource ../../.;
+    };
+
+    postPatch = ''
+      substituteInPlace ./ggml/src/ggml-metal.m \
+        --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
+      substituteInPlace ./ggml/src/ggml-metal.m \
+        --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
+    '';
+
+    # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
+    # `default.metallib` may be compiled with Metal compiler from XCode
+    # and we need to escape sandbox on MacOS to access Metal compiler.
+    # `xcrun` is used find the path of the Metal compiler, which is varible
+    # and not on $PATH
+    # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
+    __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+
+    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
+      ]
+      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
+        glibc.static
+      ] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
+        xcrunHost
       ];
-    src = lib.cleanSource ../../.;
-  };
 
-  postPatch = ''
-    substituteInPlace ./ggml/src/ggml-metal.m \
-      --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-    substituteInPlace ./ggml/src/ggml-metal.m \
-      --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$out/bin/default.metallib\";"
-  '';
+    buildInputs =
+      optionals effectiveStdenv.isDarwin darwinBuildInputs
+      ++ optionals useCuda cudaBuildInputs
+      ++ optionals useMpi [ mpi ]
+      ++ optionals useRocm rocmBuildInputs
+      ++ optionals useBlas [ blas ]
+      ++ optionals useVulkan vulkanBuildInputs
+      ++ optionals enableCurl [ curl ];
 
-  # With PR#6015 https://github.com/ggerganov/llama.cpp/pull/6015,
-  # `default.metallib` may be compiled with Metal compiler from XCode
-  # and we need to escape sandbox on MacOS to access Metal compiler.
-  # `xcrun` is used find the path of the Metal compiler, which is varible
-  # and not on $PATH
-  # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
-  __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
-
-  nativeBuildInputs =
-    [
-      cmake
-      ninja
-      pkg-config
-      git
-    ]
-    ++ optionals useCuda [
-      cudaPackages.cuda_nvcc
-
-      autoAddDriverRunpath
-    ]
-    ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
-    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
-
-  buildInputs =
-    optionals effectiveStdenv.isDarwin darwinBuildInputs
-    ++ optionals useCuda cudaBuildInputs
-    ++ optionals useMpi [ mpi ]
-    ++ optionals useRocm rocmBuildInputs
-    ++ optionals useBlas [ blas ]
-    ++ optionals useVulkan vulkanBuildInputs
-    ++ optionals enableCurl [ curl ];
-
-  cmakeFlags =
-    [
-      (cmakeBool "LLAMA_BUILD_SERVER" true)
-      (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
-      (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
-      (cmakeBool "LLAMA_CURL" enableCurl)
-      (cmakeBool "GGML_NATIVE" false)
-      (cmakeBool "GGML_BLAS" useBlas)
-      (cmakeBool "GGML_CUDA" useCuda)
-      (cmakeBool "GGML_HIPBLAS" useRocm)
-      (cmakeBool "GGML_METAL" useMetalKit)
-      (cmakeBool "GGML_VULKAN" useVulkan)
-      (cmakeBool "GGML_STATIC" enableStatic)
-    ]
-    ++ optionals useCuda [
-      (
-        with cudaPackages.flags;
-        cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
-          builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
+    cmakeFlags =
+      [
+        (cmakeBool "LLAMA_BUILD_SERVER" true)
+        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
+        (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+        (cmakeBool "LLAMA_CURL" enableCurl)
+        (cmakeBool "GGML_NATIVE" false)
+        (cmakeBool "GGML_BLAS" useBlas)
+        (cmakeBool "GGML_CUDA" useCuda)
+        (cmakeBool "GGML_HIPBLAS" useRocm)
+        (cmakeBool "GGML_METAL" useMetalKit)
+        (cmakeBool "GGML_VULKAN" useVulkan)
+        (cmakeBool "GGML_STATIC" enableStatic)
+      ]
+      ++ optionals useCuda [
+        (
+          with cudaPackages.flags;
+          cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
+            builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
+          )
         )
-      )
-    ]
-    ++ optionals useRocm [
-      (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
-      (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
-    ]
-    ++ optionals useMetalKit [
-      (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
-      (cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
-    ];
+      ]
+      ++ optionals useRocm [
+        (cmakeFeature "CMAKE_HIP_COMPILER" "${rocmPackages.llvm.clang}/bin/clang")
+        (cmakeFeature "CMAKE_HIP_ARCHITECTURES" (builtins.concatStringsSep ";" rocmPackages.clr.gpuTargets))
+      ]
+      ++ optionals useMetalKit [
+        (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
+        (cmakeBool "GGML_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
+      ];
 
-  # Environment variables needed for ROCm
-  env = optionals useRocm {
-    ROCM_PATH = "${rocmPackages.clr}";
-    HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
-  };
+    # Environment variables needed for ROCm
+    env = optionals useRocm {
+      ROCM_PATH = "${rocmPackages.clr}";
+      HIP_DEVICE_LIB_PATH = "${rocmPackages.rocm-device-libs}/amdgcn/bitcode";
+    };
 
-  # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
-  # if they haven't been added yet.
-  postInstall = ''
-    mkdir -p $out/include
-    cp $src/include/llama.h $out/include/
-  '';
+    # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
+    # if they haven't been added yet.
+    postInstall = ''
+      mkdir -p $out/include
+      cp $src/include/llama.h $out/include/
+    '';
 
-  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 lib.platforms.darwin;
+    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
+    passthru = {
+      inherit
+        useBlas
+        useCuda
+        useMetalKit
+        useMpi
+        useRocm
+        useVulkan
+        ;
 
-    # Configurations that are known to result in build failures. Can be
-    # overridden by importing Nixpkgs with `allowBroken = true`.
-    broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      shell = mkShell {
+        name = "shell-${finalAttrs.finalPackage.name}";
+        description = "contains numpy and sentencepiece";
+        buildInputs = [ llama-python ];
+        inputsFrom = [ finalAttrs.finalPackage ];
+        shellHook = ''
+          addToSearchPath "LD_LIBRARY_PATH" "${lib.getLib effectiveStdenv.cc.cc}/lib"
+        '';
+      };
 
-    description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
-    homepage = "https://github.com/ggerganov/llama.cpp/";
-    license = lib.licenses.mit;
+      shell-extra = mkShell {
+        name = "shell-extra-${finalAttrs.finalPackage.name}";
+        description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
+        buildInputs = [ llama-python-extra ];
+        inputsFrom = [ finalAttrs.finalPackage ];
+      };
+    };
 
-    # Accommodates `nix run` and `lib.getExe`
-    mainProgram = "llama-cli";
+    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 lib.platforms.darwin;
 
-    # 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.
+      # Configurations that are known to result in build failures. Can be
+      # overridden by importing Nixpkgs with `allowBroken = true`.
+      broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
-    # 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
-    ];
+      description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
+      homepage = "https://github.com/ggerganov/llama.cpp/";
+      license = lib.licenses.mit;
 
-    # Extend `badPlatforms` instead
-    platforms = lib.platforms.all;
-  };
-})
+      # Accommodates `nix run` and `lib.getExe`
+      mainProgram = "llama-cli";
+
+      # 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/python-scripts.nix

@@ -1,66 +0,0 @@
-{
-  lib,
-  stdenv,
-  buildPythonPackage,
-  poetry-core,
-  mkShell,
-  python3Packages,
-  gguf-py,
-}@inputs:
-
-let
-  llama-python-deps = with python3Packages; [
-    numpy
-    sentencepiece
-    transformers
-    protobuf
-    torchWithoutCuda
-    gguf-py
-    tqdm
-
-    # for scripts/compare-llama-bench.py
-    gitpython
-    tabulate
-
-    # for examples/pydantic-models-to-grammar-examples.py
-    docstring-parser
-    pydantic
-
-  ];
-
-  llama-python-test-deps = with python3Packages; [
-    # Server bench
-    matplotlib
-
-    # server tests
-    openai
-    behave
-    prometheus-client
-  ];
-in
-
-buildPythonPackage ({
-  pname = "llama-scripts";
-  version = "0.0.0";
-  pyproject = true;
-
-  # NOTE: The files filtered out here are not visible in the build sandbox, neither
-  # do they affect the output hash. They can be modified without triggering a rebuild.
-  src = lib.cleanSourceWith {
-    filter =
-      name: type:
-      let
-        any = builtins.any (x: x);
-        baseName = builtins.baseNameOf name;
-      in
-      any [
-        (lib.hasSuffix ".py" name)
-        (baseName == "README.md")
-        (baseName == "pyproject.toml")
-      ];
-    src = lib.cleanSource ../../.;
-  };
-  nativeBuildInputs = [ poetry-core ];
-  nativeCheckInputs = llama-python-test-deps;
-  dependencies = llama-python-deps;
-})

.devops/nix/scope.nix

@@ -1,41 +1,19 @@
 {
   lib,
   newScope,
-  python3,
   llamaVersion ? "0.0.0",
 }:
 
-let
-  pythonPackages = python3.pkgs;
-  buildPythonPackage = pythonPackages.buildPythonPackage;
-  numpy = pythonPackages.numpy;
-  tqdm = pythonPackages.tqdm;
-  sentencepiece = pythonPackages.sentencepiece;
-  pyyaml = pythonPackages.pyyaml;
-  poetry-core = pythonPackages.poetry-core;
-  pytestCheckHook = pythonPackages.pytestCheckHook;
-in
-
 # We're using `makeScope` instead of just writing out an attrset
 # because it allows users to apply overlays later using `overrideScope'`.
 # Cf. https://noogle.dev/f/lib/makeScope
 
-lib.makeScope newScope (self: {
-  inherit llamaVersion;
-  gguf-py = self.callPackage ./package-gguf-py.nix {
-    inherit
-      buildPythonPackage
-      numpy
-      tqdm
-      sentencepiece
-      poetry-core
-      pyyaml
-      pytestCheckHook
-      ;
-  };
-  python-scripts = self.callPackage ./python-scripts.nix { inherit buildPythonPackage poetry-core; };
-  llama-cpp = self.callPackage ./package.nix { };
-  docker = self.callPackage ./docker.nix { };
-  docker-min = self.callPackage ./docker.nix { interactive = false; };
-  sif = self.callPackage ./sif.nix { };
-})
+lib.makeScope newScope (
+  self: {
+    inherit llamaVersion;
+    llama-cpp = self.callPackage ./package.nix { };
+    docker = self.callPackage ./docker.nix { };
+    docker-min = self.callPackage ./docker.nix { interactive = false; };
+    sif = self.callPackage ./sif.nix { };
+  }
+)

.devops/tools.sh

@@ -8,11 +8,13 @@ arg1="$1"
 shift
 
 if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
-    python3 ./convert_hf_to_gguf.py "$@"
+    python3 ./convert-hf-to-gguf.py "$@"
 elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
     ./llama-quantize "$@"
 elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
     ./llama-cli "$@"
+elif [[ "$arg1" == '--finetune' || "$arg1" == '-f' ]]; then
+    ./llama-finetune "$@"
 elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
@@ -34,6 +36,8 @@ else
     echo "              ex: --outtype f16 \"/models/7B/\" "
     echo "  --quantize (-q): Optimize with quantization process ggml"
     echo "              ex: \"/models/7B/ggml-model-f16.bin\" \"/models/7B/ggml-model-q4_0.bin\" 2"
+    echo "  --finetune (-f): Run finetune command to create a lora finetune of the model"
+    echo "              See documentation for finetune for command-line parameters"
     echo "  --all-in-one (-a): Execute --convert & --quantize"
     echo "              ex: \"/models/\" 7B"
     echo "  --server (-s): Run a model on the server"

.ecrc

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

.github/labeler.yml

@@ -16,9 +16,7 @@ SYCL:
         - any-glob-to-any-file:
             - ggml/include/ggml-sycl.h
             - ggml/src/ggml-sycl.cpp
-            - ggml/src/ggml-sycl/**
-            - docs/backend/SYCL.md
-            - examples/sycl/**
+            - README-sycl.md
 Nvidia GPU:
     - changed-files:
         - any-glob-to-any-file:

.github/workflows/bench.yml

@@ -1,6 +1,3 @@
-# TODO: there have been some issues with the workflow, so disabling for now
-#       https://github.com/ggerganov/llama.cpp/issues/7893
-#
 # Benchmark
 name: Benchmark
 
@@ -132,8 +129,6 @@ jobs:
 
       - name: Server bench
         id: server_bench
-        env:
-            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -142,7 +137,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch $HEAD_REF \
+              --branch ${{ github.head_ref || github.ref_name }} \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -47,7 +47,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -105,7 +105,7 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -222,7 +222,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -355,10 +355,8 @@ jobs:
       - name: Dependencies
         id: depends
         run: |
-          wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | sudo apt-key add -
-          sudo wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-          sudo apt-get update -y
-          sudo apt-get install -y build-essential vulkan-sdk
+          sudo apt-get update
+          sudo apt-get install build-essential libvulkan-dev
 
       - name: Build
         id: cmake_build
@@ -375,7 +373,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -401,7 +399,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v2
 
       - name: add oneAPI to apt
         shell: bash
@@ -442,7 +440,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v2
 
       - name: add oneAPI to apt
         shell: bash
@@ -546,7 +544,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -576,7 +574,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -610,7 +608,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -696,20 +694,22 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
-          - build: 'avx2-x64'
+          - build: 'rpc-x64'
             defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
+          - build: 'noavx-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+          - build: 'avx2-x64'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
@@ -857,8 +857,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON -DGGML_RPC=ON
-          cmake --build . --config Release -j $((${env:NUMBER_OF_PROCESSORS} - 1)) -t ggml
+          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
@@ -969,14 +968,14 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Install
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP SDK installation"

.github/workflows/docker.yml

@@ -37,9 +37,9 @@ jobs:
           - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
-          #- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
-          #- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          - { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          # Note: the full-rocm image is failing due to a "no space left on device" error. It is disabled for now to allow the workflow to complete.
           #- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
@@ -96,12 +96,21 @@ jobs:
         env:
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
-      - name: Build and push Docker image (tagged + versioned)
+      - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
-        uses: docker/build-push-action@v6
+        uses: docker/build-push-action@v4
         with:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          file: ${{ matrix.config.dockerfile }}
+
+      - name: Build and push Docker image (tagged)
+        uses: docker/build-push-action@v4
+        with:
+          context: .
+          push: ${{ github.event_name == 'push' }}
+          platforms: ${{ matrix.config.platforms }}
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

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

@@ -6,13 +6,15 @@ on:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - '**/requirements*.txt'
+      - 'requirements.txt'
+      - 'requirements/*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
-      - '**/requirements*.txt'
+      - 'requirements.txt'
+      - 'requirements/*.txt'
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}

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

@@ -1,38 +0,0 @@
-name: Python Type-Check
-
-on:
-  push:
-    paths:
-      - '.github/workflows/python-type-check.yml'
-      - '**.py'
-      - '**/requirements*.txt'
-  pull_request:
-    paths:
-      - '.github/workflows/python-type-check.yml'
-      - '**.py'
-      - '**/requirements*.txt'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
-  cancel-in-progress: true
-
-jobs:
-  python-type-check:
-    runs-on: ubuntu-latest
-    name: pyright type-check
-    steps:
-      - name: Check out source repository
-        uses: actions/checkout@v4
-      - name: Set up Python environment
-        uses: actions/setup-python@v5
-        with:
-          python-version: "3.11"
-      - name: Install Python dependencies
-        # TODO: use a venv
-        run: pip install -r requirements/requirements-all.txt
-      - name: Type-check with Pyright
-        uses: jakebailey/pyright-action@v2
-        with:
-          version: 1.1.370
-          level: warning
-          warnings: true

.github/workflows/server.yml

@@ -173,7 +173,6 @@ jobs:
         if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
         run: |
           cd examples/server/tests
-          $env:PYTHONIOENCODING = ":replace"
           behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
 
       - name: Slow tests

.gitignore

@@ -47,10 +47,8 @@ build*
 !build-info.cpp.in
 !build-info.sh
 !build.zig
-!docs/build.md
 /libllama.so
 /llama-*
-/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -61,12 +59,6 @@ llama-batched-swift
 /rpc-server
 out/
 tmp/
-autogen-*.md
-
-# Deprecated
-
-/main
-/server
 
 # CI
 
@@ -80,6 +72,7 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
+
 zig-out/
 zig-cache/
 
@@ -130,6 +123,3 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
-
-# Test models for lora adapters
-/lora-tests

CMakeLists.txt

@@ -50,6 +50,9 @@ endif()
 # option list
 #
 
+# general
+option(LLAMA_CCACHE "llama: use ccache if available" ON)
+
 # debug
 option(LLAMA_ALL_WARNINGS           "llama: enable all compiler warnings"                   ON)
 option(LLAMA_ALL_WARNINGS_3RD_PARTY "llama: enable all compiler warnings in 3rd party libs" OFF)
@@ -74,6 +77,7 @@ option(LLAMA_CURL "llama: use libcurl to download model from an URL" OFF)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 
 # override ggml options
+set(GGML_CCACHE             ${LLAMA_CCACHE})
 set(GGML_SANITIZE_THREAD    ${LLAMA_SANITIZE_THREAD})
 set(GGML_SANITIZE_ADDRESS   ${LLAMA_SANITIZE_ADDRESS})
 set(GGML_SANITIZE_UNDEFINED ${LLAMA_SANITIZE_UNDEFINED})
@@ -106,16 +110,12 @@ llama_option_depr(WARNING     LLAMA_NATIVE              GGML_NATIVE)
 llama_option_depr(WARNING     LLAMA_RPC                 GGML_RPC)
 llama_option_depr(WARNING     LLAMA_SYCL                GGML_SYCL)
 llama_option_depr(WARNING     LLAMA_SYCL_F16            GGML_SYCL_F16)
-llama_option_depr(WARNING     LLAMA_CANN                GGML_CANN)
 
 #
 # build the library
 #
 
-if (NOT TARGET ggml)
-    add_subdirectory(ggml)
-    # ... otherwise assume ggml is added by a parent CMakeLists.txt
-endif()
+add_subdirectory(ggml)
 add_subdirectory(src)
 
 #
@@ -133,23 +133,7 @@ set(LLAMA_INCLUDE_INSTALL_DIR ${CMAKE_INSTALL_INCLUDEDIR} CACHE PATH "Location o
 set(LLAMA_LIB_INSTALL_DIR     ${CMAKE_INSTALL_LIBDIR}     CACHE PATH "Location of library files")
 set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location of binary  files")
 
-
-# At the moment some compile definitions are placed within the ggml/src
-# directory but not exported on the `ggml` target. This could be improved by
-# determining _precisely_ which defines are necessary for the llama-config
-# package.
-#
-set(GGML_TRANSIENT_DEFINES)
-get_target_property(GGML_DIRECTORY ggml SOURCE_DIR)
-get_directory_property(GGML_DIR_DEFINES DIRECTORY ${GGML_DIRECTORY} COMPILE_DEFINITIONS)
-if (GGML_DIR_DEFINES)
-    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_DIR_DEFINES})
-endif()
-get_target_property(GGML_TARGET_DEFINES ggml COMPILE_DEFINITIONS)
-if (GGML_TARGET_DEFINES)
-    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
-endif()
-get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
+get_directory_property(LLAMA_TRANSIENT_DEFINES COMPILE_DEFINITIONS)
 
 set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
 install(TARGETS llama LIBRARY PUBLIC_HEADER)

CMakePresets.json

@@ -28,12 +28,11 @@
     { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
     { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
     { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
-    { "name": "sycl_f16",  "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
-        "architecture": { "value": "arm64",    "strategy": "external" },
-        "toolset":      { "value": "host=x64", "strategy": "external" },
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
         "cacheVariables": {
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-msvc.cmake"
         }
@@ -41,8 +40,8 @@
 
     {
         "name": "arm64-windows-llvm", "hidden": true,
-        "architecture": { "value": "arm64",    "strategy": "external" },
-        "toolset":      { "value": "host=x64", "strategy": "external" },
+        "architecture": { "value": "arm64",       "strategy": "external" },
+        "toolset":      { "value": "host=x86_64", "strategy": "external" },
         "cacheVariables": {
             "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-windows-llvm.cmake"
         }
@@ -61,8 +60,6 @@
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
     { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
-    { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
-    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
-    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
+    { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] }
   ]
 }

CONTRIBUTING.md

@@ -1,29 +1,14 @@
-# Pull requests (for contributors)
+# Contributing Guidelines
 
-- Test your changes:
-  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
-  - Execute [the full CI locally on your machine](ci/README.md) before publishing
-- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
-  - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
-- Consider allowing write access to your branch for faster review
-- If your PR becomes stale, don't hesitate to ping the maintainers in the comments
+## Checklist
 
-# Pull requests (for collaborators)
+* Make sure your PR follows the [coding guidelines](https://github.com/ggerganov/llama.cpp/blob/master/README.md#coding-guidelines)
+* Test your changes using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
+* Execute [the full CI locally on your machine](ci/README.md) before publishing
 
-- Squash-merge PRs
-- Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
-- Optionally, pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
-
-# Coding guidelines
-
-- Avoid adding third-party dependencies, extra files, extra headers, etc.
-- Always consider cross-compatibility with other operating systems and architectures
-- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
-- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
-- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
-- Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
-- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
-
-![matmul](media/matmul.png)
+## PR formatting
 
+* Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
+    - The PR template has a series of review complexity checkboxes `[ ]` that you can mark as `[X]` for your conveience. Refer to [About task lists](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) for more information.
+* If the pull request only contains documentation changes (e.g., updating READMEs, adding new wiki pages), please add `[no ci]` to the commit title. This will skip unnecessary CI checks and help reduce build times.
+* When squashing multiple commits on merge, use the following format for your commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : Fix typo in utils.py (#1234)`

Makefile

@@ -11,15 +11,14 @@ BUILD_TARGETS = \
 	llama-embedding \
 	llama-eval-callback \
 	llama-export-lora \
+	llama-finetune \
 	llama-gbnf-validator \
 	llama-gguf \
-	llama-gguf-hash \
 	llama-gguf-split \
 	llama-gritlm \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
-	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -37,14 +36,13 @@ BUILD_TARGETS = \
 	llama-simple \
 	llama-speculative \
 	llama-tokenize \
+	llama-train-text-from-scratch \
 	llama-vdot \
 	llama-cvector-generator \
-	llama-gen-docs \
 	tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
-	tests/test-arg-parser \
 	tests/test-autorelease \
 	tests/test-backend-ops \
 	tests/test-chat-template \
@@ -65,13 +63,9 @@ TEST_TARGETS = \
 	tests/test-tokenizer-1-spm
 
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
-LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
+LEGACY_TARGETS = main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
-	retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
-
-# Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
-#  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
-LEGACY_TARGETS_BUILD = main quantize perplexity embedding server
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm
 
 # Deprecation aliases
 ifdef LLAMA_CUBLAS
@@ -198,11 +192,7 @@ ifdef GGML_RPC
 	BUILD_TARGETS += rpc-server
 endif
 
-ifdef GGML_VULKAN
-	BUILD_TARGETS += vulkan-shaders-gen
-endif
-
-default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD)
+default: $(BUILD_TARGETS)
 
 test: $(TEST_TARGETS)
 	@failures=0; \
@@ -237,7 +227,7 @@ test: $(TEST_TARGETS)
 	fi
 	@echo 'All tests passed.'
 
-all: $(BUILD_TARGETS) $(TEST_TARGETS) $(LEGACY_TARGETS_BUILD)
+all: $(BUILD_TARGETS) $(TEST_TARGETS)
 
 ifdef RISCV_CROSS_COMPILE
 CC	:= riscv64-unknown-linux-gnu-gcc
@@ -254,22 +244,17 @@ MK_CFLAGS    = -std=c11   -fPIC
 MK_CXXFLAGS  = -std=c++11 -fPIC
 MK_NVCCFLAGS = -std=c++11
 
-ifdef LLAMA_NO_CCACHE
-GGML_NO_CCACHE := 1
-DEPRECATE_WARNING := 1
-endif
-
-ifndef GGML_NO_CCACHE
+ifndef LLAMA_NO_CCACHE
 CCACHE := $(shell which ccache)
 ifdef CCACHE
 export CCACHE_SLOPPINESS = time_macros
-$(info I ccache found, compilation results will be cached. Disable with GGML_NO_CCACHE.)
+$(info I ccache found, compilation results will be cached. Disable with LLAMA_NO_CCACHE.)
 CC    := $(CCACHE) $(CC)
 CXX   := $(CCACHE) $(CXX)
 else
 $(info I ccache not found. Consider installing it for faster compilation.)
 endif # CCACHE
-endif # GGML_NO_CCACHE
+endif # LLAMA_NO_CCACHE
 
 # clock_gettime came in POSIX.1b (1993)
 # CLOCK_MONOTONIC came in POSIX.1-2001 / SUSv3 as optional
@@ -328,9 +313,9 @@ ifdef LLAMA_DEBUG
 	endif
 else
 	MK_CPPFLAGS   += -DNDEBUG
-	MK_CFLAGS     += -O3 -g
-	MK_CXXFLAGS   += -O3 -g
-	MK_NVCCFLAGS  += -O3 -g
+	MK_CFLAGS     += -O3
+	MK_CXXFLAGS   += -O3
+	MK_NVCCFLAGS  += -O3
 endif
 
 ifdef LLAMA_SANITIZE_THREAD
@@ -434,7 +419,7 @@ endif
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 
-ifndef RISCV_CROSS_COMPILE
+ifndef RISCV
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
@@ -514,12 +499,7 @@ ifneq ($(filter loongarch64%,$(UNAME_M)),)
 	MK_CXXFLAGS += -mlasx
 endif
 
-ifneq ($(filter riscv64%,$(UNAME_M)),)
-	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
-	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
-endif
-
-else # RISC-V CROSS COMPILATION
+else
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
@@ -536,21 +516,10 @@ ifndef GGML_NO_ACCELERATE
 	endif
 endif # GGML_NO_ACCELERATE
 
-ifdef GGML_MUSA
-	CC := clang
-	CXX := clang++
-	GGML_CUDA := 1
-	MK_CPPFLAGS += -DGGML_USE_MUSA
-endif
-
 ifndef GGML_NO_OPENMP
 	MK_CPPFLAGS += -DGGML_USE_OPENMP
 	MK_CFLAGS   += -fopenmp
 	MK_CXXFLAGS += -fopenmp
-	ifdef GGML_MUSA
-		MK_CPPFLAGS += -I/usr/lib/llvm-10/include/openmp
-		MK_LDFLAGS  += -L/usr/lib/llvm-10/lib
-	endif # GGML_MUSA
 endif # GGML_NO_OPENMP
 
 ifdef GGML_OPENBLAS
@@ -568,20 +537,14 @@ ifdef GGML_OPENBLAS64
 endif # GGML_OPENBLAS64
 
 ifdef GGML_BLIS
-	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_BLIS -I/usr/local/include/blis -I/usr/include/blis
+	MK_CPPFLAGS += -DGGML_USE_BLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
 	OBJ_GGML    += ggml/src/ggml-blas.o
 endif # GGML_BLIS
 
-ifdef GGML_NVPL
-	MK_CPPFLAGS += -DGGML_USE_BLAS -DGGML_BLAS_USE_NVPL -DNVPL_ILP64 -I/usr/local/include/nvpl_blas -I/usr/include/nvpl_blas
-	MK_LDFLAGS  += -L/usr/local/lib -lnvpl_blas_core -lnvpl_blas_ilp64_gomp
-	OBJ_GGML    += ggml/src/ggml-blas.o
-endif # GGML_NVPL
-
 ifndef GGML_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
-	OBJ_GGML    += ggml/src/llamafile/sgemm.o
+	OBJ_GGML    += ggml/src/sgemm.o
 endif
 
 ifdef GGML_RPC
@@ -601,27 +564,15 @@ else
 endif # GGML_CUDA_FA_ALL_QUANTS
 
 ifdef GGML_CUDA
-	ifdef GGML_MUSA
-		ifneq ('', '$(wildcard /opt/musa)')
-			CUDA_PATH ?= /opt/musa
-		else
-			CUDA_PATH ?= /usr/local/musa
-		endif
-
-		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
-		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
-		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
+	ifneq ('', '$(wildcard /opt/cuda)')
+		CUDA_PATH ?= /opt/cuda
 	else
-		ifneq ('', '$(wildcard /opt/cuda)')
-			CUDA_PATH ?= /opt/cuda
-		else
-			CUDA_PATH ?= /usr/local/cuda
-		endif
+		CUDA_PATH ?= /usr/local/cuda
+	endif
 
-		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
-		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-		MK_NVCCFLAGS += -use_fast_math
-	endif # GGML_MUSA
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
+	MK_NVCCFLAGS += -use_fast_math
 
 	OBJ_GGML += ggml/src/ggml-cuda.o
 	OBJ_GGML += $(patsubst %.cu,%.o,$(wildcard ggml/src/ggml-cuda/*.cu))
@@ -631,11 +582,9 @@ ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
 endif # LLAMA_FATAL_WARNINGS
 
-ifndef GGML_MUSA
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
-endif # GGML_MUSA
 
 ifdef LLAMA_DEBUG
 	MK_NVCCFLAGS += -lineinfo
@@ -648,12 +597,8 @@ endif # GGML_CUDA_DEBUG
 ifdef GGML_CUDA_NVCC
 	NVCC = $(CCACHE) $(GGML_CUDA_NVCC)
 else
-	ifdef GGML_MUSA
-		NVCC = $(CCACHE) mcc
-	else
-		NVCC = $(CCACHE) nvcc
-	endif # GGML_MUSA
-endif # GGML_CUDA_NVCC
+	NVCC = $(CCACHE) nvcc
+endif #GGML_CUDA_NVCC
 
 ifdef CUDA_DOCKER_ARCH
 	MK_NVCCFLAGS += -Wno-deprecated-gpu-targets -arch=$(CUDA_DOCKER_ARCH)
@@ -695,12 +640,6 @@ ifdef GGML_CUDA_DMMV_F16
 	MK_NVCCFLAGS += -DGGML_CUDA_F16
 endif # GGML_CUDA_DMMV_F16
 
-ifdef GGML_CUDA_KQUANTS_ITER
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
-else
-	MK_NVCCFLAGS += -DK_QUANTS_PER_ITERATION=2
-endif
-
 ifdef GGML_CUDA_PEER_MAX_BATCH_SIZE
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=$(GGML_CUDA_PEER_MAX_BATCH_SIZE)
 else
@@ -724,15 +663,9 @@ define NVCC_COMPILE
 	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
 else
-	ifdef GGML_MUSA
-define NVCC_COMPILE
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -c $< -o $@
-endef # NVCC_COMPILE
-	else
 define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endef # NVCC_COMPILE
-	endif # GGML_MUSA
 endif # JETSON_EOL_MODULE_DETECT
 
 ggml/src/ggml-cuda/%.o: \
@@ -755,8 +688,8 @@ endif # GGML_CUDA
 
 ifdef GGML_VULKAN
 	MK_CPPFLAGS += -DGGML_USE_VULKAN
-	MK_LDFLAGS  += $(shell pkg-config --libs vulkan)
-	OBJ_GGML    += ggml/src/ggml-vulkan.o ggml/src/ggml-vulkan-shaders.o
+	MK_LDFLAGS  += -lvulkan
+	OBJ_GGML    += ggml/src/ggml-vulkan.o
 
 ifdef GGML_VULKAN_CHECK_RESULTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_CHECK_RESULTS
@@ -770,10 +703,6 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
-ifdef GGML_VULKAN_PERF
-	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
-endif
-
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -782,28 +711,10 @@ ifdef GGML_VULKAN_RUN_TESTS
 	MK_CPPFLAGS  += -DGGML_VULKAN_RUN_TESTS
 endif
 
-GLSLC_CMD  = glslc
-_ggml_vk_genshaders_cmd = $(shell pwd)/vulkan-shaders-gen
-_ggml_vk_header = ggml/src/ggml-vulkan-shaders.hpp
-_ggml_vk_source = ggml/src/ggml-vulkan-shaders.cpp
-_ggml_vk_input_dir = ggml/src/vulkan-shaders
-_ggml_vk_shader_deps = $(echo $(_ggml_vk_input_dir)/*.comp)
-
-ggml/src/ggml-vulkan.o: ggml/src/ggml-vulkan.cpp ggml/include/ggml-vulkan.h $(_ggml_vk_header) $(_ggml_vk_source)
-	$(CXX) $(CXXFLAGS) $(shell pkg-config --cflags vulkan) -c $< -o $@
-
-$(_ggml_vk_header): $(_ggml_vk_source)
-
-$(_ggml_vk_source): $(_ggml_vk_shader_deps) vulkan-shaders-gen
-	$(_ggml_vk_genshaders_cmd) \
-		--glslc      $(GLSLC_CMD) \
-		--input-dir  $(_ggml_vk_input_dir) \
-		--target-hpp $(_ggml_vk_header) \
-		--target-cpp $(_ggml_vk_source)
-
-vulkan-shaders-gen: ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
-	$(CXX) $(CXXFLAGS) -o $@ $(LDFLAGS) ggml/src/vulkan-shaders/vulkan-shaders-gen.cpp
-
+ggml/src/ggml-vulkan.o: \
+	ggml/src/ggml-vulkan.cpp \
+	ggml/include/ggml-vulkan.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # GGML_VULKAN
 
 ifdef GGML_HIPBLAS
@@ -817,7 +728,6 @@ ifdef GGML_HIPBLAS
 
 	GGML_CUDA_DMMV_X       ?= 32
 	GGML_CUDA_MMV_Y        ?= 1
-	GGML_CUDA_KQUANTS_ITER ?= 2
 
 	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 
@@ -834,20 +744,11 @@ endif # GGML_HIP_UMA
 	HIPFLAGS += $(addprefix --offload-arch=,$(AMDGPU_TARGETS))
 	HIPFLAGS += -DGGML_CUDA_DMMV_X=$(GGML_CUDA_DMMV_X)
 	HIPFLAGS += -DGGML_CUDA_MMV_Y=$(GGML_CUDA_MMV_Y)
-	HIPFLAGS += -DK_QUANTS_PER_ITERATION=$(GGML_CUDA_KQUANTS_ITER)
 
 ifdef GGML_CUDA_FORCE_DMMV
 	HIPFLAGS += -DGGML_CUDA_FORCE_DMMV
 endif # GGML_CUDA_FORCE_DMMV
 
-ifdef GGML_CUDA_FORCE_MMQ
-	HIPFLAGS += -DGGML_CUDA_FORCE_MMQ
-endif # GGML_CUDA_FORCE_MMQ
-
-ifdef GGML_CUDA_FORCE_CUBLAS
-	HIPFLAGS += -DGGML_CUDA_FORCE_CUBLAS
-endif # GGML_CUDA_FORCE_CUBLAS
-
 ifdef GGML_CUDA_NO_PEER_COPY
 	HIPFLAGS += -DGGML_CUDA_NO_PEER_COPY
 endif # GGML_CUDA_NO_PEER_COPY
@@ -900,16 +801,15 @@ ggml/src/ggml-metal-embed.o: \
 	ggml/src/ggml-common.h
 	@echo "Embedding Metal library"
 	@sed -e '/#include "ggml-common.h"/r ggml/src/ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml/src/ggml-metal.metal > ggml/src/ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp -d))
-	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)/ggml-metal-embed.s
-	$(CC) $(CFLAGS) -c $(TEMP_ASSEMBLY)/ggml-metal-embed.s -o $@
-	@rm -f ${TEMP_ASSEMBLY}/ggml-metal-embed.s
-	@rmdir ${TEMP_ASSEMBLY}
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib"            >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"                 >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"                       >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml/src/ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"                   >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                         >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
 endif
 endif # GGML_METAL
 
@@ -917,24 +817,20 @@ OBJ_GGML += \
 	ggml/src/ggml.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
-	ggml/src/ggml-quants.o \
-	ggml/src/ggml-aarch64.o
+	ggml/src/ggml-quants.o
 
 OBJ_LLAMA = \
 	src/llama.o \
-	src/llama-vocab.o \
-	src/llama-grammar.o \
-	src/llama-sampling.o \
 	src/unicode.o \
 	src/unicode-data.o
 
 OBJ_COMMON = \
 	common/common.o \
-	common/arg.o \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
 	common/train.o \
+	common/grammar-parser.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -992,7 +888,6 @@ $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef GGML_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
-ifndef GGML_MUSA
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 
 ifndef CUDA_DOCKER_ARCH
@@ -1002,7 +897,6 @@ endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # GGML_MUSA
 endif # GGML_CUDA
 $(info )
 
@@ -1023,7 +917,6 @@ $(info   - LLAMA_NO_LLAMAFILE)
 $(info   - LLAMA_NO_ACCELERATE)
 $(info   - LLAMA_NO_OPENMP)
 $(info   - LLAMA_NO_METAL)
-$(info   - LLAMA_NO_CCACHE)
 $(info )
 endif
 
@@ -1057,22 +950,15 @@ ggml/src/ggml-quants.o: \
 	ggml/src/ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-ggml/src/ggml-aarch64.o: \
-	ggml/src/ggml-aarch64.c \
-	ggml/include/ggml.h \
-	ggml/src/ggml-aarch64.h \
-	ggml/src/ggml-common.h
-	$(CC) $(CFLAGS)    -c $< -o $@
-
 ggml/src/ggml-blas.o: \
 	ggml/src/ggml-blas.cpp \
 	ggml/include/ggml-blas.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ifndef GGML_NO_LLAMAFILE
-ggml/src/llamafile/sgemm.o: \
-	ggml/src/llamafile/sgemm.cpp \
-	ggml/src/llamafile/sgemm.h \
+ggml/src/sgemm.o: \
+	ggml/src/sgemm.cpp \
+	ggml/src/sgemm.h \
 	ggml/include/ggml.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # GGML_NO_LLAMAFILE
@@ -1106,10 +992,6 @@ src/unicode-data.o: \
 
 src/llama.o: \
 	src/llama.cpp \
-	src/llama-impl.h \
-	src/llama-vocab.h \
-	src/llama-grammar.h \
-	src/llama-sampling.h \
 	src/unicode.h \
 	include/llama.h \
 	ggml/include/ggml-cuda.h \
@@ -1119,29 +1001,6 @@ src/llama.o: \
 	ggml/include/ggml-backend.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-src/llama-vocab.o: \
-	src/llama-vocab.cpp \
-	src/llama-vocab.h \
-	src/llama-impl.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama-grammar.o: \
-	src/llama-grammar.cpp \
-	src/llama-grammar.h \
-	src/llama-impl.h \
-	src/llama-vocab.h \
-	src/llama-sampling.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-src/llama-sampling.o: \
-	src/llama-sampling.cpp \
-	src/llama-sampling.h \
-	src/llama-impl.h \
-	include/llama.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 $(LIB_LLAMA): \
 	$(OBJ_LLAMA) \
 	$(LIB_GGML)
@@ -1163,11 +1022,6 @@ common/common.o: \
 	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common/arg.o: \
-	common/arg.cpp \
-	common/arg.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 common/sampling.o: \
 	common/sampling.cpp \
 	common/sampling.h \
@@ -1179,6 +1033,11 @@ common/console.o: \
 	common/console.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+common/grammar-parser.o: \
+	common/grammar-parser.cpp \
+	common/grammar-parser.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 common/json-schema-to-grammar.o: \
 	common/json-schema-to-grammar.cpp \
 	common/json-schema-to-grammar.h
@@ -1218,15 +1077,13 @@ clean:
 	rm -rvf ggml/*.dll
 	rm -rvf ggml/*.so
 	rm -vrf ggml/src/*.o
-	rm -rvf ggml/src/llamafile/*.o
 	rm -rvf common/build-info.cpp
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
 	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
 	rm -rvf $(BUILD_TARGETS)
 	rm -rvf $(TEST_TARGETS)
-	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
-	rm -rvf $(LEGACY_TARGETS_CLEAN)
+	rm -rvf $(LEGACY_TARGETS)
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -1313,23 +1170,6 @@ llama-gguf: examples/gguf/gguf.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-examples/gguf-hash/deps/sha1/sha1.o: \
-	examples/gguf-hash/deps/sha1/sha1.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-examples/gguf-hash/deps/xxhash/xxhash.o: \
-	examples/gguf-hash/deps/xxhash/xxhash.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-examples/gguf-hash/deps/sha256/sha256.o: \
-	examples/gguf-hash/deps/sha256/sha256.c
-	$(CC) $(CFLAGS) -Iexamples/gguf-hash/deps -c $< -o $@
-
-llama-gguf-hash: examples/gguf-hash/gguf-hash.cpp examples/gguf-hash/deps/sha1/sha1.o examples/gguf-hash/deps/xxhash/xxhash.o examples/gguf-hash/deps/sha256/sha256.o\
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -Iexamples/gguf-hash/deps -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-gguf-split: examples/gguf-split/gguf-split.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1345,6 +1185,11 @@ llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
 	$(OBJ_GGML) $(OBJ_LLAMA)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1360,11 +1205,16 @@ llama-baby-llama: examples/baby-llama/baby-llama.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-export-lora: examples/export-lora/export-lora.cpp \
+llama-finetune: examples/finetune/finetune.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-export-lora: examples/export-lora/export-lora.cpp \
+	$(OBJ_GGML) common/log.h
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-retrieval: examples/retrieval/retrieval.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1440,7 +1290,6 @@ llama-server: \
 	examples/server/system-prompts.js.hpp \
 	examples/server/prompt-formats.js.hpp \
 	examples/server/json-schema-to-grammar.mjs.hpp \
-	examples/server/loading.html.hpp \
 	common/json.hpp \
 	common/stb_image.h \
 	$(OBJ_ALL)
@@ -1456,11 +1305,6 @@ examples/server/%.hpp: examples/server/public/% Makefile
 		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
 	) > $@
 
-llama-gen-docs: examples/gen-docs/gen-docs.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 libllava.a: examples/llava/llava.cpp \
 	examples/llava/llava.h \
 	examples/llava/clip.cpp \
@@ -1471,20 +1315,15 @@ libllava.a: examples/llava/llava.cpp \
 	$(CXX) $(CXXFLAGS) -static -fPIC -c $< -o $@ -Wno-cast-qual
 
 llama-llava-cli: examples/llava/llava-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
-	examples/llava/clip.cpp \
 	examples/llava/clip.h \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
-
-llama-minicpmv-cli: examples/llava/minicpmv-cli.cpp \
-	examples/llava/llava.cpp \
-	examples/llava/llava.h \
 	examples/llava/clip.cpp \
-	examples/llava/clip.h \
+	examples/llava/llava.h \
+	examples/llava/llava.cpp \
 	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) $< $(filter-out %.h $<,$^) -o $@ $(LDFLAGS) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp  -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -c examples/llava/llava.cpp -o $(call GET_OBJ_FILE, examples/llava/llava.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $< examples/llava/clip.cpp examples/llava/llava.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) $(call GET_OBJ_FILE, examples/llava/llava.cpp) -o $@ $(LDFLAGS)
 
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
@@ -1518,13 +1357,8 @@ run-benchmark-matmult: llama-benchmark-matmult
 
 .PHONY: run-benchmark-matmult swift
 
-tests/test-arg-parser: tests/test-arg-parser.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 tests/test-llama-grammar: tests/test-llama-grammar.cpp \
-	$(OBJ_ALL)
+	$(OBJ_GGML) $(OBJ_COMMON) src/unicode.o src/unicode-data.o
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -1628,50 +1462,3 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 	$(OBJ_GGML)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-#
-# Deprecated binaries that we want to keep around long enough for people to migrate to the new filenames, then these can be removed.
-#
-# Mark legacy binary targets as .PHONY so that they are always checked.
-.PHONY: main quantize perplexity embedding server
-
-# Define the object file target
-examples/deprecation-warning/deprecation-warning.o: examples/deprecation-warning/deprecation-warning.cpp
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-# NOTE: We currently will always build the deprecation-warning `main` and `server` binaries to help users migrate.
-#  Eventually we will want to remove these target from building all the time.
-main: examples/deprecation-warning/deprecation-warning.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "NOTICE: The 'main' binary is deprecated. Please use 'llama-cli' instead."
-
-server: examples/deprecation-warning/deprecation-warning.o
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
-
-quantize: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'quantize' binary is deprecated. Please use 'llama-quantize' instead."
-	@echo "  Remove the 'quantize' binary to remove this warning."
-	@echo "#########"
-endif
-
-perplexity: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'perplexity' binary is deprecated. Please use 'llama-perplexity' instead."
-	@echo "  Remove the 'perplexity' binary to remove this warning."
-	@echo "#########"
-endif
-
-embedding: examples/deprecation-warning/deprecation-warning.o
-ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
-	@echo "#########"
-	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
-	@echo "  Remove the 'embedding' binary to remove this warning."
-	@echo "#########"
-endif

Package.swift

@@ -4,16 +4,12 @@ import PackageDescription
 
 var sources = [
     "src/llama.cpp",
-    "src/llama-vocab.cpp",
-    "src/llama-grammar.cpp",
-    "src/llama-sampling.cpp",
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
     "ggml/src/ggml-alloc.c",
     "ggml/src/ggml-backend.c",
     "ggml/src/ggml-quants.c",
-    "ggml/src/ggml-aarch64.c",
 ]
 
 var resources: [Resource] = []

README-sycl.md

@@ -20,7 +20,7 @@
 **oneAPI** is an open ecosystem and a standard-based specification, supporting multiple architectures including but not limited to intel CPUs, GPUs and FPGAs. The key components of the oneAPI ecosystem include:
 
 - **DPCPP** *(Data Parallel C++)*: The primary oneAPI SYCL implementation, which includes the icpx/icx Compilers.
-- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL - Math Kernel Library)*.
 - **oneAPI LevelZero**: A high performance low level interface for fine-grained control over intel iGPUs and dGPUs.
 - **Nvidia & AMD Plugins**: These are plugins extending oneAPI's DPCPP support to SYCL on Nvidia and AMD GPU targets.
 
