tokenize : add --no-parse-special option

This should allow more easily explaining
how parse_special affects tokenization.

.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 +61,54 @@ 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
+
+      # server bench
+      ps.matplotlib
+
+      # server tests
+      ps.openai
+      ps.behave
+      ps.prometheus-client
+
+      # for examples/pydantic-models-to-grammar-examples.py
+      ps.docstring-parser
+      ps.pydantic
+
+      # for scripts/compare-llama-bench.py
+      ps.gitpython
+      ps.tabulate
+    ]
+  );
+
+  xcrunHost = runCommand "xcrunHost" {} ''
     mkdir -p $out/bin
     ln -s /usr/bin/xcrun $out/bin
   '';
@@ -84,9 +125,16 @@ 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 +146,184 @@ 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"

.dockerignore

@@ -1,7 +1,7 @@
 *.o
 *.a
 .cache/
-# Do not ignore .git directory, otherwise the reported build number will always be 0
+.git/
 .github/
 .gitignore
 .vs/

.ecrc

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

.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

@@ -23,9 +23,6 @@ env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
 
 jobs:
   macOS-latest-cmake-arm64:
@@ -50,7 +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
@@ -108,7 +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
@@ -225,7 +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
@@ -358,10 +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
@@ -378,7 +373,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v3
 
       - name: Dependencies
         id: depends
@@ -404,7 +399,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v2
 
       - name: add oneAPI to apt
         shell: bash
@@ -445,7 +440,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@v2
 
       - name: add oneAPI to apt
         shell: bash
@@ -549,7 +544,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -579,7 +574,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -613,7 +608,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v1
 
       - name: Dependencies
         id: depends
@@ -699,20 +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'
@@ -860,8 +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
@@ -956,7 +952,6 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
-          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
@@ -968,20 +963,19 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-latest-cmake-hip:
-    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
 
     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"
@@ -996,72 +990,8 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
-
-  windows-latest-cmake-hip-release:
-    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-    runs-on: windows-latest
-
-    strategy:
-      matrix:
-        gpu_target: [gfx1100, gfx1101, gfx1030]
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v4
-
-      - 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"
-          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"
-
-      - name: Verify ROCm
-        id: verify
-        run: |
-          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
-
-      - name: Build
-        id: cmake_build
-        run: |
-          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
-          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
-          md "build\bin\rocblas\library\"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        run: |
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
-
-      - name: Upload artifacts
-        uses: actions/upload-artifact@v4
-        with:
-          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
-          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
+          cmake --build build --config Release
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -1126,7 +1056,6 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
-      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

.github/workflows/docker.yml

@@ -15,17 +15,11 @@ on:
     branches:
       - master
     paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
-  workflow_dispatch: # allows manual triggering, useful for debugging
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
-# Fine-grant permission
-# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
-permissions:
-  packages: write
-
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -43,17 +37,15 @@ 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" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
-        with:
-          fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -68,34 +60,6 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
-          REPO_NAME="${{ github.event.repository.name }}"
-
-          # determine tag name postfix (build number, commit hash)
-          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
-            TAG_POSTFIX="b${BUILD_NUMBER}"
-          else
-            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
-            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
-          fi
-
-          # list all tags possible
-          TAGS=""
-          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
-          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
-
-          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
-          echo "output_tags=$TAGS"  # print out for debugging
-        env:
-          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
       # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main
@@ -113,13 +77,40 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Build and push Docker image (tagged + versioned)
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Downcase github.repository_owner
+        run: |
+          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
+      - 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 }}
-          # tag list is generated from step above
-          tags: ${{ steps.tag.outputs.output_tags }}
+          tags: "ghcr.io/${{ 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/server.yml

@@ -20,12 +20,6 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
 
-env:
-  LLAMA_LOG_COLORS: 1
-  LLAMA_LOG_PREFIX: 1
-  LLAMA_LOG_TIMESTAMPS: 1
-  LLAMA_LOG_VERBOSITY: 10
-
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
@@ -179,7 +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

@@ -50,7 +50,6 @@ build*
 !docs/build.md
 /libllama.so
 /llama-*
-/vulkan-shaders-gen
 android-ndk-*
 arm_neon.h
 cmake-build-*
@@ -61,12 +60,6 @@ llama-batched-swift
 /rpc-server
 out/
 tmp/
-autogen-*.md
-
-# Deprecated
-
-/main
-/server
 
 # CI
 
@@ -80,6 +73,7 @@ models-mnt
 !models/ggml-vocab-*.gguf*
 
 # Zig
+
 zig-out/
 zig-cache/
 
@@ -130,6 +124,3 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
-
-# Test models for lora adapters
-/lora-tests

CMakeLists.txt

@@ -62,9 +62,6 @@ option(LLAMA_SANITIZE_THREAD    "llama: enable thread sanitizer"    OFF)
 option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
 option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
-# utils
-option(LLAMA_BUILD_COMMON "llama: build common utils library" ON)
-
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
@@ -85,11 +82,11 @@ set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
-    set(GGML_LLAMAFILE_DEFAULT ON)
+    set(GGML_LLAMAFILE ON)
 endif()
 
-if (NOT DEFINED GGML_CUDA_GRAPHS)
-    set(GGML_CUDA_GRAPHS_DEFAULT ON)
+if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
+    set(GGML_CUDA_USE_GRAPHS ON)
 endif()
 