@@ -28,6 +28,10 @@
 
 The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it could support other vendor GPUs: Nvidia GPU (*AMD GPU coming*).
 
+When targeting **Intel CPU**, it is recommended to use llama.cpp for [Intel oneMKL](README.md#intel-onemkl) backend.
+
+It has the similar design of other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, etc..*. In beginning work, the oneAPI's [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) open-source migration tool (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) was used for this purpose.
+
 ## Recommended Release
 
 The SYCL backend would be broken by some PRs due to no online CI.
@@ -41,10 +45,6 @@ The following release is verified with good quality:
 
 ## News
 
-
-- 2024.8
-  - Use oneDNN as the default GEMM library, improve the compatibility for new Intel GPUs.
-
 - 2024.5
   - Performance is increased: 34 -> 37 tokens/s of llama-2-7b.Q4_0 on Arc770.
   - Arch Linux is verified successfully.
@@ -80,14 +80,7 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-SYCL backend supports Intel GPU Family:
-
-- Intel Data Center Max Series
-- Intel Flex Series, Arc Series
-- Intel Built-in Arc GPU
-- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
-
-#### Verified devices
+**Verified devices**
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -95,7 +88,7 @@ SYCL backend supports Intel GPU Family:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -196,7 +189,7 @@ Please follow the instructions for downloading and installing the Toolkit for Li
 
 Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
 
-Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI MKL for intel GPUs.
 
 - **Adding support to Nvidia GPUs**
 
@@ -244,17 +237,12 @@ Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA devic
 ### II. Build llama.cpp
 
 #### Intel GPU
-
-```
-./examples/sycl/build.sh
-```
-
-or
-
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
+# Build LLAMA with MKL BLAS acceleration for intel GPU
+
 # Option 1: Use FP32 (recommended for better performance in most cases)
 cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
@@ -288,71 +276,48 @@ cmake --build build --config Release -j -v
 
 ### III. Run the inference
 
-#### Retrieve and prepare model
+1. Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-##### Check device
-
-1. Enable oneAPI running environment
+2. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-2. List devices information
+3. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
-
-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
 ```
-found 2 SYCL devices:
-
+found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
 |ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
 |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
 | 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
-#### Choose level-zero devices
+| Attribute              | Note                                                        |
+|------------------------|-------------------------------------------------------------|
+| compute capability 1.3 | Level-zero driver/runtime, recommended                      |
+| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases |
 
-|Chosen Device ID|Setting|
-|-|-|
-|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
-|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
-
-#### Execute
-
-Choose one of following methods to run.
-
-1. Script
-
-- Use device 0:
-
-```sh
-./examples/sycl/run-llama2.sh 0
-```
-- Use multiple devices:
-
-```sh
-./examples/sycl/run-llama2.sh
-```
-
-2. Command line
-Launch inference
+4. Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user. Default device id is 0.
-- Multiple devices: Automatically choose the devices with the same backend.
-
-In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
+- Single device: Use one device target specified by the user.
+- Multiple devices: Automatically select the devices with the same largest Max compute-units.
 
 | Device selection | Parameter                              |
 |------------------|----------------------------------------|
@@ -366,6 +331,11 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
+or run by script:
+
+```sh
+./examples/sycl/run_llama2.sh 0
+```
 
 - Use multiple devices:
 
@@ -373,6 +343,12 @@ ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Bui
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
+Otherwise, you can run the script:
+
+```sh
+./examples/sycl/run_llama2.sh
+```
+
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -419,7 +395,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls.exe
+sycl-ls
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -440,18 +416,6 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
-You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
-
-Choose one of following methods to build from source code.
-
-1. Script
-
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
-2. CMake
-
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -466,8 +430,12 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Or, use CMake presets to build:
+Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
 
+Or, use CMake presets to build:
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -479,9 +447,7 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-3. Visual Studio
-
-You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -489,65 +455,52 @@ You can use Visual Studio to open llama.cpp folder as a CMake project. Choose th
 
 ### III. Run the inference
 
-#### Retrieve and prepare model
+1. Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-##### Check device
-
-1. Enable oneAPI running environment
+2. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-2. List devices information
+3. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\llama-ls-sycl-device.exe
+build\bin\ls-sycl-device.exe
 ```
 
-This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
+The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
 ```
-found 2 SYCL devices:
+found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
 |ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
 |--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
 | 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
+| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
+| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
+| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
+| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 
 ```
-#### Choose level-zero devices
 
-|Chosen Device ID|Setting|
-|-|-|
-|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
-|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
-|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+| Attribute              | Note                                                      |
+|------------------------|-----------------------------------------------------------|
+| compute capability 1.3 | Level-zero running time, recommended                      |
+| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases |
 
-#### Execute
 
-Choose one of following methods to run.
-
-1. Script
-
-```
-examples\sycl\win-run-llama2.bat
-```
-
-2. Command line
-
-Launch inference
+4. Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user. Default device id is 0.
-- Multiple devices: Automatically choose the devices with the same backend.
-
-In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
+- Single device: Use one device assigned by user.
+- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
 
 | Device selection | Parameter                              |
 |------------------|----------------------------------------|
@@ -567,7 +520,11 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
+Otherwise, run the following wrapper script:
 
+```
+.\examples\sycl\win-run-llama2.bat
+```
 
 Note:
 
@@ -581,18 +538,17 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
-
 ## Environment Variable
 
 #### Build
 
 | Name               | Value                             | Function                                    |
 |--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
 | GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
 | GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -628,18 +584,9 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
-- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
-
-  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
-
-  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
-
-  It's same for other projects including llama.cpp SYCL backend.
-
-
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- NA
+- Support row layer split for multiple card runs.

README.md

@@ -3,24 +3,73 @@
 ![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
-[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
+[![Server](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml/badge.svg?branch=master&event=schedule)](https://github.com/ggerganov/llama.cpp/actions/workflows/server.yml)
 [![Conan Center](https://shields.io/conan/v/llama-cpp)](https://conan.io/center/llama-cpp)
 
 [Roadmap](https://github.com/users/ggerganov/projects/7) / [Project status](https://github.com/ggerganov/llama.cpp/discussions/3471) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-## Recent API changes
+> [!IMPORTANT]
+[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
 
-- [Changelog for `libllama` API](https://github.com/ggerganov/llama.cpp/issues/9289)
-- [Changelog for `llama-server` REST API](https://github.com/ggerganov/llama.cpp/issues/9291)
+### Recent API changes
 
-## Hot topics
+- [2024 Jun 26] The source code and CMake build scripts have been restructured https://github.com/ggerganov/llama.cpp/pull/8006
+- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
+- [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
+- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
+- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
+- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
+- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
+- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
 
-- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+### Hot topics
+
+- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
+- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
+- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
+- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
+- Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
+- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
+- Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017
+- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
+- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
+- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
 
 ----
 
+<details>
+  <summary>Table of Contents</summary>
+  <ol>
+    <li>
+      <a href="#description">Description</a>
+    </li>
+    <li>
+      <a href="#usage">Usage</a>
+      <ul>
+        <li><a href="#get-the-code">Get the Code</a></li>
+        <li><a href="#build">Build</a></li>
+        <li><a href="#blas-build">BLAS Build</a></li>
+        <li><a href="#prepare-and-quantize">Prepare and Quantize</a></li>
+        <li><a href="#run-the-quantized-model">Run the quantized model</a></li>
+        <li><a href="#memorydisk-requirements">Memory/Disk Requirements</a></li>
+        <li><a href="#quantization">Quantization</a></li>
+        <li><a href="#interactive-mode">Interactive mode</a></li>
+        <li><a href="#constrained-output-with-grammars">Constrained output with grammars</a></li>
+        <li><a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a></li>
+        <li><a href="#seminal-papers-and-background-on-the-models">Seminal papers and background on the models</a></li>
+        <li><a href="#perplexity-measuring-model-quality">Perplexity (measuring model quality)</a></li>
+        <li><a href="#android">Android</a></li>
+        <li><a href="#docker">Docker</a></li>
+      </ul>
+    </li>
+    <li><a href="#contributing">Contributing</a></li>
+    <li><a href="#coding-guidelines">Coding guidelines</a></li>
+    <li><a href="#docs">Docs</a></li>
+  </ol>
+</details>
+
 ## Description
 
 The main goal of `llama.cpp` is to enable LLM inference with minimal setup and state-of-the-art performance on a wide
@@ -38,6 +87,14 @@ Since its [inception](https://github.com/ggerganov/llama.cpp/issues/33#issuecomm
 improved significantly thanks to many contributions. It is the main playground for developing new features for the
 [ggml](https://github.com/ggerganov/ggml) library.
 
+**Supported platforms:**
+
+- [X] Mac OS
+- [X] Linux
+- [X] Windows (via CMake)
+- [X] Docker
+- [X] FreeBSD
+
 **Supported models:**
 
 Typically finetunes of the base models below are supported as well.
@@ -77,21 +134,9 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
-- [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
-- [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
-- [x] [Smaug](https://huggingface.co/models?search=Smaug)
-- [x] [Poro 34B](https://huggingface.co/LumiOpen/Poro-34B)
-- [x] [Bitnet b1.58 models](https://huggingface.co/1bitLLM)
-- [x] [Flan T5](https://huggingface.co/models?search=flan-t5)
-- [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
-- [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
-- [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
-- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
-- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
-- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 
-(instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
+(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
 **Multimodal models:**
 
@@ -105,6 +150,12 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 - [x] [Bunny](https://github.com/BAAI-DCAI/Bunny)
 
+**HTTP server**
+
+[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
+
+[simplechat](./examples/server/public_simplechat) is a simple chat client, which can be used to chat with the model exposed using above web server (use --path to point to simplechat), from a local web browser.
+
 **Bindings:**
 
 - Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
@@ -125,20 +176,17 @@ Typically finetunes of the base models below are supported as well.
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
-- Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
 
 **UI:**
 
 Unless otherwise noted these projects are open-source with permissive licensing:
 
-- [MindWorkAI/AI-Studio](https://github.com/MindWorkAI/AI-Studio) (FSL-1.1-MIT)
 - [iohub/collama](https://github.com/iohub/coLLaMA)
 - [janhq/jan](https://github.com/janhq/jan) (AGPL)
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
 - [Layla](https://play.google.com/store/apps/details?id=com.laylalite) (proprietary)
-- [ramalama](https://github.com/containers/ramalama) (MIT)
 - [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
@@ -164,7 +212,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
 - [AIKit](https://github.com/sozercan/aikit) (MIT)
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
-- [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
@@ -172,20 +219,14 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
-- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
 
 **Infrastructure:**
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
-- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 
-**Games:**
-- [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
+---
 
-## Demo
-
-<details>
-<summary>Typical run using LLaMA v2 13B on M2 Ultra</summary>
+Here is a typical run using LLaMA v2 13B on M2 Ultra:
 
 ```
 $ make -j && ./llama-cli -m models/llama-13b-v2/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e
@@ -265,85 +306,452 @@ llama_print_timings:        eval time = 24513.59 ms /   399 runs   (   61.44 ms
 llama_print_timings:       total time = 25431.49 ms
 ```
 
-</details>
-
-<details>
-<summary>Demo of running both LLaMA-7B and whisper.cpp on a single M1 Pro MacBook</summary>
-
 And here is another demo of running both LLaMA-7B and [whisper.cpp](https://github.com/ggerganov/whisper.cpp) on a single M1 Pro MacBook:
 
 https://user-images.githubusercontent.com/1991296/224442907-7693d4be-acaa-4e01-8b4f-add84093ffff.mp4
 
-</details>
-
 ## Usage
 
 Here are the end-to-end binary build and model conversion steps for most supported models.
 
-### Basic usage
-
-Firstly, you need to get the binary. There are different methods that you can follow:
-- Method 1: Clone this repository and build locally, see [how to build](./docs/build.md)
-- Method 2: If you are using MacOS or Linux, you can install llama.cpp via [brew, flox or nix](./docs/install.md)
-- Method 3: Use a Docker image, see [documentation for Docker](./docs/docker.md)
-- Method 4: Download pre-built binary from [releases](https://github.com/ggerganov/llama.cpp/releases)
-
-You can run a basic completion using this command:
+### Get the Code
 
 ```bash
-llama-cli -m your_model.gguf -p "I believe the meaning of life is" -n 128
-
-# Output:
-# I believe the meaning of life is to find your own truth and to live in accordance with it. For me, this means being true to myself and following my passions, even if they don't align with societal expectations. I think that's what I love about yoga – it's not just a physical practice, but a spiritual one too. It's about connecting with yourself, listening to your inner voice, and honoring your own unique journey.
+git clone https://github.com/ggerganov/llama.cpp
+cd llama.cpp
 ```
 
-See [this page](./examples/main/README.md) for a full list of parameters.
+### Build
 
-### Conversation mode
+In order to build llama.cpp you have four different options.
 
-If you want a more ChatGPT-like experience, you can run in conversation mode by passing `-cnv` as a parameter:
+- Using `make`:
+  - On Linux or MacOS:
+
+      ```bash
+      make
+      ```
+
+  - On Windows:
+
+    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+    2. Extract `w64devkit` on your pc.
+    3. Run `w64devkit.exe`.
+    4. Use the `cd` command to reach the `llama.cpp` folder.
+    5. From here you can run:
+        ```bash
+        make
+        ```
+
+  - Notes:
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, run `make LLAMA_DEBUG=1`
+
+- Using `CMake`:
+
+  ```bash
+  cmake -B build
+  cmake --build build --config Release
+  ```
+
+  **Notes**:
+
+    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
+    - For faster repeated compilation, install [ccache](https://ccache.dev/).
+    - For debug builds, there are two cases:
+
+      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
+
+      ```bash
+      cmake -B build -DCMAKE_BUILD_TYPE=Debug
+      cmake --build build
+      ```
+
+      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
+
+      ```bash
+      cmake -B build -G "Xcode"
+      cmake --build build --config Debug
+      ```
+
+-   Using `gmake` (FreeBSD):
+
+    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
+    2. Add your user to **video** group
+    3. Install compilation dependencies.
+
+        ```bash
+        sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
+
+        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
+        ```
+
+### Homebrew
+
+On Mac and Linux, the homebrew package manager can be used via
+```
+brew install llama.cpp
+```
+The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
+
+### Nix
+
+On Mac and Linux, the Nix package manager can be used via
+```
+nix profile install nixpkgs#llama-cpp
+```
+For flake enabled installs.
+
+Or
+```
+nix-env --file '<nixpkgs>' --install --attr llama-cpp
+```
+For non-flake enabled installs.
+
+This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
+
+#### Flox
+
+On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
+```
+flox install llama-cpp
+```
+Flox follows the nixpkgs build of llama.cpp.
+
+### Metal Build
+
+On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
+To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
+
+When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
+argument.
+
+### BLAS Build
+
+Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
+
+- #### Accelerate Framework:
+
+  This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
+
+- #### OpenBLAS:
+
+  This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
+
+  - Using `make`:
+    - On Linux:
+      ```bash
+      make GGML_OPENBLAS=1
+      ```
+
+    - On Windows:
+
+      1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
+      2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
+      3. Extract `w64devkit` on your pc.
+      4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
+      5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
+      6. Run `w64devkit.exe`.
+      7. Use the `cd` command to reach the `llama.cpp` folder.
+      8. From here you can run:
+
+          ```bash
+          make GGML_OPENBLAS=1
+          ```
+
+  - Using `CMake` on Linux:
+
+      ```bash
+      cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
+      cmake --build build --config Release
+      ```
+
+- #### BLIS
+
+  Check [BLIS.md](docs/BLIS.md) for more information.
+
+- #### SYCL
+  SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
+
+  llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
+
+  For detailed info, please refer to [llama.cpp for SYCL](README-sycl.md).
+
+- #### Intel oneMKL
+  Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./README-sycl.md).
+
+  - Using manual oneAPI installation:
+    By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
+      ```bash
+      source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
+      cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
+      cmake --build build --config Release
+      ```
+
+  - Using oneAPI docker image:
+    If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
+
+  Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
+
+- #### CUDA
+
+  This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+
+  For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
+
+  - Using `make`:
+    ```bash
+    make GGML_CUDA=1
+    ```
+  - Using `CMake`:
+
+    ```bash
+    cmake -B build -DGGML_CUDA=ON
+    cmake --build build --config Release
+    ```
+
+  The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
+
+  | Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
+  |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | GGML_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+  | GGML_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
+  | GGML_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
+  | GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
+  | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
+  | GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
+  | GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
+  | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
+
+- #### hipBLAS
+
+  This provides BLAS acceleration on HIP-supported AMD GPUs.
+  Make sure to have ROCm installed.
+  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
+
+  - Using `make`:
+    ```bash
+    make GGML_HIPBLAS=1
+    ```
+  - Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
+    ```bash
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
+        cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build --config Release -- -j 16
+    ```
+    On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
+    However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
+
+    Note that if you get the following error:
+    ```
+    clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
+    ```
+    Try searching for a directory under `HIP_PATH` that contains the file
+    `oclc_abi_version_400.bc`. Then, add the following to the start of the
+    command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
+    like:
+    ```bash
+    HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
+    HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
+        cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
+        && cmake --build build -- -j 16
+    ```
+
+  - Using `make` (example for target gfx1030, build with 16 CPU threads):
+    ```bash
+    make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
+    ```
+
+  - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
+    ```bash
+    set PATH=%HIP_PATH%\bin;%PATH%
+    cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
+    cmake --build build
+    ```
+    Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
+    Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
+
+
+  The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
+  If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
+  The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
+
+  | Option                 | Legal values           | Default | Description                                                                                                                                                                                                                                    |
+  |------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | GGML_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | GGML_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
+
+- #### Vulkan
+
+  **With docker**:
+
+  You don't need to install Vulkan SDK. It will be installed inside the container.
+
+  ```sh
+  # Build the image
+  docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
+
+  # Then, use it:
+  docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
+  ```
+
+  **Without docker**:
+
+  Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
+
+  For example, on Ubuntu 22.04 (jammy), use the command below:
+
+  ```bash
+  wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
+  wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
+  apt update -y
+  apt-get install -y vulkan-sdk
+  # To verify the installation, use the command below:
+  vulkaninfo
+  ```
+
+  Alternatively your package manager might be able to provide the appropriate libraries.
+  For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
+  For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
+
+  Then, build llama.cpp using the cmake command below:
+
+  ```bash
+  cmake -B build -DGGML_VULKAN=1
+  cmake --build build --config Release
+  # Test the output binary (with "-ngl 33" to offload all layers to GPU)
+  ./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
+
+  # You should see in the output, ggml_vulkan detected your GPU. For example:
+  # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
+  ```
+
+### Prepare and Quantize
+
+> [!NOTE]
+> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
+
+To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
+
+Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
+It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
 
 ```bash
-llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv
+# obtain the official LLaMA model weights and place them in ./models
+ls ./models
+llama-2-7b tokenizer_checklist.chk tokenizer.model
+# [Optional] for models using BPE tokenizers
+ls ./models
+<folder containing weights and tokenizer json> vocab.json
+# [Optional] for PyTorch .bin models like Mistral-7B
+ls ./models
+<folder containing weights and tokenizer json>
 
-# Output:
-# > hi, who are you?
-# Hi there! I'm your helpful assistant! I'm an AI-powered chatbot designed to assist and provide information to users like you. I'm here to help answer your questions, provide guidance, and offer support on a wide range of topics. I'm a friendly and knowledgeable AI, and I'm always happy to help with anything you need. What's on your mind, and how can I assist you today?
-#
-# > what is 1+1?
-# Easy peasy! The answer to 1+1 is... 2!
+# install Python dependencies
+python3 -m pip install -r requirements.txt
+
+# convert the model to ggml FP16 format
+python3 convert_hf_to_gguf.py models/mymodel/
+
+# quantize the model to 4-bits (using Q4_K_M method)
+./llama-quantize ./models/mymodel/ggml-model-f16.gguf ./models/mymodel/ggml-model-Q4_K_M.gguf Q4_K_M
+
+# update the gguf filetype to current version if older version is now unsupported
+./llama-quantize ./models/mymodel/ggml-model-Q4_K_M.gguf ./models/mymodel/ggml-model-Q4_K_M-v2.gguf COPY
 ```
 
-By default, the chat template will be taken from the input model. If you want to use another chat template, pass `--chat-template NAME` as a parameter. See the list of [supported templates](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template)
+### Run the quantized model
 
 ```bash
-./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --chat-template chatml
+# start inference on a gguf model
+./llama-cli -m ./models/mymodel/ggml-model-Q4_K_M.gguf -n 128
 ```
 
-You can also use your own template via in-prefix, in-suffix and reverse-prompt parameters:
+When running the larger models, make sure you have enough disk space to store all the intermediate files.
 
-```bash
-./llama-cli -m your_model.gguf -p "You are a helpful assistant" -cnv --in-prefix 'User: ' --reverse-prompt 'User:'
+### Running on Windows with prebuilt binaries
+
+You will find prebuilt Windows binaries on the release page.
+
+Simply download and extract the latest zip package of choice: (e.g. `llama-b1380-bin-win-avx2-x64.zip`)
+
+From the unzipped folder, open a terminal/cmd window here and place a pre-converted `.gguf` model file. Test out the main example like so:
+
+```
+.\main -m llama-2-7b.Q4_0.gguf -n 128
 ```
 
-### Web server
+### Memory/Disk Requirements
 
-[llama.cpp web server](./examples/server/README.md) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
+As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
-Example usage:
+| Model | Original size | Quantized size (Q4_0) |
+|------:|--------------:|----------------------:|
+|    7B |         13 GB |                3.9 GB |
+|   13B |         24 GB |                7.8 GB |
+|   30B |         60 GB |               19.5 GB |
+|   65B |        120 GB |               38.5 GB |
 
-```bash
-./llama-server -m your_model.gguf --port 8080
+### Quantization
 
-# Basic web UI can be accessed via browser: http://localhost:8080
-# Chat completion endpoint: http://localhost:8080/v1/chat/completions
+Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
+
+*(outdated)*
+
+| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
+|------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
+|    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
+|    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |
+|    7B | ms/tok @ 4th |    127 |     55 |     54 |     76 |     83 |     72 |
+|    7B | ms/tok @ 8th |    122 |     43 |     45 |     52 |     56 |     67 |
+|    7B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
+|   13B | perplexity   | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
+|   13B | file size    |  25.0G |   6.8G |   7.6G |   8.3G |   9.1G |    13G |
+|   13B | ms/tok @ 4th |      - |    103 |    105 |    148 |    160 |    131 |
+|   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
+|   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
+
+- [k-quants](https://github.com/ggerganov/llama.cpp/pull/1684)
+- recent k-quants improvements and new i-quants
+  - [#2707](https://github.com/ggerganov/llama.cpp/pull/2707)
+  - [#2807](https://github.com/ggerganov/llama.cpp/pull/2807)
+  - [#4773 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4773)
+  - [#4856 - 2-bit i-quants (inference)](https://github.com/ggerganov/llama.cpp/pull/4856)
+  - [#4861 - importance matrix](https://github.com/ggerganov/llama.cpp/pull/4861)
+  - [#4872 - MoE models](https://github.com/ggerganov/llama.cpp/pull/4872)
+  - [#4897 - 2-bit quantization](https://github.com/ggerganov/llama.cpp/pull/4897)
+  - [#4930 - imatrix for all k-quants](https://github.com/ggerganov/llama.cpp/pull/4930)
+  - [#4951 - imatrix on the GPU](https://github.com/ggerganov/llama.cpp/pull/4957)
+  - [#4969 - imatrix for legacy quants](https://github.com/ggerganov/llama.cpp/pull/4969)
+  - [#4996 - k-qunats tuning](https://github.com/ggerganov/llama.cpp/pull/4996)
+  - [#5060 - Q3_K_XS](https://github.com/ggerganov/llama.cpp/pull/5060)
+  - [#5196 - 3-bit i-quants](https://github.com/ggerganov/llama.cpp/pull/5196)
+  - [quantization tuning](https://github.com/ggerganov/llama.cpp/pull/5320), [another one](https://github.com/ggerganov/llama.cpp/pull/5334), and [another one](https://github.com/ggerganov/llama.cpp/pull/5361)
+
+### Perplexity (measuring model quality)
+
+You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
+For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
+
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 threads.
+
+#### How to run
+
+1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+2. Run `./llama-perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
+3. Output:
 ```
+perplexity : calculating perplexity over 655 chunks
+24.43 seconds per pass - ETA 4.45 hours
+[1]4.5970,[2]5.1807,[3]6.0382,...
+```
+And after 4.45 hours, you will have the final perplexity.
 
 ### Interactive mode
 
-> [!NOTE]
-> If you prefer basic usage, please consider using conversation mode instead of interactive mode
-
+If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
 In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
 Here is an example of a few-shot interaction, invoked with the command
@@ -394,72 +802,18 @@ The `grammars/` folder contains a handful of sample grammars. To write your own,
 
 For authoring more complex JSON grammars, you can also check out https://grammar.intrinsiclabs.ai/, a browser app that lets you write TypeScript interfaces which it compiles to GBNF grammars that you can save for local use. Note that the app is built and maintained by members of the community, please file any issues or FRs on [its repo](http://github.com/intrinsiclabsai/gbnfgen) and not this one.
 
-## Build
+### Obtaining and using the Facebook LLaMA 2 model
 
-Please refer to [Build llama.cpp locally](./docs/build.md)
+- Refer to [Facebook's LLaMA download page](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) if you want to access the model data.
+- Alternatively, if you want to save time and space, you can download already converted and quantized models from [TheBloke](https://huggingface.co/TheBloke), including:
+  - [LLaMA 2 7B base](https://huggingface.co/TheBloke/Llama-2-7B-GGUF)
+  - [LLaMA 2 13B base](https://huggingface.co/TheBloke/Llama-2-13B-GGUF)
+  - [LLaMA 2 70B base](https://huggingface.co/TheBloke/Llama-2-70B-GGUF)
+  - [LLaMA 2 7B chat](https://huggingface.co/TheBloke/Llama-2-7B-chat-GGUF)
+  - [LLaMA 2 13B chat](https://huggingface.co/TheBloke/Llama-2-13B-chat-GGUF)
+  - [LLaMA 2 70B chat](https://huggingface.co/TheBloke/Llama-2-70B-chat-GGUF)
 
-## Supported backends
-
-| Backend | Target devices |
-| --- | --- |
-| [Metal](./docs/build.md#metal-build) | Apple Silicon |
-| [BLAS](./docs/build.md#blas-build) | All |
-| [BLIS](./docs/backend/BLIS.md) | All |
-| [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
-| [CUDA](./docs/build.md#cuda) | Nvidia GPU |
-| [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
-| [Vulkan](./docs/build.md#vulkan) | GPU |
-| [CANN](./docs/build.md#cann) | Ascend NPU |
-
-## Tools
-
-### Prepare and Quantize
-
-> [!NOTE]
-> You can use the [GGUF-my-repo](https://huggingface.co/spaces/ggml-org/gguf-my-repo) space on Hugging Face to quantise your model weights without any setup too. It is synced from `llama.cpp` main every 6 hours.
-
-To obtain the official LLaMA 2 weights please see the <a href="#obtaining-and-using-the-facebook-llama-2-model">Obtaining and using the Facebook LLaMA 2 model</a> section. There is also a large selection of pre-quantized `gguf` models available on Hugging Face.
-
-Note: `convert.py` has been moved to `examples/convert_legacy_llama.py` and shouldn't be used for anything other than `Llama/Llama2/Mistral` models and their derivatives.
-It does not support LLaMA 3, you can use `convert_hf_to_gguf.py` with LLaMA 3 downloaded from Hugging Face.
-
-To learn more about quantizing model, [read this documentation](./examples/quantize/README.md)
-
-### Perplexity (measuring model quality)
-
-You can use the `perplexity` example to measure perplexity over a given prompt (lower perplexity is better).
-For more information, see [https://huggingface.co/docs/transformers/perplexity](https://huggingface.co/docs/transformers/perplexity).
-
-To learn more how to measure perplexity using llama.cpp, [read this documentation](./examples/perplexity/README.md)
-
-## Contributing
-
-- Contributors can open PRs
-- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
-- Collaborators will be invited based on contributions
-- Any help with managing issues and PRs is very appreciated!
-- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
-- Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
-- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
-- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
-
-## Other documentations
-
-- [main (cli)](./examples/main/README.md)
-- [server](./examples/server/README.md)
-- [jeopardy](./examples/jeopardy/README.md)
-- [GBNF grammars](./grammars/README.md)
-
-**Development documentations**
-
-- [How to build](./docs/build.md)
-- [Running on Docker](./docs/docker.md)
-- [Build on Android](./docs/android.md)
-- [Performance troubleshooting](./docs/development/token_generation_performance_tips.md)
-- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
-
-**Seminal papers and background on the models**
+### Seminal papers and background on the models
 
 If your issue is with model generation quality, then please at least scan the following links and papers to understand the limitations of LLaMA models. This is especially important when choosing an appropriate model size and appreciating both the significant and subtle differences between LLaMA models and ChatGPT:
 - LLaMA:
@@ -470,3 +824,178 @@ If your issue is with model generation quality, then please at least scan the fo
 - GPT-3.5 / InstructGPT / ChatGPT:
     - [Aligning language models to follow instructions](https://openai.com/research/instruction-following)
     - [Training language models to follow instructions with human feedback](https://arxiv.org/abs/2203.02155)
+
+### Android
+
+#### Build on Android using Termux
+[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+```
+$ mkdir build-android
+$ cd build-android
+$ export NDK=<your_ndk_directory>
+$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+$ make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+```
+
+Here's a demo of an interactive session running on Pixel 5 phone:
+
+https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
+
+### Docker
+
+#### Prerequisites
+* Docker must be installed and running on your system.
+* Create a folder to store big models & intermediate files (ex. /llama/models)
+
+#### Images
+We have three Docker images available for this project:
+
+1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
+2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+
+Additionally, there the following images, similar to the above:
+
+- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+
+#### Usage
+
+The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
+
+Replace `/path/to/models` below with the actual path where you downloaded the models.
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
+```
+
+On completion, you are ready to play!
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a light image:
+
+```bash
+docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
+```
+
+or with a server image:
+
+```bash
+docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
+```
+
+### Docker With CUDA
+
+Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
+
+#### Building Locally
+
+```bash
+docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
+docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
+docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `CUDA_VERSION` set to `11.7.1`
+- `CUDA_DOCKER_ARCH` set to `all`
+
+The resulting images, are essentially the same as the non-CUDA images:
+
+1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
+3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
+
+#### Usage
+
+After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+```
+
+### Contributing
+
+- Contributors can open PRs
+- Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
+- Collaborators will be invited based on contributions
+- Any help with managing issues and PRs is very appreciated!
+- Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)
+- A bit of backstory for those who are interested: [Changelog podcast](https://changelog.com/podcast/532)
+
+### Coding guidelines
+
+- Avoid adding third-party dependencies, extra files, extra headers, etc.
+- Always consider cross-compatibility with other operating systems and architectures
+- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
+- There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
+- Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
+- See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
+- Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+
+![matmul](media/matmul.png)
+
+### Docs
+
+- [main (cli)](./examples/main/README.md)
+- [server](./examples/server/README.md)
+- [jeopardy](./examples/jeopardy/README.md)
+- [BLIS](./docs/BLIS.md)
+- [Performance troubleshooting](./docs/token_generation_performance_tips.md)
+- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)
+- [GBNF grammars](./grammars/README.md)

ci/run.sh

@@ -13,9 +13,6 @@
 # # with SYCL support
 # GG_BUILD_SYCL=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
 #
-# # with VULKAN support
-# GG_BUILD_VULKAN=1 bash ./ci/run.sh ./tmp/results ./tmp/mnt
-#
 
 if [ -z "$2" ]; then
     echo "usage: $0 <output-dir> <mnt-dir>"
@@ -43,7 +40,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -55,10 +52,6 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
 
     CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
-
-if [ ! -z ${GG_BUILD_VULKAN} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_VULKAN=1"
-fi
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -110,11 +103,8 @@ function gg_run_ctest_debug {
 
     set -e
 
-    # Check cmake, make and ctest are installed
-    gg_check_build_requirements
-
     (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                  ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     (time ctest --output-on-failure -L main -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
@@ -141,11 +131,8 @@ function gg_run_ctest_release {
 
     set -e
 
-    # Check cmake, make and ctest are installed
-    gg_check_build_requirements
-
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure -L main ) 2>&1 | tee -a $OUT/${ci}-ctest.log
@@ -273,6 +260,7 @@ function gg_sum_ctest_with_model_release {
 }
 
 # open_llama_7b_v2
+# requires: GG_BUILD_CUDA
 
 function gg_run_open_llama_7b_v2 {
     cd ${SRC}
@@ -296,8 +284,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -431,7 +419,7 @@ function gg_run_pythia_1_4b {
     set -e
 
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -541,6 +529,7 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
+# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -561,8 +550,8 @@ function gg_run_pythia_2_8b {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DGGML_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -697,7 +686,7 @@ function gg_run_embd_bge_small {
     set -e
 
     (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -712,20 +701,6 @@ function gg_run_embd_bge_small {
     set +e
 }
 
-function gg_check_build_requirements {
-    if ! command -v cmake &> /dev/null; then
-        gg_printf 'cmake not found, please install'
-    fi
-
-    if ! command -v make &> /dev/null; then
-        gg_printf 'make not found, please install'
-    fi
-
-    if ! command -v ctest &> /dev/null; then
-        gg_printf 'ctest not found, please install'
-    fi
-}
-
 function gg_sum_embd_bge_small {
     gg_printf '### %s\n\n' "${ci}"
 
@@ -766,7 +741,7 @@ if [ -z ${GG_BUILD_LOW_PERF} ]; then
     fi
 
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
-        if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
+        if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run pythia_1_4b
         else
             test $ret -eq 0 && gg_run pythia_2_8b

cmake/llama-config.cmake.in

@@ -8,13 +8,6 @@ set(GGML_CUDA       @GGML_CUDA@)
 set(GGML_METAL      @GGML_METAL@)
 set(GGML_HIPBLAS    @GGML_HIPBLAS@)
 set(GGML_ACCELERATE @GGML_ACCELERATE@)
-set(GGML_VULKAN @GGML_VULKAN@)
-set(GGML_VULKAN_CHECK_RESULTS @GGML_VULKAN_CHECK_RESULTS@)
-set(GGML_VULKAN_DEBUG @GGML_VULKAN_DEBUG@)
-set(GGML_VULKAN_MEMORY_DEBUG @GGML_VULKAN_MEMORY_DEBUG@)
-set(GGML_VULKAN_VALIDATE @GGML_VULKAN_VALIDATE@)
-set(GGML_SYCL @GGML_SYCL@)
-set(GGML_OPENMP @GGML_OPENMP@)
 
 @PACKAGE_INIT@
 
@@ -44,36 +37,18 @@ if (GGML_METAL)
     find_library(METALKIT_FRAMEWORK MetalKit REQUIRED)
 endif()
 
-if (GGML_VULKAN)
-    find_package(Vulkan REQUIRED)
-endif()
-
 if (GGML_HIPBLAS)
     find_package(hip REQUIRED)
     find_package(hipblas REQUIRED)
     find_package(rocblas REQUIRED)
 endif()
 
-if (GGML_SYCL)
-    find_package(IntelSYCL REQUIRED)
-    find_package(MKL REQUIRED)
-endif()
-
-if (GGML_OPENMP)
-    find_package(OpenMP REQUIRED)
-endif()
-
-
-find_library(ggml_LIBRARY ggml
-    REQUIRED
-    HINTS ${LLAMA_LIB_DIR})
-
 find_library(llama_LIBRARY llama
     REQUIRED
     HINTS ${LLAMA_LIB_DIR})
 
-set(_llama_link_deps "${ggml_LIBRARY}" "@GGML_LINK_LIBRARIES@")
-set(_llama_transient_defines "@GGML_TRANSIENT_DEFINES@")
+set(_llama_link_deps "Threads::Threads" "@LLAMA_EXTRA_LIBS@")
+set(_llama_transient_defines "@LLAMA_TRANSIENT_DEFINES@")
 
 add_library(llama UNKNOWN IMPORTED)
 

common/CMakeLists.txt

@@ -54,12 +54,12 @@ add_library(${TARGET} STATIC
     base64.hpp
     common.h
     common.cpp
-    arg.h
-    arg.cpp
     sampling.h
     sampling.cpp
     console.h
     console.cpp
+    grammar-parser.h
+    grammar-parser.cpp
     json.hpp
     json-schema-to-grammar.cpp
     train.h

common/arg.h

@@ -1,77 +0,0 @@
-#pragma once
-
-#include "common.h"
-
-#include <set>
-#include <string>
-#include <vector>
-
-//
-// CLI argument parsing
-//
-
-struct llama_arg {
-    std::set<enum llama_example> examples = {LLAMA_EXAMPLE_COMMON};
-    std::vector<const char *> args;
-    const char * value_hint   = nullptr; // help text or example for arg value
-    const char * value_hint_2 = nullptr; // for second arg value
-    const char * env          = nullptr;
-    std::string help;
-    bool is_sparam = false; // is current arg a sampling param?
-    void (*handler_void)   (gpt_params & params) = nullptr;
-    void (*handler_string) (gpt_params & params, const std::string &) = nullptr;
-    void (*handler_str_str)(gpt_params & params, const std::string &, const std::string &) = nullptr;
-    void (*handler_int)    (gpt_params & params, int) = nullptr;
-
-    llama_arg(
-        const std::initializer_list<const char *> & args,
-        const char * value_hint,
-        const std::string & help,
-        void (*handler)(gpt_params & params, const std::string &)
-    ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
-
-    llama_arg(
-        const std::initializer_list<const char *> & args,
-        const char * value_hint,
-        const std::string & help,
-        void (*handler)(gpt_params & params, int)
-    ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
-
-    llama_arg(
-        const std::initializer_list<const char *> & args,
-        const std::string & help,
-        void (*handler)(gpt_params & params)
-    ) : args(args), help(help), handler_void(handler) {}
-
-    // support 2 values for arg
-    llama_arg(
-        const std::initializer_list<const char *> & args,
-        const char * value_hint,
-        const char * value_hint_2,
-        const std::string & help,
-        void (*handler)(gpt_params & params, const std::string &, const std::string &)
-    ) : args(args), value_hint(value_hint), value_hint_2(value_hint_2), help(help), handler_str_str(handler) {}
-
-    llama_arg & set_examples(std::initializer_list<enum llama_example> examples);
-    llama_arg & set_env(const char * env);
-    llama_arg & set_sparam();
-    bool in_example(enum llama_example ex);
-    bool get_value_from_env(std::string & output);
-    bool has_value_from_env();
-    std::string to_string();
-};
-
-struct gpt_params_context {
-    enum llama_example ex = LLAMA_EXAMPLE_COMMON;
-    gpt_params & params;
-    std::vector<llama_arg> options;
-    void(*print_usage)(int, char **) = nullptr;
-    gpt_params_context(gpt_params & params) : params(params) {}
-};
-
-// parse input arguments from CLI
-// if one argument has invalid value, it will automatically display usage of the specific argument (and not the full usage message)
-bool gpt_params_parse(int argc, char ** argv, gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
-
-// function to be used by test-arg-parser
-gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/common.h