 # transition helpers
@@ -109,7 +106,6 @@ 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
@@ -136,23 +132,7 @@ 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 ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
 install(TARGETS llama LIBRARY PUBLIC_HEADER)
@@ -194,17 +174,15 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
 
 #
-# utils, programs, examples and tests
+# programs, examples and tests
 #
 
-if (LLAMA_BUILD_COMMON)
-    add_subdirectory(common)
-endif()
+add_subdirectory(common)
 
 if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     include(CTest)
     add_subdirectory(tests)
-endif()
+endif ()
 
 if (LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)

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

@@ -1,18 +1,13 @@
-# Pull requests (for contributors)
+# Pull requests
 
+- Always squash-merge the PR before merging
+- Use the following format for your final commit: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
 - 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
+- If the pull request contains only 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
 - 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
-
-# Pull requests (for collaborators)
-
-- 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
+  - 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 conveience
 
 # Coding guidelines
 
@@ -27,8 +22,3 @@
 
 ![matmul](media/matmul.png)
 
-# Resources
-
-The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
-
-https://github.com/ggerganov/llama.cpp/projects

Makefile

@@ -11,6 +11,7 @@ BUILD_TARGETS = \
 	llama-embedding \
 	llama-eval-callback \
 	llama-export-lora \
+	llama-finetune \
 	llama-gbnf-validator \
 	llama-gguf \
 	llama-gguf-hash \
@@ -19,7 +20,6 @@ BUILD_TARGETS = \
 	llama-imatrix \
 	llama-infill \
 	llama-llava-cli \
-	llama-minicpmv-cli\
 	llama-lookahead \
 	llama-lookup \
 	llama-lookup-create \
@@ -37,14 +37,13 @@ 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 \
@@ -54,7 +53,6 @@ TEST_TARGETS = \
 	tests/test-grammar-parser \
 	tests/test-json-schema-to-grammar \
 	tests/test-llama-grammar \
-	tests/test-log \
 	tests/test-model-load-cancel \
 	tests/test-opt \
 	tests/test-quantize-fns \
@@ -66,13 +64,13 @@ 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_CLEAN = 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
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune 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
+LEGACY_TARGETS_BUILD = main quantize perplexity embedding server finetune
 
 # Deprecation aliases
 ifdef LLAMA_CUBLAS
@@ -149,14 +147,6 @@ GGML_NO_METAL := 1
 DEPRECATE_WARNING := 1
 endif
 
-ifdef LLAMA_DISABLE_LOGS
-REMOVE_WARNING := 1
-endif
-
-ifdef LLAMA_SERVER_VERBOSE
-REMOVE_WARNING := 1
-endif
-
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -207,10 +197,6 @@ ifdef GGML_RPC
 	BUILD_TARGETS += rpc-server
 endif
 
-ifdef GGML_VULKAN
-	BUILD_TARGETS += vulkan-shaders-gen
-endif
-
 default: $(BUILD_TARGETS) $(LEGACY_TARGETS_BUILD)
 
 test: $(TEST_TARGETS)
@@ -337,9 +323,9 @@ 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
@@ -360,11 +346,19 @@ ifdef LLAMA_SANITIZE_UNDEFINED
 	MK_LDFLAGS  += -fsanitize=undefined -g
 endif
 
+ifdef LLAMA_SERVER_VERBOSE
+	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
+endif
+
 ifdef LLAMA_SERVER_SSL
 	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
 	MK_LDFLAGS += -lssl -lcrypto
 endif
 
+ifdef LLAMA_DISABLE_LOGS
+	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
+endif # LLAMA_DISABLE_LOGS
+
 # warnings
 WARN_FLAGS = \
 	-Wall \
@@ -435,7 +429,7 @@ 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 $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
@@ -515,12 +509,7 @@ ifneq ($(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
@@ -537,21 +526,10 @@ 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
@@ -569,17 +547,11 @@ 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
@@ -602,27 +574,15 @@ 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$(CUDA_PATH)/lib -L/usr/lib64
-		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
+	ifneq ('', '$(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 -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
-		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
-		MK_NVCCFLAGS += -use_fast_math
-	endif # GGML_MUSA
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(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))
@@ -632,11 +592,9 @@ 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
@@ -649,12 +607,8 @@ 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)
@@ -725,15 +679,9 @@ 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: \
@@ -756,8 +704,8 @@ 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
@@ -771,10 +719,6 @@ 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
@@ -783,28 +727,10 @@ 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
@@ -841,14 +767,6 @@ 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
@@ -901,16 +819,15 @@ 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
 
@@ -923,20 +840,16 @@ OBJ_GGML += \
 
 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/log.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
 
@@ -994,7 +907,6 @@ $(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
@@ -1004,7 +916,6 @@ 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 )
 
@@ -1029,14 +940,6 @@ $(info   - LLAMA_NO_CCACHE)
 $(info )
 endif
 
-ifdef REMOVE_WARNING
-$(info !!! REMOVAL WARNING !!!)
-$(info The following LLAMA_ options have been removed and are no longer supported)
-$(info   - LLAMA_DISABLE_LOGS   (https://github.com/ggerganov/llama.cpp/pull/9418))
-$(info   - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418))
-$(info )
-endif
-
 #
 # Build libraries
 #
@@ -1116,10 +1019,6 @@ 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 \
@@ -1129,29 +1028,6 @@ 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)
@@ -1173,16 +1049,6 @@ common/common.o: \
 	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common/arg.o: \
-	common/arg.cpp \
-	common/arg.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
-common/log.o: \
-	common/log.cpp \
-	common/log.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 common/sampling.o: \
 	common/sampling.cpp \
 	common/sampling.h \
@@ -1194,6 +1060,11 @@ 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
@@ -1233,14 +1104,12 @@ 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)
 	find examples pocs -type f -name "*.o" -delete
 