@@ -4,11 +4,18 @@
 
 #include "llama.h"
 
+#include "sampling.h"
+
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
+#include <cmath>
 #include <string>
 #include <vector>
+#include <random>
+#include <thread>
+#include <unordered_map>
+#include <tuple>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -26,15 +33,6 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct llama_lora_adapter_info {
-    std::string path;
-    float scale;
-};
-
-struct llama_lora_adapter_container : llama_lora_adapter_info {
-    struct llama_lora_adapter * adapter;
-};
-
 // build info
 extern int LLAMA_BUILD_NUMBER;
 extern char const * LLAMA_COMMIT;
@@ -47,103 +45,26 @@ struct llama_control_vector_load_info;
 // CPU utils
 //
 
-struct cpu_params {
-    int      n_threads                   = -1;
-    bool     cpumask[GGML_MAX_N_THREADS] = {false}; // CPU affinity mask.
-    bool     mask_valid                  = false;   // Default: any CPU
-    enum ggml_sched_priority  priority   = GGML_SCHED_PRIO_NORMAL;  // Scheduling prio : (0 - normal, 1 - medium, 2 - high, 3 - realtime)
-    bool     strict_cpu                  = false;   // Use strict CPU placement
-    uint32_t poll                        = 50;      // Polling (busywait) level (0 - no polling, 100 - mostly polling)
-};
-
 int32_t cpu_get_num_physical_cores();
 int32_t cpu_get_num_math();
 
 //
-// Common params
+// CLI argument parsing
 //
 
-enum llama_example {
-    LLAMA_EXAMPLE_COMMON,
-    LLAMA_EXAMPLE_SPECULATIVE,
-    LLAMA_EXAMPLE_MAIN,
-    LLAMA_EXAMPLE_INFILL,
-    LLAMA_EXAMPLE_EMBEDDING,
-    LLAMA_EXAMPLE_PERPLEXITY,
-    LLAMA_EXAMPLE_RETRIEVAL,
-    LLAMA_EXAMPLE_PASSKEY,
-    LLAMA_EXAMPLE_IMATRIX,
-    LLAMA_EXAMPLE_BENCH,
-    LLAMA_EXAMPLE_SERVER,
-    LLAMA_EXAMPLE_CVECTOR_GENERATOR,
-    LLAMA_EXAMPLE_EXPORT_LORA,
-    LLAMA_EXAMPLE_LLAVA,
-    LLAMA_EXAMPLE_LOOKUP,
-    LLAMA_EXAMPLE_PARALLEL,
-
-    LLAMA_EXAMPLE_COUNT,
-};
-
-enum gpt_sampler_type {
-    GPT_SAMPLER_TYPE_NONE        = 0,
-    GPT_SAMPLER_TYPE_TOP_K       = 1,
-    GPT_SAMPLER_TYPE_TOP_P       = 2,
-    GPT_SAMPLER_TYPE_MIN_P       = 3,
-    GPT_SAMPLER_TYPE_TFS_Z       = 4,
-    GPT_SAMPLER_TYPE_TYPICAL_P   = 5,
-    GPT_SAMPLER_TYPE_TEMPERATURE = 6,
-};
-
 // dimensionality reduction methods, used by cvector-generator
 enum dimre_method {
     DIMRE_METHOD_PCA,
     DIMRE_METHOD_MEAN,
 };
 
-// sampler parameters
-struct gpt_sampler_params {
-    uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
-
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
-    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range    = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.00f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = false; // consider newlines as a repeatable token
-    bool    ignore_eos        = false;
-    bool    no_perf           = false; // disable performance metrics
-
-    std::vector<enum gpt_sampler_type> samplers = {
-        GPT_SAMPLER_TYPE_TOP_K,
-        GPT_SAMPLER_TYPE_TFS_Z,
-        GPT_SAMPLER_TYPE_TYPICAL_P,
-        GPT_SAMPLER_TYPE_TOP_P,
-        GPT_SAMPLER_TYPE_MIN_P,
-        GPT_SAMPLER_TYPE_TEMPERATURE
-    };
-
-    std::string grammar; // optional BNF-like grammar to constrain sampling
-
-    std::vector<llama_logit_bias> logit_bias; // logit biases to apply
-
-    // print the parameters into a string
-    std::string print() const;
-};
-
 struct gpt_params {
+    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+
+    int32_t n_threads             = cpu_get_num_math();
+    int32_t n_threads_draft       =    -1;
+    int32_t n_threads_batch       =    -1; // number of threads to use for batch processing (-1 = use n_threads)
+    int32_t n_threads_batch_draft =    -1;
     int32_t n_predict             =    -1; // new tokens to predict
     int32_t n_ctx                 =     0; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
@@ -170,11 +91,6 @@ struct gpt_params {
     int32_t yarn_orig_ctx         =     0; // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
-    struct cpu_params cpuparams;
-    struct cpu_params cpuparams_batch;
-    struct cpu_params draft_cpuparams;
-    struct cpu_params draft_cpuparams_batch;
-
     ggml_backend_sched_eval_callback cb_eval = nullptr;
     void * cb_eval_user_data                 = nullptr;
 
@@ -183,34 +99,34 @@ struct gpt_params {
     enum llama_split_mode        split_mode        = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
     enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
-    enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    struct gpt_sampler_params sparams;
+    // // sampling parameters
+    struct llama_sampling_params sparams;
 
-    std::string model                = ""; // model path                                                    // NOLINT
-    std::string model_draft          = ""; // draft model for speculative decoding                          // NOLINT
-    std::string model_alias          = "unknown"; // model alias                                            // NOLINT
-    std::string model_url            = ""; // model url to download                                         // NOLINT
-    std::string hf_token             = ""; // HF token                                                      // NOLINT
-    std::string hf_repo              = ""; // HF repo                                                       // NOLINT
-    std::string hf_file              = ""; // HF file                                                       // NOLINT
-    std::string prompt               = "";                                                                  // NOLINT
-    std::string prompt_file          = ""; // store the external prompt file name                           // NOLINT
-    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state             // NOLINT
-    std::string input_prefix         = ""; // string to prefix user inputs with                             // NOLINT
-    std::string input_suffix         = ""; // string to suffix user inputs with                             // NOLINT
-    std::string logdir               = ""; // directory in which to save YAML log files                     // NOLINT
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding           // NOLINT
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding          // NOLINT
-    std::string logits_file          = ""; // file for saving *all* logits                                  // NOLINT
-    std::string rpc_servers          = ""; // comma separated list of RPC servers                           // NOLINT
+    std::string model                = ""; // model path
+    std::string model_draft          = ""; // draft model for speculative decoding
+    std::string model_alias          = "unknown"; // model alias
+    std::string model_url            = ""; // model url to download
+    std::string hf_repo              = ""; // HF repo
+    std::string hf_file              = ""; // HF file
+    std::string prompt               = "";
+    std::string prompt_file          = ""; // store the external prompt file name
+    std::string path_prompt_cache    = ""; // path to file for saving/loading prompt eval state
+    std::string input_prefix         = ""; // string to prefix user inputs with
+    std::string input_suffix         = ""; // string to suffix user inputs with
+    std::string logdir               = ""; // directory in which to save YAML log files
+    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
+    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
+    std::string logits_file          = ""; // file for saving *all* logits
+    std::string rpc_servers          = ""; // comma separated list of RPC servers
 
     std::vector<std::string> in_files;   // all input files
     std::vector<std::string> antiprompt; // strings upon which more user input is prompted (a.k.a. reverse prompts)
     std::vector<llama_model_kv_override> kv_overrides;
 
-    bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
-    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
+    // TODO: avoid tuple, use struct
+    std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
+    std::string lora_base  = "";                              // base model path for the lora adapter
 
     std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
 
@@ -247,14 +163,15 @@ struct gpt_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
-    bool no_perf           = false; // disable performance metrics
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
+    bool ignore_eos        = false; // ignore generated EOS tokens
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
+    bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
@@ -264,7 +181,7 @@ struct gpt_params {
     std::string cache_type_v = "f16"; // KV cache data type for the V
 
     // multimodal models (see examples/llava)
-    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
+    std::string mmproj = "";        // path to multimodal projector
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -277,18 +194,18 @@ struct gpt_params {
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
 
     std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";                                                                         // NOLINT
-    std::string chat_template = "";                                                                         // NOLINT
-    std::string system_prompt = "";                                                                         // NOLINT
+    std::string public_path   = "";
+    std::string chat_template = "";
+    std::string system_prompt = "";
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
 
-    std::string ssl_file_key  = "";                                                                         // NOLINT
-    std::string ssl_file_cert = "";                                                                         // NOLINT
+    std::string ssl_file_key  = "";
+    std::string ssl_file_cert = "";
 
     bool endpoint_slots   = true;
     bool endpoint_metrics = false;
@@ -336,19 +253,16 @@ struct gpt_params {
     std::string cvector_negative_file = "examples/cvector-generator/negative.txt";
 
     bool spm_infill = false; // suffix/prefix/middle pattern for infill
-
-    std::string lora_outfile = "ggml-lora-merged-f16.gguf";
-
-    // batched-bench params
-    bool batched_bench_output_jsonl = false;
 };
 
-std::string gpt_params_get_system_info(const gpt_params & params);
+void gpt_params_handle_model_default(gpt_params & params);
 
-bool parse_cpu_range(const std::string& range, bool(&boolmask)[GGML_MAX_N_THREADS]);
-bool parse_cpu_mask(const std::string& mask, bool(&boolmask)[GGML_MAX_N_THREADS]);
-void postprocess_cpu_params(cpu_params& cpuparams, const cpu_params* role_model = nullptr);
-bool set_process_priority(enum ggml_sched_priority prio);
+bool gpt_params_parse_ex   (int argc, char ** argv, gpt_params & params);
+bool gpt_params_parse      (int argc, char ** argv, gpt_params & params);
+bool gpt_params_find_arg   (int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+void gpt_params_print_usage(int argc, char ** argv, const gpt_params & params);
+
+std::string gpt_params_get_system_info(const gpt_params & params);
 
 //
 // String utils
@@ -359,8 +273,6 @@ std::vector<std::string> string_split(std::string input, char separator);
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
-void string_replace_all(std::string & s, const std::string & search, const std::string & replace);
-
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
     std::vector<T> values;
@@ -392,23 +304,14 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-struct llama_init_result {
-    struct llama_model   * model   = nullptr;
-    struct llama_context * context = nullptr;
-    std::vector<llama_lora_adapter_container> lora_adapters;
-};
+// TODO: avoid tuplue, use struct
+std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params);
 
-struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
+struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
+struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
-struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
-struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
-struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
-
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
-
-// clear LoRA adapters from context, then apply new list of adapters
-void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
 
 // Batch utils
 
@@ -446,13 +349,25 @@ std::string llama_token_to_piece(
                        llama_token   token,
                        bool          special = true);
 
+// TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
+//       that takes into account the tokenizer type and decides how to handle the leading space
+//
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
-// optionally renders special/control tokens
-std::string llama_detokenize(
+// removes the leading space from the first non-BOS token
+std::string llama_detokenize_spm(
                          llama_context * ctx,
-        const std::vector<llama_token> & tokens,
-                                  bool   special = true);
+        const std::vector<llama_token> & tokens);
+
+// detokenizes a vector of tokens into a string
+// should work similar to Python's `tokenizer.decode`
+std::string llama_detokenize_bpe(
+                         llama_context * ctx,
+        const std::vector<llama_token> & tokens);
+
+// Uses the value from the model metadata if possible, otherwise
+// defaults to true when model type is SPM, otherwise false.
+bool llama_should_add_bos_token(const llama_model * model);
 
 //
 // Chat template utils

common/grammar-parser.cpp

@@ -0,0 +1,536 @@
+#include "grammar-parser.h"
+#include <cstdint>
+#include <cwchar>
+#include <string>
+#include <utility>
+#include <stdexcept>
+#include <exception>
+
+namespace grammar_parser {
+    // NOTE: assumes valid utf8 (but checks for overrun)
+    // copied from llama.cpp
+    static std::pair<uint32_t, const char *> decode_utf8(const char * src) {
+        static const int lookup[] = { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 2, 2, 3, 4 };
+        uint8_t  first_byte = static_cast<uint8_t>(*src);
+        uint8_t  highbits   = first_byte >> 4;
+        int      len        = lookup[highbits];
+        uint8_t  mask       = (1 << (8 - len)) - 1;
+        uint32_t value      = first_byte & mask;
+        const char * end    = src + len; // may overrun!
+        const char * pos    = src + 1;
+        for ( ; pos < end && *pos; pos++) {
+            value = (value << 6) + (static_cast<uint8_t>(*pos) & 0x3F);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static uint32_t get_symbol_id(parse_state & state, const char * src, size_t len) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        auto result = state.symbol_ids.emplace(std::string(src, len), next_id);
+        return result.first->second;
+    }
+
+    static uint32_t generate_symbol_id(parse_state & state, const std::string & base_name) {
+        uint32_t next_id = static_cast<uint32_t>(state.symbol_ids.size());
+        state.symbol_ids[base_name + '_' + std::to_string(next_id)] = next_id;
+        return next_id;
+    }
+
+    static void add_rule(
+            parse_state & state,
+            uint32_t      rule_id,
+            const std::vector<llama_grammar_element> & rule) {
+        if (state.rules.size() <= rule_id) {
+            state.rules.resize(rule_id + 1);
+        }
+        state.rules[rule_id] = rule;
+    }
+
+    static bool is_digit_char(char c) {
+        return '0' <= c && c <= '9';
+    }
+
+    static bool is_word_char(char c) {
+        return ('a' <= c && c <= 'z') || ('A' <= c && c <= 'Z') || c == '-' || is_digit_char(c);
+    }
+
+    static std::pair<uint32_t, const char *> parse_hex(const char * src, int size) {
+        const char * pos   = src;
+        const char * end   = src + size;
+        uint32_t     value = 0;
+        for ( ; pos < end && *pos; pos++) {
+            value <<= 4;
+            char c = *pos;
+            if ('a' <= c && c <= 'f') {
+                value += c - 'a' + 10;
+            } else if ('A' <= c && c <= 'F') {
+                value += c - 'A' + 10;
+            } else if ('0' <= c && c <= '9') {
+                value += c - '0';
+            } else {
+                break;
+            }
+        }
+        if (pos != end) {
+            throw std::runtime_error("expecting " + std::to_string(size) + " hex chars at " + src);
+        }
+        return std::make_pair(value, pos);
+    }
+
+    static const char * parse_space(const char * src, bool newline_ok) {
+        const char * pos = src;
+        while (*pos == ' ' || *pos == '\t' || *pos == '#' ||
+                (newline_ok && (*pos == '\r' || *pos == '\n'))) {
+            if (*pos == '#') {
+                while (*pos && *pos != '\r' && *pos != '\n') {
+                    pos++;
+                }
+            } else {
+                pos++;
+            }
+        }
+        return pos;
+    }
+
+    static const char * parse_name(const char * src) {
+        const char * pos = src;
+        while (is_word_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting name at ") + src);
+        }
+        return pos;
+    }
+
+    static const char * parse_int(const char * src) {
+        const char * pos = src;
+        while (is_digit_char(*pos)) {
+            pos++;
+        }
+        if (pos == src) {
+            throw std::runtime_error(std::string("expecting integer at ") + src);
+        }
+        return pos;
+    }
+
+    static std::pair<uint32_t, const char *> parse_char(const char * src) {
+        if (*src == '\\') {
+            switch (src[1]) {
+                case 'x': return parse_hex(src + 2, 2);
+                case 'u': return parse_hex(src + 2, 4);
+                case 'U': return parse_hex(src + 2, 8);
+                case 't': return std::make_pair('\t', src + 2);
+                case 'r': return std::make_pair('\r', src + 2);
+                case 'n': return std::make_pair('\n', src + 2);
+                case '\\':
+                case '"':
+                case '[':
+                case ']':
+                    return std::make_pair(src[1], src + 2);
+                default:
+                    throw std::runtime_error(std::string("unknown escape at ") + src);
+            }
+        } else if (*src) {
+            return decode_utf8(src);
+        }
+        throw std::runtime_error("unexpected end of input");
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested);
+
+    static const char * parse_sequence(
+            parse_state                        & state,
+            const char                         * src,
+            const std::string                  & rule_name,
+            std::vector<llama_grammar_element> & out_elements,
+            bool                                 is_nested) {
+        size_t last_sym_start = out_elements.size();
+        const char * pos = src;
+
+        auto handle_repetitions = [&](int min_times, int max_times) {
+
+            if (last_sym_start == out_elements.size()) {
+                throw std::runtime_error(std::string("expecting preceding item to */+/?/{ at ") + pos);
+            }
+
+            // apply transformation to previous symbol (last_sym_start to end) according to
+            // the following rewrite rules:
+            // S{m,n} --> S S S (m times) S'(n-m)
+            //            S'(x)   ::= S S'(x-1) |
+            //            (... n-m definitions of these S' rules ...)
+            //            S'(1)   ::= S |
+            // S{m,} -->  S S S (m times) S'
+            //            S'     ::= S S' |
+            // S*     --> S{0,}
+            //        --> S'     ::= S S' |
+            // S+     --> S{1,}
+            //        --> S S'
+            //            S'     ::= S S' |
+            // S?     --> S{0,1}
+            //        --> S'
+            //            S'     ::= S |
+
+            std::vector<llama_grammar_element> previous_elements(out_elements.begin() + last_sym_start, out_elements.end());
+            if (min_times == 0) {
+                out_elements.resize(last_sym_start);
+            } else {
+                // Repeat the previous elements (min_times - 1) times
+                for (int i = 1; i < min_times; i++) {
+                    out_elements.insert(out_elements.end(), previous_elements.begin(), previous_elements.end());
+                }
+            }
+
+            uint32_t last_rec_rule_id = 0;
+            auto n_opt = max_times < 0 ? 1 : max_times - min_times;
+
+            std::vector<llama_grammar_element> rec_rule(previous_elements);
+            for (int i = 0; i < n_opt; i++) {
+                rec_rule.resize(previous_elements.size());
+                uint32_t rec_rule_id = generate_symbol_id(state, rule_name);
+                if (i > 0 || max_times < 0) {
+                    rec_rule.push_back({LLAMA_GRETYPE_RULE_REF, max_times < 0 ? rec_rule_id : last_rec_rule_id});
+                }
+                rec_rule.push_back({LLAMA_GRETYPE_ALT, 0});
+                rec_rule.push_back({LLAMA_GRETYPE_END, 0});
+                add_rule(state, rec_rule_id, rec_rule);
+                last_rec_rule_id = rec_rule_id;
+            }
+            if (n_opt > 0) {
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, last_rec_rule_id});
+            }
+        };
+
+        while (*pos) {
+            if (*pos == '"') { // literal string
+                pos++;
+                last_sym_start = out_elements.size();
+                while (*pos != '"') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    out_elements.push_back({LLAMA_GRETYPE_CHAR, char_pair.first});
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '[') { // char range(s)
+                pos++;
+                enum llama_gretype start_type = LLAMA_GRETYPE_CHAR;
+                if (*pos == '^') {
+                    pos++;
+                    start_type = LLAMA_GRETYPE_CHAR_NOT;
+                }
+                last_sym_start = out_elements.size();
+                while (*pos != ']') {
+                    if (!*pos) {
+                        throw std::runtime_error("unexpected end of input");
+                    }
+                    auto char_pair = parse_char(pos);
+                         pos       = char_pair.second;
+                    enum llama_gretype type = last_sym_start < out_elements.size()
+                        ? LLAMA_GRETYPE_CHAR_ALT
+                        : start_type;
+
+                    out_elements.push_back({type, char_pair.first});
+                    if (pos[0] == '-' && pos[1] != ']') {
+                        if (!pos[1]) {
+                            throw std::runtime_error("unexpected end of input");
+                        }
+                        auto endchar_pair = parse_char(pos + 1);
+                             pos          = endchar_pair.second;
+                        out_elements.push_back({LLAMA_GRETYPE_CHAR_RNG_UPPER, endchar_pair.first});
+                    }
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (is_word_char(*pos)) { // rule reference
+                const char * name_end    = parse_name(pos);
+                uint32_t     ref_rule_id = get_symbol_id(state, pos, name_end - pos);
+                pos = parse_space(name_end, is_nested);
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, ref_rule_id});
+            } else if (*pos == '(') { // grouping
+                // parse nested alternates into synthesized rule
+                pos = parse_space(pos + 1, true);
+                uint32_t sub_rule_id = generate_symbol_id(state, rule_name);
+                pos = parse_alternates(state, pos, rule_name, sub_rule_id, true);
+                last_sym_start = out_elements.size();
+                // output reference to synthesized rule
+                out_elements.push_back({LLAMA_GRETYPE_RULE_REF, sub_rule_id});
+                if (*pos != ')') {
+                    throw std::runtime_error(std::string("expecting ')' at ") + pos);
+                }
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '.') { // any char
+                last_sym_start = out_elements.size();
+                out_elements.push_back({LLAMA_GRETYPE_CHAR_ANY, 0});
+                pos = parse_space(pos + 1, is_nested);
+            } else if (*pos == '*') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(0, -1);
+            } else if (*pos == '+') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(1, -1);
+            } else if (*pos == '?') {
+                pos = parse_space(pos + 1, is_nested);
+                handle_repetitions(0, 1);
+            } else if (*pos == '{') {
+                pos = parse_space(pos + 1, is_nested);
+
+                if (!is_digit_char(*pos)) {
+                    throw std::runtime_error(std::string("expecting an int at ") + pos);
+                }
+                const char * int_end = parse_int(pos);
+                int min_times = std::stoul(std::string(pos, int_end - pos));
+                pos = parse_space(int_end, is_nested);
+
+                int max_times = -1;
+
+                if (*pos == '}') {
+                    max_times = min_times;
+                    pos = parse_space(pos + 1, is_nested);
+                } else if (*pos == ',') {
+                    pos = parse_space(pos + 1, is_nested);
+
+                    if (is_digit_char(*pos)) {
+                        const char * int_end = parse_int(pos);
+                        max_times = std::stoul(std::string(pos, int_end - pos));
+                        pos = parse_space(int_end, is_nested);
+                    }
+
+                    if (*pos != '}') {
+                        throw std::runtime_error(std::string("expecting '}' at ") + pos);
+                    }
+                    pos = parse_space(pos + 1, is_nested);
+                } else {
+                    throw std::runtime_error(std::string("expecting ',' at ") + pos);
+                }
+                handle_repetitions(min_times, max_times);
+            } else {
+                break;
+            }
+        }
+        return pos;
+    }
+
+    const char * parse_alternates(
+            parse_state       & state,
+            const char        * src,
+            const std::string & rule_name,
+            uint32_t            rule_id,
+            bool                is_nested) {
+        std::vector<llama_grammar_element> rule;
+        const char * pos = parse_sequence(state, src, rule_name, rule, is_nested);
+        while (*pos == '|') {
+            rule.push_back({LLAMA_GRETYPE_ALT, 0});
+            pos = parse_space(pos + 1, true);
+            pos = parse_sequence(state, pos, rule_name, rule, is_nested);
+        }
+        rule.push_back({LLAMA_GRETYPE_END, 0});
+        add_rule(state, rule_id, rule);
+        return pos;
+    }
+
+    static const char * parse_rule(parse_state & state, const char * src) {
+        const char * name_end = parse_name(src);
+        const char * pos      = parse_space(name_end, false);
+        size_t       name_len = name_end - src;
+        uint32_t     rule_id  = get_symbol_id(state, src, name_len);
+        const std::string name(src, name_len);
+
+        if (!(pos[0] == ':' && pos[1] == ':' && pos[2] == '=')) {
+            throw std::runtime_error(std::string("expecting ::= at ") + pos);
+        }
+        pos = parse_space(pos + 3, true);
+
+        pos = parse_alternates(state, pos, name, rule_id, false);
+
+        if (*pos == '\r') {
+            pos += pos[1] == '\n' ? 2 : 1;
+        } else if (*pos == '\n') {
+            pos++;
+        } else if (*pos) {
+            throw std::runtime_error(std::string("expecting newline or end at ") + pos);
+        }
+        return parse_space(pos, true);
+    }
+
+    parse_state parse(const char * src) {
+        try {
+            parse_state state;
+            const char * pos = parse_space(src, true);
+            while (*pos) {
+                pos = parse_rule(state, pos);
+            }
+            // Validate the state to ensure that all rules are defined
+            for (const auto & rule : state.rules) {
+                for (const auto & elem : rule) {
+                    if (elem.type == LLAMA_GRETYPE_RULE_REF) {
+                        // Ensure that the rule at that location exists
+                        if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
+                            // Get the name of the rule that is missing
+                            for (const auto & kv : state.symbol_ids) {
+                                if (kv.second == elem.value) {
+                                    throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+            return state;
+        } catch (const std::exception & err) {
+            fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());
+            return parse_state();
+        }
+    }
+
+    static void print_grammar_char(FILE * file, uint32_t c) {
+        if (0x20 <= c && c <= 0x7f) {
+            fprintf(file, "%c", static_cast<char>(c));
+        } else {
+            // cop out of encoding UTF-8
+            fprintf(file, "<U+%04X>", c);
+        }
+    }
+
+    static bool is_char_element(llama_grammar_element elem) {
+        switch (elem.type) {
+            case LLAMA_GRETYPE_CHAR:           return true;
+            case LLAMA_GRETYPE_CHAR_NOT:       return true;
+            case LLAMA_GRETYPE_CHAR_ALT:       return true;
+            case LLAMA_GRETYPE_CHAR_RNG_UPPER: return true;
+            case LLAMA_GRETYPE_CHAR_ANY:       return true;
+            default:                           return false;
+        }
+    }
+
+    static void print_rule_binary(FILE * file, const std::vector<llama_grammar_element> & rule) {
+        for (auto elem : rule) {
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:            fprintf(file, "END");            break;
+                case LLAMA_GRETYPE_ALT:            fprintf(file, "ALT");            break;
+                case LLAMA_GRETYPE_RULE_REF:       fprintf(file, "RULE_REF");       break;
+                case LLAMA_GRETYPE_CHAR:           fprintf(file, "CHAR");           break;
+                case LLAMA_GRETYPE_CHAR_NOT:       fprintf(file, "CHAR_NOT");       break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER: fprintf(file, "CHAR_RNG_UPPER"); break;
+                case LLAMA_GRETYPE_CHAR_ALT:       fprintf(file, "CHAR_ALT");       break;
+                case LLAMA_GRETYPE_CHAR_ANY:       fprintf(file, "CHAR_ANY");       break;
+            }
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                case LLAMA_GRETYPE_ALT:
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "(%u) ", elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                case LLAMA_GRETYPE_CHAR_NOT:
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                case LLAMA_GRETYPE_CHAR_ALT:
+                case LLAMA_GRETYPE_CHAR_ANY:
+                    fprintf(file, "(\"");
+                    print_grammar_char(file, elem.value);
+                    fprintf(file, "\") ");
+                    break;
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    static void print_rule(
+            FILE     * file,
+            uint32_t   rule_id,
+            const std::vector<llama_grammar_element> & rule,
+            const std::map<uint32_t, std::string>    & symbol_id_names) {
+        if (rule.empty() || rule.back().type != LLAMA_GRETYPE_END) {
+            throw std::runtime_error(
+                "malformed rule, does not end with LLAMA_GRETYPE_END: " + std::to_string(rule_id));
+        }
+        fprintf(file, "%s ::= ", symbol_id_names.at(rule_id).c_str());
+        for (size_t i = 0, end = rule.size() - 1; i < end; i++) {
+            llama_grammar_element elem = rule[i];
+            switch (elem.type) {
+                case LLAMA_GRETYPE_END:
+                    throw std::runtime_error(
+                        "unexpected end of rule: " + std::to_string(rule_id) + "," +
+                        std::to_string(i));
+                case LLAMA_GRETYPE_ALT:
+                    fprintf(file, "| ");
+                    break;
+                case LLAMA_GRETYPE_RULE_REF:
+                    fprintf(file, "%s ", symbol_id_names.at(elem.value).c_str());
+                    break;
+                case LLAMA_GRETYPE_CHAR:
+                    fprintf(file, "[");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_NOT:
+                    fprintf(file, "[^");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_RNG_UPPER without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    fprintf(file, "-");
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_ALT:
+                    if (i == 0 || !is_char_element(rule[i - 1])) {
+                        throw std::runtime_error(
+                            "LLAMA_GRETYPE_CHAR_ALT without preceding char: " +
+                            std::to_string(rule_id) + "," + std::to_string(i));
+                    }
+                    print_grammar_char(file, elem.value);
+                    break;
+                case LLAMA_GRETYPE_CHAR_ANY:
+                    fprintf(file, ".");
+                    break;
+            }
+            if (is_char_element(elem)) {
+                switch (rule[i + 1].type) {
+                    case LLAMA_GRETYPE_CHAR_ALT:
+                    case LLAMA_GRETYPE_CHAR_RNG_UPPER:
+                    case LLAMA_GRETYPE_CHAR_ANY:
+                        break;
+                    default:
+                        fprintf(file, "] ");
+                }
+            }
+        }
+        fprintf(file, "\n");
+    }
+
+    void print_grammar(FILE * file, const parse_state & state) {
+        try {
+            std::map<uint32_t, std::string> symbol_id_names;
+            for (const auto & kv : state.symbol_ids) {
+                symbol_id_names[kv.second] = kv.first;
+            }
+            for (size_t i = 0, end = state.rules.size(); i < end; i++) {
+                // fprintf(file, "%zu: ", i);
+                // print_rule_binary(file, state.rules[i]);
+                print_rule(file, uint32_t(i), state.rules[i], symbol_id_names);
+                // fprintf(file, "\n");
+            }
+        } catch (const std::exception & err) {
+            fprintf(stderr, "\n%s: error printing grammar: %s\n", __func__, err.what());
+        }
+    }
+
+    std::vector<const llama_grammar_element *> parse_state::c_rules() {
+        std::vector<const llama_grammar_element *> ret;
+        ret.reserve(rules.size());
+        for (const auto & rule : rules) {
+            ret.push_back(rule.data());
+        }
+        return ret;
+    }
+}

common/grammar-parser.h

@@ -0,0 +1,29 @@
+// Implements a parser for an extended Backus-Naur form (BNF), producing the
+// binary context-free grammar format specified by llama.h. Supports character
+// ranges, grouping, and repetition operators. As an example, a grammar for
+// arithmetic might look like:
+//
+// root  ::= expr
+// expr  ::= term ([-+*/] term)*
+// term  ::= num | "(" space expr ")" space
+// num   ::= [0-9]+ space
+// space ::= [ \t\n]*
+
+#pragma once
+#include "llama.h"
+#include <vector>
+#include <map>
+#include <cstdint>
+#include <string>
+
+namespace grammar_parser {
+    struct parse_state {
+        std::map<std::string, uint32_t>                 symbol_ids;
+        std::vector<std::vector<llama_grammar_element>> rules;
+
+        std::vector<const llama_grammar_element *> c_rules();
+    };
+
+    parse_state parse(const char * src);
+    void print_grammar(FILE * file, const parse_state & state);
+}

common/log.h

@@ -630,7 +630,7 @@ inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
     buf << "[ ";
 
     bool first = true;
-    for (const auto & token : tokens)
+    for (const auto &token : tokens)
     {
         if (!first) {
             buf << ", ";

common/ngram-cache.h

@@ -37,18 +37,11 @@ struct llama_ngram {
     }
 };
 
-struct llama_token_hash_function {
-    size_t operator()(const llama_token token) const {
-        // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
-        return token * 11400714819323198485llu;
-    }
-};
-
 struct llama_ngram_hash_function {
     size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
-        for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
+        size_t hash = 0;
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
         }
         return hash;
     }

common/sampling.cpp

@@ -1,450 +1,459 @@
+#define LLAMA_API_INTERNAL
 #include "sampling.h"
+#include <random>
 
-#include "common.h"
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
+    struct llama_sampling_context * result = new llama_sampling_context();
 
-#include <cmath>
-#include <unordered_map>
+    result->params  = params;
+    result->grammar = nullptr;
 
-// the ring buffer works similarly to std::deque, but with a fixed capacity
-// TODO: deduplicate with llama-impl.h
-template<typename T>
-struct ring_buffer {
-    ring_buffer(size_t cap) : capacity(cap), data(cap) {}
+    // if there is a grammar, parse it
+    if (!params.grammar.empty()) {
+        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
 
-    T & front() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[first];
-    }
-
-    const T & front() const {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[first];
-    }
-
-    T & back() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[pos];
-    }
-
-    const T & back() const {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        return data[pos];
-    }
-
-    void push_back(const T & value) {
-        if (sz == capacity) {
-            // advance the start when buffer is full
-            first = (first + 1) % capacity;
-        } else {
-            sz++;
-        }
-        data[pos] = value;
-        pos = (pos + 1) % capacity;
-    }
-
-    T pop_front() {
-        if (sz == 0) {
-            throw std::runtime_error("ring buffer is empty");
-        }
-        T value = data[first];
-        first = (first + 1) % capacity;
-        sz--;
-        return value;
-    }
-
-    const T & rat(size_t i) const {
-        if (i >= sz) {
-            throw std::runtime_error("ring buffer: index out of bounds");
-        }
-        return data[(first + sz - i - 1) % capacity];
-    }
-
-    std::vector<T> to_vector() const {
-        std::vector<T> result;
-        result.reserve(sz);
-        for (size_t i = 0; i < sz; i++) {
-            result.push_back(data[(first + i) % capacity]);
-        }
-        return result;
-    }
-
-    void clear() {
-        // here only reset the status of the buffer
-        sz = 0;
-        first = 0;
-        pos = 0;
-    }
-
-    bool empty() const {
-        return sz == 0;
-    }
-
-    size_t size() const {
-        return sz;
-    }
-
-    size_t capacity = 0;
-    size_t sz = 0;
-    size_t first = 0;
-    size_t pos = 0;
-    std::vector<T> data;
-};
-
-struct gpt_sampler {
-    gpt_sampler_params params;
-
-    struct llama_sampler * grmr;
-    struct llama_sampler * chain;
-
-    ring_buffer<llama_token> prev;
-
-    std::vector<llama_token_data> cur;
-
-    llama_token_data_array cur_p;
-
-    void set_logits(struct llama_context * ctx, int idx) {
-        const auto * logits = llama_get_logits_ith(ctx, idx);
-
-        const int n_vocab = llama_n_vocab(llama_get_model(ctx));
-
-        cur.resize(n_vocab);
-
-        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-            cur[token_id] = llama_token_data{token_id, logits[token_id], 0.0f};
+        // will be empty (default) if there are parse errors
+        if (result->parsed_grammar.rules.empty()) {
+            fprintf(stderr, "%s: failed to parse grammar\n", __func__);
+            delete result;
+            return nullptr;
         }
 
-        cur_p = { cur.data(), cur.size(), -1, false };
-    }
-};
+        // Ensure that there is a "root" node.
+        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
+            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
+            delete result;
+            return nullptr;
+        }
 
-std::string gpt_sampler_params::print() const {
+        std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
+
+        struct llama_grammar * grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), result->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        result->grammar = grammar;
+    }
+
+    result->prev.resize(params.n_prev);
+
+    result->n_valid = 0;
+
+    llama_sampling_set_rng_seed(result, params.seed);
+
+    return result;
+}
+
+void llama_sampling_free(struct llama_sampling_context * ctx) {
+    if (ctx->grammar != NULL) {
+        llama_grammar_free(ctx->grammar);
+    }
+
+    delete ctx;
+}
+
+void llama_sampling_reset(llama_sampling_context * ctx) {
+    if (ctx->grammar != NULL) {
+        llama_grammar_free(ctx->grammar);
+        ctx->grammar = NULL;
+    }
+
+    if (!ctx->parsed_grammar.rules.empty()) {
+        std::vector<const llama_grammar_element *> grammar_rules(ctx->parsed_grammar.c_rules());
+
+        struct llama_grammar * grammar = llama_grammar_init(
+                grammar_rules.data(),
+                grammar_rules.size(), ctx->parsed_grammar.symbol_ids.at("root"));
+        if (grammar == nullptr) {
+            throw std::runtime_error("Failed to initialize llama_grammar");
+        }
+        ctx->grammar = grammar;
+    }
+
+    std::fill(ctx->prev.begin(), ctx->prev.end(), 0);
+    ctx->cur.clear();
+    ctx->n_valid = 0;
+}
+
+void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
+    if (seed == LLAMA_DEFAULT_SEED) {
+        seed = std::random_device{}();
+    }
+    ctx->rng.seed(seed);
+}
+
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
+    if (dst->grammar) {
+        llama_grammar_free(dst->grammar);
+        dst->grammar = nullptr;
+    }
+
+    if (src->grammar) {
+        dst->grammar = llama_grammar_copy(src->grammar);
+    }
+
+    dst->prev = src->prev;
+}
+
+llama_token llama_sampling_last(llama_sampling_context * ctx) {
+    return ctx->prev.back();
+}
+
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n) {
+    const int size = ctx_sampling->prev.size();
+
+    n = std::min(n, size);
+
+    std::string result;
+
+    for (int i = size - n; i < size; i++) {
+        result += llama_token_to_piece(ctx_main, ctx_sampling->prev[i]);
+    }
+
+    return result;
+}
+
+std::string llama_sampling_print(const llama_sampling_params & params) {
     char result[1024];
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
             "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
-            penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
-            top_k, tfs_z, top_p, min_p, typ_p, temp,
-            mirostat, mirostat_eta, mirostat_tau);
+            params.penalty_last_n, params.penalty_repeat, params.penalty_freq, params.penalty_present,
+            params.top_k, params.tfs_z, params.top_p, params.min_p, params.typical_p, params.temp,
+            params.mirostat, params.mirostat_eta, params.mirostat_tau);
 
     return std::string(result);
 }
 
-struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params) {
-    llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
-
-    lparams.no_perf = params.no_perf;
-
-    auto * result = new gpt_sampler {
-        /* .params = */ params,
-        /* .grmr   = */ llama_sampler_init_grammar(model, params.grammar.c_str(), "root"),
-        /* .chain  = */ llama_sampler_chain_init(lparams),
-        /* .prev   = */ ring_buffer<llama_token>(std::max(32, params.n_prev)),
-        /* .cur    = */ {},
-        /* .cur_p  = */ {},
-    };
-
-    llama_sampler_chain_add(result->chain,
-            llama_sampler_init_logit_bias(
-                llama_n_vocab(model),
-                params.logit_bias.size(),
-                params.logit_bias.data()));
-
-    llama_sampler_chain_add(result->chain,
-            llama_sampler_init_penalties(
-                llama_n_vocab  (model),
-                llama_token_eos(model),
-                llama_token_nl (model),
-                params.penalty_last_n,
-                params.penalty_repeat,
-                params.penalty_freq,
-                params.penalty_present,
-                params.penalize_nl,
-                params.ignore_eos));
-
-    if (params.temp > 0.0f) {
-        if (params.mirostat == 0) {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case GPT_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
-                        break;
-                    case GPT_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TFS_Z:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case GPT_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+std::string llama_sampling_order_print(const llama_sampling_params & params) {
+    std::string result = "CFG -> Penalties ";
+    if (params.mirostat == 0) {
+        for (auto sampler_type : params.samplers_sequence) {
+            const auto sampler_type_name = llama_sampling_type_to_str(sampler_type);
+            if (!sampler_type_name.empty()) {
+                result += "-> " + sampler_type_name + " ";
             }
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-            llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
-        } else if (params.mirostat == 1) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
-        } else if (params.mirostat == 2) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
-        } else {
-            GGML_ASSERT(false && "unknown mirostat version");
         }
     } else {
-        llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
+        result += "-> mirostat ";
     }
 
     return result;
 }
 
-void gpt_sampler_free(struct gpt_sampler * gsmpl) {
-    if (gsmpl) {
-        llama_sampler_free(gsmpl->grmr);
-
-        llama_sampler_free(gsmpl->chain);
-
-        delete gsmpl;
-    }
-}
-
-void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar) {
-    if (accept_grammar) {
-        llama_sampler_accept(gsmpl->grmr, token);
-    }
-
-    llama_sampler_accept(gsmpl->chain, token);
-
-    gsmpl->prev.push_back(token);
-}
-
-void gpt_sampler_reset(struct gpt_sampler * gsmpl) {
-    llama_sampler_reset(gsmpl->grmr);
-
-    llama_sampler_reset(gsmpl->chain);
-}
-
-struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
-    return new gpt_sampler {
-        /* .params = */ gsmpl->params,
-        /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
-        /* .chain  = */ llama_sampler_clone(gsmpl->chain),
-        /* .prev   = */ gsmpl->prev,
-        /* .cur    = */ gsmpl->cur,
-        /* .cur_p  = */ gsmpl->cur_p,
-    };
-}
-
-void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl) {
-    // TODO: measure grammar performance
-
-    if (gsmpl) {
-        llama_perf_sampler_print(gsmpl->chain);
-    }
-    if (ctx) {
-        llama_perf_context_print(ctx);
-    }
-}
-
-llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
-    gsmpl->set_logits(ctx, idx);
-
-    auto & grmr  = gsmpl->grmr;
-    auto & chain = gsmpl->chain;
-    auto & cur_p = gsmpl->cur_p; // initialized by set_logits
-
-    if (grammar_first) {
-        llama_sampler_apply(grmr, &cur_p);
-    }
-
-    llama_sampler_apply(chain, &cur_p);
-
-    GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
-
-    const llama_token id = cur_p.data[cur_p.selected].id;
-
-    if (grammar_first) {
-        return id;
-    }
-
-    // check if it the sampled token fits the grammar
-    {
-        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
-
-        llama_sampler_apply(grmr, &single_token_data_array);
-
-        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-        if (is_valid) {
-            return id;
-        }
-    }
-
-    // resampling:
-    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
-    gsmpl->set_logits(ctx, idx);
-
-    llama_sampler_apply(grmr,  &cur_p);
-    llama_sampler_apply(chain, &cur_p);
-
-    GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
-
-    return cur_p.data[cur_p.selected].id;
-}
-
-uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
-    return llama_sampler_get_seed(gsmpl->chain);
-}
-
-// helpers
-
-llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {
-    return &gsmpl->cur_p;
-}
-
-llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
-    return gsmpl->prev.rat(0);
-}
-
-std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
-    std::string result = "\tlogits ";
-
-    for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
-        const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
-        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
-    }
-
-    return result;
-}
-
-std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main, int n) {
-    n = std::min(n, (int) gsmpl->prev.size());
-
-    if (n <= 0) {
-        return "";
-    }
-
-    std::string result;
-    result.reserve(8*n); // 8 is the average length of a token [citation needed], TODO: compute this from the vocab
-
-    for (int i = n - 1; i >= 0; i--) {
-        const llama_token id = gsmpl->prev.rat(i);
-
-        GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
-
-        result += llama_token_to_piece(ctx_main, id);
-    }
-
-    return result;
-}
-
-char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) {
-    switch (cnstr) {
-        case GPT_SAMPLER_TYPE_TOP_K:       return 'k';
-        case GPT_SAMPLER_TYPE_TFS_Z:       return 'f';
-        case GPT_SAMPLER_TYPE_TYPICAL_P:   return 'y';
-        case GPT_SAMPLER_TYPE_TOP_P:       return 'p';
-        case GPT_SAMPLER_TYPE_MIN_P:       return 'm';
-        case GPT_SAMPLER_TYPE_TEMPERATURE: return 't';
-        default : return '?';
-    }
-}
-
-std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) {
-    switch (cnstr) {
-        case GPT_SAMPLER_TYPE_TOP_K:       return "top_k";
-        case GPT_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
-        case GPT_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
-        case GPT_SAMPLER_TYPE_TOP_P:       return "top_p";
-        case GPT_SAMPLER_TYPE_MIN_P:       return "min_p";
-        case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature";
+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type) {
+    switch (sampler_type) {
+        case llama_sampler_type::TOP_K:       return "top_k";
+        case llama_sampler_type::TFS_Z:       return "tfs_z";
+        case llama_sampler_type::TYPICAL_P:   return "typical_p";
+        case llama_sampler_type::TOP_P:       return "top_p";
+        case llama_sampler_type::MIN_P:       return "min_p";
+        case llama_sampler_type::TEMPERATURE: return "temperature";
         default : return "";
     }
 }
 