@@ -1360,11 +1229,16 @@ 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-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
+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, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-bench: examples/llama-bench/llama-bench.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1375,11 +1249,16 @@ 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, $<)
@@ -1455,7 +1334,6 @@ 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)
@@ -1471,11 +1349,6 @@ 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 \
@@ -1486,20 +1359,15 @@ 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
@@ -1533,18 +1401,8 @@ 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)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-log: tests/test-log.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)
 
@@ -1653,45 +1511,56 @@ llama-q8dot: pocs/vdot/q8dot.cpp ggml/src/ggml.o \
 # 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 $@
+.PHONY: main quantize perplexity embedding server finetune
 
 # 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)
+main: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out $<,$^) $(call GET_OBJ_FILE, $<) -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)
+server: examples/deprecation-warning/deprecation-warning.cpp
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo "NOTICE: The 'server' binary is deprecated. Please use 'llama-server' instead."
 
-quantize: examples/deprecation-warning/deprecation-warning.o
+quantize: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard quantize))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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
+perplexity: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard perplexity))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -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
+embedding: examples/deprecation-warning/deprecation-warning.cpp
 ifneq (,$(wildcard embedding))
-	$(CXX) $(CXXFLAGS) $< -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 	@echo "#########"
 	@echo "WARNING: The 'embedding' binary is deprecated. Please use 'llama-embedding' instead."
 	@echo "  Remove the 'embedding' binary to remove this warning."
 	@echo "#########"
 endif
+
+finetune: examples/deprecation-warning/deprecation-warning.cpp
+ifneq (,$(wildcard finetune))
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	@echo "#########"
+	@echo "WARNING: The 'finetune' binary is deprecated. Please use 'llama-finetune' instead."
+	@echo "  Remove the 'finetune' binary to remove this warning."
+	@echo "#########"
+endif

Package.swift

@@ -4,9 +4,6 @@ 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",

README.md

@@ -3,22 +3,39 @@
 ![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++
 
+> [!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)
+
 ## Recent API changes
 
-- [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)
+- [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
 
 ## Hot topics
 
-- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
-- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
+- **`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
 
 ----
 
@@ -78,20 +95,8 @@ 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] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
-- [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))
 
@@ -113,7 +118,6 @@ Typically finetunes of the base models below are supported as well.
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
-- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
@@ -134,14 +138,12 @@ Typically finetunes of the base models below are supported as well.
 
 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)
@@ -167,25 +169,17 @@ 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`)*
 
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
-- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [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
-- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
 
 **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
 
@@ -411,11 +405,9 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [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
 
@@ -443,7 +435,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - 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, PRs and projects is very appreciated!
+- 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)

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
@@ -114,7 +107,7 @@ function gg_run_ctest_debug {
     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
 
@@ -145,7 +138,7 @@ function gg_run_ctest_release {
     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 +266,7 @@ 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 +290,8 @@ 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 +425,7 @@ 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 +535,7 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
+# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -561,8 +556,8 @@ 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 +692,7 @@ 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
 
@@ -712,81 +707,6 @@ function gg_run_embd_bge_small {
     set +e
 }
 
-function gg_sum_embd_bge_small {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'BGE Small (BERT):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-}
-
-# rerank_tiny
-
-function gg_run_rerank_tiny {
-    cd ${SRC}
-
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
-    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
-
-    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
-
-    path_models="../models-mnt/rerank-tiny"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-
-    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s><s>hi\nwhat is panda?</s><s>it's a bear\nwhat is panda?</s><s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
-
-    # sample output
-    # rerank score 0:    0.029
-    # rerank score 1:    0.029
-    # rerank score 2:    0.135
-
-    # check that the score is in the range [$3, $4]
-    function check_score {
-        qnt="$1"
-        score=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
-
-        if [ $(echo "$score < $3" | bc) -eq 1 ] || [ $(echo "$score > $4" | bc) -eq 1 ]; then
-            printf '  - %s @ %s (FAIL: score not in range [%s, %s])\n' "$qnt" "$score" "$3" "$4"
-            return 20
-        fi
-
-        printf '  - %s @ %s OK\n' "$qnt" "$score"
-        return 0
-    }
-
-    check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
-    check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
-    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.15" | tee -a $OUT/${ci}-rk-f16.log
-
-    set +e
-}
-
-function gg_sum_rerank_tiny {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'Rerank Tiny (Jina):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-rk-f16.log)"
-}
-
 function gg_check_build_requirements {
     if ! command -v cmake &> /dev/null; then
         gg_printf 'cmake not found, please install'
@@ -801,10 +721,16 @@ function gg_check_build_requirements {
     fi
 }
 
-## main
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
 
-export LLAMA_LOG_PREFIX=1
-export LLAMA_LOG_TIMESTAMPS=1
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
+## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
@@ -828,7 +754,6 @@ test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
-    test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
         test $ret -eq 0 && gg_run test_scripts_debug
@@ -836,7 +761,7 @@ 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

@@ -51,23 +51,21 @@ endif()
 set(TARGET common)
 
 add_library(${TARGET} STATIC
-    arg.cpp
-    arg.h
     base64.hpp
-    common.cpp
     common.h
-    console.cpp
-    console.h
-    json-schema-to-grammar.cpp
-    json.hpp
-    log.cpp
-    log.h
-    ngram-cache.cpp
-    ngram-cache.h
-    sampling.cpp
+    common.cpp
     sampling.h
-    train.cpp
+    sampling.cpp
+    console.h
+    console.cpp
+    grammar-parser.h
+    grammar-parser.cpp
+    json.hpp
+    json-schema-to-grammar.cpp
     train.h
+    train.cpp
+    ngram-cache.h
+    ngram-cache.cpp
     )
 
 if (BUILD_SHARED_LIBS)

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

@@ -4,9 +4,18 @@
 
 #include "llama.h"
 
+#include "sampling.h"
+
+#define LOG_NO_FILE_LINE_FUNCTION
+#include "log.h"
+
+#include <cmath>
 #include <string>
 #include <vector>
-#include <sstream>
+#include <random>
+#include <thread>
+#include <unordered_map>
+#include <tuple>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -24,15 +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;
@@ -45,103 +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)
@@ -168,11 +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;
 
@@ -183,32 +101,34 @@ struct gpt_params {
     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_token             = ""; // HF token
+    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
 
@@ -245,15 +165,15 @@ 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 ctx_shift         = true;  // context shift on inifinite text generation
 
     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
@@ -263,7 +183,7 @@ 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
@@ -271,24 +191,23 @@ struct gpt_params {
     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embendings
-    bool reranking         = false; // enable reranking support on server
 
     // server 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;
@@ -336,24 +255,18 @@ 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;
 };
 
-// call once at the start of a program if it uses libcommon
-// initializes the logging system and prints info about the build
-void gpt_init();
+void gpt_params_handle_hf_token(gpt_params & params);
+void gpt_params_handle_model_default(gpt_params & params);
+
+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);
 
-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);
-
 //
 // String utils
 //
@@ -363,8 +276,6 @@ 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;
@@ -382,11 +293,6 @@ static std::vector<T> string_split(const std::string & str, char delim) {
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
-std::string string_from(bool value);
-std::string string_from(const std::vector<int> & values);
-std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
-std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
-
 //
 // Filesystem utils
 //
@@ -401,24 +307,15 @@ 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 ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_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 * 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);
-
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -463,6 +360,10 @@ std::string llama_detokenize(
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
+// 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.cpp