-std::vector<gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
-    std::unordered_map<std::string, gpt_sampler_type> sampler_canonical_name_map {
-        { "top_k",       GPT_SAMPLER_TYPE_TOP_K },
-        { "top_p",       GPT_SAMPLER_TYPE_TOP_P },
-        { "typ_p",       GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "min_p",       GPT_SAMPLER_TYPE_MIN_P },
-        { "tfs_z",       GPT_SAMPLER_TYPE_TFS_Z },
-        { "temperature", GPT_SAMPLER_TYPE_TEMPERATURE },
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
+        {"top_k",       llama_sampler_type::TOP_K},
+        {"top_p",       llama_sampler_type::TOP_P},
+        {"typical_p",   llama_sampler_type::TYPICAL_P},
+        {"min_p",       llama_sampler_type::MIN_P},
+        {"tfs_z",       llama_sampler_type::TFS_Z},
+        {"temperature", llama_sampler_type::TEMPERATURE}
     };
 
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    std::unordered_map<std::string, gpt_sampler_type> sampler_alt_name_map {
-        { "top-k",       GPT_SAMPLER_TYPE_TOP_K },
-        { "top-p",       GPT_SAMPLER_TYPE_TOP_P },
-        { "nucleus",     GPT_SAMPLER_TYPE_TOP_P },
-        { "typical-p",   GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typical",     GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typ-p",       GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "typ",         GPT_SAMPLER_TYPE_TYPICAL_P },
-        { "min-p",       GPT_SAMPLER_TYPE_MIN_P },
-        { "tfs-z",       GPT_SAMPLER_TYPE_TFS_Z },
-        { "tfs",         GPT_SAMPLER_TYPE_TFS_Z },
-        { "temp",        GPT_SAMPLER_TYPE_TEMPERATURE },
+    std::unordered_map<std::string, llama_sampler_type> sampler_alt_name_map {
+        {"top-k",       llama_sampler_type::TOP_K},
+        {"top-p",       llama_sampler_type::TOP_P},
+        {"nucleus",     llama_sampler_type::TOP_P},
+        {"typical-p",   llama_sampler_type::TYPICAL_P},
+        {"typical",     llama_sampler_type::TYPICAL_P},
+        {"min-p",       llama_sampler_type::MIN_P},
+        {"tfs-z",       llama_sampler_type::TFS_Z},
+        {"tfs",         llama_sampler_type::TFS_Z},
+        {"temp",        llama_sampler_type::TEMPERATURE}
     };
 
-    std::vector<gpt_sampler_type> samplers;
-    samplers.reserve(names.size());
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names.size());
+    for (const auto & name : names)
+    {
+        auto sampler_item = sampler_canonical_name_map.find(name);
+        if (sampler_item != sampler_canonical_name_map.end())
+        {
+            sampler_types.push_back(sampler_item->second);
+        }
+        else
+        {
+            if (allow_alt_names)
+            {
+                sampler_item = sampler_alt_name_map.find(name);
+                if (sampler_item != sampler_alt_name_map.end())
+                {
+                    sampler_types.push_back(sampler_item->second);
+                }
+            }
+        }
+    }
+    return sampler_types;
+}
 
-    for (const auto & name : names) {
-        auto sampler = sampler_canonical_name_map.find(name);
-        if (sampler != sampler_canonical_name_map.end()) {
-            samplers.push_back(sampler->second);
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string) {
+    std::unordered_map<char, llama_sampler_type> sampler_name_map {
+        {'k', llama_sampler_type::TOP_K},
+        {'p', llama_sampler_type::TOP_P},
+        {'y', llama_sampler_type::TYPICAL_P},
+        {'m', llama_sampler_type::MIN_P},
+        {'f', llama_sampler_type::TFS_Z},
+        {'t', llama_sampler_type::TEMPERATURE}
+    };
+
+    std::vector<llama_sampler_type> sampler_types;
+    sampler_types.reserve(names_string.size());
+    for (const auto & c : names_string) {
+        const auto sampler_item = sampler_name_map.find(c);
+        if (sampler_item != sampler_name_map.end()) {
+            sampler_types.push_back(sampler_item->second);
+        }
+    }
+    return sampler_types;
+}
+
+// no reasons to expose this function in header
+static void sampler_queue(
+                   struct llama_context * ctx_main,
+            const llama_sampling_params & params,
+                 llama_token_data_array & cur_p,
+                                 size_t   min_keep) {
+    const float         temp              = params.temp;
+    const float         dynatemp_range    = params.dynatemp_range;
+    const float         dynatemp_exponent = params.dynatemp_exponent;
+    const int32_t       top_k             = params.top_k;
+    const float         top_p             = params.top_p;
+    const float         min_p             = params.min_p;
+    const float         tfs_z             = params.tfs_z;
+    const float         typical_p         = params.typical_p;
+    const std::vector<llama_sampler_type> & samplers_sequence = params.samplers_sequence;
+
+    for (auto sampler_type : samplers_sequence) {
+        switch (sampler_type) {
+            case llama_sampler_type::TOP_K    : llama_sample_top_k    (ctx_main, &cur_p, top_k,     min_keep); break;
+            case llama_sampler_type::TFS_Z    : llama_sample_tail_free(ctx_main, &cur_p, tfs_z,     min_keep); break;
+            case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
+            case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
+            case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
+            case llama_sampler_type::TEMPERATURE:
+                if (dynatemp_range > 0) {
+                    float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
+                    float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
+                    llama_sample_entropy(ctx_main, &cur_p, dynatemp_min, dynatemp_max, dynatemp_exponent);
+                } else {
+                    llama_sample_temp(ctx_main, &cur_p, temp);
+                }
+                break;
+            default : break;
+        }
+    }
+}
+
+static llama_token llama_sampling_sample_impl(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx,
+                  bool is_resampling) {
+    const llama_sampling_params & params = ctx_sampling->params;
+
+    const float   temp            = params.temp;
+    const int     mirostat        = params.mirostat;
+    const float   mirostat_tau    = params.mirostat_tau;
+    const float   mirostat_eta    = params.mirostat_eta;
+
+    std::vector<float> original_logits;
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, /* apply_grammar= */ is_resampling, &original_logits);
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
+        GGML_ASSERT(!original_logits.empty());
+    }
+    llama_token id = 0;
+    // Get a pointer to the logits
+    float * logits = llama_get_logits_ith(ctx_main, idx);
+
+    if (temp < 0.0) {
+        // greedy sampling, with probs
+        llama_sample_softmax(ctx_main, &cur_p);
+        id = cur_p.data[0].id;
+    } else if (temp == 0.0) {
+        // greedy sampling, no probs
+        id = llama_sample_token_greedy(ctx_main, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
+        } else if (mirostat == 2) {
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
         } else {
-            if (allow_alt_names) {
-                sampler = sampler_alt_name_map.find(name);
-                if (sampler != sampler_alt_name_map.end()) {
-                    samplers.push_back(sampler->second);
+            // temperature sampling
+            size_t min_keep = std::max(1, params.min_keep);
+
+            sampler_queue(ctx_main, params, cur_p, min_keep);
+
+            id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
+
+            //{
+            //    const int n_top = 10;
+            //    LOG("top %d candidates:\n", n_top);
+
+            //    for (int i = 0; i < n_top; i++) {
+            //        const llama_token id = cur_p.data[i].id;
+            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
+            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
+            //    }
+            //}
+
+            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+        }
+    }
+
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
+        // Create an array with a single token data element for the sampled id
+        llama_token_data single_token_data = {id, logits[id], 0.0f};
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
+
+        // Apply grammar constraints to the single token
+        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
+
+        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
+        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+
+        // If the token is not valid according to the grammar, perform resampling
+        if (!is_valid) {
+            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
+
+            // Restore logits from the copy
+            std::copy(original_logits.begin(), original_logits.end(), logits);
+
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ true);
+        }
+    }
+
+    ctx_sampling->n_valid = temp == 0.0f ? 0 : cur_p.size;
+
+    return id;
+}
+
+static llama_token_data_array llama_sampling_prepare_impl(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
+    const llama_sampling_params & params = ctx_sampling->params;
+
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
+
+    const int32_t penalty_last_n  = params.penalty_last_n < 0 ? params.n_prev : params.penalty_last_n;
+    const float   penalty_repeat  = params.penalty_repeat;
+    const float   penalty_freq    = params.penalty_freq;
+    const float   penalty_present = params.penalty_present;
+
+    const bool    penalize_nl     = params.penalize_nl;
+
+    auto & prev = ctx_sampling->prev;
+    auto & cur  = ctx_sampling->cur;
+
+    // Get a pointer to the logits
+    float * logits = llama_get_logits_ith(ctx_main, idx);
+
+    if (ctx_sampling->grammar != NULL && !apply_grammar) {
+        GGML_ASSERT(original_logits != NULL);
+        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
+        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+    }
+
+    // apply params.logit_bias map
+    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
+        logits[it->first] += it->second;
+    }
+
+    if (ctx_cfg) {
+        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
+        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
+    }
+
+    cur.clear();
+
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+    }
+
+    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
+
+    // apply penalties
+    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
+    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
+    if (penalty_tokens_used_size) {
+        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
+
+        llama_sample_repetition_penalties(ctx_main, &cur_p,
+                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
+                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
+
+        if (!penalize_nl) {
+            for (size_t idx = 0; idx < cur_p.size; idx++) {
+                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
+                    cur_p.data[idx].logit = nl_logit;
+                    break;
                 }
             }
         }
     }
 
-    return samplers;
-}
-
-std::vector<gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars) {
-    std::unordered_map<char, gpt_sampler_type> sampler_name_map = {
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_K),       GPT_SAMPLER_TYPE_TOP_K },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TFS_Z),       GPT_SAMPLER_TYPE_TFS_Z },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TYPICAL_P),   GPT_SAMPLER_TYPE_TYPICAL_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TOP_P),       GPT_SAMPLER_TYPE_TOP_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_MIN_P),       GPT_SAMPLER_TYPE_MIN_P },
-        { gpt_sampler_type_to_chr(GPT_SAMPLER_TYPE_TEMPERATURE), GPT_SAMPLER_TYPE_TEMPERATURE }
-    };
-
-    std::vector<gpt_sampler_type> samplers;
-    samplers.reserve(chars.size());
-
-    for (const auto & c : chars) {
-        const auto sampler = sampler_name_map.find(c);
-        if (sampler != sampler_name_map.end()) {
-            samplers.push_back(sampler->second);
-        }
+    // apply grammar checks before sampling logic
+    if (apply_grammar && ctx_sampling->grammar != NULL) {
+        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
-    return samplers;
+    return cur_p;
+}
+
+llama_token llama_sampling_sample(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx) {
+    // Call the implementation function with is_resampling set to false by default
+    return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, /* is_resampling= */ false);
+}
+
+llama_token_data_array llama_sampling_prepare(
+                  struct llama_sampling_context * ctx_sampling,
+                  struct llama_context * ctx_main,
+                  struct llama_context * ctx_cfg,
+                  const int idx,
+                  bool apply_grammar,
+                  std::vector<float> * original_logits) {
+    return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
+}
+
+void llama_sampling_accept(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar) {
+    ctx_sampling->prev.erase(ctx_sampling->prev.begin());
+    ctx_sampling->prev.push_back(id);
+
+    if (ctx_sampling->grammar != NULL && apply_grammar) {
+        llama_grammar_accept_token(ctx_main, ctx_sampling->grammar, id);
+    }
 }

common/sampling.h

@@ -2,82 +2,159 @@
 
 #include "llama.h"
 
-#include "common.h"
+#include "grammar-parser.h"
 
+#include <random>
 #include <string>
+#include <unordered_map>
 #include <vector>
 
-// gpt_sampler extends llama_sampler with additional functionality:
+// sampler types
+enum class llama_sampler_type : char {
+    TOP_K       = 'k',
+    TOP_P       = 'p',
+    MIN_P       = 'm',
+    TFS_Z       = 'f',
+    TYPICAL_P   = 'y',
+    TEMPERATURE = 't'
+};
+
+// sampling parameters
+typedef struct llama_sampling_params {
+    int32_t     n_prev                = 64;                 // number of previous tokens to remember
+    int32_t     n_probs               = 0;                  // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;                  // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t     top_k                 = 40;                 // <= 0 to use vocab size
+    float       top_p                 = 0.95f;              // 1.0 = disabled
+    float       min_p                 = 0.05f;              // 0.0 = disabled
+    float       tfs_z                 = 1.00f;              // 1.0 = disabled
+    float       typical_p             = 1.00f;              // 1.0 = disabled
+    float       temp                  = 0.80f;              // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float       dynatemp_range        = 0.00f;              // 0.0 = disabled
+    float       dynatemp_exponent     = 1.00f;              // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t     penalty_last_n        = 64;                 // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float       penalty_repeat        = 1.00f;              // 1.0 = disabled
+    float       penalty_freq          = 0.00f;              // 0.0 = disabled
+    float       penalty_present       = 0.00f;              // 0.0 = disabled
+    int32_t     mirostat              = 0;                  // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float       mirostat_tau          = 5.00f;              // target entropy
+    float       mirostat_eta          = 0.10f;              // learning rate
+    bool        penalize_nl           = false;              // consider newlines as a repeatable token
+    uint32_t    seed                  = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
+
+    std::vector<llama_sampler_type> samplers_sequence = {
+        llama_sampler_type::TOP_K,
+        llama_sampler_type::TFS_Z,
+        llama_sampler_type::TYPICAL_P,
+        llama_sampler_type::TOP_P,
+        llama_sampler_type::MIN_P,
+        llama_sampler_type::TEMPERATURE
+    };
+
+    std::string grammar;  // optional BNF-like grammar to constrain sampling
+
+    // Classifier-Free Guidance
+    // https://arxiv.org/abs/2306.17806
+    std::string cfg_negative_prompt; // string to help guidance
+    float       cfg_scale     = 1.f; // how strong is guidance
+
+    std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
+
+    std::vector<llama_token> penalty_prompt_tokens;
+    bool                     use_penalty_prompt_tokens = false;
+} llama_sampling_params;
+
+// general sampler context
+// TODO: move to llama.h
+struct llama_sampling_context {
+    // parameters that will be used for sampling
+    llama_sampling_params params;
+
+    // mirostat sampler state
+    float mirostat_mu;
+
+    llama_grammar * grammar;
+
+    // internal
+    grammar_parser::parse_state parsed_grammar;
+
+    // TODO: replace with ring-buffer
+    std::vector<llama_token>      prev;
+    std::vector<llama_token_data> cur;
+    size_t n_valid; // Number of correct top tokens with correct probabilities.
+
+    std::mt19937 rng;
+};
+
+#include "common.h"
+
+// Create a new sampling context instance.
+struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params);
+
+void llama_sampling_free(struct llama_sampling_context * ctx);
+
+// Reset the sampler context
+// - clear prev tokens
+// - reset grammar
+void llama_sampling_reset(llama_sampling_context * ctx);
+
+// Set the sampler seed
+void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
+
+// Copy the sampler context
+void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
+
+// Get the last sampled token
+llama_token llama_sampling_last(llama_sampling_context * ctx);
+
+// Get a string representation of the last sampled tokens
+std::string llama_sampling_prev_str(llama_sampling_context * ctx_sampling, llama_context * ctx_main, int n);
+
+// Print sampling parameters into a string
+std::string llama_sampling_print(const llama_sampling_params & params);
+
+// Print sampling order into a string
+std::string llama_sampling_order_print(const llama_sampling_params & params);
+
+std::string llama_sampling_type_to_str(llama_sampler_type sampler_type);
+
+std::vector<llama_sampler_type> llama_sampling_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<llama_sampler_type> llama_sampling_types_from_chars(const std::string & names_string);
+
+// this is a common sampling function used across the examples for convenience
+// it can serve as a starting point for implementing your own sampling function
+// Note: When using multiple sequences, it is the caller's responsibility to call
+//       llama_sampling_reset when a sequence ends
 //
-//  - grammar support
-//  - custom sampler logic based on the parameters
-//  - history of the last accepted tokens
-//  - performance metrics
+// required:
+//  - ctx_main:     context to use for sampling
+//  - ctx_sampling: sampling-specific context
 //
-// This goal is to have a common implementation of the sampling logic shared across the examples.
-// For example, depending on the temperature, the sampling chain can be very simple (greedy) or more
-// complex (top-k, top-p, etc).
+// optional:
+//  - ctx_cfg:      context to use for classifier-free guidance
+//  - idx:          sample from llama_get_logits_ith(ctx, idx)
 //
-// Another example is related to the grammar. In general, the grammar constraints applied on the full
-// vocabulary can be very taxing. To improve performance, the grammar can be applied only to the sampled
-// token in order to verify if it fits the grammar. And only if the token doesn't fit the grammar, the
-// grammar constraints are applied to the full vocabulary and the token is resampled.
-//
-// The gpt_sampler also maintains a container with the last accepted tokens. In the future, this can
-// be moved into the core llama library.
-//
-// For convenience, the gpt_sampler also maintains a container with the current candidate tokens.
-// This can be used to access the probabilities of the rest of the non-sampled tokens.
-//
-// TODO: measure grammar performance
+// returns:
+//  - token:      sampled token
+//  - candidates: vector of candidate tokens
 //
+llama_token llama_sampling_sample(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        struct llama_context * ctx_cfg,
+        int idx = -1);
 
-struct gpt_sampler;
+// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
+llama_token_data_array llama_sampling_prepare(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        struct llama_context * ctx_cfg,
+        int idx = 0,
+        bool apply_grammar = true,
+        std::vector<float> * original_logits = nullptr);
 
-// llama_sampler API overloads
-
-struct gpt_sampler * gpt_sampler_init(const struct llama_model * model, const struct gpt_sampler_params & params);
-
-void gpt_sampler_free(struct gpt_sampler * gsmpl);
-
-// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
-void                 gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar);
-void                 gpt_sampler_reset (struct gpt_sampler * gsmpl);
-struct gpt_sampler * gpt_sampler_clone (struct gpt_sampler * gsmpl);
-
-// arguments can be nullptr to skip printing
-void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl);
-
-// extended sampling implementation:
-//
-// - set logits
-// - apply the configured sampler chain
-// - check if the token fits the grammar (if any)
-// - if not: resample by first applying the grammar constraints and then sampling again (slower path)
-//
-// if grammar_first is true, the grammar is applied before the samplers (slower)
-// useful in cases where all the resulting candidates (not just the sampled one) must fit the grammar
-//
-llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
-
-uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl);
-
-// helpers
-
-// access the internal list of current candidate tokens
-llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl);
-
-// get the last accepted token
-llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl);
-
-// print the sampler chain into a string
-std::string gpt_sampler_print(const struct gpt_sampler * gsmpl);
-
-// get a string representation of the last accepted tokens
-std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx, int n);
-
-char        gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr);
-std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr);
-
-std::vector<enum gpt_sampler_type> gpt_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
-std::vector<enum gpt_sampler_type> gpt_sampler_types_from_chars(const std::string & chars);
+void llama_sampling_accept(
+        struct llama_sampling_context * ctx_sampling,
+        struct llama_context * ctx_main,
+        llama_token id,
+        bool apply_grammar);

convert_hf_to_gguf_update.py

@@ -31,7 +31,6 @@ import re
 import requests
 import sys
 import json
-import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -51,7 +50,7 @@ class TOKENIZER_TYPE(IntEnum):
 
 # TODO: this string has to exercise as much pre-tokenizer functionality as possible
 #       will be updated with time - contributions welcome
-CHK_TXT = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
+chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
 
 if len(sys.argv) == 2:
     token = sys.argv[1]
@@ -92,13 +91,6 @@ models = [
     {"name": "gemma-2",        "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/google/gemma-2-9b", },
     {"name": "jais",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/core42/jais-13b", },
     {"name": "t5",             "tokt": TOKENIZER_TYPE.UGM, "repo": "https://huggingface.co/google-t5/t5-small", },
-    {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
-    {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
-    {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
-    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
-    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
-    {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
-    {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
 ]
 
 
@@ -107,8 +99,8 @@ def download_file_with_auth(url, token, save_path):
     response = sess.get(url, headers=headers)
     response.raise_for_status()
     os.makedirs(os.path.dirname(save_path), exist_ok=True)
-    with open(save_path, 'wb') as downloaded_file:
-        downloaded_file.write(response.content)
+    with open(save_path, 'wb') as f:
+        f.write(response.content)
     logger.info(f"File {save_path} downloaded successfully")
 
 
@@ -127,27 +119,12 @@ def download_model(model):
     if tokt == TOKENIZER_TYPE.UGM:
         files.append("spiece.model")
 
-    if os.path.isdir(repo):
-        # If repo is a path on the file system, copy the directory
-        for file in files:
-            src_path = os.path.join(repo, file)
-            dst_path = f"models/tokenizers/{name}/{file}"
-            if os.path.isfile(dst_path):
-                logger.info(f"{name}: File {dst_path} already exists - skipping")
-                continue
-            if os.path.isfile(src_path):
-                shutil.copy2(src_path, dst_path)
-                logger.info(f"{name}: Copied {src_path} to {dst_path}")
-            else:
-                logger.warning(f"{name}: Source file {src_path} does not exist")
-    else:
-        # If repo is a URL, download the files
-        for file in files:
-            save_path = f"models/tokenizers/{name}/{file}"
-            if os.path.isfile(save_path):
-                logger.info(f"{name}: File {save_path} already exists - skipping")
-                continue
-            download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
+    for file in files:
+        save_path = f"models/tokenizers/{name}/{file}"
+        if os.path.isfile(save_path):
+            logger.info(f"{name}: File {save_path} already exists - skipping")
+            continue
+        download_file_with_auth(f"{repo}/resolve/main/{file}", token, save_path)
 
 
 for model in models:
@@ -182,7 +159,7 @@ for model in models:
         logger.error(f"Error loading tokenizer for model {name}. The model may not exist or is not accessible with the provided token. Error: {e}")
         continue  # Skip to the next model if the tokenizer can't be loaded
 
-    chktok = tokenizer.encode(CHK_TXT)
+    chktok = tokenizer.encode(chktxt)
     chkhsh = sha256(str(chktok).encode()).hexdigest()
 
     logger.info(f"model: {name}")
@@ -214,7 +191,7 @@ src_func = f"""
         # we will use this unique identifier to write a "tokenizer.ggml.pre" entry in the GGUF file which we can
         # use in llama.cpp to implement the same pre-tokenizer
 
-        chktxt = {repr(CHK_TXT)}
+        chktxt = {repr(chktxt)}
 
         chktok = tokenizer.encode(chktxt)
         chkhsh = sha256(str(chktok).encode()).hexdigest()
@@ -310,7 +287,7 @@ tests = [
     "333333333",
     "Cửa Việt", # llama-bpe fails on this
     " discards",
-    CHK_TXT,
+    chktxt,
 ]
 
 # write the tests to ./models/ggml-vocab-{name}.gguf.inp

convert_llama_ggml_to_gguf.py

@@ -116,7 +116,7 @@ class Tensor:
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
-        self.dtype= gguf.GGMLQuantizationType(dtype)
+        self.dtype= dtype
         self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])
@@ -132,10 +132,6 @@ class Tensor:
 
 
 class GGMLModel:
-
-    file_format: GGMLFormat
-    format_version: int
-
     def __init__(self):
         self.hyperparameters = None
         self.vocab = None
@@ -294,7 +290,7 @@ class GGMLToGGUF:
         if self.vocab_override is not None:
             vo = self.vocab_override
             logger.info('* Adding vocab item(s)')
-            for (_, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
+            for (idx, (vbytes, score, ttype)) in enumerate(vo.all_tokens()):
                 tokens.append(vbytes)
                 scores.append(score)
                 toktypes.append(ttype)
@@ -358,8 +354,7 @@ class GGMLToGGUF:
 
 
 def handle_metadata(cfg, hp):
-    import examples.convert_legacy_llama as convert
-
+    import convert
     assert cfg.model_metadata_dir.is_dir(), 'Metadata dir is not a directory'
     hf_config_path   = cfg.model_metadata_dir / "config.json"
     orig_config_path = cfg.model_metadata_dir / "params.json"

convert_lora_to_gguf.py

@@ -1,400 +0,0 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-
-from __future__ import annotations
-
-from dataclasses import dataclass
-import logging
-import argparse
-import os
-import sys
-import json
-from math import prod
-from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
-
-import torch
-
-if TYPE_CHECKING:
-    from torch import Tensor
-
-if 'NO_LOCAL_GGUF' not in os.environ:
-    sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
-import gguf
-
-# reuse model definitions from convert_hf_to_gguf.py
-from convert_hf_to_gguf import LazyTorchTensor, Model
-
-logger = logging.getLogger("lora-to-gguf")
-
-
-@dataclass
-class PartialLoraTensor:
-    A: Tensor | None = None
-    B: Tensor | None = None
-
-
-# magic to support tensor shape modifications and splitting
-class LoraTorchTensor:
-    _lora_A: Tensor  # (n_rank, row_size)
-    _lora_B: Tensor  # (col_size, n_rank)
-    _rank: int
-
-    def __init__(self, A: Tensor, B: Tensor):
-        assert len(A.shape) == len(B.shape)
-        assert A.shape[-2] == B.shape[-1]
-        if A.dtype != B.dtype:
-            A = A.to(torch.float32)
-            B = B.to(torch.float32)
-        self._lora_A = A
-        self._lora_B = B
-        self._rank = B.shape[-1]
-
-    def get_lora_A_B(self) -> tuple[Tensor, Tensor]:
-        return (self._lora_A, self._lora_B)
-
-    def __getitem__(
-        self,
-        indices: (
-            SupportsIndex
-            | slice
-            | tuple[SupportsIndex | slice | Tensor, ...]  # TODO: add ellipsis in the type signature
-        ),
-    ) -> LoraTorchTensor:
-        shape = self.shape
-        if isinstance(indices, SupportsIndex):
-            if len(shape) > 2:
-                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
-            else:
-                raise NotImplementedError  # can't return a vector
-        elif isinstance(indices, slice):
-            if len(shape) > 2:
-                return LoraTorchTensor(self._lora_A[indices], self._lora_B[indices])
-            else:
-                return LoraTorchTensor(self._lora_A, self._lora_B[indices])
-        elif isinstance(indices, tuple):
-            assert len(indices) > 0
-            if indices[-1] is Ellipsis:
-                return self[indices[:-1]]
-            # expand ellipsis
-            indices = tuple(
-                u
-                for v in (
-                    (
-                        (slice(None, None) for _ in range(len(indices) - 1))
-                        if i is Ellipsis
-                        else (i,)
-                    )
-                    for i in indices
-                )
-                for u in v
-            )
-
-            if len(indices) < len(shape):
-                indices = (*indices, *(slice(None, None) for _ in range(len(indices), len(shape))))
-
-            # TODO: make sure this is correct
-            indices_A = (
-                *(
-                    (
-                        j.__index__() % self._lora_A.shape[i]
-                        if isinstance(j, SupportsIndex)
-                        else slice(None, None)
-                    )
-                    for i, j in enumerate(indices[:-2])
-                ),
-                slice(None, None),
-                indices[-1],
-            )
-            indices_B = indices[:-1]
-            return LoraTorchTensor(self._lora_A[indices_A], self._lora_B[indices_B])
-        else:
-            raise NotImplementedError  # unknown indice type
-
-    @property
-    def dtype(self) -> torch.dtype:
-        assert self._lora_A.dtype == self._lora_B.dtype
-        return self._lora_A.dtype
-
-    @property
-    def shape(self) -> tuple[int, ...]:
-        assert len(self._lora_A.shape) == len(self._lora_B.shape)
-        return (*self._lora_B.shape[:-1], self._lora_A.shape[-1])
-
-    def size(self, dim=None):
-        assert dim is None
-        return self.shape
-
-    def reshape(self, *shape: int | tuple[int, ...]) -> LoraTorchTensor:
-        if isinstance(shape[0], tuple):
-            new_shape: tuple[int, ...] = shape[0]
-        else:
-            new_shape = cast(tuple[int, ...], shape)
-        orig_shape = self.shape
-        if len(new_shape) < 2:
-            raise NotImplementedError  # can't become a vector
-
-        # expand -1 in the shape
-        if any(dim == -1 for dim in new_shape):
-            n_elems = prod(orig_shape)
-            n_new_elems = prod(dim if dim != -1 else 1 for dim in new_shape)
-            assert n_elems % n_new_elems == 0
-            new_shape = (*(dim if dim != -1 else n_elems // n_new_elems for dim in new_shape),)
-
-        if new_shape[-1] != orig_shape[-1]:
-            raise NotImplementedError  # can't reshape the row size trivially
-
-        shape_A = (*(1 for _ in new_shape[:-2]), self._rank, orig_shape[-1])
-        shape_B = (*new_shape[:-1], self._rank)
-        return LoraTorchTensor(
-            self._lora_A.reshape(shape_A),
-            self._lora_B.reshape(shape_B),
-        )
-
-    def reshape_as(self, other: Tensor) -> LoraTorchTensor:
-        return self.reshape(*other.shape)
-
-    def view(self, *size: int) -> LoraTorchTensor:
-        return self.reshape(*size)
-
-    def permute(self, *dims: int) -> LoraTorchTensor:
-        shape = self.shape
-        dims = tuple(dim - len(shape) if dim >= 0 else dim for dim in dims)
-        if dims[-1] == -1:
-            # TODO: support higher dimensional A shapes bigger than 1
-            assert all(dim == 1 for dim in self._lora_A.shape[:-2])
-            return LoraTorchTensor(self._lora_A, self._lora_B.permute(*dims))
-        if len(shape) == 2 and dims[-1] == -2 and dims[-2] == -1:
-            return LoraTorchTensor(self._lora_B.permute(*dims), self._lora_A.permute(*dims))
-        else:
-            # TODO: compose the above two
-            raise NotImplementedError
-
-    def transpose(self, dim0: int, dim1: int) -> LoraTorchTensor:
-        shape = self.shape
-        dims = [i for i in range(len(shape))]
-        dims[dim0], dims[dim1] = dims[dim1], dims[dim0]
-        return self.permute(*dims)
-
-    def swapaxes(self, axis0: int, axis1: int) -> LoraTorchTensor:
-        return self.transpose(axis0, axis1)
-
-    def to(self, *args, **kwargs):
-        return LoraTorchTensor(self._lora_A.to(*args, **kwargs), self._lora_B.to(*args, **kwargs))
-
-    @classmethod
-    def __torch_function__(cls, func: Callable, types, args=(), kwargs=None):
-        del types  # unused
-
-        if kwargs is None:
-            kwargs = {}
-
-        if func is torch.permute:
-            return type(args[0]).permute(*args, **kwargs)
-        elif func is torch.reshape:
-            return type(args[0]).reshape(*args, **kwargs)
-        elif func is torch.stack:
-            assert isinstance(args[0], Sequence)
-            dim = kwargs.get("dim", 0)
-            assert dim == 0
-            return LoraTorchTensor(
-                torch.stack([a._lora_A for a in args[0]], dim),
-                torch.stack([b._lora_B for b in args[0]], dim),
-            )
-        elif func is torch.cat:
-            assert isinstance(args[0], Sequence)
-            dim = kwargs.get("dim", 0)
-            assert dim == 0
-            if len(args[0][0].shape) > 2:
-                return LoraTorchTensor(
-                    torch.cat([a._lora_A for a in args[0]], dim),
-                    torch.cat([b._lora_B for b in args[0]], dim),
-                )
-            elif all(torch.equal(args[0][0]._lora_A, t._lora_A) for t in args[0][1:]):
-                return LoraTorchTensor(
-                    args[0][0]._lora_A,
-                    torch.cat([b._lora_B for b in args[0]], dim),
-                )
-            else:
-                raise NotImplementedError
-        else:
-            raise NotImplementedError
-
-
-def get_base_tensor_name(lora_tensor_name: str) -> str:
-    base_name = lora_tensor_name.replace("base_model.model.", "")
-    base_name = base_name.replace(".lora_A.weight", ".weight")
-    base_name = base_name.replace(".lora_B.weight", ".weight")
-    return base_name
-
-
-def parse_args() -> argparse.Namespace:
-    parser = argparse.ArgumentParser(
-        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
-    parser.add_argument(
-        "--outfile", type=Path,
-        help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
-    )
-    parser.add_argument(
-        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "auto"], default="f16",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
-    )
-    parser.add_argument(
-        "--bigendian", action="store_true",
-        help="model is executed on big endian machine",
-    )
-    parser.add_argument(
-        "--no-lazy", action="store_true",
-        help="use more RAM by computing all outputs before writing (use in case lazy evaluation is broken)",
-    )
-    parser.add_argument(
-        "--verbose", action="store_true",
-        help="increase output verbosity",
-    )
-    parser.add_argument(
-        "--dry-run", action="store_true",
-        help="only print out what will be done, without writing any new files",
-    )
-    parser.add_argument(
-        "--base", type=Path, required=True,
-        help="directory containing base model file",
-    )
-    parser.add_argument(
-        "lora_path", type=Path,
-        help="directory containing LoRA adapter file",
-    )
-
-    return parser.parse_args()
-
-
-if __name__ == '__main__':
-    args = parse_args()
-    logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
-
-    ftype_map: dict[str, gguf.LlamaFileType] = {
-        "f32": gguf.LlamaFileType.ALL_F32,
-        "f16": gguf.LlamaFileType.MOSTLY_F16,
-        "bf16": gguf.LlamaFileType.MOSTLY_BF16,
-        "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
-        "auto": gguf.LlamaFileType.GUESSED,
-    }
-
-    ftype = ftype_map[args.outtype]
-
-    dir_base_model: Path = args.base
-    dir_lora: Path = args.lora_path
-    lora_config = dir_lora / "adapter_config.json"
-    input_model = dir_lora / "adapter_model.safetensors"
-
-    if args.outfile is not None:
-        fname_out = args.outfile
-    else:
-        # output in the same directory as the model by default
-        fname_out = dir_lora
-
-    if os.path.exists(input_model):
-        # lazy import load_file only if lora is in safetensors format.
-        from safetensors.torch import load_file
-
-        lora_model = load_file(input_model, device="cpu")
-    else:
-        input_model = os.path.join(dir_lora, "adapter_model.bin")
-        lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
-
-    # load base model
-    logger.info(f"Loading base model: {dir_base_model.name}")
-    hparams = Model.load_hparams(dir_base_model)
-    with torch.inference_mode():
-        try:
-            model_class = Model.from_model_architecture(hparams["architectures"][0])
-        except NotImplementedError:
-            logger.error(f"Model {hparams['architectures'][0]} is not supported")
-            sys.exit(1)
-
-        class LoraModel(model_class):
-            model_arch = model_class.model_arch
-
-            lora_alpha: float
-
-            def __init__(self, *args, dir_lora_model: Path, lora_alpha: float, **kwargs):
-
-                super().__init__(*args, **kwargs)
-
-                self.dir_model_card = dir_lora_model
-                self.lora_alpha = float(lora_alpha)
-
-            def set_type(self):
-                self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
-                self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
-
-            def set_gguf_parameters(self):
-                self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
-                super().set_gguf_parameters()
-
-            def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
-                tensor_map: dict[str, PartialLoraTensor] = {}
-
-                for name, tensor in lora_model.items():
-                    if self.lazy:
-                        tensor = LazyTorchTensor.from_eager(tensor)
-                    base_name = get_base_tensor_name(name)
-                    is_lora_a = ".lora_A.weight" in name
-                    is_lora_b = ".lora_B.weight" in name
-                    if not is_lora_a and not is_lora_b:
-                        if ".base_layer.weight" in name:
-                            continue
-                        logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
-                        sys.exit(1)
-
-                    if base_name in tensor_map:
-                        if is_lora_a:
-                            tensor_map[base_name].A = tensor
-                        else:
-                            tensor_map[base_name].B = tensor
-                    else:
-                        if is_lora_a:
-                            tensor_map[base_name] = PartialLoraTensor(A=tensor)
-                        else:
-                            tensor_map[base_name] = PartialLoraTensor(B=tensor)
-
-                for name, tensor in tensor_map.items():
-                    assert tensor.A is not None
-                    assert tensor.B is not None
-                    yield (name, cast(torch.Tensor, LoraTorchTensor(tensor.A, tensor.B)))
-
-            def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
-                dest = list(super().modify_tensors(data_torch, name, bid))
-                # some archs may have the same tensor for lm_head and output (tie word embeddings)
-                # in this case, adapters targeting lm_head will fail when using llama-export-lora
-                # therefore, we ignore them for now
-                # see: https://github.com/ggerganov/llama.cpp/issues/9065
-                if name == "lm_head.weight" and len(dest) == 0:
-                    raise ValueError("lm_head is present in adapter, but is ignored in base model")
-                for dest_name, dest_data in dest:
-                    assert isinstance(dest_data, LoraTorchTensor)
-                    lora_a, lora_b = dest_data.get_lora_A_B()
-
-                    yield (dest_name + ".lora_a", lora_a)
-                    yield (dest_name + ".lora_b", lora_b)
-
-        with open(lora_config, "r") as f:
-            lparams: dict[str, Any] = json.load(f)
-
-        alpha: float = lparams["lora_alpha"]
-
-        model_instance = LoraModel(
-            dir_base_model,
-            ftype,
-            fname_out,
-            is_big_endian=args.bigendian,
-            use_temp_file=False,
-            eager=args.no_lazy,
-            dry_run=args.dry_run,
-            dir_lora_model=dir_lora,
-            lora_alpha=alpha,
-            is_lora=True,
-        )
-
-        logger.info("Exporting model...")
-        model_instance.write()
-        logger.info(f"Model successfully exported to {model_instance.fname_out}")

docs/HOWTO-add-model.md

@@ -1,4 +1,4 @@
-# Add a new model architecture to `llama.cpp`
+## Add a new model architecture to `llama.cpp`
 
 Adding a model requires few steps:
 
@@ -9,15 +9,15 @@ Adding a model requires few steps:
 After following these steps, you can open PR.
 
 Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
-- [main](/examples/main/)
-- [imatrix](/examples/imatrix/)
-- [quantize](/examples/quantize/)
-- [server](/examples/server/)
+- [main](../examples/main)
+- [imatrix](../examples/imatrix)
+- [quantize](../examples/quantize)
+- [server](../examples/server)
 
 ### 1. Convert the model to GGUF
 
 This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
-Depending on the model architecture, you can use either [convert_hf_to_gguf.py](/convert_hf_to_gguf.py) or [examples/convert_legacy_llama.py](/examples/convert_legacy_llama.py) (for `llama/llama2` models in `.pth` format).
+Depending on the model architecture, you can use either [convert_hf_to_gguf.py](../convert_hf_to_gguf.py) or [examples/convert_legacy_llama.py](../examples/convert_legacy_llama.py) (for `llama/llama2` models in `.pth` format).
 
 The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
 
@@ -31,7 +31,7 @@ class MyModel(Model):
     model_arch = gguf.MODEL_ARCH.GROK
 ```
 
-2. Define the layout of the GGUF tensors in [constants.py](/gguf-py/gguf/constants.py)
+2. Define the layout of the GGUF tensors in [constants.py](../gguf-py/gguf/constants.py)
 
 Add an enum entry in `MODEL_ARCH`, the model human friendly name in `MODEL_ARCH_NAMES` and the GGUF tensor names in `MODEL_TENSORS`.
 
@@ -54,7 +54,7 @@ Example for `falcon` model:
 
 As a general rule, before adding a new tensor name to GGUF, be sure the equivalent naming does not already exist.
 
-Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](/gguf-py/gguf/tensor_mapping.py) file.
+Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](../gguf-py/gguf/tensor_mapping.py) file.
 
 If the tensor name is part of a repetitive layer/block, the key word `bid` substitutes it.
 
@@ -100,7 +100,7 @@ Have a look at existing implementation like `build_llama`, `build_dbrx` or `buil
 
 When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support for missing backend operations can be added in another PR.
 
-Note: to debug the inference graph: you can use [llama-eval-callback](/examples/eval-callback/).
+Note: to debug the inference graph: you can use [llama-eval-callback](../examples/eval-callback).
 
 ## GGUF specification
 

docs/android.md

@@ -1,56 +0,0 @@
-
-# Android
-
-## Build on Android using Termux
-[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
-
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
-
-[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
-
-## Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-```
-$ mkdir build-android
-$ cd build-android
-$ export NDK=<your_ndk_directory>
-$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-$ make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
-```
-
-Here's a demo of an interactive session running on Pixel 5 phone:
-
-https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4

docs/backend/CANN.md

@@ -1,259 +0,0 @@
-# llama.cpp for CANN
-
- - [Background](#background)
- - [News](#news)
- - [OS](#os)
- - [Hardware](#hardware)
- - [Model Supports](#model-supports)
- - [DataType Supports](#datatype-supports)
- - [Docker](#docker)
- - [Linux](#linux)
- - [TODO](#todo)
-
-
-## Background
-
-**Ascend NPU** is a range of AI processors using Neural Processing Unit. It will efficiently handle matrix-matrix multiplication, dot-product and scalars.
-
-**CANN** (Compute Architecture for Neural Networks) is a heterogeneous computing architecture for AI scenarios, providing support for multiple AI frameworks on the top and serving AI processors and programming at the bottom. It plays a crucial role in bridging the gap between upper and lower layers, and is a key platform for improving the computing efficiency of Ascend AI processors. Meanwhile, it offers a highly efficient and easy-to-use programming interface for diverse application scenarios, allowing users to rapidly build AI applications and services based on the Ascend platform.
-
-**Llama.cpp + CANN**
-
-The llama.cpp CANN backend is designed to support Ascend NPU. It utilize the ability of AscendC and ACLNN which are intergrated to CANN Toolkit and kernels to using Ascend NPU directly.
-
-## News
-
-- 2024.8
-  - Support `Q4_0` and `Q8_0` data type for Ascend NPU.
-- 2024.7
-  - Create CANN backend for Ascend NPU.
-
-## OS
-
-| OS      | Status  | Verified                                       |
-|:-------:|:-------:|:----------------------------------------------:|
-| Linux   | Support | Ubuntu 22.04, OpenEuler22.03                   |
-
-
-## Hardware
-
-### Ascend NPU
-
-**Verified devices**
-| Ascend NPU                    | Status  |
-|:-----------------------------:|:-------:|
-| Atlas 300T A2                 | Support |
-
-*Notes:*
-
-- If you have trouble with Ascend NPU device, please create a issue with **[CANN]** prefix/tag.
-- If you run successfully with your Ascend NPU device, please help update the upper table.
-
-
-## Model Supports
-
-| Model Name                  | FP16  | Q8_0 | Q4_0 |
-|:----------------------------|:-----:|:----:|:----:|
-| AquilaChat2-7B              |   √   |   √  |   √  |
-| Baichuan-7b                 |   √   |   √  |   √  |
-| Baichuan2-7B-Chat           |   √   |   √  |   √  |
-| bitnet_b1_58-large          |   √   |   √  |   √  |
-| bloom-560m                  |   √   |   x  |   √  |
-| bloomz-alpaca-560m          |   √   |   x  |   √  |
-| c4ai-command-r-35B-v01      |   x   |   x  |   x  |
-| chatglm3-6B                 |   x   |   x  |   x  |
-| chinese-alpaca-2-1.3b       |   √   |   √  |   √  |
-| CodeShell-7B                |   √   |   √  |   √  |
-| deepseek-ai_deepseek-coder-1.3B-base | x |   x  |   x  |
-| deepseek-ai_DeepSeek-V2-Lite | x   |   x  |   x   |
-| deepseek-coder-6.7B-instruct | x   |   x  |   x   |
-| DeepSeek-V2-Lite-64x1.5B    |   x   |   x  |   x  |
-| falcon-7b-instruct          |   √   |   √  |   √  |
-| flan-t5-large               |   √   |   √  |   √  |
-| gemma-2-9b-it               |   √   |   √  |   √  |
-| glm-4-9B                    |   x   |   x  |   x  |
-| gpt2                        |   √   |   √  |   √  |
-| Gpt2-163M                   |   √   |   √  |   √  |
-| granite-3B-code-instruct    |   √   |   √  |   √  |
-| GritLM-7B                   |   √   |   √  |   √  |
-| internlm2_5-7b-chat         |   √   |   √  |   √  |
-| koala-7B-HF                 |   √   |   √  |   √  |
-| Llama-2-7b-chat-hf          |   √   |   √  |   √  |
-| Llama-3-Smaug-8B            |   √   |   √  |   √  |
-| Llama2-Chinese-7b-Chat      |   √   |   √  |   √  |
-| Llama3-8B                   |   √   |   √  |   √  |
-| Llama3-8b-chinese           |   √   |   √  |   √  |
-| mamba-130m-hf               |   √   |   √  |   √  |
-| Mistral-7B-Instruct-v0.2    |   √   |   √  |   √  |
-| Mixtral-8x7B-Instruct-v0.1  |   x   |   √  |   √  |
-| mpt-7B                      |   √   |   √  |   √  |
-| OLMo-1B-hf                  |   √   |   √  |   √  |
-| OpenELM-3B-Instruct         |   √   |   √  |   √  |
-| Orion-14b-base              |   √   |   √  |   √  |
-| phi1                        |   x   |   x  |   x  |
-| phi2                        |   x   |   x  |   x  |
-| Phi-3-mini-4k-instruct      |   √   |   √  |   √  |
-| plamo-13b                   |   √   |   √  |   √  |
-| pythia-70M                  |   x   |   x  |   x  |
-| Qwen-7B                     |   √   |   √  |   √  |
-| Qwen2-1.5B-Instruct         |   √   |   x  |   √  |
-| Refact-1_6B-fim             |   √   |   √  |   √  |
-| SmolLM-135M                 |   √   |   √  |   √  |
-| stablelm-zephyr             |   x   |   x  |   x  |
-| stablelm-2-zephyr-1_6b      |   x   |   x  |   x  |
-| starcoderbase-1b            |   √   |   √  |   √  |
-| starcoder2-3b               |   √   |   √  |   √  |
-| vigogne-7b-chat             |   √   |   √  |   √  |
-| xverse-7b-chat              |   √   |   √  |   √  |
-| Yi-6b-Chat                  |   √   |   √  |   √  |
-
-
-
-## DataType Supports
-
-| DataType               | Status  |
-|:----------------------:|:-------:|
-| FP16                   | Support |
-| Q8_0                   | Support |
-| Q4_0                   | Support |
-
-## Docker
-
-### Build Images
-You can get a image with llama.cpp in one command.
-```sh
-docker build -t llama-cpp-cann -f .devops/llama-cli-cann.Dockerfile .
-```
-
-### Run container
-
-```sh
-# Find all cards.
-npu-smi info
-
-# Select the cards that you want to use, make sure these cards are not used by someone.
-# Following using cards of device0.
-docker run --name llamacpp --device /dev/davinci0  --device /dev/davinci_manager --device /dev/devmm_svm --device /dev/hisi_hdc -v /usr/local/dcmi:/usr/local/dcmi -v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi -v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ -v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info -v /PATH_TO_YOUR_MODELS/:/app/models -it llama-cpp-cann -m /app/models/MODEL_PATH -ngl 32 -p "Building a website can be done in 10 simple steps:"
-```
-
-*Notes:*
-
-- You may need to install Ascend Driver and firmware on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
-
-## Linux
-
-### I. Setup Environment
-
-1. **Install Ascend Driver and firmware**
-
-    ```sh
-    # create driver running user.
-    sudo groupadd -g HwHiAiUser
-    sudo useradd -g HwHiAiUser -d /home/HwHiAiUser -m HwHiAiUser -s /bin/bash
-    sudo usermod -aG HwHiAiUser $USER
-
-    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
-    # and install driver.
-    sudo sh Ascend-hdk-910b-npu-driver_x.x.x_linux-{arch}.run --full --install-for-all
-    ```
-
-    Once installed, run `npu-smi info` to check whether driver is installed successfully.
-    ```sh
-    +-------------------------------------------------------------------------------------------+
-    | npu-smi 24.1.rc2               Version: 24.1.rc2                                          |
-    +----------------------+---------------+----------------------------------------------------+
-    | NPU   Name           | Health        | Power(W)    Temp(C)           Hugepages-Usage(page)|
-    | Chip                 | Bus-Id        | AICore(%)   Memory-Usage(MB)  HBM-Usage(MB)        |
-    +======================+===============+====================================================+
-    | 2     xxx            | OK            | 64.4        51                15   / 15            |
-    | 0                    | 0000:01:00.0  | 0           1873 / 15077      0    / 32768         |
-    +======================+===============+====================================================+
-    | 5     xxx            | OK            | 64.0        52                15   / 15            |
-    | 0                    | 0000:81:00.0  | 0           1874 / 15077      0    / 32768         |
-    +======================+===============+====================================================+
-    | No running processes found in NPU 2                                                       |
-    +======================+===============+====================================================+
-    | No running processes found in NPU 5                                                       |
-    +======================+===============+====================================================+
-    ```
-
-2. **Install Ascend Firmware**
-    ```sh
-    # download driver from https://www.hiascend.com/hardware/firmware-drivers/community according to your system
-    # and install driver.
-    sudo sh Ascend-hdk-910b-npu-firmware_x.x.x.x.X.run --full
-    ```
-    If the following messaage appers, firmware is installed successfully.
-    ```sh
-    Firmware package installed successfully!
-    ```
-
-
-3. **Install CANN toolkit and kernels**
-
-    CANN toolkit and kernels can be obtained from the official [CANN Toolkit](https://www.hiascend.com/zh/developer/download/community/result?module=cann) page.
-
-    Please download the corresponding version that satified your system. The minimum version required is 8.0.RC2.alpha002 and here is the install command.
-    ```sh
-    pip3 install attrs numpy decorator sympy cffi pyyaml pathlib2 psutil protobuf scipy requests absl-py wheel typing_extensions
-    sh Ascend-cann-toolkit_8.0.RC2.alpha002_linux-aarch64.run --install
-    sh Ascend-cann-kernels-910b_8.0.RC2.alpha002_linux.run --install
-    ```
-
-    Set Ascend Variables:
-    ```sh
-    echo "source ~/Ascend/ascend-toolkit/set_env.sh" >> ~/.bashrc
-    source ~/.bashrc
-    ```
-
-Upon a successful installation, CANN is enabled for the available ascend devices.
-
-### II. Build llama.cpp
-
-```sh
-cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
-cmake --build build --config release
-```
-
-### III. Run the inference
-
-1. **Retrieve and prepare model**
-
-    You can refer to the general [*Prepare and Quantize*](../../README.md#prepare-and-quantize) guide for model prepration.
-
-    **Notes**:
-
-      - CANN backend only supports FP16/Q4_0/Q8_0 models currently.
-
-2. **Launch inference**
-
-    There are two device selection modes:
-
-    - Single device: Use one device target specified by the user.
-    - Multiple devices: Automatically choose the devices with the same backend.
-
-    | Device selection | Parameter                              |
-    |:----------------:|:--------------------------------------:|
-    | Single device    | --split-mode none --main-gpu DEVICE_ID |
-    | Multiple devices | --split-mode layer (default)           |
-
-    Examples:
-
-    - Use device 0:
-
-    ```sh
-    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
-    ```
-
-    - Use multiple devices:
-
-    ```sh
-    ./build/bin/llama-cli -m path_to_model -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
-    ```
-
-### **GitHub contribution**:
-Please add the **[CANN]** prefix/tag in issues/PRs titles to help the CANN-team check/address them without delay.
-
-
-## TODO
-- Support more models and data types.

docs/build.md

@@ -1,388 +0,0 @@
-# Build llama.cpp locally
-
-**To get the Code:**
-
-```bash
-git clone https://github.com/ggerganov/llama.cpp
-cd llama.cpp
-```
-
-In order to build llama.cpp you have four different options.
-
-- Using `make`:
-  - On Linux or MacOS:
-
-      ```bash
-      make
-      ```
-
-  - On Windows (x86/x64 only, arm64 requires cmake):
-
-    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-    2. Extract `w64devkit` on your pc.
-    3. Run `w64devkit.exe`.
-    4. Use the `cd` command to reach the `llama.cpp` folder.
-    5. From here you can run:
-        ```bash
-        make
-        ```
-
-  - Notes:
-    - For `Q4_0_4_4` quantization type build, add the `GGML_NO_LLAMAFILE=1` flag. For example, use `make GGML_NO_LLAMAFILE=1`.
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `make -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, run `make LLAMA_DEBUG=1`
-
-- Using `CMake`:
-
-  ```bash
-  cmake -B build
-  cmake --build build --config Release
-  ```
-
-  **Notes**:
-
-    - For `Q4_0_4_4` quantization type build, add the `-DGGML_LLAMAFILE=OFF` cmake option. For example, use `cmake -B build -DGGML_LLAMAFILE=OFF`.
-    - For faster compilation, add the `-j` argument to run multiple jobs in parallel. For example, `cmake --build build --config Release -j 8` will run 8 jobs in parallel.
-    - For faster repeated compilation, install [ccache](https://ccache.dev/).
-    - For debug builds, there are two cases:
-
-      1. Single-config generators (e.g. default = `Unix Makefiles`; note that they just ignore the `--config` flag):
-
-      ```bash
-      cmake -B build -DCMAKE_BUILD_TYPE=Debug
-      cmake --build build
-      ```
-
-      2. Multi-config generators (`-G` param set to Visual Studio, XCode...):
-
-      ```bash
-      cmake -B build -G "Xcode"
-      cmake --build build --config Debug
-      ```
-    - Building for Windows (x86, x64 and arm64) with MSVC or clang as compilers:
-      - Install Visual Studio 2022, e.g. via the [Community Edition](https://visualstudio.microsoft.com/de/vs/community/). In the installer, select at least the following options (this also automatically installs the required additional tools like CMake,...):
-        - Tab Workload: Desktop-development with C++
-        - Tab Components (select quickly via search): C++-_CMake_ Tools for Windows, _Git_ for Windows, C++-_Clang_ Compiler for Windows, MS-Build Support for LLVM-Toolset (clang)
-      - Please remember to always use a Developer Command Prompt / PowerShell for VS2022 for git, build, test
-      - For Windows on ARM (arm64, WoA) build with:
-        ```bash
-        cmake --preset arm64-windows-llvm-release -D GGML_OPENMP=OFF
-        cmake --build build-arm64-windows-llvm-release
-        ```
-        Note: Building for arm64 could also be done just with MSVC (with the build-arm64-windows-MSVC preset, or the standard CMake build instructions). But MSVC does not support inline ARM assembly-code, used e.g. for the accelerated Q4_0_4_8 CPU kernels.
-
--   Using `gmake` (FreeBSD):
-
-    1. Install and activate [DRM in FreeBSD](https://wiki.freebsd.org/Graphics)
-    2. Add your user to **video** group
-    3. Install compilation dependencies.
-
-        ```bash
-        sudo pkg install gmake automake autoconf pkgconf llvm15 openblas
-
-        gmake CC=/usr/local/bin/clang15 CXX=/usr/local/bin/clang++15 -j4
-        ```
-
-## Metal Build
-
-On MacOS, Metal is enabled by default. Using Metal makes the computation run on the GPU.
-To disable the Metal build at compile time use the `GGML_NO_METAL=1` flag or the `GGML_METAL=OFF` cmake option.
-
-When built with Metal support, you can explicitly disable GPU inference with the `--n-gpu-layers|-ngl 0` command-line
-argument.
-
-## BLAS Build
-
-Building the program with BLAS support may lead to some performance improvements in prompt processing using batch sizes higher than 32 (the default is 512). Support with CPU-only BLAS implementations doesn't affect the normal generation performance. We may see generation performance improvements with GPU-involved BLAS implementations, e.g. cuBLAS, hipBLAS. There are currently several different BLAS implementations available for build and use:
-
-### Accelerate Framework:
-
-This is only available on Mac PCs and it's enabled by default. You can just build using the normal instructions.
-
-### OpenBLAS:
-
-This provides BLAS acceleration using only the CPU. Make sure to have OpenBLAS installed on your machine.
-
-- Using `make`:
-  - On Linux:
-    ```bash
-    make GGML_OPENBLAS=1
-    ```
-
-  - On Windows:
-
-    1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
-    2. Download the latest version of [OpenBLAS for Windows](https://github.com/xianyi/OpenBLAS/releases).
-    3. Extract `w64devkit` on your pc.
-    4. From the OpenBLAS zip that you just downloaded copy `libopenblas.a`, located inside the `lib` folder, inside `w64devkit\x86_64-w64-mingw32\lib`.
-    5. From the same OpenBLAS zip copy the content of the `include` folder inside `w64devkit\x86_64-w64-mingw32\include`.
-    6. Run `w64devkit.exe`.
-    7. Use the `cd` command to reach the `llama.cpp` folder.
-    8. From here you can run:
-
-        ```bash
-        make GGML_OPENBLAS=1
-        ```
-
-- Using `CMake` on Linux:
-
-    ```bash
-    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=OpenBLAS
-    cmake --build build --config Release
-    ```
-
-### BLIS
-
-Check [BLIS.md](./backend/BLIS.md) for more information.
-
-### SYCL
-
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators.
-
-llama.cpp based on SYCL is used to **support Intel GPU** (Data Center Max series, Flex series, Arc series, Built-in GPU and iGPU).
-
-For detailed info, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
-
-### Intel oneMKL
-
-Building through oneAPI compilers will make avx_vnni instruction set available for intel processors that do not support avx512 and avx512_vnni. Please note that this build config **does not support Intel GPU**. For Intel GPU support, please refer to [llama.cpp for SYCL](./backend/SYCL.md).
-
-- Using manual oneAPI installation:
-  By default, `GGML_BLAS_VENDOR` is set to `Generic`, so if you already sourced intel environment script and assign `-DGGML_BLAS=ON` in cmake, the mkl version of Blas will automatically been selected. Otherwise please install oneAPI and follow the below steps:
-    ```bash
-    source /opt/intel/oneapi/setvars.sh # You can skip this step if  in oneapi-basekit docker image, only required for manual installation
-    cmake -B build -DGGML_BLAS=ON -DGGML_BLAS_VENDOR=Intel10_64lp -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_NATIVE=ON
-    cmake --build build --config Release
-    ```
-
-- Using oneAPI docker image:
-  If you do not want to source the environment vars and install oneAPI manually, you can also build the code using intel docker container: [oneAPI-basekit](https://hub.docker.com/r/intel/oneapi-basekit). Then, you can use the commands given above.
-
-Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
-
-### CUDA
-
-This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
-
-For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
-
-- Using `make`:
-  ```bash
-  make GGML_CUDA=1
-  ```
-- Using `CMake`:
-
-  ```bash
-  cmake -B build -DGGML_CUDA=ON
-  cmake --build build --config Release
-  ```
-
-The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used.
-
-The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted. In Windows this setting is available in the NVIDIA control panel as `System Memory Fallback`.
-
-The following compilation options are also available to tweak performance:
-
-| Option                        | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
-|-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-| GGML_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
-| GGML_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
-| GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, RDNA3). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower.                       |
-| GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models                                                                                                                                                                                       |
-| GGML_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
-| GGML_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
-| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
-| GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                  |
-
-### MUSA
-
-- Using `make`:
-  ```bash
-  make GGML_MUSA=1
-  ```
-- Using `CMake`:
-
-  ```bash
-  cmake -B build -DGGML_MUSA=ON
-  cmake --build build --config Release
-  ```
-
-### hipBLAS
-
-This provides BLAS acceleration on HIP-supported AMD GPUs.
-Make sure to have ROCm installed.
-You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
-
-- Using `make`:
-  ```bash
-  make GGML_HIPBLAS=1
-  ```
-- Using `CMake` for Linux (assuming a gfx1030-compatible AMD GPU):
-  ```bash
-  HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -R)" \
-      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-      && cmake --build build --config Release -- -j 16
-  ```
-  On Linux it is also possible to use unified memory architecture (UMA) to share main memory between the CPU and integrated GPU by setting `-DGGML_HIP_UMA=ON`.
-  However, this hurts performance for non-integrated GPUs (but enables working with integrated GPUs).
-
-  Note that if you get the following error:
-  ```
-  clang: error: cannot find ROCm device library; provide its path via '--rocm-path' or '--rocm-device-lib-path', or pass '-nogpulib' to build without ROCm device library
-  ```
-  Try searching for a directory under `HIP_PATH` that contains the file
-  `oclc_abi_version_400.bc`. Then, add the following to the start of the
-  command: `HIP_DEVICE_LIB_PATH=<directory-you-just-found>`, so something
-  like:
-  ```bash
-  HIPCXX="$(hipconfig -l)/clang" HIP_PATH="$(hipconfig -p)" \
-  HIP_DEVICE_LIB_PATH=<directory-you-just-found> \
-      cmake -S . -B build -DGGML_HIPBLAS=ON -DAMDGPU_TARGETS=gfx1030 -DCMAKE_BUILD_TYPE=Release \
-      && cmake --build build -- -j 16
-  ```
-
-- Using `make` (example for target gfx1030, build with 16 CPU threads):
-  ```bash
-  make -j16 GGML_HIPBLAS=1 GGML_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
-  ```
-
-- Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
-  ```bash
-  set PATH=%HIP_PATH%\bin;%PATH%
-  cmake -S . -B build -G Ninja -DAMDGPU_TARGETS=gfx1100 -DGGML_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release
-  cmake --build build
-  ```
-  Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
-  Find your gpu version string by matching the most significant version information from `rocminfo | grep gfx | head -1 | awk '{print $2}'` with the list of processors, e.g. `gfx1035` maps to `gfx1030`.
-
-
-The environment variable [`HIP_VISIBLE_DEVICES`](https://rocm.docs.amd.com/en/latest/understand/gpu_isolation.html#hip-visible-devices) can be used to specify which GPU(s) will be used.
-If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
-The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
-
-| Option                 | Legal values           | Default | Description                                                                                                                                                                                                                                    |
-|------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
-| GGML_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-| GGML_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
-| GGML_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
-
-### Vulkan
-
-**Windows**
-
-#### w64devkit
-
-Download and extract [w64devkit](https://github.com/skeeto/w64devkit/releases).
-
-Download and install the [Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows). When selecting components, only the Vulkan SDK Core is required.
-
-Launch `w64devkit.exe` and run the following commands to copy Vulkan dependencies:
-```sh
-SDK_VERSION=1.3.283.0
-cp /VulkanSDK/$SDK_VERSION/Bin/glslc.exe $W64DEVKIT_HOME/bin/
-cp /VulkanSDK/$SDK_VERSION/Lib/vulkan-1.lib $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/
-cp -r /VulkanSDK/$SDK_VERSION/Include/* $W64DEVKIT_HOME/x86_64-w64-mingw32/include/
-cat > $W64DEVKIT_HOME/x86_64-w64-mingw32/lib/pkgconfig/vulkan.pc <<EOF
-Name: Vulkan-Loader
-Description: Vulkan Loader
-Version: $SDK_VERSION
-Libs: -lvulkan-1
-EOF
-
-```
-Switch into the `llama.cpp` directory and run `make GGML_VULKAN=1`.
-
-#### MSYS2
-Install [MSYS2](https://www.msys2.org/) and then run the following commands in a UCRT terminal to install dependencies.
-  ```sh
-  pacman -S git \
-      mingw-w64-ucrt-x86_64-gcc \
-      mingw-w64-ucrt-x86_64-cmake \
-      mingw-w64-ucrt-x86_64-vulkan-devel \
-      mingw-w64-ucrt-x86_64-shaderc
-  ```
-Switch into `llama.cpp` directory and build using CMake.
-```sh
-cmake -B build -DGGML_VULKAN=ON
-cmake --build build --config Release
-```
-
-**With docker**:
-
-You don't need to install Vulkan SDK. It will be installed inside the container.
-
-```sh
-# Build the image
-docker build -t llama-cpp-vulkan -f .devops/llama-cli-vulkan.Dockerfile .
-
-# Then, use it:
-docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-vulkan -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
-```
-
-**Without docker**:
-
-Firstly, you need to make sure you have installed [Vulkan SDK](https://vulkan.lunarg.com/doc/view/latest/linux/getting_started_ubuntu.html)
-
-For example, on Ubuntu 22.04 (jammy), use the command below:
-
-```bash
-wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key add -
-wget -qO /etc/apt/sources.list.d/lunarg-vulkan-jammy.list https://packages.lunarg.com/vulkan/lunarg-vulkan-jammy.list
-apt update -y
-apt-get install -y vulkan-sdk
-# To verify the installation, use the command below:
-vulkaninfo
-```
-
-Alternatively your package manager might be able to provide the appropriate libraries.
-For example for Ubuntu 22.04 you can install `libvulkan-dev` instead.
-For Fedora 40, you can install `vulkan-devel`, `glslc` and `glslang` packages.
-
-Then, build llama.cpp using the cmake command below:
-
-```bash
-cmake -B build -DGGML_VULKAN=1
-cmake --build build --config Release
-# Test the output binary (with "-ngl 33" to offload all layers to GPU)
-./bin/llama-cli -m "PATH_TO_MODEL" -p "Hi you how are you" -n 50 -e -ngl 33 -t 4
-
-# You should see in the output, ggml_vulkan detected your GPU. For example:
-# ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
-```
-
-### CANN
-This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
-
-For more information about Ascend NPU in [Ascend Community](https://www.hiascend.com/en/).
-
-Make sure to have the CANN toolkit installed. You can download it from here: [CANN Toolkit](https://www.hiascend.com/developer/download/community/result?module=cann)
-
-Go to `llama.cpp` directory and build using CMake.
-```bash
-cmake -B build -DGGML_CANN=on -DCMAKE_BUILD_TYPE=release
-cmake --build build --config release
-```
-
-You can test with:
-
-`./build/llama-cli -m PATH_TO_MODEL -p "Building a website can be done in 10 steps:" -ngl 32`
-
-If the fllowing info is output on screen, you are using `llama.cpp by CANN backend`:
-```bash
-llm_load_tensors:       CANN buffer size = 13313.00 MiB
-llama_new_context_with_model:       CANN compute buffer size =  1260.81 MiB
-```
-
-For detailed info, such as model/device supports, CANN install, please refer to [llama.cpp for CANN](./backend/CANN.md).
-
-### Android
-
-To read documentation for how to build on Android, [click here](./android.md)
-
-### Arm CPU optimized mulmat kernels
-
-Llama.cpp includes a set of optimized mulmat kernels for the Arm architecture, leveraging Arm® Neon™, int8mm and SVE instructions. These kernels are enabled at build time through the appropriate compiler cpu-type flags, such as `-DCMAKE_C_FLAGS=-march=armv8.2a+i8mm+sve`. Note that these optimized kernels require the model to be quantized into one of the formats: `Q4_0_4_4` (Arm Neon), `Q4_0_4_8` (int8mm) or `Q4_0_8_8` (SVE). The SVE mulmat kernel specifically requires a vector width of 256 bits. When running on devices with a different vector width, it is recommended to use the `Q4_0_4_8` (int8mm) or `Q4_0_4_4` (Arm Neon) formats for better performance. Refer to [examples/quantize/README.md](../examples/quantize/README.md) for more information on the quantization formats.
-
-To support `Q4_0_4_4`, you must build with `GGML_NO_LLAMAFILE=1` (`make`) or `-DGGML_LLAMAFILE=OFF` (`cmake`).

docs/docker.md

@@ -1,86 +0,0 @@
-# Docker
-
-## Prerequisites
-* Docker must be installed and running on your system.
-* Create a folder to store big models & intermediate files (ex. /llama/models)
-
-## Images
-We have three Docker images available for this project:
-
-1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
-2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
-
-Additionally, there the following images, similar to the above:
-
-- `ghcr.io/ggerganov/llama.cpp:full-cuda`: Same as `full` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:light-cuda`: Same as `light` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:server-cuda`: Same as `server` but compiled with CUDA support. (platforms: `linux/amd64`)
-- `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-- `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
-
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
-
-## Usage
-
-The easiest way to download the models, convert them to ggml and optimize them is with the --all-in-one command which includes the full docker image.
-
-Replace `/path/to/models` below with the actual path where you downloaded the models.
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --all-in-one "/models/" 7B
-```
-
-On completion, you are ready to play!
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:full --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a light image:
-
-```bash
-docker run -v /path/to/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512
-```
-
-or with a server image:
-
-```bash
-docker run -v /path/to/models:/models -p 8000:8000 ghcr.io/ggerganov/llama.cpp:server -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512
-```
-
-## Docker With CUDA
-
-Assuming one has the [nvidia-container-toolkit](https://github.com/NVIDIA/nvidia-container-toolkit) properly installed on Linux, or is using a GPU enabled cloud, `cuBLAS` should be accessible inside the container.
-
-## Building Docker locally
-
-```bash
-docker build -t local/llama.cpp:full-cuda -f .devops/full-cuda.Dockerfile .
-docker build -t local/llama.cpp:light-cuda -f .devops/llama-cli-cuda.Dockerfile .
-docker build -t local/llama.cpp:server-cuda -f .devops/llama-server-cuda.Dockerfile .
-```
-
-You may want to pass in some different `ARGS`, depending on the CUDA environment supported by your container host, as well as the GPU architecture.
-
-The defaults are:
-
-- `CUDA_VERSION` set to `12.6.0`
-- `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
-
-The resulting images, are essentially the same as the non-CUDA images:
-
-1. `local/llama.cpp:full-cuda`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
-2. `local/llama.cpp:light-cuda`: This image only includes the main executable file.
-3. `local/llama.cpp:server-cuda`: This image only includes the server executable file.
-
-## Usage
-
-After building locally, Usage is similar to the non-CUDA examples, but you'll need to add the `--gpus` flag. You will also want to use the `--n-gpu-layers` flag.
-
-```bash
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
-docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
-```

docs/install.md

@@ -1,39 +0,0 @@
-# Install pre-built version of llama.cpp
-
-## Homebrew
-
-On Mac and Linux, the homebrew package manager can be used via
-
-```sh
-brew install llama.cpp
-```
-The formula is automatically updated with new `llama.cpp` releases. More info: https://github.com/ggerganov/llama.cpp/discussions/7668
-
-## Nix
-
-On Mac and Linux, the Nix package manager can be used via
-
-```sh
-nix profile install nixpkgs#llama-cpp
-```
-For flake enabled installs.
-
-Or
-
-```sh
-nix-env --file '<nixpkgs>' --install --attr llama-cpp
-```
-
-For non-flake enabled installs.
-
-This expression is automatically updated within the [nixpkgs repo](https://github.com/NixOS/nixpkgs/blob/nixos-24.05/pkgs/by-name/ll/llama-cpp/package.nix#L164).
-
-## Flox
-
-On Mac and Linux, Flox can be used to install llama.cpp within a Flox environment via
-
-```sh
-flox install llama-cpp
-```
-
-Flox follows the nixpkgs build of llama.cpp.

docs/token_generation_performance_tips.md

@@ -1,7 +1,7 @@
 # Token generation performance troubleshooting
 
 ## Verifying that the model is running on the GPU with CUDA
-Make sure you compiled llama with the correct env variables according to [this guide](/docs/build.md#cuda), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
+Make sure you compiled llama with the correct env variables according to [this guide](../README.md#CUDA), so that llama accepts the `-ngl N` (or `--n-gpu-layers N`) flag. When running llama, you may configure `N` to be very large, and llama will offload the maximum possible number of layers to the GPU, even if it's less than the number you configured. For example:
 ```shell
 ./llama-cli -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
 ```

examples/CMakeLists.txt

@@ -21,8 +21,8 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)
     add_subdirectory(export-lora)
+    add_subdirectory(finetune)
     add_subdirectory(gbnf-validator)
-    add_subdirectory(gguf-hash)
     add_subdirectory(gguf-split)
     add_subdirectory(gguf)
     add_subdirectory(gritlm)
@@ -52,4 +52,5 @@ else()
     add_subdirectory(simple)
     add_subdirectory(speculative)
     add_subdirectory(tokenize)
+    add_subdirectory(train-text-from-scratch)
 endif()

examples/baby-llama/baby-llama.cpp

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "train.h"
 
+#include <vector>
 #include <cassert>
 #include <cstdlib>
 #include <cstring>
@@ -18,7 +19,7 @@ constexpr float rms_norm_eps = 5e-6f;
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);

examples/batched-bench/README.md

@@ -49,12 +49,3 @@ There are 2 modes of operation:
 |   128 |    256 |    8 |   3072 |    0.751 |  1363.92 |   15.110 |   135.54 |   15.861 |   193.69 |
 |   128 |    256 |   16 |   6144 |    1.569 |  1304.93 |   18.073 |   226.64 |   19.642 |   312.80 |
 |   128 |    256 |   32 |  12288 |    3.409 |  1201.35 |   19.223 |   426.15 |   22.633 |   542.93 |
-
-### JSONL output
-
-Pass `--output-format jsonl` to output JSONL instead of Markdown, á la
-
-```json lines
-{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 1, "n_kv": 256, "t_pp": 0.233810, "speed_pp": 547.453064, "t_tg": 3.503684, "speed_tg": 36.532974, "t": 3.737494, "speed": 68.495094}
-{"n_kv_max": 2048, "n_batch": 2048, "n_ubatch": 512, "flash_attn": 0, "is_pp_shared": 0, "n_gpu_layers": 99, "n_threads": 8, "n_threads_batch": 8, "pp": 128, "tg": 128, "pl": 2, "n_kv": 512, "t_pp": 0.422602, "speed_pp": 605.770935, "t_tg": 11.106112, "speed_tg": 23.050371, "t": 11.528713, "speed": 44.410854}
-```

examples/batched-bench/batched-bench.cpp

@@ -1,13 +1,36 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
+#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int, char ** argv) {
+// mutates the input string
+static std::vector<int> parse_list(char * p) {
+    std::vector<int> ret;
+
+    char * q = p;
+
+    while (*p) {
+        if (*p == ',') {
+            *p = '\0';
+            ret.push_back(std::atoi(q));
+            q = p + 1;
+        }
+
+        ++p;
+    }
+
+    ret.push_back(std::atoi(q));
+
+    return ret;
+}
+
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
     LOG_TEE("\n");
@@ -16,7 +39,8 @@ static void print_usage(int, char ** argv) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
         return 1;
     }
 
@@ -45,7 +69,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
 
     // ensure enough sequences are available
-    ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
+    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
@@ -98,13 +122,12 @@ int main(int argc, char ** argv) {
         }
     }
 
-    if (!params.batched_bench_output_jsonl) {
-        LOG_TEE("\n");
-        LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
-        LOG_TEE("\n");
-        LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
-        LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
-    }
+    LOG_TEE("\n");
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("\n");
+
+    LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP",     "TG",     "B",    "N_KV",     "T_PP s",   "S_PP t/s", "T_TG s",   "S_TG t/s", "T s",      "S t/s");
+    LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
 
     for (        int i_pp = 0; i_pp < (int) n_pp.size(); ++i_pp) {
         for (    int i_tg = 0; i_tg < (int) n_tg.size(); ++i_tg) {
@@ -172,22 +195,12 @@ int main(int argc, char ** argv) {
                 const float speed_tg = pl*tg / t_tg;
                 const float speed    = n_kv / t;
 
-                if(params.batched_bench_output_jsonl) {
-                    LOG_TEE(
-                        "{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
-                        "\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
-                        n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
-                        pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
-                    );
-                } else {
-                    LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
-                }
+                LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
             }
         }
     }
 
-    LOG_TEE("\n");
-    llama_perf_context_print(ctx);
+    llama_print_timings(ctx);
 
     llama_batch_free(batch);
 

examples/batched.swift/Sources/main.swift

@@ -27,6 +27,7 @@ guard let model = llama_load_model_from_file(modelPath.cString(using: .utf8), mo
     print("Failed to load model")
     exit(1)
 }
+
 defer {
     llama_free_model(model)
 }
@@ -36,6 +37,7 @@ var tokens = tokenize(text: prompt, add_bos: true)
 let n_kv_req = UInt32(tokens.count) + UInt32((n_len - Int(tokens.count)) * n_parallel)
 
 var context_params = llama_context_default_params()
+context_params.seed = 1234
 context_params.n_ctx = n_kv_req
 context_params.n_batch = UInt32(max(n_len, n_parallel))
 context_params.n_threads = 8
@@ -46,26 +48,11 @@ guard context != nil else {
     print("Failed to initialize context")
     exit(1)
 }
+
 defer {
     llama_free(context)
 }
 
-var sparams = llama_sampler_chain_default_params()
-
-let smpl = llama_sampler_chain_init(sparams)
-guard smpl != nil else {
-    print("Failed to initialize sampling")
-    exit(1)
-}
-defer {
-    llama_sampler_free(smpl)
-}
-
-llama_sampler_chain_add(smpl, llama_sampler_init_top_k(40));
-llama_sampler_chain_add(smpl, llama_sampler_init_top_p(0.9, 1));
-llama_sampler_chain_add(smpl, llama_sampler_init_temp (0.4));
-llama_sampler_chain_add(smpl, llama_sampler_init_dist (1234));
-
 let n_ctx = llama_n_ctx(context)
 
 print("\nn_len = \(n_len), n_ctx = \(n_ctx), n_batch = \(context_params.n_batch), n_parallel = \(n_parallel), n_kv_req = \(n_kv_req)\n")
@@ -138,7 +125,32 @@ while n_cur <= n_len {
             continue
         }
 
-        let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
+        var n_vocab = llama_n_vocab(model)
+        var logits = llama_get_logits_ith(context, i_batch[i])
+
+        var candidates: [llama_token_data] = .init(repeating: llama_token_data(), count: Int(n_vocab))
+
+        for token_id in 0 ..< n_vocab {
+            candidates.append(llama_token_data(id: token_id, logit: logits![Int(token_id)], p: 0.0))
+        }
+
+        var candidates_p: llama_token_data_array = .init(
+            data: &candidates,
+            size: candidates.count,
+            sorted: false
+        )
+
+        let top_k: Int32 = 40
+        let top_p: Float = 0.9
+        let temp: Float = 0.4
+
+        llama_sample_top_k(context, &candidates_p, top_k, 1)
+        llama_sample_top_p(context, &candidates_p, top_p, 1)
+        llama_sample_temp(context, &candidates_p, temp)
+
+        let new_token_id = llama_sample_token(context, &candidates_p)
+
+        // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
@@ -198,10 +210,9 @@ if n_parallel > 1 {
 
 let t_main_end = ggml_time_us()
 
-print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n\n")
+print("decoded \(n_decode) tokens in \(String(format: "%.2f", Double(t_main_end - t_main_start) / 1_000_000.0)) s, speed: \(String(format: "%.2f", Double(n_decode) / (Double(t_main_end - t_main_start) / 1_000_000.0))) t/s\n")
 
-llama_perf_sampler_print(smpl)
-llama_perf_context_print(context)
+llama_print_timings(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let utf8Count = text.utf8.count
@@ -218,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
 
 private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
     var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), 0, false)
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
     if nTokens < 0 {
         let actualTokensCount = -Int(nTokens)
         result = .init(repeating: 0, count: actualTokensCount)
@@ -227,7 +238,6 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
             token,
             &result,
             Int32(result.count),
-            0,
             false
         )
         assert(check == actualTokensCount)

examples/batched/batched.cpp

@@ -1,13 +1,15 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
+#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int, char ** argv) {
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
     LOG_TEE("\n");
@@ -19,7 +21,8 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
         return 1;
     }
 
@@ -28,7 +31,7 @@ int main(int argc, char ** argv) {
     int n_parallel = params.n_parallel;
 
     // total length of the sequences including the prompt
-    int n_predict = params.n_predict;
+    int n_predict = 32;
 
     // init LLM
 
@@ -62,15 +65,6 @@ int main(int argc, char ** argv) {
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
-    auto sparams = llama_sampler_chain_default_params();
-
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
-    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
-    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
-    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
-
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
@@ -170,7 +164,29 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, i_batch[i]);
+
+            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 };
+
+            const int   top_k = 40;
+            const float top_p = 0.9f;
+            const float temp  = 0.4f;
+
+            llama_sample_top_k(ctx, &candidates_p, top_k, 1);
+            llama_sample_top_p(ctx, &candidates_p, top_p, 1);
+            llama_sample_temp (ctx, &candidates_p, temp);
+
+            const llama_token new_token_id = llama_sample_token(ctx, &candidates_p);
+
+            //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
             // is it an end of generation? -> mark the stream as finished
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
@@ -228,15 +244,12 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    LOG_TEE("\n");
-    llama_perf_sampler_print(smpl);
-    llama_perf_context_print(ctx);
+    llama_print_timings(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
 
-    llama_sampler_free(smpl);
     llama_free(ctx);
     llama_free_model(model);
 

examples/benchmark/benchmark-matmult.cpp

@@ -21,7 +21,7 @@
 #endif
 
 static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
-    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);
@@ -54,7 +54,7 @@ static void tensor_dump(const ggml_tensor * tensor, const char * name) {
 #define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
 
 struct benchmark_params_struct {
-    int     n_threads     = 1;
+    int32_t n_threads     = 1;
     int32_t n_iterations  = 10;
 };
 
@@ -183,7 +183,7 @@ int main(int argc, char ** argv)  {
 
     ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
 
-    TENSOR_DUMP(ggml_graph_node(gf, 0));
+    TENSOR_DUMP(gf->nodes[0]);
 
     printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
 
@@ -224,7 +224,7 @@ int main(int argc, char ** argv)  {
 
 
     // Let's use the F32 result from above as a reference for the quantized multiplication
-    float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 0));
+    float sum_of_F32_reference = tensor_sum_elements(gf->nodes[0]);
 
     printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
     printf("=====================================================================================\n");
@@ -252,7 +252,7 @@ int main(int argc, char ** argv)  {
 
         // Check that the matrix multiplication result is in the right ballpark
         // We cannot use the exact value from the F32 multiplication because the quantizuation will be slightly different
-        float sum_of_Q4_result = tensor_sum_elements(ggml_graph_node(gf31, 0));
+        float sum_of_Q4_result = tensor_sum_elements(gf31->nodes[0]);
         float delta = std::abs(sum_of_Q4_result - sum_of_F32_reference);
         float allowed_delta = (sum_of_F32_reference) / 1000 / 1000; //  Let's accept an epsilon of 10^-6
 

examples/convert_legacy_llama.py

@@ -24,7 +24,7 @@ from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar
+from typing import TYPE_CHECKING, Any, Callable, IO, Iterable, Literal, TypeVar, Optional
 
 import numpy as np
 
@@ -346,6 +346,42 @@ class Params:
         return params
 
 
+@dataclass
+class Metadata:
+    name: Optional[str] = None
+    author: Optional[str] = None
+    version: Optional[str] = None
+    url: Optional[str] = None
+    description: Optional[str] = None
+    licence: Optional[str] = None
+    source_url: Optional[str] = None
+    source_hf_repo: Optional[str] = None
+
+    @staticmethod
+    def load(metadata_path: Path) -> Metadata:
+        if metadata_path is None or not metadata_path.exists():
+            return Metadata()
+
+        with open(metadata_path, 'r') as file:
+            data = json.load(file)
+
+        # Create a new Metadata instance
+        metadata = Metadata()
+
+        # Assigning values to Metadata attributes if they exist in the JSON file
+        # This is based on LLM_KV_NAMES mapping in llama.cpp
+        metadata.name = data.get("general.name")
+        metadata.author = data.get("general.author")
+        metadata.version = data.get("general.version")
+        metadata.url = data.get("general.url")
+        metadata.description = data.get("general.description")
+        metadata.license = data.get("general.license")
+        metadata.source_url = data.get("general.source.url")
+        metadata.source_hf_repo = data.get("general.source.huggingface.repository")
+
+        return metadata
+
+
 #
 # data loading
 # TODO: reuse (probably move to gguf.py?)
@@ -456,13 +492,12 @@ class LazyTensor:
 
 LazyModel: TypeAlias = 'dict[str, LazyTensor]'
 
-ModelFormat: TypeAlias = Literal['ggml', 'torch', 'safetensors', 'none']
 
 @dataclass
 class ModelPlus:
     model: LazyModel
     paths: list[Path]  # Where this was read from.
-    format: ModelFormat
+    format: Literal['ggml', 'torch', 'safetensors', 'none']
     vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
@@ -501,7 +536,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
 
 
 def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
-    formats: set[ModelFormat] = set(mp.format for mp in models_plus)
+    formats = set(mp.format for mp in models_plus)
     assert len(formats) == 1, "different formats?"
     format = formats.pop()
     paths = [path for mp in models_plus for path in mp.paths]
@@ -520,7 +555,7 @@ def merge_multifile_models(models_plus: list[ModelPlus]) -> ModelPlus:
     else:
         model = merge_sharded([mp.model for mp in models_plus])
 
-    return ModelPlus(model, paths, format, vocab)
+    return ModelPlus(model, paths, format, vocab)  # pytype: disable=wrong-arg-types
 
 
 def permute_lazy(lazy_tensor: LazyTensor, n_head: int, n_head_kv: int) -> LazyTensor:
@@ -770,7 +805,7 @@ class OutputFile:
     def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
-    def add_meta_model(self, params: Params, metadata: gguf.Metadata | None) -> None:
+    def add_meta_model(self, params: Params, metadata: Metadata) -> None:
         # Metadata About The Model And Its Provenence
         name = "LLaMA"
         if metadata is not None and metadata.name is not None:
@@ -788,73 +823,16 @@ class OutputFile:
                 self.gguf.add_author(metadata.author)
             if metadata.version is not None:
                 self.gguf.add_version(metadata.version)
-            if metadata.organization is not None:
-                self.gguf.add_organization(metadata.organization)
-
-            if metadata.finetune is not None:
-                self.gguf.add_finetune(metadata.finetune)
-            if metadata.basename is not None:
-                self.gguf.add_basename(metadata.basename)
-
-            if metadata.description is not None:
-                self.gguf.add_description(metadata.description)
-            if metadata.quantized_by is not None:
-                self.gguf.add_quantized_by(metadata.quantized_by)
-
-            if metadata.size_label is not None:
-                self.gguf.add_size_label(metadata.size_label)
-
-            if metadata.license is not None:
-                self.gguf.add_license(metadata.license)
-            if metadata.license_name is not None:
-                self.gguf.add_license_name(metadata.license_name)
-            if metadata.license_link is not None:
-                self.gguf.add_license_link(metadata.license_link)
-
             if metadata.url is not None:
                 self.gguf.add_url(metadata.url)
-            if metadata.doi is not None:
-                self.gguf.add_doi(metadata.doi)
-            if metadata.uuid is not None:
-                self.gguf.add_uuid(metadata.uuid)
-            if metadata.repo_url is not None:
-                self.gguf.add_repo_url(metadata.repo_url)
-
+            if metadata.description is not None:
+                self.gguf.add_description(metadata.description)
+            if metadata.licence is not None:
+                self.gguf.add_licence(metadata.licence)
             if metadata.source_url is not None:
                 self.gguf.add_source_url(metadata.source_url)
-            if metadata.source_doi is not None:
-                self.gguf.add_source_doi(metadata.source_doi)
-            if metadata.source_uuid is not None:
-                self.gguf.add_source_uuid(metadata.source_uuid)
-            if metadata.source_repo_url is not None:
-                self.gguf.add_source_repo_url(metadata.source_repo_url)
-
-            if metadata.base_models is not None:
-                self.gguf.add_base_model_count(len(metadata.base_models))
-                for key, base_model_entry in enumerate(metadata.base_models):
-                    if "name" in base_model_entry:
-                        self.gguf.add_base_model_name(key, base_model_entry["name"])
-                    if "author" in base_model_entry:
-                        self.gguf.add_base_model_author(key, base_model_entry["author"])
-                    if "version" in base_model_entry:
-                        self.gguf.add_base_model_version(key, base_model_entry["version"])
-                    if "organization" in base_model_entry:
-                        self.gguf.add_base_model_organization(key, base_model_entry["organization"])
-                    if "url" in base_model_entry:
-                        self.gguf.add_base_model_url(key, base_model_entry["url"])
-                    if "doi" in base_model_entry:
-                        self.gguf.add_base_model_doi(key, base_model_entry["doi"])
-                    if "uuid" in base_model_entry:
-                        self.gguf.add_base_model_uuid(key, base_model_entry["uuid"])
-                    if "repo_url" in base_model_entry:
-                        self.gguf.add_base_model_repo_url(key, base_model_entry["repo_url"])
-
-            if metadata.tags is not None:
-                self.gguf.add_tags(metadata.tags)
-            if metadata.languages is not None:
-                self.gguf.add_languages(metadata.languages)
-            if metadata.datasets is not None:
-                self.gguf.add_datasets(metadata.datasets)
+            if metadata.source_hf_repo is not None:
+                self.gguf.add_source_hf_repo(metadata.source_hf_repo)
 
     def add_meta_arch(self, params: Params) -> None:
         # Metadata About The Neural Architecture Itself
@@ -965,7 +943,7 @@ class OutputFile:
     @staticmethod
     def write_vocab_only(
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
-        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: gguf.Metadata | None = None,
+        endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False, metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -999,7 +977,7 @@ class OutputFile:
         fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
-        metadata: gguf.Metadata | None = None,
+        metadata: Metadata = None,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -1042,32 +1020,35 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
     raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
-def per_model_weight_count_estimation(tensors: Iterable[tuple[str, LazyTensor]]) -> tuple[int, int, int]:
-    total_params = 0
-    shared_params = 0
-    expert_params = 0
-
-    for name, lazy_tensor in tensors:
-        # We don't need these
-        if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
-            continue
-
-        # Got A Tensor
-        sum_weights_in_tensor: int = 1
-
-        # Tensor Volume
+def model_parameter_count(model: LazyModel) -> int:
+    total_model_parameters = 0
+    for i, (name, lazy_tensor) in enumerate(model.items()):
+        sum_weights_in_tensor = 1
         for dim in lazy_tensor.shape:
             sum_weights_in_tensor *= dim
+        total_model_parameters += sum_weights_in_tensor
+    return total_model_parameters
 
-        if ".experts." in name:
-            if ".experts.0." in name:
-                expert_params += sum_weights_in_tensor
-        else:
-            shared_params += sum_weights_in_tensor
 
-        total_params += sum_weights_in_tensor
+def model_parameter_count_rounded_notation(model_params_count: int) -> str:
+    if model_params_count > 1e12 :
+        # Trillions Of Parameters
+        scaled_model_params = model_params_count * 1e-12
+        scale_suffix = "T"
+    elif model_params_count > 1e9 :
+        # Billions Of Parameters
+        scaled_model_params = model_params_count * 1e-9
+        scale_suffix = "B"
+    elif model_params_count > 1e6 :
+        # Millions Of Parameters
+        scaled_model_params = model_params_count * 1e-6
+        scale_suffix = "M"
+    else:
+        # Thousands Of Parameters
+        scaled_model_params = model_params_count * 1e-3
+        scale_suffix = "K"
 
-    return total_params, shared_params, expert_params
+    return f"{round(scaled_model_params)}{scale_suffix}"
 
 
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1249,24 +1230,34 @@ class VocabFactory:
         return vocab, special_vocab
 
 
-def default_convention_outfile(file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> str:
-    name = metadata.name if metadata.name is not None else None
-    basename = metadata.basename if metadata.basename is not None else None
-    finetune = metadata.finetune if metadata.finetune is not None else None
-    version = metadata.version if metadata.version is not None else None
-    size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(*model_params_count, expert_count=expert_count or 0)
-
-    output_type = {
+def default_convention_outfile(file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> str:
+    quantization = {
         GGMLFileType.AllF32:    "F32",
         GGMLFileType.MostlyF16: "F16",
         GGMLFileType.MostlyQ8_0: "Q8_0",
     }[file_type]
 
-    return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
+    parameters = model_parameter_count_rounded_notation(model_params_count)
+
+    expert_count = ""
+    if params.n_experts is not None:
+        expert_count = f"{params.n_experts}x"
+
+    version = ""
+    if metadata is not None and metadata.version is not None:
+        version = f"-{metadata.version}"
+
+    name = "ggml-model"
+    if metadata is not None and metadata.name is not None:
+        name = metadata.name
+    elif params.path_model is not None:
+        name = params.path_model.name
+
+    return f"{name}{version}-{expert_count}{parameters}-{quantization}"
 
 
-def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count: int | None, model_params_count: tuple[int, int, int], metadata: gguf.Metadata) -> Path:
-    default_filename = default_convention_outfile(file_type, expert_count, model_params_count, metadata)
+def default_outfile(model_paths: list[Path], file_type: GGMLFileType, params: Params, model_params_count: int, metadata: Metadata) -> Path:
+    default_filename = default_convention_outfile(file_type, params, model_params_count, metadata)
     ret = model_paths[0].parent / f"{default_filename}.gguf"
     if ret in model_paths:
         logger.error(
@@ -1305,9 +1296,8 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
     parser.add_argument("--verbose",      action="store_true",    help="increase output verbosity")
-    parser.add_argument("--metadata",     type=Path,              help="Specify the path for an authorship metadata override file")
+    parser.add_argument("--metadata",     type=Path,              help="Specify the path for a metadata file")
     parser.add_argument("--get-outfile",  action="store_true",    help="get calculated default outfile name")
-    parser.add_argument("--model-name",   type=str, default=None, help="name of the model")
 
     args = parser.parse_args(args_in)
 
@@ -1319,36 +1309,32 @@ def main(args_in: list[str] | None = None) -> None:
     else:
         logging.basicConfig(level=logging.INFO)
 
-    model_name = args.model_name
-    dir_model = args.model
-
-    metadata = gguf.Metadata.load(args.metadata, dir_model, model_name)
+    metadata = Metadata.load(args.metadata)
 
     if args.get_outfile:
-        model_plus = load_some_model(dir_model)
+        model_plus = load_some_model(args.model)
         params = Params.load(model_plus)
-        model = convert_model_names(model_plus.model, params, args.skip_unknown)
-        model_params_count = per_model_weight_count_estimation(model_plus.model.items())
-        ftype = pick_output_type(model, args.outtype)
-
-        if (metadata is None or metadata.name is None) and params.path_model is not None:
-            metadata.name = params.path_model.name
-
-        print(f"{default_convention_outfile(ftype, params.n_experts, model_params_count, metadata)}") # noqa: NP100
+        model   = convert_model_names(model_plus.model, params, args.skip_unknown)
+        model_params_count = model_parameter_count(model_plus.model)
+        ftype   = pick_output_type(model, args.outtype)
+        print(f"{default_convention_outfile(ftype, params, model_params_count, metadata)}") # noqa: NP100
         return
 
     if args.no_vocab and args.vocab_only:
         raise ValueError("--vocab-only does not make sense with --no-vocab")
 
     if args.dump_single:
-        model_plus = lazy_load_file(dir_model)
+        model_plus = lazy_load_file(args.model)
         do_dump_model(model_plus)
         return
 
     if not args.vocab_only:
-        model_plus = load_some_model(dir_model)
+        model_plus = load_some_model(args.model)
     else:
-        model_plus = ModelPlus(model = {}, paths = [dir_model / 'dummy'], format = 'none', vocab = None)
+        model_plus = ModelPlus(model = {}, paths = [args.model / 'dummy'], format = 'none', vocab = None)
+
+    model_params_count = model_parameter_count(model_plus.model)
+    logger.info(f"model parameters count : {model_params_count} ({model_parameter_count_rounded_notation(model_params_count)})")
 
     if args.dump:
         do_dump_model(model_plus)
@@ -1381,7 +1367,7 @@ def main(args_in: list[str] | None = None) -> None:
         logger.info(f"params = {params}")
 
     model_parent_path = model_plus.paths[0].parent
-    vocab_path = Path(args.vocab_dir or dir_model or model_parent_path)
+    vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
     vocab_types = None if args.no_vocab else args.vocab_type.split(",")
     vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
@@ -1410,23 +1396,13 @@ def main(args_in: list[str] | None = None) -> None:
     if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
         vocab = model_plus.vocab
 
-    assert params is not None
-
-    if metadata.name is None and params.path_model is not None:
-        metadata.name = params.path_model.name
-
-    model_params_count = per_model_weight_count_estimation(model_plus.model.items())
-    logger.info(f"model parameters count : {model_params_count} ({gguf.model_weight_count_rounded_notation(model_params_count[0])})")
-
     logger.info(f"Vocab info: {vocab}")
     logger.info(f"Special vocab info: {special_vocab}")
     model   = model_plus.model
     model   = convert_model_names(model, params, args.skip_unknown)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
-    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
-
-    metadata.size_label = gguf.size_label(*model_params_count, expert_count=params.n_experts or 0)
+    outfile = args.outfile or default_outfile(model_plus.paths, ftype, params, model_params_count, metadata)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")

examples/cvector-generator/cvector-generator.cpp

@@ -1,4 +1,3 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -36,7 +35,9 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     return ret;
 }
 
-static void print_usage(int, char ** argv) {
+static void print_usage(int argc, char ** argv, const gpt_params & params) {
+    gpt_params_print_usage(argc, argv, params);
+
     printf("\nexample usage:\n");
     printf("\n    CPU only:   %s -m ./llama-3.Q4_K_M.gguf\n", argv[0]);
     printf("\n    with GPU:   %s -m ./llama-3.Q4_K_M.gguf -ngl 99\n", argv[0]);
@@ -270,7 +271,7 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
         tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
         tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
@@ -389,7 +390,8 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
         return 1;
     }
 
@@ -412,10 +414,9 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // int n_ctx = llama_n_ctx(ctx);
     int n_layers = llama_n_layer(model);
@@ -484,8 +485,8 @@ int main(int argc, char ** argv) {
     if (use_pca) {
         // run PCA
         PCA::pca_params pca_params;
-        pca_params.n_threads    = params.cpuparams.n_threads;
-        pca_params.n_batch      = params.n_pca_batch;
+        pca_params.n_threads = params.n_threads;
+        pca_params.n_batch = params.n_pca_batch;
         pca_params.n_iterations = params.n_pca_iterations;
         PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final);
     } else {

examples/cvector-generator/pca.hpp

@@ -12,9 +12,12 @@
 
 #include <cstdio>
 #include <ctime>
-#include <random>
 #include <string>
+#include <tuple>
 #include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
 
 #define DEBUG_POS 5
 
@@ -226,8 +229,8 @@ static ggml_status compute_piter(
         result.eigenvectors.resize(params.n_batch);
         result.distances.resize(params.n_batch);
         // get output nodes
-        for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
-            auto node = ggml_graph_node(gf, i);
+        for (int i = 0; i < gf->n_nodes; ++i) {
+            auto node = gf->nodes[i];
             int iter = -1;
             // find b_tensor (without copying data from device)
             if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) {

examples/deprecation-warning/README.md

@@ -1,49 +0,0 @@
-# Migration notice for binary filenames
-
-> [!IMPORTANT]
-[2024 Jun 12] Binaries have been renamed w/ a `llama-` prefix. `main` is now `llama-cli`, `server` is `llama-server`, etc (https://github.com/ggerganov/llama.cpp/pull/7809)
-
-This migration was important, but it is a breaking change that may not always be immediately obvious to users.
-
-Please update all scripts and workflows to use the new binary names.
-
-| Old Filename | New Filename |
-| ---- | ---- |
-| main | llama-cli |
-| server | llama-server |
-| llama-bench | llama-bench |
-| embedding | llama-embedding |
-| quantize | llama-quantize |
-| tokenize | llama-tokenize |
-| export-lora | llama-export-lora |
-| libllava.a | libllava.a |
-| baby-llama | llama-baby-llama |
-| batched | llama-batched |
-| batched-bench | llama-batched-bench |
-| benchmark-matmult | llama-benchmark-matmult |
-| convert-llama2c-to-ggml | llama-convert-llama2c-to-ggml |
-| eval-callback | llama-eval-callback |
-| gbnf-validator | llama-gbnf-validator |
-| gguf | llama-gguf |
-| gguf-split | llama-gguf-split |
-| gritlm | llama-gritlm |
-| imatrix | llama-imatrix |
-| infill | llama-infill |
-| llava-cli | llama-llava-cli |
-| lookahead | llama-lookahead |
-| lookup | llama-lookup |
-| lookup-create | llama-lookup-create |
-| lookup-merge | llama-lookup-merge |
-| lookup-stats | llama-lookup-stats |
-| parallel | llama-parallel |
-| passkey | llama-passkey |
-| perplexity | llama-perplexity |
-| q8dot | llama-q8dot |
-| quantize-stats | llama-quantize-stats |
-| retrieval | llama-retrieval |
-| save-load-state | llama-save-load-state |
-| simple | llama-simple |
-| speculative | llama-speculative |
-| vdot | llama-vdot |
-| tests/test-c.o | tests/test-c.o |
-

examples/deprecation-warning/deprecation-warning.cpp

@@ -1,35 +0,0 @@
-// Warns users that this filename was deprecated, and provides a link for more information.
-
-#include <cstdio>
-#include <string>
-#include <unordered_map>
-
-// Main
-int main(int argc, char** argv) {
-    std::string filename = "main";
-    if (argc >= 1) {
-        filename = argv[0];
-    }
-
-    // Get only the program name from the full path
-    auto pos = filename.find_last_of('/');
-    if (pos != std::string::npos) {
-        filename = filename.substr(pos+1);
-    }
-
-    // Append "llama-" to the beginning of filename to get the replacemnt filename
-    auto replacement_filename = "llama-" + filename;
-
-    // The exception is if the filename is "main", then our replacement filename is "llama-cli"
-    if (filename == "main") {
-        replacement_filename = "llama-cli";
-    }
-
-    fprintf(stdout, "\n");
-    fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
-    fprintf(stdout, " Please use '%s' instead.\n", replacement_filename.c_str());
-    fprintf(stdout, " See https://github.com/ggerganov/llama.cpp/tree/master/examples/deprecation-warning/README.md for more information.\n");
-    fprintf(stdout, "\n");
-
-    return EXIT_FAILURE;
-}

examples/embedding/README.md

@@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null
+./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
+llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
 ```
 