@@ -1,401 +0,0 @@
-#include "log.h"
-
-#include <condition_variable>
-#include <cstdarg>
-#include <cstdio>
-#include <mutex>
-#include <sstream>
-#include <thread>
-#include <vector>
-
-int gpt_log_verbosity_thold = LOG_DEFAULT_LLAMA;
-
-void gpt_log_set_verbosity_thold(int verbosity) {
-    gpt_log_verbosity_thold = verbosity;
-}
-
-#define LOG_COL_DEFAULT "\033[0m"
-#define LOG_COL_BOLD    "\033[1m"
-#define LOG_COL_RED     "\033[31m"
-#define LOG_COL_GREEN   "\033[32m"
-#define LOG_COL_YELLOW  "\033[33m"
-#define LOG_COL_BLUE    "\033[34m"
-#define LOG_COL_MAGENTA "\033[35m"
-#define LOG_COL_CYAN    "\033[36m"
-#define LOG_COL_WHITE   "\033[37m"
-
-static int64_t t_us() {
-    return std::chrono::duration_cast<std::chrono::microseconds>(std::chrono::system_clock::now().time_since_epoch()).count();
-}
-
-// colors
-enum gpt_log_col : int {
-    GPT_LOG_COL_DEFAULT = 0,
-    GPT_LOG_COL_BOLD,
-    GPT_LOG_COL_RED,
-    GPT_LOG_COL_GREEN,
-    GPT_LOG_COL_YELLOW,
-    GPT_LOG_COL_BLUE,
-    GPT_LOG_COL_MAGENTA,
-    GPT_LOG_COL_CYAN,
-    GPT_LOG_COL_WHITE,
-};
-
-// disable colors by default
-static std::vector<const char *> g_col = {
-    "",
-    "",
-    "",
-    "",
-    "",
-    "",
-    "",
-    "",
-    "",
-};
-
-struct gpt_log_entry {
-    enum ggml_log_level level;
-
-    bool prefix;
-
-    int64_t timestamp;
-
-    std::vector<char> msg;
-
-    // signals the worker thread to stop
-    bool is_end;
-
-    void print(FILE * file = nullptr) const {
-        FILE * fcur = file;
-        if (!fcur) {
-            // stderr displays DBG messages only when their verbosity level is not higher than the threshold
-            // these messages will still be logged to a file
-            if (level == GGML_LOG_LEVEL_DEBUG && gpt_log_verbosity_thold < LOG_DEFAULT_DEBUG) {
-                return;
-            }
-
-            fcur = stdout;
-
-            if (level != GGML_LOG_LEVEL_NONE) {
-                fcur = stderr;
-            }
-        }
-
-        if (level != GGML_LOG_LEVEL_NONE && level != GGML_LOG_LEVEL_CONT && prefix) {
-            if (timestamp) {
-                // [M.s.ms.us]
-                fprintf(fcur, "%s%d.%02d.%03d.%03d%s ",
-                        g_col[GPT_LOG_COL_BLUE],
-                        (int) (timestamp / 1000000 / 60),
-                        (int) (timestamp / 1000000 % 60),
-                        (int) (timestamp / 1000 % 1000),
-                        (int) (timestamp % 1000),
-                        g_col[GPT_LOG_COL_DEFAULT]);
-            }
-
-            switch (level) {
-                case GGML_LOG_LEVEL_INFO:  fprintf(fcur, "%sI %s", g_col[GPT_LOG_COL_GREEN],   g_col[GPT_LOG_COL_DEFAULT]); break;
-                case GGML_LOG_LEVEL_WARN:  fprintf(fcur, "%sW %s", g_col[GPT_LOG_COL_MAGENTA], ""                        ); break;
-                case GGML_LOG_LEVEL_ERROR: fprintf(fcur, "%sE %s", g_col[GPT_LOG_COL_RED],     ""                        ); break;
-                case GGML_LOG_LEVEL_DEBUG: fprintf(fcur, "%sD %s", g_col[GPT_LOG_COL_YELLOW],  ""                        ); break;
-                default:
-                    break;
-            }
-        }
-
-        fprintf(fcur, "%s", msg.data());
-
-        if (level == GGML_LOG_LEVEL_WARN || level == GGML_LOG_LEVEL_ERROR || level == GGML_LOG_LEVEL_DEBUG) {
-            fprintf(fcur, "%s", g_col[GPT_LOG_COL_DEFAULT]);
-        }
-
-        fflush(fcur);
-    }
-};
-
-struct gpt_log {
-    // default capacity - will be expanded if needed
-    gpt_log() : gpt_log(256) {}
-
-    gpt_log(size_t capacity) {
-        file = nullptr;
-        prefix = false;
-        timestamps = false;
-        running = false;
-        t_start = t_us();
-
-        // initial message size - will be expanded if longer messages arrive
-        entries.resize(capacity);
-        for (auto & entry : entries) {
-            entry.msg.resize(256);
-        }
-
-        head = 0;
-        tail = 0;
-
-        resume();
-    }
-
-    ~gpt_log() {
-        pause();
-        if (file) {
-            fclose(file);
-        }
-    }
-
-private:
-    std::mutex mtx;
-    std::thread thrd;
-    std::condition_variable cv;
-
-    FILE * file;
-
-    bool prefix;
-    bool timestamps;
-    bool running;
-
-    int64_t t_start;
-
-    // ring buffer of entries
-    std::vector<gpt_log_entry> entries;
-    size_t head;
-    size_t tail;
-
-    // worker thread copies into this
-    gpt_log_entry cur;
-
-public:
-    void add(enum ggml_log_level level, const char * fmt, va_list args) {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        if (!running) {
-            // discard messages while the worker thread is paused
-            return;
-        }
-
-        auto & entry = entries[tail];
-
-        {
-            // cannot use args twice, so make a copy in case we need to expand the buffer
-            va_list args_copy;
-            va_copy(args_copy, args);
-
-#if 1
-            const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args);
-            if (n >= entry.msg.size()) {
-                entry.msg.resize(n + 1);
-                vsnprintf(entry.msg.data(), entry.msg.size(), fmt, args_copy);
-            }
-#else
-            // hack for bolding arguments
-
-            std::stringstream ss;
-            for (int i = 0; fmt[i] != 0; i++) {
-                if (fmt[i] == '%') {
-                    ss << LOG_COL_BOLD;
-                    while (fmt[i] != ' ' && fmt[i] != ')' && fmt[i] != ']' && fmt[i] != 0) ss << fmt[i++];
-                    ss << LOG_COL_DEFAULT;
-                    if (fmt[i] == 0) break;
-                }
-                ss << fmt[i];
-            }
-            const size_t n = vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args);
-            if (n >= entry.msg.size()) {
-                entry.msg.resize(n + 1);
-                vsnprintf(entry.msg.data(), entry.msg.size(), ss.str().c_str(), args_copy);
-            }
-#endif
-        }
-
-        entry.level = level;
-        entry.prefix = prefix;
-        entry.timestamp = 0;
-        if (timestamps) {
-            entry.timestamp = t_us() - t_start;
-        }
-        entry.is_end = false;
-
-        tail = (tail + 1) % entries.size();
-        if (tail == head) {
-            // expand the buffer
-            std::vector<gpt_log_entry> new_entries(2*entries.size());
-
-            size_t new_tail = 0;
-
-            do {
-                new_entries[new_tail] = std::move(entries[head]);
-
-                head     = (head     + 1) % entries.size();
-                new_tail = (new_tail + 1);
-            } while (head != tail);
-
-            head = 0;
-            tail = new_tail;
-
-            for (size_t i = tail; i < new_entries.size(); i++) {
-                new_entries[i].msg.resize(256);
-            }
-
-            entries = std::move(new_entries);
-        }
-
-        cv.notify_one();
-    }
-
-    void resume() {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        if (running) {
-            return;
-        }
-
-        running = true;
-
-        thrd = std::thread([this]() {
-            while (true) {
-                {
-                    std::unique_lock<std::mutex> lock(mtx);
-                    cv.wait(lock, [this]() { return head != tail; });
-
-                    cur = entries[head];
-
-                    head = (head + 1) % entries.size();
-                }
-
-                if (cur.is_end) {
-                    break;
-                }
-
-                cur.print(); // stdout and stderr
-
-                if (file) {
-                    cur.print(file);
-                }
-            }
-        });
-    }
-
-    void pause() {
-        {
-            std::lock_guard<std::mutex> lock(mtx);
-
-            if (!running) {
-                return;
-            }
-
-            running = false;
-
-            // push an entry to signal the worker thread to stop
-            {
-                auto & entry = entries[tail];
-                entry.is_end = true;
-
-                tail = (tail + 1) % entries.size();
-            }
-
-            cv.notify_one();
-        }
-
-        thrd.join();
-    }
-
-    void set_file(const char * path) {
-        pause();
-
-        if (file) {
-            fclose(file);
-        }
-
-        if (path) {
-            file = fopen(path, "w");
-        } else {
-            file = nullptr;
-        }
-
-        resume();
-    }
-
-    void set_colors(bool colors) {
-        pause();
-
-        if (colors) {
-            g_col[GPT_LOG_COL_DEFAULT] = LOG_COL_DEFAULT;
-            g_col[GPT_LOG_COL_BOLD]    = LOG_COL_BOLD;
-            g_col[GPT_LOG_COL_RED]     = LOG_COL_RED;
-            g_col[GPT_LOG_COL_GREEN]   = LOG_COL_GREEN;
-            g_col[GPT_LOG_COL_YELLOW]  = LOG_COL_YELLOW;
-            g_col[GPT_LOG_COL_BLUE]    = LOG_COL_BLUE;
-            g_col[GPT_LOG_COL_MAGENTA] = LOG_COL_MAGENTA;
-            g_col[GPT_LOG_COL_CYAN]    = LOG_COL_CYAN;
-            g_col[GPT_LOG_COL_WHITE]   = LOG_COL_WHITE;
-        } else {
-            for (size_t i = 0; i < g_col.size(); i++) {
-                g_col[i] = "";
-            }
-        }
-
-        resume();
-    }
-
-    void set_prefix(bool prefix) {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        this->prefix = prefix;
-    }
-
-    void set_timestamps(bool timestamps) {
-        std::lock_guard<std::mutex> lock(mtx);
-
-        this->timestamps = timestamps;
-    }
-};
-
-//
-// public API
-//
-
-struct gpt_log * gpt_log_init() {
-    return new gpt_log;
-}
-
-struct gpt_log * gpt_log_main() {
-    static struct gpt_log log;
-
-    return &log;
-}
-
-void gpt_log_pause(struct gpt_log * log) {
-    log->pause();
-}
-
-void gpt_log_resume(struct gpt_log * log) {
-    log->resume();
-}
-
-void gpt_log_free(struct gpt_log * log) {
-    delete log;
-}
-
-void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...) {
-    va_list args;
-    va_start(args, fmt);
-    log->add(level, fmt, args);
-    va_end(args);
-}
-
-void gpt_log_set_file(struct gpt_log * log, const char * file) {
-    log->set_file(file);
-}
-
-void gpt_log_set_colors(struct gpt_log * log, bool colors) {
-    log->set_colors(colors);
-}
-
-void gpt_log_set_prefix(struct gpt_log * log, bool prefix) {
-    log->set_prefix(prefix);
-}
-
-void gpt_log_set_timestamps(struct gpt_log * log, bool timestamps) {
-    log->set_timestamps(timestamps);
-}