 The above command will output space-separated float values.
@@ -50,11 +50,11 @@ The above command will output space-separated float values.
 ### Unix-based systems (Linux, macOS, etc.):
 
 ```bash
-./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```
 
 ### Windows:
 
 ```powershell
-llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
+embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2  --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
 ```

examples/embedding/embedding.cpp

@@ -1,4 +1,3 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -32,24 +31,13 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
-    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    const struct llama_model * model = llama_get_model(ctx);
-
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
     fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
-        // encoder-only model
-        if (llama_encode(ctx, batch) < 0) {
-            fprintf(stderr, "%s : failed to encode\n", __func__);
-        }
-    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
-        // decoder-only model
-        if (llama_decode(ctx, batch) < 0) {
-            fprintf(stderr, "%s : failed to decode\n", __func__);
-        }
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -57,22 +45,11 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        const float * embd = nullptr;
-        int embd_pos = 0;
+        // try to get sequence embeddings - supported only when pooling_type is not NONE
+        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
 
-        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-            // try to get token embeddings
-            embd = llama_get_embeddings_ith(ctx, i);
-            embd_pos = i;
-            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
-        } else {
-            // try to get sequence embeddings - supported only when pooling_type is not NONE
-            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-            embd_pos = batch.seq_id[i][0];
-            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
-        }
-
-        float * out = output + embd_pos * n_embd;
+        float * out = output + batch.seq_id[i][0] * n_embd;
         llama_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
@@ -80,7 +57,8 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
@@ -90,14 +68,22 @@ int main(int argc, char ** argv) {
 
     print_build_info();
 
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    llama_model * model;
+    llama_context * ctx;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == NULL) {
         fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
@@ -107,9 +93,8 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-
-    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
-        fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
         return 1;
     }
 
@@ -168,23 +153,13 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
-    // count number of embeddings
-    int n_embd_count = 0;
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        for (int k = 0; k < n_prompts; k++) {
-            n_embd_count += inputs[k].size();
-        }
-    } else {
-        n_embd_count = n_prompts;
-    }
-
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_embd_count * n_embd, 0);
+    std::vector<float> embeddings(n_prompts * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int e = 0; // number of embeddings already stored
+    int p = 0; // number of prompts processed already
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -194,11 +169,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + e * n_embd;
+            float * out = emb + p * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
-            s = 0;
             llama_batch_clear(batch);
+            p += s;
+            s = 0;
         }
 
         // add to batch
@@ -207,62 +182,39 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + e * n_embd;
+    float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
+        // print the first part of the embeddings or for a single prompt, the full embedding
         fprintf(stdout, "\n");
-
-        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-            for (int j = 0; j < n_embd_count; j++) {
-                fprintf(stdout, "embedding %d: ", j);
-                for (int i = 0; i < std::min(3, n_embd); i++) {
-                    if (params.embd_normalize == 0) {
-                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                    }
+        for (int j = 0; j < n_prompts; j++) {
+            fprintf(stdout, "embedding %d: ", j);
+            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                if (params.embd_normalize == 0) {
+                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
+                } else {
+                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
                 }
-                fprintf(stdout, " ... ");
-                for (int i = n_embd - 3; i < n_embd; i++) {
-                    if (params.embd_normalize == 0) {
-                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                    }
-                }
-                fprintf(stdout, "\n");
-            }
-        } else {
-            // print the first part of the embeddings or for a single prompt, the full embedding
-            for (int j = 0; j < n_prompts; j++) {
-                fprintf(stdout, "embedding %d: ", j);
-                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                    if (params.embd_normalize == 0) {
-                        fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                    }
-                }
-                fprintf(stdout, "\n");
             }
+            fprintf(stdout, "\n");
+        }
 
-            // print cosine similarity matrix
-            if (n_prompts > 1) {
-                fprintf(stdout, "\n");
-                printf("cosine similarity matrix:\n\n");
-                for (int i = 0; i < n_prompts; i++) {
-                    fprintf(stdout, "%6.6s ", prompts[i].c_str());
+        // print cosine similarity matrix
+        if (n_prompts > 1) {
+            fprintf(stdout, "\n");
+            printf("cosine similarity matrix:\n\n");
+            for (int i = 0; i < n_prompts; i++) {
+                fprintf(stdout, "%6.6s ", prompts[i].c_str());
+            }
+            fprintf(stdout, "\n");
+            for (int i = 0; i < n_prompts; i++) {
+                for (int j = 0; j < n_prompts; j++) {
+                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                    fprintf(stdout, "%6.2f ", sim);
                 }
+                fprintf(stdout, "%1.10s", prompts[i].c_str());
                 fprintf(stdout, "\n");
-                for (int i = 0; i < n_prompts; i++) {
-                    for (int j = 0; j < n_prompts; j++) {
-                        float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                        fprintf(stdout, "%6.2f ", sim);
-                    }
-                    fprintf(stdout, "%1.10s", prompts[i].c_str());
-                    fprintf(stdout, "\n");
-                }
             }
         }
     }
@@ -281,23 +233,23 @@ int main(int argc, char ** argv) {
             }
             fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
         fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
             fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
+            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
                 fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
+                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
                     float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
                     fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_embd_count) fprintf(stdout, ", "); else break;
+                    if (j < n_prompts) fprintf(stdout, ", "); else break;
                 }
                 fprintf(stdout, " ]");
                 i++;
-                if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
+                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
             }
             fprintf(stdout, "\n  ]");
         }
@@ -305,10 +257,8 @@ int main(int argc, char ** argv) {
         if (notArray) fprintf(stdout, "\n}\n");
     }
 
-    LOG_TEE("\n");
-    llama_perf_context_print(ctx);
-
     // clean up
+    llama_print_timings(ctx);
     llama_batch_free(batch);
     llama_free(ctx);
     llama_free_model(model);

examples/eval-callback/eval-callback.cpp

@@ -1,4 +1,3 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -63,7 +62,7 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else if (type == GGML_TYPE_I8) {
                         v = (float) *(int8_t *) &data[i];
                     } else {
-                        GGML_ABORT("fatal error");
+                        GGML_ASSERT(false);
                     }
                     printf("%12.4f", v);
                     sum += v;
@@ -100,7 +99,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 
     char src1_str[128] = {0};
     if (src1) {
-        snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
+        sprintf(src1_str, "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
     }
 
     printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
@@ -128,7 +127,7 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
 }
 
 static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
 
@@ -145,12 +144,15 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
     print_build_info();
 
+    std::mt19937 rng(params.seed);
+
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -161,10 +163,9 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
         fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
@@ -181,8 +182,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    LOG_TEE("\n");
-    llama_perf_context_print(ctx);
+    llama_print_timings(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/export-lora/README.md

@@ -6,28 +6,21 @@ Apply LORA adapters to base model and export the resulting model.
 usage: llama-export-lora [options]
 
 options:
-  -m,    --model                  model path from which to load base model (default '')
-         --lora FNAME             path to LoRA adapter  (can be repeated to use multiple adapters)
-         --lora-scaled FNAME S    path to LoRA adapter with user defined scaling S  (can be repeated to use multiple adapters)
-  -t,    --threads N              number of threads to use during computation (default: 4)
-  -o,    --output FNAME           output file (default: 'ggml-lora-merged-f16.gguf')
+  -h, --help                         show this help message and exit
+  -m FNAME, --model-base FNAME       model path from which to load base model (default '')
+  -o FNAME, --model-out FNAME        path to save exported model (default '')
+  -l FNAME, --lora FNAME             apply LoRA adapter
+  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S
+  -t N, --threads N                  number of threads to use during computation (default: 4)
 ```
 