common/log.h

@@ -1,92 +1,724 @@
 #pragma once
 
-#include "ggml.h" // for ggml_log_level
+#include <chrono>
+#include <cstring>
+#include <sstream>
+#include <iostream>
+#include <thread>
+#include <vector>
+#include <algorithm>
+#include <cinttypes>
 
-#ifndef __GNUC__
-#    define LOG_ATTRIBUTE_FORMAT(...)
-#elif defined(__MINGW32__)
-#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
-#else
-#    define LOG_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+// --------------------------------
+//
+// Basic usage:
+//
+// --------
+//
+//  The LOG() and LOG_TEE() macros are ready to go by default
+//   they do not require any initialization.
+//
+//  LOGLN() and LOG_TEELN() are variants which automatically
+//   include \n character at the end of the log string.
+//
+//  LOG() behaves exactly like printf, by default writing to a logfile.
+//  LOG_TEE() additionally, prints to the screen too ( mimics Unix tee command ).
+//
+//  Default logfile is named
+//   "llama.<threadID>.log"
+//  Default LOG_TEE() secondary output target is
+//   stderr
+//
+//  Logs can be dynamically disabled or enabled using functions:
+//   log_disable()
+//  and
+//   log_enable()
+//
+//  A log target can be changed with:
+//   log_set_target( string )
+//    creating and opening, or re-opening a file by string filename
+//  or
+//   log_set_target( FILE* )
+//    allowing to point at stderr, stdout, or any valid FILE* file handler.
+//
+// --------
+//
+// End of Basic usage.
+//
+// --------------------------------
+
+// Specifies a log target.
+//  default uses log_handler() with "llama.log" log file
+//  this can be changed, by defining LOG_TARGET
+//  like so:
+//
+//  #define LOG_TARGET (a valid FILE*)
+//  #include "log.h"
+//
+//  or it can be simply redirected to stdout or stderr
+//  like so:
+//
+//  #define LOG_TARGET stderr
+//  #include "log.h"
+//
+//  The log target can also be redirected to a different function
+//  like so:
+//
+//  #define LOG_TARGET log_handler_different()
+//  #include "log.h"
+//
+//  FILE* log_handler_different()
+//  {
+//      return stderr;
+//  }
+//
+//  or:
+//
+//  #define LOG_TARGET log_handler_another_one("somelog.log")
+//  #include "log.h"
+//
+//  FILE* log_handler_another_one(char*filename)
+//  {
+//      static FILE* logfile = nullptr;
+//      (...)
+//      if( !logfile )
+//      {
+//          fopen(...)
+//      }
+//      (...)
+//      return logfile
+//  }
+//
+#ifndef LOG_TARGET
+    #define LOG_TARGET log_handler()
 #endif
 
-#define LOG_DEFAULT_DEBUG 1
-#define LOG_DEFAULT_LLAMA 0
+#ifndef LOG_TEE_TARGET
+    #define LOG_TEE_TARGET stderr
+#endif
 
-// needed by the LOG_TMPL macro to avoid computing log arguments if the verbosity lower
-// set via gpt_log_set_verbosity()
-extern int gpt_log_verbosity_thold;
+// Utility for synchronizing log configuration state
+//  since std::optional was introduced only in c++17
+enum LogTriState
+{
+    LogTriStateSame,
+    LogTriStateFalse,
+    LogTriStateTrue
+};
 
-void gpt_log_set_verbosity_thold(int verbosity); // not thread-safe
+// Utility to obtain "pid" like unique process id and use it when creating log files.
+inline std::string log_get_pid()
+{
+   static std::string pid;
+   if (pid.empty())
+   {
+       // std::this_thread::get_id() is the most portable way of obtaining a "process id"
+       //  it's not the same as "pid" but is unique enough to solve multiple instances
+       //  trying to write to the same log.
+       std::stringstream ss;
+       ss << std::this_thread::get_id();
+       pid = ss.str();
+   }
 
-// the gpt_log uses an internal worker thread to print/write log messages
-// when the worker thread is paused, incoming log messages are discarded
-struct gpt_log;
+   return pid;
+}
 
-struct gpt_log * gpt_log_init();
-struct gpt_log * gpt_log_main(); // singleton, automatically destroys itself on exit
-void             gpt_log_pause (struct gpt_log * log); // pause  the worker thread, not thread-safe
-void             gpt_log_resume(struct gpt_log * log); // resume the worker thread, not thread-safe
-void             gpt_log_free  (struct gpt_log * log);
+// Utility function for generating log file names with unique id based on thread id.
+//  invocation with log_filename_generator( "llama", "log" ) creates a string "llama.<number>.log"
+//  where the number is a runtime id of the current thread.
 
-LOG_ATTRIBUTE_FORMAT(3, 4)
-void gpt_log_add(struct gpt_log * log, enum ggml_log_level level, const char * fmt, ...);
+#define log_filename_generator(log_file_basename, log_file_extension) log_filename_generator_impl(LogTriStateSame, log_file_basename, log_file_extension)
 
-// defaults: file = NULL, colors = false, prefix = false, timestamps = false
-//
-// regular log output:
-//
-//   ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
-//   llm_load_tensors: ggml ctx size =    0.27 MiB
-//   llm_load_tensors: offloading 32 repeating layers to GPU
-//   llm_load_tensors: offloading non-repeating layers to GPU
-//
-// with prefix = true, timestamps = true, the log output will look like this:
-//
-//   0.00.035.060 D ggml_backend_metal_log_allocated_size: allocated buffer, size =  6695.84 MiB, ( 6695.91 / 21845.34)
-//   0.00.035.064 I llm_load_tensors: ggml ctx size =    0.27 MiB
-//   0.00.090.578 I llm_load_tensors: offloading 32 repeating layers to GPU
-//   0.00.090.579 I llm_load_tensors: offloading non-repeating layers to GPU
-//
-// I - info    (stdout, V = 0)
-// W - warning (stderr, V = 0)
-// E - error   (stderr, V = 0)
-// D - debug   (stderr, V = LOG_DEFAULT_DEBUG)
-//
+// INTERNAL, DO NOT USE
+inline std::string log_filename_generator_impl(LogTriState multilog, const std::string & log_file_basename, const std::string & log_file_extension)
+{
+    static bool _multilog = false;
 
-void gpt_log_set_file      (struct gpt_log * log, const char * file);       // not thread-safe
-void gpt_log_set_colors    (struct gpt_log * log,       bool   colors);     // not thread-safe
-void gpt_log_set_prefix    (struct gpt_log * log,       bool   prefix);     // whether to output prefix to each log
-void gpt_log_set_timestamps(struct gpt_log * log,       bool   timestamps); // whether to output timestamps in the prefix
+    if (multilog != LogTriStateSame)
+    {
+        _multilog = multilog == LogTriStateTrue;
+    }
 
-// helper macros for logging
-// use these to avoid computing log arguments if the verbosity of the log is higher than the threshold
-//
-// for example:
-//
-//   LOG_DBG("this is a debug message: %d\n", expensive_function());
-//
-// this will avoid calling expensive_function() if LOG_DEFAULT_DEBUG > gpt_log_verbosity_thold
-//
+    std::stringstream buf;
 