 For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2.gguf \
-    -o open-llama-3b-v2-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2-q8_0.gguf \
+    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
+    -l lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.bin
 ```
 
-Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:
-
-```bash
-./bin/llama-export-lora \
-    -m your_base_model.gguf \
-    -o your_merged_model.gguf \
-    --lora-scaled lora_task_A.gguf 0.5 \
-    --lora-scaled lora_task_B.gguf 0.5
-```
+Multiple LORA adapters can be applied by passing multiple `-l FN` or `-s FN S` command line parameters.

examples/export-lora/export-lora.cpp

@@ -1,421 +1,462 @@
-#include "arg.h"
+
 #include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
 
-#include <map>
 #include <vector>
 #include <string>
 #include <thread>
-#include <fstream>
 
-static bool g_verbose = false;
-
-struct tensor_transformation {
-    struct ggml_tensor * in;
-    struct ggml_tensor * out;
-    bool is_copy;
-};
-
-static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
-    int id = gguf_find_key(ctx_gguf, key.c_str());
-    return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
-}
-
-static float get_kv_f32(struct gguf_context * ctx_gguf, const std::string & key) {
-    int id = gguf_find_key(ctx_gguf, key.c_str());
-    return id < 0 ? 0.0f : gguf_get_val_f32(ctx_gguf, id);
-}
-
-static void zeros(std::ofstream & file, size_t n) {
-    char zero = 0;
-    for (size_t i = 0; i < n; ++i) {
-        file.write(&zero, 1);
-    }
-}
-
-static std::string ggml_ne_string(const ggml_tensor * t) {
-    std::string str;
-    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
-        str += std::to_string(t->ne[i]);
-        if (i + 1 < GGML_MAX_DIMS) {
-            str += ", ";
-        }
-    }
-    return str;
-}
-
-static struct gguf_context * load_gguf(std::string & fname, struct ggml_context ** ctx_ggml) {
-    struct gguf_init_params params = {
-        /*.no_alloc = */ true,
-        /*.ctx      = */ ctx_ggml,
-    };
-    struct gguf_context * ctx_gguf = gguf_init_from_file(fname.c_str(), params);
-    if (!ctx_gguf) {
-        throw std::runtime_error("failed to load input GGUF from " + fname);
-    }
-    return ctx_gguf;
-}
-
-struct file_input {
-    struct ggml_context * ctx_meta = nullptr;
-    struct gguf_context * ctx_gguf = nullptr;
-    std::ifstream f_in;
-    std::map<std::string, ggml_tensor *> tensors;
-    float alpha;
+struct lora_info {
+    std::string filename;
     float scale;
-
-    file_input(std::string & fname, float scale): f_in(fname, std::ios::binary), scale(scale) {
-        if (!f_in.is_open()) {
-            throw std::runtime_error("failed to open input gguf from " + fname);
-        }
-
-        ctx_gguf = load_gguf(fname, &ctx_meta);
-        alpha = get_kv_f32(ctx_gguf, "adapter.lora.alpha");
-        printf("%s: loaded gguf from %s\n", __func__, fname.c_str());
-
-        for (ggml_tensor * cur = ggml_get_first_tensor(ctx_meta); cur; cur = ggml_get_next_tensor(ctx_meta, cur)) {
-            std::string name(cur->name);
-            tensors[name] = cur;
-            if (g_verbose) {
-                printf("%s: %s\n", __func__, cur->name);
-            }
-        }
-    }
-
-    ggml_tensor * get_tensor(std::string name) {
-        if (tensors.find(name) == tensors.end()) {
-            return nullptr;
-        }
-        return tensors[name];
-    }
-
-    void read_tensor_data(std::string name, std::vector<uint8_t> & buf) {
-        if (tensors.find(name) == tensors.end()) {
-            throw std::runtime_error("cannot find tensor with name: " + name);
-        }
-        auto len = ggml_nbytes(tensors[name]);
-        if (buf.size() < len) {
-            buf.resize(len);
-        }
-        auto i_tensor_in = gguf_find_tensor(ctx_gguf, name.c_str()); // idx of tensor in the input file
-        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
-        f_in.seekg(offset);
-        f_in.read((char* )buf.data(), len);
-    }
-
-    ~file_input() {
-        gguf_free(ctx_gguf);
-        ggml_free(ctx_meta);
-    }
 };
 
-struct lora_merge_ctx {
-    // input base model + adapters
-    file_input base_model;
-    std::vector<std::unique_ptr<file_input>> adapters;
-
-    // for computing merged tensor
+struct export_lora_params {
+    std::string fn_model_base;
+    std::string fn_model_out;
+    std::vector<struct lora_info> lora;
     int n_threads;
-    ggml_backend_t backend = nullptr;
-    ggml_gallocr_t allocr = nullptr;
-    std::vector<uint8_t> read_buf;
+};
 
-    // output file
-    struct gguf_context * ctx_out;
-    struct ggml_context * ctx_out_ggml;
-    std::ofstream fout;
+struct lora_data {
+    struct lora_info     info;
+    std::vector<uint8_t> data;
+    struct ggml_context * ctx;
 
-    lora_merge_ctx(
-            std::string & base_fname,
-            std::vector<llama_lora_adapter_info> & lora_files,
-            std::string & outfile,
-            int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
-        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+    uint32_t lora_r;
+    uint32_t lora_alpha;
+};
 
-        if (gguf_find_key(base_model.ctx_gguf, LLM_KV_SPLIT_COUNT) >= 0) {
-            throw std::runtime_error("split model is not yet supported");
-        }
+struct llama_file {
+    // use FILE * so we don't have to re-open the file to mmap
+    FILE * fp;
+    size_t size;
 
-        for (auto & lora_inp : lora_files) {
-            auto fname = lora_inp.path;
-            auto scale = lora_inp.scale;
-            std::unique_ptr<file_input> adapter(new file_input(fname, scale));
-            check_metadata_lora(adapter.get());
-            adapters.push_back(std::move(adapter));
-        }
-
-        ctx_out = gguf_init_empty();
-        struct ggml_init_params params = {
-            /*.mem_size   =*/ gguf_get_n_tensors(base_model.ctx_gguf)*ggml_tensor_overhead(),
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ true,
-        };
-        ctx_out_ggml = ggml_init(params);
-        backend = ggml_backend_cpu_init();
-        allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
-    }
-
-    void check_metadata_lora(file_input * adapter) {
-        auto general_type = get_kv_str(adapter->ctx_gguf, "general.type");
-        if (general_type != "adapter") {
-            throw std::runtime_error("expect general.type to be 'adapter', but got: " + general_type);
-        }
-
-        auto adapter_type = get_kv_str(adapter->ctx_gguf, "adapter.type");
-        if (adapter_type != "lora") {
-            throw std::runtime_error("expect adapter.type to be 'lora', but got: " + adapter_type);
-        }
-
-        auto general_arch_base = get_kv_str(base_model.ctx_gguf, "general.architecture");
-        auto general_arch_lora = get_kv_str(adapter->ctx_gguf,   "general.architecture");
-        if (general_arch_base != general_arch_lora) {
-            throw std::runtime_error("model arch and LoRA arch mismatch");
-        }
-    }
-
-    ggml_type get_out_tensor_type(struct ggml_tensor * t) {
-        if (t->type == GGML_TYPE_F32) {
-            return GGML_TYPE_F32;
+    llama_file(const char * fname, const char * mode) {
+        fp = std::fopen(fname, mode);
+        if (fp == NULL) {
+            size = 0;
         } else {
-            return GGML_TYPE_F16;
+            seek(0, SEEK_END);
+            size = tell();
+            seek(0, SEEK_SET);
         }
     }
 
-    void run_merge() {
-        // prepare metadata
-        gguf_set_kv(ctx_out, base_model.ctx_gguf);
-        // output is forced to f16 for now
-        gguf_set_val_u32(ctx_out, "general.file_type", LLAMA_FTYPE_MOSTLY_F16);
-
-        // check if all lora adapters have the same tensors
-        // TODO: remove this when we can support merging subset of adapters. Ref: https://github.com/ggerganov/llama.cpp/pull/8607#discussion_r1686027777
-        static const char * err_no_subset_adapter = "Input adapters do not have the same list of tensors. This is not yet supported. Please merge the adapter one-by-one instead of merging all at once.";
-        if (adapters.size() > 1) {
-            for (size_t i = 1; i < adapters.size(); ++i) {
-                if (adapters[0]->tensors.size() != adapters[i]->tensors.size()) {
-                    throw std::runtime_error(err_no_subset_adapter);
-                }
-                for (auto & it : adapters[i]->tensors) {
-                    if (adapters[0]->get_tensor(it.first) == nullptr) {
-                        throw std::runtime_error(err_no_subset_adapter);
-                    }
-                }
-            }
-        }
-
-        // mapping base tensor to out tensor (same shape with base, but different type)
-        std::vector<tensor_transformation> trans;
-        for (auto & it : base_model.tensors) {
-            bool t_a = true;
-            bool t_b = true;
-            for (auto & adapter : adapters) {
-                t_a &= nullptr != adapter->get_tensor(it.first + ".lora_a");
-                t_b &= nullptr != adapter->get_tensor(it.first + ".lora_b");
-            }
-            auto base_tensor = it.second;
-            if (!t_a && !t_b) {
-                // only copy
-                struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
-                ggml_set_name(cpy_tensor, base_tensor->name);
-                trans.push_back({
-                    cpy_tensor,
-                    cpy_tensor,
-                    true,
-                });
-                gguf_add_tensor(ctx_out, cpy_tensor);
-            } else if (t_a && t_b) {
-                // need merging
-                struct ggml_tensor * out_tensor = ggml_new_tensor(
-                    ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
-                ggml_set_name(out_tensor, base_tensor->name);
-                trans.push_back({
-                    base_tensor,
-                    out_tensor,
-                    false,
-                });
-                gguf_add_tensor(ctx_out, out_tensor);
-            } else {
-                throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
-            }
-        }
-
-        // placeholder for the meta data
-        {
-            size_t meta_size = gguf_get_meta_size(ctx_out);
-            zeros(fout, meta_size);
-        }
-
-        // process base model tensors
-        size_t n_merged = 0;
-        for (auto & it : trans) {
-            if (!it.is_copy) {
-                merge_tensor(it.in, it.out);
-                n_merged++;
-            } else {
-                copy_tensor(it.in);
-            }
-        }
-
-        // write output metadata
-        {
-            std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
-            gguf_get_meta_data(ctx_out, data.data());
-            fout.seekp(0);
-            fout.write((const char *)data.data(), data.size());
-        }
-
-        printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
+    size_t tell() const {
+#ifdef _WIN32
+        __int64 ret = _ftelli64(fp);
+#else
+        long ret = std::ftell(fp);
+#endif
+        GGML_ASSERT(ret != -1); // this really shouldn't fail
+        return (size_t) ret;
     }
 
-    void copy_tensor(struct ggml_tensor * base) {
-        printf("%s :  %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
-        size_t len = ggml_nbytes(base);
-        base_model.read_tensor_data(base->name, read_buf);
-        fout.write((char* )read_buf.data(), len);
-        zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
+    void seek(size_t offset, int whence) {
+#ifdef _WIN32
+        int ret = _fseeki64(fp, (__int64) offset, whence);
+#else
+        int ret = std::fseek(fp, (long) offset, whence);
+#endif
+        GGML_ASSERT(ret == 0); // same
     }
 
-    void merge_tensor(struct ggml_tensor * base, struct ggml_tensor * out) {
-        std::string name_base(base->name);
-        std::string name_lora_a = name_base + ".lora_a";
-        std::string name_lora_b = name_base + ".lora_b";
-
-        printf("%s : %s [%s]\n", __func__, base->name, ggml_ne_string(base).c_str());
-
-        // context for input tensor
-        std::vector<struct ggml_tensor *> inp_a(adapters.size());
-        std::vector<struct ggml_tensor *> inp_b(adapters.size());
-        struct ggml_init_params params {
-            /*.mem_size   =*/ ggml_tensor_overhead()*(2+adapters.size()*2),
-            /*.mem_buffer =*/ NULL,
-            /*.no_alloc   =*/ true,
-        };
-        struct ggml_context * ctx = ggml_init(params);
-
-        // alloc tensors
-        struct ggml_tensor * inp_base = ggml_new_tensor(ctx, GGML_TYPE_F32, GGML_MAX_DIMS, base->ne);
-        for (size_t i = 0; i < adapters.size(); ++i) {
-            auto t_a = adapters[i]->get_tensor(name_lora_a);
-            auto t_b = adapters[i]->get_tensor(name_lora_b);
-            // TODO: add support for quantized lora
-            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
-                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
-            }
-            inp_a[i] = ggml_dup_tensor(ctx, t_a);
-            inp_b[i] = ggml_dup_tensor(ctx, t_b);
+    void read_raw(void * ptr, size_t size) {
+        if (size == 0) {
+            return;
         }
-        ggml_backend_buffer_t buffer = ggml_backend_alloc_ctx_tensors(ctx, backend);
-
-        // load base tensor to backend buffer
-        base_model.read_tensor_data(name_base, read_buf);
-        if (base->type != GGML_TYPE_F32) {
-            // optionally dequantize it
-            printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
-            auto nels = ggml_nelements(inp_base);
-            ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
-            std::vector<uint8_t> dequant_buf(nels * sizeof(float));
-            qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
-            ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
-        } else {
-            ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
+        errno = 0;
+        std::size_t ret = std::fread(ptr, size, 1, fp);
+        if (ferror(fp)) {
+            die_fmt("read error: %s", strerror(errno));
         }
-
-        // load lora tensors to backend buffer
-        for (size_t i = 0; i < adapters.size(); ++i) {
-            adapters[i]->read_tensor_data(name_lora_a, read_buf);
-            ggml_backend_tensor_set(inp_a[i], read_buf.data(), 0, ggml_nbytes(inp_a[i]));
-            adapters[i]->read_tensor_data(name_lora_b, read_buf);
-            ggml_backend_tensor_set(inp_b[i], read_buf.data(), 0, ggml_nbytes(inp_b[i]));
+        if (ret != 1) {
+            die("unexpectedly reached end of file");
         }
-
-        // build graph
-        struct ggml_cgraph * gf;
-        {
-            static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead();
-            static std::vector<uint8_t> buf(buf_size);
-            struct ggml_init_params params0 = {
-                /*.mem_size   =*/ buf_size,
-                /*.mem_buffer =*/ buf.data(),
-                /*.no_alloc   =*/ true,
-            };
-            struct ggml_context * ctx0 = ggml_init(params0);
-            gf = ggml_new_graph(ctx0);
-            struct ggml_tensor * cur = inp_base;
-            for (size_t i = 0; i < adapters.size(); ++i) {
-                struct ggml_tensor * a_T = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_cast(ctx0, inp_a[i], GGML_TYPE_F32)));
-                struct ggml_tensor * delta = ggml_mul_mat(ctx0, a_T, ggml_cast(ctx0, inp_b[i], GGML_TYPE_F32));
-                // scale
-                const float alpha = adapters[i]->alpha;
-                const float rank  = (float) inp_b[i]->ne[0];
-                const float scale = alpha ? adapters[i]->scale * alpha / rank : adapters[i]->scale;
-                delta = ggml_scale(ctx0, delta, scale);
-                cur = ggml_add(ctx0, delta, cur);
-                printf("%s :   + merging from adapter[%ld] type=%s\n", __func__, i, ggml_type_name(inp_a[i]->type));
-                printf("%s :     input_scale=%f calculated_scale=%f rank=%d\n", __func__, adapters[i]->scale, scale, (int) inp_b[i]->ne[0]);
-            }
-            cur = ggml_cast(ctx0, cur, out->type);
-            printf("%s :   + output type is %s\n", __func__, ggml_type_name(out->type));
-            ggml_build_forward_expand(gf, cur);
-            ggml_free(ctx0);
-        }
-
-        // compute
-        {
-            ggml_gallocr_alloc_graph(allocr, gf);
-            ggml_backend_cpu_set_n_threads(backend, n_threads);
-            ggml_backend_graph_compute(backend, gf);
-        }
-
-        // write data to output file
-        {
-            auto * result = ggml_graph_node(gf, -1);
-            size_t len = ggml_nbytes(result);
-            if (read_buf.size() < len) {
-                read_buf.resize(len);
-            }
-            ggml_backend_tensor_get(result, read_buf.data(), 0, len);
-            fout.write((char* )read_buf.data(), len);
-            zeros(fout, GGML_PAD(len, GGUF_DEFAULT_ALIGNMENT) - len);
-        }
-
-        ggml_free(ctx);
-        ggml_backend_buffer_free(buffer);
     }
 
-    ~lora_merge_ctx() {
-        ggml_gallocr_free(allocr);
-        ggml_backend_free(backend);
-        gguf_free(ctx_out);
-        ggml_free(ctx_out_ggml);
+    std::uint32_t read_u32() {
+        std::uint32_t ret;
+        read_raw(&ret, sizeof(ret));
+        return ret;
+    }
+
+    std::string read_string(std::uint32_t len) {
+        std::vector<char> chars(len);
+        read_raw(chars.data(), len);
+        return std::string(chars.data(), len);
+    }
+
+    void write_raw(const void * ptr, size_t size) {
+        if (size == 0) {
+            return;
+        }
+        errno = 0;
+        size_t ret = std::fwrite(ptr, size, 1, fp);
+        if (ret != 1) {
+            die_fmt("write error: %s", strerror(errno));
+        }
+    }
+
+    void write_u32(std::uint32_t val) {
+        write_raw(&val, sizeof(val));
+    }
+
+    bool eof() {
+        return tell() >= size;
+    }
+
+    ~llama_file() {
+        if (fp) {
+            std::fclose(fp);
+        }
     }
 };
 
-static void print_usage(int, char ** argv) {
-    printf("\nexample usage:\n");
-    printf("\n  %s -m base-model.gguf --lora lora-file.gguf -o merged-model-f16.gguf\n", argv[0]);
-    printf("\nNOTE: output model is F16\n");
+static struct export_lora_params get_default_export_lora_params() {
+    struct export_lora_params result;
+    result.fn_model_base = "";
+    result.fn_model_out  = "";
+    result.n_threads = GGML_DEFAULT_N_THREADS;
+    return result;
+}
+
+static void export_lora_print_usage(int /*argc*/, char ** argv, const struct export_lora_params * params) {
+    fprintf(stderr, "usage: %s [options]\n", argv[0]);
+    fprintf(stderr, "\n");
+    fprintf(stderr, "options:\n");
+    fprintf(stderr, "  -h, --help                         show this help message and exit\n");
+    fprintf(stderr, "  -m FNAME, --model-base FNAME       model path from which to load base model (default '%s')\n", params->fn_model_base.c_str());
+    fprintf(stderr, "  -o FNAME, --model-out FNAME        path to save exported model (default '%s')\n", params->fn_model_out.c_str());
+    fprintf(stderr, "  -l FNAME, --lora FNAME             apply LoRA adapter\n");
+    fprintf(stderr, "  -s FNAME S, --lora-scaled FNAME S  apply LoRA adapter with user defined scaling S\n");
+    fprintf(stderr, "  -t N, --threads N                  number of threads to use during computation (default: %d)\n", params->n_threads);
+}
+
+static bool export_lora_params_parse(int argc, char ** argv, struct export_lora_params * params) {
+    bool invalid_param = false;
+    std::string arg;
+    struct export_lora_params default_params = get_default_export_lora_params();
+    const std::string arg_prefix = "--";
+
+    for (int i = 1; i < argc; i++) {
+        arg = argv[i];
+        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
+            std::replace(arg.begin(), arg.end(), '_', '-');
+        }
+
+        if (arg == "-m" || arg == "--model-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_base = argv[i];
+        } else if (arg == "-o" || arg == "--model-out") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->fn_model_out = argv[i];
+        } else if (arg == "-l" || arg == "--lora") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            lora.scale = 1.0f;
+            params->lora.push_back(lora);
+        } else if (arg == "-s" || arg == "--lora-scaled") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            struct lora_info lora;
+            lora.filename = argv[i];
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            lora.scale = std::stof(argv[i]);
+            params->lora.push_back(lora);
+        } else if (arg == "-t" || arg == "--threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params->n_threads = std::stoi(argv[i]);
+            if (params->n_threads <= 0) {
+                params->n_threads = std::thread::hardware_concurrency();
+            }
+        } else {
+            fprintf(stderr, "error: unknown argument: '%s'\n", arg.c_str());
+            export_lora_print_usage(argc, argv, &default_params);
+            exit(1);
+        }
+    }
+
+    if (params->fn_model_base == default_params.fn_model_base) {
+        fprintf(stderr, "error: please specify a filename for model-base.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (params->fn_model_out == default_params.fn_model_out) {
+        fprintf(stderr, "error: please specify a filename for model-out.\n");
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: '%s'\n", arg.c_str());
+        export_lora_print_usage(argc, argv, &default_params);
+        exit(1);
+    }
+    return true;
+}
+
+static void free_lora(struct lora_data * lora) {
+    if (lora->ctx != NULL) {
+        ggml_free(lora->ctx);
+    }
+    delete lora;
+}
+
+static struct lora_data * load_lora(struct lora_info * info) {
+    struct lora_data * result = new struct lora_data;
+    result->info = *info;
+    result->ctx = NULL;
+    result->lora_r     = 1;
+    result->lora_alpha = 1;
+
+    struct llama_file file(info->filename.c_str(), "rb");
+    if (file.fp == NULL) {
+        fprintf(stderr, "warning: Could not open lora adapter '%s'. Ignoring this adapter.\n",
+            info->filename.c_str());
+        free_lora(result);
+        return NULL;
+    }
+
+    struct ggml_init_params params_ggml;
+    params_ggml.mem_size   = ggml_tensor_overhead() * GGML_DEFAULT_GRAPH_SIZE;
+    params_ggml.mem_buffer = NULL;
+    params_ggml.no_alloc   = true;
+    result->ctx = ggml_init(params_ggml);
+
+    uint32_t magic   = file.read_u32();
+    if (magic != LLAMA_FILE_MAGIC_GGLA) {
+        die_fmt("unexpected lora header file magic in '%s'", info->filename.c_str());
+    }
+    uint32_t version = file.read_u32();
+    if (version != 1) {
+        die_fmt("unexpected lora file version '%u' in '%s'", (unsigned) version, info->filename.c_str());
+    }
+    result->lora_r     = file.read_u32();
+    result->lora_alpha = file.read_u32();
+    // read tensor infos from file
+    std::vector<char> name_buf;
+    std::vector<struct ggml_tensor *> tensors;
+    std::vector<size_t> tensors_offset;
+    size_t total_nbytes_pad = 0;
+    while(!file.eof()) {
+        int64_t ne[4]   = {1,1,1,1};
+        uint32_t n_dims  = file.read_u32();
+        uint32_t namelen = file.read_u32();
+        uint32_t type    = file.read_u32();
+        for (uint32_t k = 0; k < n_dims; ++k) {
+            ne[k] = (int64_t)file.read_u32();
+        }
+        name_buf.clear();
+        name_buf.resize(namelen + 1, '\0');
+        file.read_raw(name_buf.data(), namelen);
+        file.seek((0-file.tell()) & 31, SEEK_CUR);
+        size_t offset = file.tell();
+        struct ggml_tensor * tensor = ggml_new_tensor(result->ctx, (enum ggml_type) type, n_dims, ne);
+        ggml_set_name(tensor, name_buf.data());
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        total_nbytes_pad += nbytes_pad;
+        tensors.push_back(tensor);
+        tensors_offset.push_back(offset);
+        file.seek(nbytes, SEEK_CUR);
+    }
+    // read tensor data
+    result->data.resize(total_nbytes_pad);
+    size_t data_offset = 0;
+    for (size_t i = 0; i < tensors.size(); ++i) {
+        struct ggml_tensor * tensor = tensors[i];
+        size_t offset     = tensors_offset[i];
+        size_t nbytes     = ggml_nbytes(tensor);
+        size_t nbytes_pad = ggml_nbytes_pad(tensor);
+        file.seek(offset, SEEK_SET);
+        tensor->data = result->data.data() + data_offset;
+        file.read_raw(tensor->data, nbytes);
+        data_offset += nbytes_pad;
+    }
+    return result;
+}
+
+
+static struct ggml_cgraph * build_graph_lora(
+    struct ggml_context * ctx,
+    struct ggml_tensor * tensor,
+    struct ggml_tensor * lora_a,
+    struct ggml_tensor * lora_b,
+    float scaling
+) {
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
+    if (scaling != 1.0f) {
+        ab = ggml_scale(ctx, ab, scaling);
+    }
+    struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);
+
+    struct ggml_cgraph * gf = ggml_new_graph(ctx);
+    ggml_build_forward_expand (gf, res);
+    return gf;
+}
+
+static bool apply_lora(struct ggml_tensor * tensor, struct lora_data * lora, int n_threads) {
+    if (lora->ctx == NULL) {
+        return false;
+    }
+    std::string name = ggml_get_name(tensor);
+    std::string name_a = name + std::string(".loraA");
+    std::string name_b = name + std::string(".loraB");
+    struct ggml_tensor * lora_a = ggml_get_tensor(lora->ctx, name_a.c_str());
+    struct ggml_tensor * lora_b = ggml_get_tensor(lora->ctx, name_b.c_str());
+    if (lora_a == NULL || lora_b == NULL) {
+        return false;
+    }
+
+    float scaling = lora->info.scale * (float)lora->lora_alpha / (float)lora->lora_r;
+
+    struct ggml_init_params params;
+    params.mem_size   = GGML_OBJECT_SIZE + ggml_graph_overhead() + ggml_tensor_overhead()*4 + GGML_MEM_ALIGN*5;
+    params.mem_buffer = NULL;
+    params.no_alloc   = true;
+    struct ggml_context * ctx = NULL;
+    struct ggml_gallocr * alloc = NULL;
+    struct ggml_cgraph  * gf = NULL;
+
+    ctx   = ggml_init(params);
+    alloc = ggml_gallocr_new(ggml_backend_cpu_buffer_type());
+    gf    = build_graph_lora(ctx, tensor, lora_a, lora_b, scaling);
+
+    ggml_gallocr_alloc_graph(alloc, gf);
+
+    struct ggml_cplan cplan = ggml_graph_plan(gf, n_threads);
+    static std::vector<uint8_t> data_work;
+    data_work.resize(cplan.work_size);
+    cplan.work_data = data_work.data();
+
+    ggml_graph_compute(gf, &cplan);
+
+    ggml_gallocr_free(alloc);
+    ggml_free(ctx);
+    return true;
+}
+
+static void export_lora(struct export_lora_params * params) {
+    // load all loras
+    std::vector<struct lora_data *> loras;
+    for (size_t i = 0; i < params->lora.size(); ++i) {
+        struct lora_data * lora = load_lora(&params->lora[i]);
+        if (lora != NULL) {
+            loras.push_back(lora);
+        }
+    }
+    if (loras.size() == 0) {
+        fprintf(stderr, "warning: no lora adapters will be applied.\n");
+    }
+
+    // open input file
+    struct llama_file fin(params->fn_model_base.c_str(), "rb");
+    if (!fin.fp) {
+        die_fmt("Could not open file '%s'\n", params->fn_model_base.c_str());
+    }
+
+    // open base model gguf, read tensors without their data
+    struct ggml_context * ctx_in;
+    struct gguf_init_params params_gguf;
+    params_gguf.no_alloc = true;
+    params_gguf.ctx      = &ctx_in;
+    struct gguf_context * gguf_in = gguf_init_from_file(params->fn_model_base.c_str(), params_gguf);
+
+    // create new gguf
+    struct gguf_context * gguf_out = gguf_init_empty();
+
+    // copy meta data from base model: kv and tensors
+    gguf_set_kv(gguf_out, gguf_in);
+    int n_tensors = gguf_get_n_tensors(gguf_in);
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+        gguf_add_tensor(gguf_out, tensor);
+    }
+
+    // create output file
+    struct llama_file fout(params->fn_model_out.c_str(), "wb");
+    if (!fout.fp) {
+        die_fmt("Could not create file '%s'\n", params->fn_model_out.c_str());
+    }
+
+    // write gguf meta data
+    std::vector<uint8_t> meta;
+    meta.resize(gguf_get_meta_size(gguf_out));
+    gguf_get_meta_data(gguf_out, meta.data());
+    fout.write_raw(meta.data(), meta.size());
+
+    std::vector<uint8_t> data;
+    std::vector<uint8_t> padding;
+    for (int i=0; i < n_tensors; ++i) {
+        const char * name = gguf_get_tensor_name(gguf_in, i);
+        struct ggml_tensor * tensor = ggml_get_tensor(ctx_in, name);
+
+        // read tensor data
+        data.resize(ggml_nbytes(tensor));
+        tensor->data = data.data();
+        size_t offset = gguf_get_tensor_offset(gguf_in, i);
+        fin.seek(offset + meta.size(), SEEK_SET);
+        fin.read_raw(data.data(), data.size());
+
+        // apply all loras
+        for (size_t k = 0; k < loras.size(); ++k) {
+            apply_lora(tensor, loras[k], params->n_threads);
+        }
+
+        // write tensor data + padding
+        padding.clear();
+        padding.resize(GGML_PAD(data.size(), gguf_get_alignment(gguf_out)) - data.size(), 0);
+
+        GGML_ASSERT(fout.tell() == offset + meta.size());
+        // fout.seek(offset + meta.size(), SEEK_SET);
+        fout.write_raw(data.data(), data.size());
+        fout.write_raw(padding.data(), padding.size());
+
+        if (i % 2 == 0) {
+            printf(".");
+        }
+    }
     printf("\n");
+
+    // close gguf
+    gguf_free(gguf_out);
+    gguf_free(gguf_in);
+
+    // free loras
+    for (size_t i = 0; i < loras.size(); ++i) {
+        free_lora(loras[i]);
+    }
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    struct export_lora_params params = get_default_export_lora_params();
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
+    if (!export_lora_params_parse(argc, argv, &params)) {
         return 1;
     }
 
-    g_verbose = (params.verbosity == 1);
-    try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.cpuparams.n_threads);
-        ctx.run_merge();
-    } catch (const std::exception & err) {
-        fprintf(stderr, "%s\n", err.what());
-        exit(EXIT_FAILURE);
-    }
-
-    printf("done, output file is %s\n", params.lora_outfile.c_str());
+    export_lora(&params);
 
     return 0;
 }

examples/finetune/CMakeLists.txt

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

examples/finetune/README.md

@@ -0,0 +1,90 @@
+# finetune
+
+Basic usage instructions:
+
+```bash
+# get training data
+wget https://raw.githubusercontent.com/brunoklein99/deep-learning-notes/master/shakespeare.txt
+
+# finetune LORA adapter
+./bin/llama-finetune \
+        --model-base open-llama-3b-v2-q8_0.gguf \
+        --checkpoint-in  chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf \
+        --checkpoint-out chk-lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.gguf \
+        --lora-out lora-open-llama-3b-v2-q8_0-shakespeare-ITERATION.bin \
+        --train-data "shakespeare.txt" \
+        --save-every 10 \
+        --threads 6 --adam-iter 30 --batch 4 --ctx 64 \
+        --use-checkpointing
+
+# predict
+./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+```
+
+**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
+The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
+So in above example after 10 iterations these files will be written:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-10.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+After 10 more iterations:
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-20.gguf
+- chk-lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.gguf
+- lora-open-llama-3b-v2-q8_0-shakespeare-20.bin
+- lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
+
+Checkpoint files (`--checkpoint-in FN`, `--checkpoint-out FN`) store the training process. When the input checkpoint file does not exist, it will begin finetuning a new randomly initialized adapter.
+
+llama.cpp compatible LORA adapters will be saved with filename specified by `--lora-out FN`.
+These LORA adapters can then be used by `llama-cli` together with the base model, like in the 'predict' example command above.
+
+In `llama-cli` you can also load multiple LORA adapters, which will then be mixed together.
+
+For example if you have two LORA adapters `lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin` and `lora-open-llama-3b-v2-q8_0-bible-LATEST.bin`, you can mix them together like this:
+
+```bash
+./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \
+  --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin \
+  --lora lora-open-llama-3b-v2-q8_0-bible-LATEST.bin
+```
+
+You can change how strong each LORA adapter is applied to the base model by using `--lora-scaled FN SCALE` instead of `--lora FN`.
+
+For example to apply 40% of the 'shakespeare' LORA adapter, 80% of the 'bible' LORA adapter and 100% of yet another one:
+
+```bash
+./bin/llama-cli -m open-llama-3b-v2-q8_0.gguf \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin 0.4 \
+  --lora-scaled lora-open-llama-3b-v2-q8_0-bible-LATEST.bin 0.8 \
+  --lora lora-open-llama-3b-v2-q8_0-yet-another-one-LATEST.bin
+```
+
+The scale numbers don't need to add up to one, and you can also use numbers greater than 1 to further increase the influence of an adapter. But making the values too big will sometimes result in worse output. Play around to find good values.
+
+Gradient checkpointing reduces the memory requirements by ~50% but increases the runtime.
+If you have enough RAM, you can make finetuning a bit faster by disabling checkpointing with `--no-checkpointing`.
+
+The default LORA rank can be specified with `--lora-r N`.
+The LORA rank can be configured for each model tensor type separately with these command line options:
+
+```bash
+  --lora-r N                 LORA r: default rank. Also specifies resulting scaling together with lora-alpha. (default 4)
+  --rank-att-norm N          LORA rank for attention norm tensor (default 1)
+  --rank-ffn-norm N          LORA rank for feed-forward norm tensor (default 1)
+  --rank-out-norm N          LORA rank for output norm tensor (default 1)
+  --rank-tok-embd N          LORA rank for token embeddings tensor (default 4)
+  --rank-out N               LORA rank for output tensor (default 4)
+  --rank-wq N                LORA rank for wq tensor (default 4)
+  --rank-wk N                LORA rank for wk tensor (default 4)
+  --rank-wv N                LORA rank for wv tensor (default 4)
+  --rank-wo N                LORA rank for wo tensor (default 4)
+  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
+  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
+  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
+```
+
+The LORA rank of 'norm' tensors should always be 1.
+
+To see all available options use `finetune --help`.

examples/finetune/convert_finetune_checkpoint_to_gguf.py

@@ -0,0 +1,487 @@
+#!/usr/bin/env python3
+# finetune checkpoint --> gguf conversion
+
+import argparse
+import gguf
+import struct
+import numpy as np
+from pathlib import Path
+
+# gguf constants
+LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
+LLM_KV_OPTIMIZER_TYPE_ADAM  = "adam"
+LLM_KV_OPTIMIZER_TYPE_LBFGS = "lbfgs"
+LLM_KV_OPTIMIZER_FILE_VERSION               = "optimizer.file_version"
+LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT     = "optimizer.convergence_past_count"
+LLM_KV_OPTIMIZER_PARAMETER_COUNT            = "optimizer.parameter_count"
+LLM_KV_OPTIMIZER_ITERATION_COUNT            = "optimizer.iteration_count"
+LLM_KV_OPTIMIZER_JUST_INITIALIZED           = "optimizer.just_initialized"
+LLM_KV_OPTIMIZER_ADAM_BEST_LOSS             = "optimizer.adam.best_loss"
+LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS         = "optimizer.adam.previous_loss"
+LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT  = "optimizer.adam.no_improvement_count"
+LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT = "optimizer.lbfgs.approx_hessian_count"
+LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS            = "optimizer.lbfgs.best_loss"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP     = "optimizer.lbfgs.line_search_step"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J        = "optimizer.lbfgs.line_search_j"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K        = "optimizer.lbfgs.line_search_k"
+LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END      = "optimizer.lbfgs.line_search_end"
+LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT = "optimizer.lbfgs.no_improvement_count"
+
+LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS    = "optimizer.adam.first_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS   = "optimizer.adam.second_moments"
+LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES = "optimizer.adam.past_loss_values"
+
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS  = "optimizer.lbfgs.current_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS = "optimizer.lbfgs.previous_parameters"
+LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS   = "optimizer.lbfgs.current_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS  = "optimizer.lbfgs.previous_gradients"
+LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION    = "optimizer.lbfgs.search_direction"
+LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES    = "optimizer.lbfgs.past_loss_values"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA        = "optimizer.lbfgs.memory_alpha"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS           = "optimizer.lbfgs.memory_ys"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S            = "optimizer.lbfgs.memory_s"
+LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y            = "optimizer.lbfgs.memory_y"
+
+LLM_KV_TRAINING_TYPE_TRAIN_MODEL   = "train_model"
+LLM_KV_TRAINING_TYPE_FINETUNE_LORA = "finetune_lora"
+LLM_KV_TRAINING_TYPE               = "training.type"
+LLM_KV_TRAINING_FILE_VERSION       = "training.file_version"
+LLM_KV_TRAINING_ITERATION_COUNT    = "training.iteration_count"
+LLM_KV_TRAINING_SAMPLE_COUNT       = "training.sample_count"
+LLM_KV_TRAINING_TOKEN_COUNT        = "training.token_count"
+
+LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD  = "training.lora.rank.token_embd"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM = "training.lora.rank.output_norm"
+LLM_KV_TRAINING_LORA_RANK_OUTPUT      = "training.lora.rank.output"
+LLM_KV_TRAINING_LORA_RANK_ATTN_NORM   = "training.lora.rank.attn_norm"
+LLM_KV_TRAINING_LORA_RANK_ATTN_Q      = "training.lora.rank.attn_q"
+LLM_KV_TRAINING_LORA_RANK_ATTN_K      = "training.lora.rank.attn_k"
+LLM_KV_TRAINING_LORA_RANK_ATTN_V      = "training.lora.rank.attn_v"
+LLM_KV_TRAINING_LORA_RANK_ATTN_OUT    = "training.lora.rank.attn_output"
+LLM_KV_TRAINING_LORA_RANK_FFN_NORM    = "training.lora.rank.ffn_norm"
+LLM_KV_TRAINING_LORA_RANK_FFN_GATE    = "training.lora.rank.ffn_gate"
+LLM_KV_TRAINING_LORA_RANK_FFN_DOWN    = "training.lora.rank.ffn_down"
+LLM_KV_TRAINING_LORA_RANK_FFN_UP      = "training.lora.rank.ffn_up"
+
+class Tensor:
+    def __init__(self, dtype='f', ne=None):
+        if ne is None:
+            ne = []
+        self.dtype = dtype
+        self.ne = ne
+        self.nbytes = 0
+        if self.dtype == 'f':
+            if len(self.ne) == 0:
+                self.nbytes = 0
+            else:
+                self.nbytes = int(np.product(self.ne)) * 4
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+    def load(self, data, offset):
+        nd = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        namelen = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        dtype = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        assert(nd == len(self.ne))
+        ne = []
+        for d in range(nd):
+            n = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+            ne.append(n)
+
+        if tuple(ne) != tuple(self.ne):
+            raise ValueError(f"Tensor.load: Expected number of elements {str(self.ne)} does not match what is read from file {str(ne)}")
+
+        if self.dtype == 'f':
+            assert(dtype == 0)
+        else:
+            raise ValueError(f"Unhandled data type '{self.dtype}'")
+
+        self.name = bytes(data[offset:offset+namelen]); offset += namelen
+        # 32-byte alignment
+        offset += (0 - offset) & 31
+        self.data = data[offset:offset+self.nbytes]
+        offset += self.nbytes
+        return offset
+
+    def max_storage_size(self):
+        result = 0
+        result += 4 # nd
+        result += 4 # namelen
+        result += 4 # dtype
+        result += len(self.ne)*8 # ne
+        result += 48 # name (maximum as of commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9)
+        result += 31 # 32-byte alignment
+        result += self.nbytes
+        return result
+
+    def save_gguf(self, gguf_writer, name):
+        gguf_writer.add_tensor(
+            name=name,
+            tensor=self.data,
+            raw_shape=np.array(list(reversed(self.ne))),
+            raw_dtype=gguf.GGMLQuantizationType.F32)
+
+class OptimizationContext:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]
+        offset += 4
+
+        if self.version != 1:
+            raise ValueError('Invalid version of optimization context in checkpoint file')
+
+        self.past    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.lbfgs_m = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.nx      = struct.unpack('N',  bytes(data[offset:offset + 8]))[0];  offset += 8
+        self.iter    = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.just_initialized = bool(struct.unpack('<i', bytes(data[offset:offset + 4]))[0]);  offset += 4
+
+        self.adam_m  = Tensor('f', [self.nx])
+        self.adam_v  = Tensor('f', [self.nx])
+        self.adam_pf = Tensor('f', [self.past] if self.past > 0 else [])
+
+        self.lbfgs_x    = Tensor('f', [self.nx])
+        self.lbfgs_xp   = Tensor('f', [self.nx])
+        self.lbfgs_g    = Tensor('f', [self.nx])
+        self.lbfgs_gp   = Tensor('f', [self.nx])
+        self.lbfgs_d    = Tensor('f', [self.nx])
+        self.lbfgs_pf   = Tensor('f', [self.past] if self.past > 0 else [])
+        self.lbfgs_lmal = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lmys = Tensor('f', [self.lbfgs_m])
+        self.lbfgs_lms  = Tensor('f', [self.nx, self.lbfgs_m])
+        self.lbfgs_lmy  = Tensor('f', [self.nx, self.lbfgs_m])
+
+        # forgot to save type in version 1:
+        # guess self.type from number of remaining bytes
+        size_type_0 = 12 + sum([t.max_storage_size() for t in
+                                [self.adam_m, self.adam_v]
+                                +([self.adam_pf] if (self.past > 0) else [])])
+        size_type_1 = 24 + sum([t.max_storage_size() for t in
+                                [self.lbfgs_x, self.lbfgs_xp, self.lbfgs_g,
+                                 self.lbfgs_gp, self.lbfgs_d, self.lbfgs_pf,
+                                 self.lbfgs_lmal, self.lbfgs_lmys,
+                                 self.lbfgs_lms, self.lbfgs_lmy]
+                                 +([self.lbfgs_pf] if (self.past > 0) else [])])
+        # due to alignment padding the size might not by exact
+        # but the difference in size for both types is significant,
+        # so we can just use whichever is closest
+        remaining = len(data) - offset
+        if abs(remaining - size_type_0) < abs(remaining - size_type_1):
+            self.type = 0
+        else:
+            self.type = 1
+
+        if self.type == 0:
+            offset = self.adam_m.load(data, offset)
+            offset = self.adam_v.load(data, offset)
+            offset = self.adam_pf.load(data,offset)
+
+            self.adam_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_fx_prev          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.adam_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        elif self.type == 1:
+            offset = self.lbfgs_x.load(data, offset)
+            offset = self.lbfgs_xp.load(data, offset)
+            offset = self.lbfgs_g.load(data, offset)
+            offset = self.lbfgs_gp.load(data, offset)
+            offset = self.lbfgs_d.load(data, offset)
+            offset = self.lbfgs_pf.load(data, offset)
+            offset = self.lbfgs_lmal.load(data, offset)
+            offset = self.lbfgs_lmys.load(data, offset)
+            offset = self.lbfgs_lms.load(data, offset)
+            offset = self.lbfgs_lmy.load(data, offset)
+
+            self.lbfgs_fx_best          = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_step             = struct.unpack('<f', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_j                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_k                = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_end              = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+            self.lbfgs_n_no_improvement = struct.unpack('<i', bytes(data[offset:offset + 4]))[0];  offset += 4
+
+        else:
+            raise ValueError(f"Invalid optimizer type '{self.type}'")
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_FILE_VERSION, 0)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_CONVERGENCE_PAST_COUNT, self.past)
+        gguf_writer.add_uint64(LLM_KV_OPTIMIZER_PARAMETER_COUNT, self.nx)
+        gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ITERATION_COUNT, self.iter)
+        gguf_writer.add_bool(LLM_KV_OPTIMIZER_JUST_INITIALIZED, self.just_initialized)
+
+        if self.type == 0:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_BEST_LOSS, self.adam_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS, self.adam_fx_prev)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_ADAM_NO_IMPROVEMENT_COUNT, self.adam_n_no_improvement)
+
+            self.adam_m.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_FIRST_MOMENTS)
+            self.adam_v.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS)
+            if self.past > 0:
+                self.adam_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES)
+
+        elif self.type == 1:
+            gguf_writer.add_string(LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, self.lbfgs_m)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS, self.lbfgs_fx_best)
+            gguf_writer.add_float32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_STEP, self.lbfgs_step)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_J, self.lbfgs_j)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_K, self.lbfgs_k)
+            gguf_writer.add_int32(LLM_KV_OPTIMIZER_LBFGS_LINE_SEARCH_END, self.lbfgs_end)
+            gguf_writer.add_uint32(LLM_KV_OPTIMIZER_LBFGS_NO_IMPROVEMENT_COUNT, self.lbfgs_n_no_improvement)
+
+            self.lbfgs_x.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_PARAMETERS)
+            self.lbfgs_xp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_PARAMETERS)
+            self.lbfgs_g.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_CURRENT_GRADIENTS)
+            self.lbfgs_gp.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PREVIOUS_GRADIENTS)
+            self.lbfgs_d.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_SEARCH_DIRECTION)
+            if self.past > 0:
+                self.lbfgs_pf.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_PAST_LOSS_VALUES)
+            self.lbfgs_lmal.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_ALPHA)
+            self.lbfgs_lmys.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_YS)
+            self.lbfgs_lms.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_S)
+            self.lbfgs_lmy.save_gguf(gguf_writer, name=LLM_TENSOR_OPTIMIZER_LBFGS_MEMORY_Y)
+        else:
+            raise ValueError('Unknown optimizer type')
+
+class LoraParams:
+    def __init__(self):
+        pass
+
+    def load(self, data, offset):
+        self.n_rank_attention_norm  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wq              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wk              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wv              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_wo              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_ffn_norm        = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w1              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w2              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_w3              = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_tok_embeddings  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_norm            = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rank_output          = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_TOKEN_EMBD,  self.n_rank_tok_embeddings)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT_NORM, self.n_rank_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_OUTPUT,      self.n_rank_output)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_NORM,   self.n_rank_attention_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_Q,      self.n_rank_wq)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_K,      self.n_rank_wk)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_V,      self.n_rank_wv)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,    self.n_rank_wo)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_NORM,    self.n_rank_ffn_norm)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_GATE,    self.n_rank_w1)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,    self.n_rank_w2)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_LORA_RANK_FFN_UP,      self.n_rank_w3)
+
+class ModelParams:
+    def __init__(self, n_ff = None):
+        self.n_ff = n_ff
+
+    def load(self, data, offset):
+        self.n_vocab = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_embd  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_mult  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_head  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_layer = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        self.n_rot   = struct.unpack('<I', bytes(data[offset:offset + 4]))[0];  offset += 4
+        return offset
+
+    def get_n_ff(self):
+        if self.n_ff is None:
+            # struct my_llama_model::get_n_ff in train-text-from-scratch.cpp commit 3b5515bbe0e2224425986ba24f1f5d84aa38dce9
+            return ((2*(4*self.n_embd)//3 + self.n_mult - 1)//self.n_mult)*self.n_mult
+        else:
+            return self.n_ff
+
+    def save_gguf(self, gguf_writer):
+        # self.n_vocab not saved
+        gguf_writer.add_embedding_length(self.n_embd)
+        gguf_writer.add_head_count(self.n_head)
+        gguf_writer.add_block_count(self.n_layer)
+        gguf_writer.add_rope_dimension_count(self.n_rot)
+        gguf_writer.add_feed_forward_length(self.get_n_ff())
+
+def tensor_name(key, bid=None, suffix=".weight"):
+    return gguf.TENSOR_NAMES[key].format(bid=bid) + suffix
+
+class Layer:
+    def __init__(self, params, lora_params, bid):
+        self.bid = bid
+        self.att_norm_a = Tensor('f', [lora_params.n_rank_attention_norm, params.n_embd])
+        self.att_norm_b = Tensor('f', [lora_params.n_rank_attention_norm, 1])
+        self.wq_a       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wq_b       = Tensor('f', [lora_params.n_rank_wq, params.n_embd])
+        self.wk_a       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wk_b       = Tensor('f', [lora_params.n_rank_wk, params.n_embd])
+        self.wv_a       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wv_b       = Tensor('f', [lora_params.n_rank_wv, params.n_embd])
+        self.wo_a       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.wo_b       = Tensor('f', [lora_params.n_rank_wo, params.n_embd])
+        self.ffn_norm_a = Tensor('f', [lora_params.n_rank_ffn_norm, params.n_embd])
+        self.ffn_norm_b = Tensor('f', [lora_params.n_rank_ffn_norm, 1])
+        self.w1_a       = Tensor('f', [lora_params.n_rank_w1, params.n_embd])
+        self.w1_b       = Tensor('f', [lora_params.n_rank_w1, params.get_n_ff()])
+        self.w2_a       = Tensor('f', [lora_params.n_rank_w2, params.get_n_ff()])
+        self.w2_b       = Tensor('f', [lora_params.n_rank_w2, params.n_embd])
+        self.w3_a       = Tensor('f', [lora_params.n_rank_w3, params.n_embd])
+        self.w3_b       = Tensor('f', [lora_params.n_rank_w3, params.get_n_ff()])
+
+    def load(self, data, offset):
+        offset = self.att_norm_a.load(data, offset)
+        offset = self.att_norm_b.load(data, offset)
+        offset = self.wq_a.load(data, offset)
+        offset = self.wq_b.load(data, offset)
+        offset = self.wk_a.load(data, offset)
+        offset = self.wk_b.load(data, offset)
+        offset = self.wv_a.load(data, offset)
+        offset = self.wv_b.load(data, offset)
+        offset = self.wo_a.load(data, offset)
+        offset = self.wo_b.load(data, offset)
+        offset = self.ffn_norm_a.load(data, offset)
+        offset = self.ffn_norm_b.load(data, offset)
+        offset = self.w1_a.load(data, offset)
+        offset = self.w1_b.load(data, offset)
+        offset = self.w2_a.load(data, offset)
+        offset = self.w2_b.load(data, offset)
+        offset = self.w3_a.load(data, offset)
+        offset = self.w3_b.load(data, offset)
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.att_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_a"))
+        self.att_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_NORM, self.bid, ".weight.lora_b"))
+        self.wq_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_a"))
+        self.wq_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_Q,    self.bid, ".weight.lora_b"))
+        self.wk_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_a"))
+        self.wk_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_K,    self.bid, ".weight.lora_b"))
+        self.wv_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_a"))
+        self.wv_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_V,    self.bid, ".weight.lora_b"))
+        self.wo_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_a"))
+        self.wo_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.ATTN_OUT,  self.bid, ".weight.lora_b"))
+        self.ffn_norm_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_a"))
+        self.ffn_norm_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_NORM,  self.bid, ".weight.lora_b"))
+        self.w1_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_a"))
+        self.w1_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_GATE,  self.bid, ".weight.lora_b"))
+        self.w2_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_a"))
+        self.w2_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_DOWN,  self.bid, ".weight.lora_b"))
+        self.w3_a.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_a"))
+        self.w3_b.save_gguf      (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.FFN_UP,    self.bid, ".weight.lora_b"))
+
+class LoraModel:
+    def __init__(self, n_ff = None):
+        self.params = ModelParams(n_ff = n_ff)
+        self.lora_params = LoraParams()
+        self.layers = []
+
+    def load(self, data, offset):
+        offset = self.params.load(data, offset)
+        offset = self.lora_params.load(data, offset)
+
+        self.tok_embd_a = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_embd])
+        self.tok_embd_b = Tensor('f', [self.lora_params.n_rank_tok_embeddings, self.params.n_vocab])
+        self.norm_a     = Tensor('f', [self.lora_params.n_rank_norm, self.params.n_embd])
+        self.norm_b     = Tensor('f', [self.lora_params.n_rank_norm, 1])
+        self.output_a   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_embd])
+        self.output_b   = Tensor('f', [self.lora_params.n_rank_output, self.params.n_vocab])
+
+        offset = self.tok_embd_a.load(data, offset)
+        offset = self.tok_embd_b.load(data, offset)
+        offset = self.norm_a.load(data, offset)
+        offset = self.norm_b.load(data, offset)
+        offset = self.output_a.load(data, offset)
+        offset = self.output_b.load(data, offset)
+
+        self.layers.clear()
+        for bid in range(self.params.n_layer):
+            layer = Layer(self.params, self.lora_params, bid)
+            offset = layer.load(data, offset)
+            self.layers.append(layer)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        self.params.save_gguf(gguf_writer)
+        self.lora_params.save_gguf(gguf_writer)
+
+        self.tok_embd_a.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_a"))
+        self.tok_embd_b.save_gguf(gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD,  suffix=".weight.lora_b"))
+        self.norm_a.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_a"))
+        self.norm_b.save_gguf    (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT_NORM, suffix=".weight.lora_b"))
+        self.output_a.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_a"))
+        self.output_b.save_gguf  (gguf_writer, name=tensor_name(gguf.MODEL_TENSOR.OUTPUT,      suffix=".weight.lora_b"))
+
+        for layer in self.layers:
+            layer.save_gguf(gguf_writer)
+
+class LoraCheckpoint:
+    def __init__(self, n_ff = None):
+        self.model = LoraModel(n_ff = n_ff)
+        self.opt_ctx = OptimizationContext()
+
+    def load(self, data, offset):
+        magic   = bytes(reversed(data[offset:offset + 4])); offset += 4
+        if magic != b'ggcl':
+            raise ValueError(f"File header magic indicates, that this is no finetune-lora checkpoint file. Expected 'ggcl', Got '{str(magic)}'")
+
+        self.version = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        if self.version != 0:
+            raise ValueError('Invalid version of checkpoint file')
+
+        self.train_its     = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_samples = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+        self.train_tokens  = struct.unpack('<I', bytes(data[offset:offset + 4]))[0]; offset += 4
+
+        offset = self.model.load(data, offset)
+        offset = self.opt_ctx.load(data, offset)
+
+        return offset
+
+    def save_gguf(self, gguf_writer):
+        gguf_writer.add_file_type(gguf.GGMLQuantizationType.F32)
+        gguf_writer.add_layer_norm_rms_eps(1e-5)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_FILE_VERSION,    0)
+        gguf_writer.add_string(LLM_KV_TRAINING_TYPE,            LLM_KV_TRAINING_TYPE_FINETUNE_LORA)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_ITERATION_COUNT, self.train_its)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_SAMPLE_COUNT,    self.train_samples)
+        gguf_writer.add_uint32(LLM_KV_TRAINING_TOKEN_COUNT,     self.train_tokens)
+        self.model.save_gguf(gguf_writer)
+        self.opt_ctx.save_gguf(gguf_writer)
+
+def handle_args():
+    parser = argparse.ArgumentParser(description = 'Convert finetune checkpoints to GGUF')
+    parser.add_argument('--input',  '-i', type = Path, help = 'Input finetune checkpoint filename', required=True)
+    parser.add_argument('--output', '-o', type = Path, help = 'Output GGUF filename', required=True)
+    parser.add_argument('--ff', type = int, help = "Feedforward size, if not provided compute from n_mult. Provide this if you get 'ValueError: Tensor.load: Expected number of elements does not match what is read from file'", required=False)
+    return parser.parse_args()
+
+def main():
+    cfg = handle_args()
+    print(cfg)
+    data = np.memmap(cfg.input, mode = 'r')
+    chk = LoraCheckpoint(n_ff = cfg.ff)
+    offset = 0
+    offset = chk.load(data, offset)
+    # we should have read all available data
+    assert(offset == len(data))
+
+    gguf_writer = gguf.GGUFWriter(cfg.output, gguf.MODEL_ARCH_NAMES[gguf.MODEL_ARCH.LLAMA], use_temp_file = False)
+    chk.save_gguf(gguf_writer)
+    print("    gguf: write header")
+    gguf_writer.write_header_to_file()
+    print("    gguf: write metadata")
+    gguf_writer.write_kv_data_to_file()
+    print("    gguf: write tensors")
+    gguf_writer.write_tensors_to_file()
+    gguf_writer.close()
+
+if __name__ == '__main__':
+    main()

examples/finetune/finetune.sh

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

examples/gbnf-validator/gbnf-validator.cpp

@@ -1,5 +1,9 @@
+#define LLAMA_API_INTERNAL
+
+#include "grammar-parser.h"
+#include "ggml.h"
+#include "llama.h"
 #include "unicode.h"
-#include "llama-grammar.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -8,28 +12,24 @@
 #include <string>
 #include <vector>
 
-static bool llama_grammar_validate(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
-    const auto cpts = unicode_cpts_from_utf8(input_str);
-
-    const llama_grammar_rules  & rules      = llama_grammar_get_rules (grammar);
-          llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
+static bool llama_sample_grammar_string(struct llama_grammar * grammar, const std::string & input_str, size_t & error_pos, std::string & error_msg) {
+    auto decoded = decode_utf8(input_str, {});
+    const auto & code_points = decoded.first;
 
     size_t pos = 0;
-    for (const auto & cpt : cpts) {
-        const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
-
-        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
-
-        if (stacks_cur.empty()) {
+    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+        auto prev_stacks = grammar->stacks;
+        llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
+        if (grammar->stacks.empty()) {
             error_pos = pos;
-            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
-            stacks_cur = stacks_prev;
+            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
+            grammar->stacks = prev_stacks;
             return false;
         }
         ++pos;
     }
 
-    for (const auto & stack : stacks_cur) {
+    for (const auto & stack : grammar->stacks) {
         if (stack.empty()) {
             return true;
         }
@@ -80,7 +80,27 @@ int main(int argc, char** argv) {
         grammar_str = buffer.str();
     }
 
-    llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
+    // Parse the GBNF grammar
+    auto parsed_grammar = grammar_parser::parse(grammar_str.c_str());
+
+    // will be empty (default) if there are parse errors
+    if (parsed_grammar.rules.empty()) {
+        fprintf(stdout, "%s: failed to parse grammar\n", __func__);
+        return 1;
+    }
+
+    // Ensure that there is a "root" node.
+    if (parsed_grammar.symbol_ids.find("root") == parsed_grammar.symbol_ids.end()) {
+        fprintf(stdout, "%s: grammar does not contain a 'root' symbol\n", __func__);
+        return 1;
+    }
+
+    std::vector<const llama_grammar_element *> grammar_rules(parsed_grammar.c_rules());
+
+    // Create the LLAMA grammar
+    auto grammar = llama_grammar_init(
+            grammar_rules.data(),
+            grammar_rules.size(), parsed_grammar.symbol_ids.at("root"));
     if (grammar == nullptr) {
         throw std::runtime_error("Failed to initialize llama_grammar");
     }
@@ -97,7 +117,7 @@ int main(int argc, char** argv) {
     // Validate the input string against the grammar
     size_t error_pos;
     std::string error_msg;
-    bool is_valid = llama_grammar_validate(grammar, input_str, error_pos, error_msg);
+    bool is_valid = llama_sample_grammar_string(grammar, input_str, error_pos, error_msg);
 
     if (is_valid) {
         fprintf(stdout, "Input string is valid according to the grammar.\n");
@@ -106,7 +126,7 @@ int main(int argc, char** argv) {
     }
 
     // Clean up
-    llama_grammar_free_impl(grammar);
+    llama_grammar_free(grammar);
 
     return 0;
 }

examples/gen-docs/gen-docs.cpp

@@ -1,52 +0,0 @@
-#include "arg.h"
-#include "common.h"
-
-#include <fstream>
-#include <string>
-
-// Export usage message (-h) to markdown format
-
-static void export_md(std::string fname, llama_example ex) {
-    std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
-
-    gpt_params params;
-    auto ctx_arg = gpt_params_parser_init(params, ex);
-
-    file << "| Argument | Explanation |\n";
-    file << "| -------- | ----------- |\n";
-    for (auto & opt : ctx_arg.options) {
-        file << "| `";
-        // args
-        for (const auto & arg : opt.args) {
-        if (arg == opt.args.front()) {
-                file << arg;
-                if (opt.args.size() > 1) file << ", ";
-            } else {
-                file << arg << (arg != opt.args.back() ? ", " : "");
-            }
-        }
-        // value hint
-        if (opt.value_hint) {
-            std::string md_value_hint(opt.value_hint);
-            string_replace_all(md_value_hint, "|", "\\|");
-            file << " " << md_value_hint;
-        }
-        if (opt.value_hint_2) {
-            std::string md_value_hint_2(opt.value_hint_2);
-            string_replace_all(md_value_hint_2, "|", "\\|");
-            file << " " << md_value_hint_2;
-        }
-        // help text
-        std::string md_help(opt.help);
-        string_replace_all(md_help, "\n", "<br/>");
-        string_replace_all(md_help, "|", "\\|");
-        file << "` | " << md_help << " |\n";
-    }
-}
-
-int main(int, char **) {
-    export_md("autogen-main.md", LLAMA_EXAMPLE_MAIN);
-    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER);
-
-    return 0;
-}

examples/gguf-hash/CMakeLists.txt

@@ -1,15 +0,0 @@
-set(TARGET llama-gguf-hash)
-add_executable(${TARGET} gguf-hash.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-
-# clibs dependencies
-include_directories(deps/)
-add_library(xxhash OBJECT deps/xxhash/xxhash.c deps/xxhash/xxhash.h)
-target_link_libraries(${TARGET} PRIVATE xxhash)
-add_library(sha1 OBJECT deps/sha1/sha1.c deps/sha1/sha1.h)
-target_link_libraries(${TARGET} PRIVATE sha1)
-add_library(sha256 OBJECT deps/sha256/sha256.c deps/sha256/sha256.h)
-target_link_libraries(${TARGET} PRIVATE sha256)
-
-target_link_libraries(${TARGET} PRIVATE ggml ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gguf-hash/README.md

@@ -1,206 +0,0 @@
-
-# llama-gguf-hash
-
-CLI to hash GGUF files to detect difference on a per model and per tensor level.
-
-**Command line options:**
-
-- `--help`: display help message
-- `--xxh64`: use xhash 64bit hash mode (default)
-- `--sha1`: use sha1
-- `--uuid`: use uuid
-- `--sha256`: use sha256
-- `--all`: use all hash
-- `--no-layer`: exclude per layer hash
-- `--uuid`: generate UUIDv5 ID
-- `-c`, `--check <manifest>`:  verify against a manifest
-
-## About
-
-While most POSIX systems already have hash checking programs like sha256sum, it
-is designed to check entire files. This is not ideal for our purpose if we want
-to check for consistency of the tensor data even if the metadata content of the
-gguf KV store has been updated.
-
-This program is designed to hash a gguf tensor payload on a 'per tensor layer'
-in addition to a 'entire tensor model' hash. The intent is that the entire
-tensor layer can be checked first but if there is any detected inconsistencies,
-then the per tensor hash can be used to narrow down the specific tensor layer
-that has inconsistencies.
-
-For Maintainers:
-- Detection of tensor inconsistency during development and automated tests
-    - This is served by xxh64 which is fast
-    - This is also served by having per tensor layer to assist in narrowing down
-      the location of the faulty tensor layer
-    - This is also served by sha1 which is much slower but more widely supported
-
-For Model Creators:
-- Optional consistent UUID generation based on model tensor content
-    - This is served by UUIDv5 which is useful for databases keys
-        - llama.cpp UUIDv5 Namespace: `ef001206-dadc-5f6d-a15f-3359e577d4e5`
-            - Made via UUIDv5 URL namespace of `en.wikipedia.org/wiki/Llama.cpp`
-
-For Model Users:
-- Assurance of tensor layer integrity even if metadata was updated
-    - This is served by sha256 which is still considered very secure as of 2024
-
-### Design Note
-
-- The default behavior of this program if no arguments is provided is to hash
-  using xxhash's xxh32 mode because it is very fast and is primarily targeted
-  towards maintainers who may want to use this in automated tests.
-- xxhash support xxh32 and xxh128 for 32bit hash and 128bit hash respectively
-  however we picked 64bit xxhash as most computers are 64bit as of 2024 and thus
-  would have a better affinity to calculating hash that is 64bit in size.
-
-## Compile Example
-
-```bash
-cmake -B build -DCMAKE_BUILD_TYPE=Debug -DLLAMA_FATAL_WARNINGS=ON
-make -C build clean
-make -C build llama-gguf-hash VERBOSE=1
-./build/bin/llama-gguf-hash test.gguf
-./build/bin/llama-gguf-hash --xxh64 test.gguf
-./build/bin/llama-gguf-hash --sha1 test.gguf
-./build/bin/llama-gguf-hash --uuid test.gguf
-./build/bin/llama-gguf-hash --sha256 test.gguf
-```
-
-## Generation and Verification Example
-
-To generate we may use this command
-
-```bash
-./llama-gguf-hash --all test.gguf > test.gguf.manifest
-```
-
-Which would generate a manifest that looks like below, which contains multiple hash type and per tensor layer hashes as well
-(This excludes UUID as that is an ID not a hash)
-
-```bash
-xxh64     f66e9cd66a4396a0  test.gguf:tensor_0
-sha1      59f79ecefd8125a996fdf419239051a7e99e5f20  test.gguf:tensor_0
-sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0
-xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1
-sha1      4765f592eacf096df4628ba59476af94d767080a  test.gguf:tensor_1
-sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1
-xxh64     a0af5d700049693b  test.gguf:tensor_2
-sha1      25cbfbad4513cc348e2c95ebdee69d6ff2fd8753  test.gguf:tensor_2
-sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2
-xxh64     e83fddf559d7b6a6  test.gguf:tensor_3
-sha1      a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c  test.gguf:tensor_3
-sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3
-xxh64     1257733306b7992d  test.gguf:tensor_4
-sha1      d7bc61db93bb685ce9d598da89717c66729b7543  test.gguf:tensor_4
-sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4
-xxh64     d238d16ba4711e58  test.gguf:tensor_5
-sha1      0706566c198fe1072f37e0a5135b4b5f23654c52  test.gguf:tensor_5
-sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5
-xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6
-sha1      73922a0727226a409049f6fc3172a52219ca6f00  test.gguf:tensor_6
-sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6
-xxh64     c22021c29854f093  test.gguf:tensor_7
-sha1      efc39cece6a951188fc41e354c73bbfe6813d447  test.gguf:tensor_7
-sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7
-xxh64     936df61f5d64261f  test.gguf:tensor_8
-sha1      c2490296d789a4f34398a337fed8377d943d9f06  test.gguf:tensor_8
-sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8
-xxh64     93fd20c64421c081  test.gguf:tensor_9
-sha1      7047ce1e78437a6884337a3751c7ee0421918a65  test.gguf:tensor_9
-sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9
-xxh64     5a54d3aad816f302  test.gguf
-sha1      d15be52c4ff213e823cb6dd13af7ee2f978e7042  test.gguf
-sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf
-```
-
-We can then use the normal check command which will by default check for the highest security strength hash and verify against that:
-
-```bash
-$ ./llama-gguf-hash --check test.gguf.manifest test.gguf
-manifest  test.gguf.manifest  sha256  sha1  xxh64
-sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0  -  Ok
-sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1  -  Ok
-sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2  -  Ok
-sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3  -  Ok
-sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4  -  Ok
-sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5  -  Ok
-sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6  -  Ok
-sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7  -  Ok
-sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8  -  Ok
-sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9  -  Ok
-sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf  -  Ok
-
-Verification results for test.gguf.manifest - Success
-```
-
-Or we may explicitly ask for a faster hash like:
-
-```bash
-$ ./llama-gguf-hash --check test.gguf.manifest --xxh64 test.gguf
-manifest  test.gguf.manifest  sha256  sha1  xxh64
-xxh64     f66e9cd66a4396a0  test.gguf:tensor_0  -  Ok
-xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1  -  Ok
-xxh64     a0af5d700049693b  test.gguf:tensor_2  -  Ok
-xxh64     e83fddf559d7b6a6  test.gguf:tensor_3  -  Ok
-xxh64     1257733306b7992d  test.gguf:tensor_4  -  Ok
-xxh64     d238d16ba4711e58  test.gguf:tensor_5  -  Ok
-xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6  -  Ok
-xxh64     c22021c29854f093  test.gguf:tensor_7  -  Ok
-xxh64     936df61f5d64261f  test.gguf:tensor_8  -  Ok
-xxh64     93fd20c64421c081  test.gguf:tensor_9  -  Ok
-xxh64     5a54d3aad816f302  test.gguf  -  Ok
-
-Verification results for test.gguf.manifest - Success
-```
-
-Or maybe we want to just check that all the hash is valid:
-
-```bash
-$./llama-gguf-hash --check test.gguf.manifest --all test.gguf.manifest
-manifest  test.gguf.manifest  sha256  sha1  xxh64
-xxh64     f66e9cd66a4396a0  test.gguf:tensor_0  -  Ok
-sha1      59f79ecefd8125a996fdf419239051a7e99e5f20  test.gguf:tensor_0  -  Ok
-sha256    c0510d38fa060c46265e0160a85c7243096b01dd31c2f355bdbb5516b20de1bd  test.gguf:tensor_0  -  Ok
-xxh64     7d3a1f9ac04d0537  test.gguf:tensor_1  -  Ok
-sha1      4765f592eacf096df4628ba59476af94d767080a  test.gguf:tensor_1  -  Ok
-sha256    8514cbcc73692a2c56bd7a33a022edd5ff819614bd23b19915d7224387f397a7  test.gguf:tensor_1  -  Ok
-xxh64     a0af5d700049693b  test.gguf:tensor_2  -  Ok
-sha1      25cbfbad4513cc348e2c95ebdee69d6ff2fd8753  test.gguf:tensor_2  -  Ok
-sha256    947e6b36e20f2cc95e1d2ce1c1669d813d574657ac6b5ac5196158d454d35180  test.gguf:tensor_2  -  Ok
-xxh64     e83fddf559d7b6a6  test.gguf:tensor_3  -  Ok
-sha1      a9cba73e2d90f2ee3dae2548caa42bef3fe6a96c  test.gguf:tensor_3  -  Ok
-sha256    423b044e016d8ac73c39f23f60bf01bedef5ecb03c0230accd824c91fe86f1a1  test.gguf:tensor_3  -  Ok
-xxh64     1257733306b7992d  test.gguf:tensor_4  -  Ok
-sha1      d7bc61db93bb685ce9d598da89717c66729b7543  test.gguf:tensor_4  -  Ok
-sha256    79737cb3912d4201384cf7f16a1a37ff7823f23ea796cb205b6ca361ab9e3ebf  test.gguf:tensor_4  -  Ok
-xxh64     d238d16ba4711e58  test.gguf:tensor_5  -  Ok
-sha1      0706566c198fe1072f37e0a5135b4b5f23654c52  test.gguf:tensor_5  -  Ok
-sha256    60949be8298eced0ecdde64487643d018407bd261691e061d9e9c3dbc9fd358b  test.gguf:tensor_5  -  Ok
-xxh64     3fbc3b65ab8c7f39  test.gguf:tensor_6  -  Ok
-sha1      73922a0727226a409049f6fc3172a52219ca6f00  test.gguf:tensor_6  -  Ok
-sha256    574f4c46ff384a3b9a225eb955d2a871847a2e8b3fa59387a8252832e92ef7b0  test.gguf:tensor_6  -  Ok
-xxh64     c22021c29854f093  test.gguf:tensor_7  -  Ok
-sha1      efc39cece6a951188fc41e354c73bbfe6813d447  test.gguf:tensor_7  -  Ok
-sha256    4c0410cd3c500f078ae5b21e8dc9eb79e29112713b2ab58a882f82a3868d4d75  test.gguf:tensor_7  -  Ok
-xxh64     936df61f5d64261f  test.gguf:tensor_8  -  Ok
-sha1      c2490296d789a4f34398a337fed8377d943d9f06  test.gguf:tensor_8  -  Ok
-sha256    c4401313feeba0261275c3b25bd2d8fe40ce04e0f440c2980ed0e9674c30ff01  test.gguf:tensor_8  -  Ok
-xxh64     93fd20c64421c081  test.gguf:tensor_9  -  Ok
-sha1      7047ce1e78437a6884337a3751c7ee0421918a65  test.gguf:tensor_9  -  Ok
-sha256    23d57cf0d7a6e90b0b3616b41300e0cd354781e812add854a5f95aa55f2bc514  test.gguf:tensor_9  -  Ok
-xxh64     5a54d3aad816f302  test.gguf  -  Ok
-sha1      d15be52c4ff213e823cb6dd13af7ee2f978e7042  test.gguf  -  Ok
-sha256    7dd641b32f59b60dbd4b5420c4b0f6321ccf48f58f6ae201a3dbc4a58a27c6e4  test.gguf  -  Ok
-
-Verification results for test.gguf.manifest - Success
-```
-
-
-## Crypto/Hash Libraries Used
-
-These micro c libraries dependencies was installed via the [clib c package manager](https://github.com/clibs)
-
-- https://github.com/Cyan4973/xxHash
-- https://github.com/clibs/sha1/
-- https://github.com/jb55/sha256.c

examples/gguf-hash/deps/rotate-bits/package.json

@@ -1,13 +0,0 @@
-{
-  "name": "rotate-bits",
-  "version": "0.1.1",
-  "repo": "jb55/rotate-bits.h",
-  "description": "rotate bits",
-  "keywords": ["rotl", "rotr"],
-  "src": ["rotate-bits.h"],
-  "license": "Public Domain",
-  "development": {
-    "thlorenz/tap.c": "*"
-  }
-}
-

examples/gguf-hash/deps/rotate-bits/rotate-bits.h

@@ -1,46 +0,0 @@
-
-
-#ifndef __ROTATE_DEFS_H
-#define __ROTATE_DEFS_H
-
-#ifdef _MSC_VER
-
-#include <stdlib.h>
-
-#define ROTL32(v, n) _rotl((v), (n))
-#define ROTL64(v, n) _rotl64((v), (n))
-
-#define ROTR32(v, n) _rotr((v), (n))
-#define ROTR64(v, n) _rotr64((v), (n))
-
-#else
-
-#include <stdint.h>
-
-#define U8V(v) ((uint8_t)(v) & 0xFFU)
-#define U16V(v) ((uint16_t)(v) & 0xFFFFU)
-#define U32V(v) ((uint32_t)(v) & 0xFFFFFFFFU)
-#define U64V(v) ((uint64_t)(v) & 0xFFFFFFFFFFFFFFFFU)
-
-#define ROTL32(v, n) \
-  (U32V((uint32_t)(v) << (n)) | ((uint32_t)(v) >> (32 - (n))))
-
-// tests fail if we don't have this cast...
-#define ROTL64(v, n) \
-  (U64V((uint64_t)(v) << (n)) | ((uint64_t)(v) >> (64 - (n))))
-
-#define ROTR32(v, n) ROTL32(v, 32 - (n))
-#define ROTR64(v, n) ROTL64(v, 64 - (n))
-
-#endif
-
-#define ROTL8(v, n) \
-  (U8V((uint8_t)(v) << (n)) | ((uint8_t)(v) >> (8 - (n))))
-
-#define ROTL16(v, n) \
-  (U16V((uint16_t)(v) << (n)) | ((uint16_t)(v) >> (16 - (n))))
-
-#define ROTR8(v, n) ROTL8(v, 8 - (n))
-#define ROTR16(v, n) ROTL16(v, 16 - (n))
-
-#endif

examples/gguf-hash/deps/sha1/package.json

@@ -1,9 +0,0 @@
-{
-  "name": "sha1",
-  "version": "0.0.1",
-  "repo": "clibs/sha1",
-  "description": "sha1 hash algorithm",
-  "keywords": ["sha1", "hash"],
-  "license": "public domain",
-  "src": ["sha1.c", "sha1.h"]
-}

examples/gguf-hash/deps/sha1/sha1.c

@@ -1,295 +0,0 @@
-/*
-SHA-1 in C
-By Steve Reid <steve@edmweb.com>
-100% Public Domain
-
-Test Vectors (from FIPS PUB 180-1)
-"abc"
-  A9993E36 4706816A BA3E2571 7850C26C 9CD0D89D
-"abcdbcdecdefdefgefghfghighijhijkijkljklmklmnlmnomnopnopq"
-  84983E44 1C3BD26E BAAE4AA1 F95129E5 E54670F1
-A million repetitions of "a"
-  34AA973C D4C4DAA4 F61EEB2B DBAD2731 6534016F
-*/
-
-/* #define LITTLE_ENDIAN * This should be #define'd already, if true. */
-/* #define SHA1HANDSOFF * Copies data before messing with it. */
-
-#define SHA1HANDSOFF
-
-#include <stdio.h>
-#include <string.h>
-
-/* for uint32_t */
-#include <stdint.h>
-
-#include "sha1.h"
-
-
-#define rol(value, bits) (((value) << (bits)) | ((value) >> (32 - (bits))))
-
-/* blk0() and blk() perform the initial expand. */
-/* I got the idea of expanding during the round function from SSLeay */
-#if BYTE_ORDER == LITTLE_ENDIAN
-#define blk0(i) (block->l[i] = (rol(block->l[i],24)&0xFF00FF00) \
-    |(rol(block->l[i],8)&0x00FF00FF))
-#elif BYTE_ORDER == BIG_ENDIAN
-#define blk0(i) block->l[i]
-#else
-#error "Endianness not defined!"
-#endif
-#define blk(i) (block->l[i&15] = rol(block->l[(i+13)&15]^block->l[(i+8)&15] \
-    ^block->l[(i+2)&15]^block->l[i&15],1))
-
-/* (R0+R1), R2, R3, R4 are the different operations used in SHA1 */
-#define R0(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk0(i)+0x5A827999+rol(v,5);w=rol(w,30);
-#define R1(v,w,x,y,z,i) z+=((w&(x^y))^y)+blk(i)+0x5A827999+rol(v,5);w=rol(w,30);
-#define R2(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0x6ED9EBA1+rol(v,5);w=rol(w,30);
-#define R3(v,w,x,y,z,i) z+=(((w|x)&y)|(w&x))+blk(i)+0x8F1BBCDC+rol(v,5);w=rol(w,30);
-#define R4(v,w,x,y,z,i) z+=(w^x^y)+blk(i)+0xCA62C1D6+rol(v,5);w=rol(w,30);
-
-
-/* Hash a single 512-bit block. This is the core of the algorithm. */
-
-void SHA1Transform(
-    uint32_t state[5],
-    const unsigned char buffer[64]
-)
-{
-    uint32_t a, b, c, d, e;
-
-    typedef union
-    {
-        unsigned char c[64];
-        uint32_t l[16];
-    } CHAR64LONG16;
-
-#ifdef SHA1HANDSOFF
-    CHAR64LONG16 block[1];      /* use array to appear as a pointer */
-
-    memcpy(block, buffer, 64);
-#else
-    /* The following had better never be used because it causes the
-     * pointer-to-const buffer to be cast into a pointer to non-const.
-     * And the result is written through.  I threw a "const" in, hoping
-     * this will cause a diagnostic.
-     */
-    CHAR64LONG16 *block = (const CHAR64LONG16 *) buffer;
-#endif
-    /* Copy context->state[] to working vars */
-    a = state[0];
-    b = state[1];
-    c = state[2];
-    d = state[3];
-    e = state[4];
-    /* 4 rounds of 20 operations each. Loop unrolled. */
-    R0(a, b, c, d, e, 0);
-    R0(e, a, b, c, d, 1);
-    R0(d, e, a, b, c, 2);
-    R0(c, d, e, a, b, 3);
-    R0(b, c, d, e, a, 4);
-    R0(a, b, c, d, e, 5);
-    R0(e, a, b, c, d, 6);
-    R0(d, e, a, b, c, 7);
-    R0(c, d, e, a, b, 8);
-    R0(b, c, d, e, a, 9);
-    R0(a, b, c, d, e, 10);
-    R0(e, a, b, c, d, 11);
-    R0(d, e, a, b, c, 12);
-    R0(c, d, e, a, b, 13);
-    R0(b, c, d, e, a, 14);
-    R0(a, b, c, d, e, 15);
-    R1(e, a, b, c, d, 16);
-    R1(d, e, a, b, c, 17);
-    R1(c, d, e, a, b, 18);
-    R1(b, c, d, e, a, 19);
-    R2(a, b, c, d, e, 20);
-    R2(e, a, b, c, d, 21);
-    R2(d, e, a, b, c, 22);
-    R2(c, d, e, a, b, 23);
-    R2(b, c, d, e, a, 24);
-    R2(a, b, c, d, e, 25);
-    R2(e, a, b, c, d, 26);
-    R2(d, e, a, b, c, 27);
-    R2(c, d, e, a, b, 28);
-    R2(b, c, d, e, a, 29);
-    R2(a, b, c, d, e, 30);
-    R2(e, a, b, c, d, 31);
-    R2(d, e, a, b, c, 32);
-    R2(c, d, e, a, b, 33);
-    R2(b, c, d, e, a, 34);
-    R2(a, b, c, d, e, 35);
-    R2(e, a, b, c, d, 36);
-    R2(d, e, a, b, c, 37);
-    R2(c, d, e, a, b, 38);
-    R2(b, c, d, e, a, 39);
-    R3(a, b, c, d, e, 40);
-    R3(e, a, b, c, d, 41);
-    R3(d, e, a, b, c, 42);
-    R3(c, d, e, a, b, 43);
-    R3(b, c, d, e, a, 44);
-    R3(a, b, c, d, e, 45);
-    R3(e, a, b, c, d, 46);
-    R3(d, e, a, b, c, 47);
-    R3(c, d, e, a, b, 48);
-    R3(b, c, d, e, a, 49);
-    R3(a, b, c, d, e, 50);
-    R3(e, a, b, c, d, 51);
-    R3(d, e, a, b, c, 52);
-    R3(c, d, e, a, b, 53);
-    R3(b, c, d, e, a, 54);
-    R3(a, b, c, d, e, 55);
-    R3(e, a, b, c, d, 56);
-    R3(d, e, a, b, c, 57);
-    R3(c, d, e, a, b, 58);
-    R3(b, c, d, e, a, 59);
-    R4(a, b, c, d, e, 60);
-    R4(e, a, b, c, d, 61);
-    R4(d, e, a, b, c, 62);
-    R4(c, d, e, a, b, 63);
-    R4(b, c, d, e, a, 64);
-    R4(a, b, c, d, e, 65);
-    R4(e, a, b, c, d, 66);
-    R4(d, e, a, b, c, 67);
-    R4(c, d, e, a, b, 68);
-    R4(b, c, d, e, a, 69);
-    R4(a, b, c, d, e, 70);
-    R4(e, a, b, c, d, 71);
-    R4(d, e, a, b, c, 72);
-    R4(c, d, e, a, b, 73);
-    R4(b, c, d, e, a, 74);
-    R4(a, b, c, d, e, 75);
-    R4(e, a, b, c, d, 76);
-    R4(d, e, a, b, c, 77);
-    R4(c, d, e, a, b, 78);
-    R4(b, c, d, e, a, 79);
-    /* Add the working vars back into context.state[] */
-    state[0] += a;
-    state[1] += b;
-    state[2] += c;
-    state[3] += d;
-    state[4] += e;
-    /* Wipe variables */
-    a = b = c = d = e = 0;
-#ifdef SHA1HANDSOFF
-    memset(block, '\0', sizeof(block));
-#endif
-}
-
-
-/* SHA1Init - Initialize new context */
-
-void SHA1Init(
-    SHA1_CTX * context
-)
-{
-    /* SHA1 initialization constants */
-    context->state[0] = 0x67452301;
-    context->state[1] = 0xEFCDAB89;
-    context->state[2] = 0x98BADCFE;
-    context->state[3] = 0x10325476;
-    context->state[4] = 0xC3D2E1F0;
-    context->count[0] = context->count[1] = 0;
-}
-
-
-/* Run your data through this. */
-
-void SHA1Update(
-    SHA1_CTX * context,
-    const unsigned char *data,
-    uint32_t len
-)
-{
-    uint32_t i;
-
-    uint32_t j;
-
-    j = context->count[0];
-    if ((context->count[0] += len << 3) < j)
-        context->count[1]++;
-    context->count[1] += (len >> 29);
-    j = (j >> 3) & 63;
-    if ((j + len) > 63)
-    {
-        memcpy(&context->buffer[j], data, (i = 64 - j));
-        SHA1Transform(context->state, context->buffer);
-        for (; i + 63 < len; i += 64)
-        {
-            SHA1Transform(context->state, &data[i]);
-        }
-        j = 0;
-    }
-    else
-        i = 0;
-    memcpy(&context->buffer[j], &data[i], len - i);
-}
-
-
-/* Add padding and return the message digest. */
-
-void SHA1Final(
-    unsigned char digest[20],
-    SHA1_CTX * context
-)
-{
-    unsigned i;
-
-    unsigned char finalcount[8];
-
-    unsigned char c;
-
-#if 0    /* untested "improvement" by DHR */
-    /* Convert context->count to a sequence of bytes
-     * in finalcount.  Second element first, but
-     * big-endian order within element.
-     * But we do it all backwards.
-     */
-    unsigned char *fcp = &finalcount[8];
-
-    for (i = 0; i < 2; i++)
-    {
-        uint32_t t = context->count[i];
-
-        int j;
-
-        for (j = 0; j < 4; t >>= 8, j++)
-            *--fcp = (unsigned char) t}
-#else
-    for (i = 0; i < 8; i++)
-    {
-        finalcount[i] = (unsigned char) ((context->count[(i >= 4 ? 0 : 1)] >> ((3 - (i & 3)) * 8)) & 255);      /* Endian independent */
-    }
-#endif
-    c = 0200;
-    SHA1Update(context, &c, 1);
-    while ((context->count[0] & 504) != 448)
-    {
-        c = 0000;
-        SHA1Update(context, &c, 1);
-    }
-    SHA1Update(context, finalcount, 8); /* Should cause a SHA1Transform() */
-    for (i = 0; i < 20; i++)
-    {
-        digest[i] = (unsigned char)
-            ((context->state[i >> 2] >> ((3 - (i & 3)) * 8)) & 255);
-    }
-    /* Wipe variables */
-    memset(context, '\0', sizeof(*context));
-    memset(&finalcount, '\0', sizeof(finalcount));
-}
-
-void SHA1(
-    char *hash_out,
-    const char *str,
-    uint32_t len)
-{
-    SHA1_CTX ctx;
-    unsigned int ii;
-
-    SHA1Init(&ctx);
-    for (ii=0; ii<len; ii+=1)
-        SHA1Update(&ctx, (const unsigned char*)str + ii, 1);
-    SHA1Final((unsigned char *)hash_out, &ctx);
-}
-

examples/gguf-hash/deps/sha1/sha1.h

@@ -1,52 +0,0 @@
-#ifndef SHA1_H
-#define SHA1_H
-
-/*
-   SHA-1 in C
-   By Steve Reid <steve@edmweb.com>
-   100% Public Domain
- */
-
-#include "stdint.h"
-
-#if defined(__cplusplus)
-extern "C" {
-#endif
-
-typedef struct
-{
-    uint32_t state[5];
-    uint32_t count[2];
-    unsigned char buffer[64];
-} SHA1_CTX;
-
-void SHA1Transform(
-    uint32_t state[5],
-    const unsigned char buffer[64]
-    );
-
-void SHA1Init(
-    SHA1_CTX * context
-    );
-
-void SHA1Update(
-    SHA1_CTX * context,
-    const unsigned char *data,
-    uint32_t len
-    );
-
-void SHA1Final(
-    unsigned char digest[20],
-    SHA1_CTX * context
-    );
-
-void SHA1(
-    char *hash_out,
-    const char *str,
-    uint32_t len);
-
-#if defined(__cplusplus)
-}
-#endif
-
-#endif /* SHA1_H */

examples/gguf-hash/deps/sha256/package.json

@@ -1,15 +0,0 @@
-{
-  "name": "sha256",
-  "version": "0.0.2",
-  "repo": "jb55/sha256.c",
-  "description": "sha256 in c",
-  "keywords": ["sha256", "sha2"],
-  "src": ["sha256.c", "sha256.h"],
-  "dependencies": {
-    "jb55/rotate-bits.h": "0.1.1"
-  },
-  "development": {
-    "thlorenz/tap.c": "*"
-  }
-}
-

examples/gguf-hash/deps/sha256/sha256.c

@@ -1,221 +0,0 @@
-/* Crypto/Sha256.c -- SHA-256 Hash
-2010-06-11 : Igor Pavlov : Public domain
-This code is based on public domain code from Wei Dai's Crypto++ library. */
-
-#include "rotate-bits/rotate-bits.h"
-#include "sha256.h"
-
-/* define it for speed optimization */
-#define _SHA256_UNROLL
-#define _SHA256_UNROLL2
-
-void
-sha256_init(sha256_t *p)
-{
-  p->state[0] = 0x6a09e667;
-  p->state[1] = 0xbb67ae85;
-  p->state[2] = 0x3c6ef372;
-  p->state[3] = 0xa54ff53a;
-  p->state[4] = 0x510e527f;
-  p->state[5] = 0x9b05688c;
-  p->state[6] = 0x1f83d9ab;
-  p->state[7] = 0x5be0cd19;
-  p->count = 0;
-}
-
-#define S0(x) (ROTR32(x, 2) ^ ROTR32(x,13) ^ ROTR32(x, 22))
-#define S1(x) (ROTR32(x, 6) ^ ROTR32(x,11) ^ ROTR32(x, 25))
-#define s0(x) (ROTR32(x, 7) ^ ROTR32(x,18) ^ (x >> 3))
-#define s1(x) (ROTR32(x,17) ^ ROTR32(x,19) ^ (x >> 10))
-
-#define blk0(i) (W[i] = data[i])
-#define blk2(i) (W[i&15] += s1(W[(i-2)&15]) + W[(i-7)&15] + s0(W[(i-15)&15]))
-
-#define Ch(x,y,z) (z^(x&(y^z)))
-#define Maj(x,y,z) ((x&y)|(z&(x|y)))
-
-#define a(i) T[(0-(i))&7]
-#define b(i) T[(1-(i))&7]
-#define c(i) T[(2-(i))&7]
-#define d(i) T[(3-(i))&7]
-#define e(i) T[(4-(i))&7]
-#define f(i) T[(5-(i))&7]
-#define g(i) T[(6-(i))&7]
-#define h(i) T[(7-(i))&7]
-
-
-#ifdef _SHA256_UNROLL2
-
-#define R(a,b,c,d,e,f,g,h, i) h += S1(e) + Ch(e,f,g) + K[i+j] + (j?blk2(i):blk0(i));\
-  d += h; h += S0(a) + Maj(a, b, c)
-
-#define RX_8(i) \
-  R(a,b,c,d,e,f,g,h, i); \
-  R(h,a,b,c,d,e,f,g, (i+1)); \
-  R(g,h,a,b,c,d,e,f, (i+2)); \
-  R(f,g,h,a,b,c,d,e, (i+3)); \
-  R(e,f,g,h,a,b,c,d, (i+4)); \
-  R(d,e,f,g,h,a,b,c, (i+5)); \
-  R(c,d,e,f,g,h,a,b, (i+6)); \
-  R(b,c,d,e,f,g,h,a, (i+7))
-
-#else
-
-#define R(i) h(i) += S1(e(i)) + Ch(e(i),f(i),g(i)) + K[i+j] + (j?blk2(i):blk0(i));\
-  d(i) += h(i); h(i) += S0(a(i)) + Maj(a(i), b(i), c(i))
-
-#ifdef _SHA256_UNROLL
-
-#define RX_8(i) R(i+0); R(i+1); R(i+2); R(i+3); R(i+4); R(i+5); R(i+6); R(i+7);
-
-#endif
-
-#endif
-
-static const uint32_t K[64] = {
-  0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
-  0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
-  0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3,
-  0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
-  0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc,
-  0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
-  0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7,
-  0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
-  0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13,
-  0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
-  0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3,
-  0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
-  0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5,
-  0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
-  0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208,
-  0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2
-};
-
-static void
-sha256_transform(uint32_t *state, const uint32_t *data)
-{
-  uint32_t W[16] = {0};
-  unsigned j;
-  #ifdef _SHA256_UNROLL2
-  uint32_t a,b,c,d,e,f,g,h;
-  a = state[0];
-  b = state[1];
-  c = state[2];
-  d = state[3];
-  e = state[4];
-  f = state[5];
-  g = state[6];
-  h = state[7];
-  #else
-  uint32_t T[8];
-  for (j = 0; j < 8; j++)
-    T[j] = state[j];
-  #endif
-
-  for (j = 0; j < 64; j += 16)
-  {
-    #if defined(_SHA256_UNROLL) || defined(_SHA256_UNROLL2)
-    RX_8(0); RX_8(8);
-    #else
-    unsigned i;
-    for (i = 0; i < 16; i++) { R(i); }
-    #endif
-  }
-
-  #ifdef _SHA256_UNROLL2
-  state[0] += a;
-  state[1] += b;
-  state[2] += c;
-  state[3] += d;
-  state[4] += e;
-  state[5] += f;
-  state[6] += g;
-  state[7] += h;
-  #else
-  for (j = 0; j < 8; j++)
-    state[j] += T[j];
-  #endif
-
-  /* Wipe variables */
-  /* memset(W, 0, sizeof(W)); */
-  /* memset(T, 0, sizeof(T)); */
-}
-
-#undef S0
-#undef S1
-#undef s0
-#undef s1
-
-static void
-sha256_write_byte_block(sha256_t *p)
-{
-  uint32_t data32[16];
-  unsigned i;
-  for (i = 0; i < 16; i++)
-    data32[i] =
-      ((uint32_t)(p->buffer[i * 4    ]) << 24) +
-      ((uint32_t)(p->buffer[i * 4 + 1]) << 16) +
-      ((uint32_t)(p->buffer[i * 4 + 2]) <<  8) +
-      ((uint32_t)(p->buffer[i * 4 + 3]));
-  sha256_transform(p->state, data32);
-}
-
-
-void
-sha256_hash(unsigned char *buf, const unsigned char *data, size_t size)
-{
-  sha256_t hash;
-  sha256_init(&hash);
-  sha256_update(&hash, data, size);
-  sha256_final(&hash, buf);
-}
-
-
-void
-sha256_update(sha256_t *p, const unsigned char *data, size_t size)
-{
-  uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
-  while (size > 0)
-  {
-    p->buffer[curBufferPos++] = *data++;
-    p->count++;
-    size--;
-    if (curBufferPos == 64)
-    {
-      curBufferPos = 0;
-      sha256_write_byte_block(p);
-    }
-  }
-}
-
-
-void
-sha256_final(sha256_t *p, unsigned char *digest)
-{
-  uint64_t lenInBits = (p->count << 3);
-  uint32_t curBufferPos = (uint32_t)p->count & 0x3F;
-  unsigned i;
-  p->buffer[curBufferPos++] = 0x80;
-  while (curBufferPos != (64 - 8))
-  {
-    curBufferPos &= 0x3F;
-    if (curBufferPos == 0)
-      sha256_write_byte_block(p);
-    p->buffer[curBufferPos++] = 0;
-  }
-  for (i = 0; i < 8; i++)
-  {
-    p->buffer[curBufferPos++] = (unsigned char)(lenInBits >> 56);
-    lenInBits <<= 8;
-  }
-  sha256_write_byte_block(p);
-
-  for (i = 0; i < 8; i++)
-  {
-    *digest++ = (unsigned char)(p->state[i] >> 24);
-    *digest++ = (unsigned char)(p->state[i] >> 16);
-    *digest++ = (unsigned char)(p->state[i] >> 8);
-    *digest++ = (unsigned char)(p->state[i]);
-  }
-  sha256_init(p);
-}