-#define LOG_TMPL(level, verbosity, ...) \
-    do { \
-        if ((verbosity) <= gpt_log_verbosity_thold) { \
-            gpt_log_add(gpt_log_main(), (level), __VA_ARGS__); \
-        } \
+    buf << log_file_basename;
+    if (_multilog)
+    {
+        buf << ".";
+        buf << log_get_pid();
+    }
+    buf << ".";
+    buf << log_file_extension;
+
+    return buf.str();
+}
+
+#ifndef LOG_DEFAULT_FILE_NAME
+    #define LOG_DEFAULT_FILE_NAME log_filename_generator("llama", "log")
+#endif
+
+// Utility for turning #define values into string literals
+//  so we can have a define for stderr and
+//  we can print "stderr" instead of literal stderr, etc.
+#define LOG_STRINGIZE1(s) #s
+#define LOG_STRINGIZE(s) LOG_STRINGIZE1(s)
+
+#define LOG_TEE_TARGET_STRING LOG_STRINGIZE(LOG_TEE_TARGET)
+
+// Allows disabling timestamps.
+//  in order to disable, define LOG_NO_TIMESTAMPS
+//  like so:
+//
+//  #define LOG_NO_TIMESTAMPS
+//  #include "log.h"
+//
+#ifndef LOG_NO_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TIMESTAMP_FMT "%s"
+    #define LOG_TIMESTAMP_VAL ,""
+#endif
+
+#ifdef LOG_TEE_TIMESTAMPS
+    #ifndef _MSC_VER
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #else
+        #define LOG_TEE_TIMESTAMP_FMT "[%" PRIu64 "] "
+        #define LOG_TEE_TIMESTAMP_VAL , (std::chrono::duration_cast<std::chrono::duration<std::uint64_t>>(std::chrono::system_clock::now().time_since_epoch())).count()
+    #endif
+#else
+    #define LOG_TEE_TIMESTAMP_FMT "%s"
+    #define LOG_TEE_TIMESTAMP_VAL ,""
+#endif
+
+// Allows disabling file/line/function prefix
+//  in order to disable, define LOG_NO_FILE_LINE_FUNCTION
+//  like so:
+//
+//  #define LOG_NO_FILE_LINE_FUNCTION
+//  #include "log.h"
+//
+#ifndef LOG_NO_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_FLF_FMT "%s"
+    #define LOG_FLF_VAL ,""
+#endif
+
+#ifdef LOG_TEE_FILE_LINE_FUNCTION
+    #ifndef _MSC_VER
+        #define LOG_TEE_FLF_FMT "[%24s:%5d][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, __LINE__, __FUNCTION__
+    #else
+        #define LOG_TEE_FLF_FMT "[%24s:%5ld][%24s] "
+        #define LOG_TEE_FLF_VAL , __FILE__, (long)__LINE__, __FUNCTION__
+    #endif
+#else
+    #define LOG_TEE_FLF_FMT "%s"
+    #define LOG_TEE_FLF_VAL ,""
+#endif
+
+// INTERNAL, DO NOT USE
+//  USE LOG() INSTEAD
+//
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
+    #define LOG_IMPL(str, ...)                                                                                      \
+    do {                                                                                                            \
+        if (LOG_TARGET != nullptr)                                                                                  \
+        {                                                                                                           \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                     \
+        }                                                                                                           \
     } while (0)
+#else
+    #define LOG_IMPL(str, ...)                                                                                           \
+    do {                                                                                                                 \
+        if (LOG_TARGET != nullptr)                                                                                       \
+        {                                                                                                                \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TARGET);                                                                                          \
+        }                                                                                                                \
+    } while (0)
+#endif
 
-#define LOG(...)             LOG_TMPL(GGML_LOG_LEVEL_NONE, 0,         __VA_ARGS__)
-#define LOGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_NONE, verbosity, __VA_ARGS__)
+// INTERNAL, DO NOT USE
+//  USE LOG_TEE() INSTEAD
+//
+#if !defined(_MSC_VER) || defined(__INTEL_LLVM_COMPILER) || defined(__clang__)
+    #define LOG_TEE_IMPL(str, ...)                                                                                                      \
+    do {                                                                                                                                \
+        if (LOG_TARGET != nullptr)                                                                                                      \
+        {                                                                                                                               \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL, __VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                         \
+        }                                                                                                                               \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                         \
+        {                                                                                                                               \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL, __VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                     \
+        }                                                                                                                               \
+    } while (0)
+#else
+    #define LOG_TEE_IMPL(str, ...)                                                                                                           \
+    do {                                                                                                                                     \
+        if (LOG_TARGET != nullptr)                                                                                                           \
+        {                                                                                                                                    \
+            fprintf(LOG_TARGET, LOG_TIMESTAMP_FMT LOG_FLF_FMT str "%s" LOG_TIMESTAMP_VAL LOG_FLF_VAL "", ##__VA_ARGS__);                     \
+            fflush(LOG_TARGET);                                                                                                              \
+        }                                                                                                                                    \
+        if (LOG_TARGET != nullptr && LOG_TARGET != stdout && LOG_TARGET != stderr && LOG_TEE_TARGET != nullptr)                              \
+        {                                                                                                                                    \
+            fprintf(LOG_TEE_TARGET, LOG_TEE_TIMESTAMP_FMT LOG_TEE_FLF_FMT str "%s" LOG_TEE_TIMESTAMP_VAL LOG_TEE_FLF_VAL "", ##__VA_ARGS__); \
+            fflush(LOG_TEE_TARGET);                                                                                                          \
+        }                                                                                                                                    \
+    } while (0)
+#endif
 
-#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  0,                 __VA_ARGS__)
-#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  0,                 __VA_ARGS__)
-#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, 0,                 __VA_ARGS__)
-#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, LOG_DEFAULT_DEBUG, __VA_ARGS__)
-#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  0,                 __VA_ARGS__)
+// The '\0' as a last argument, is a trick to bypass the silly
+//  "warning: ISO C++11 requires at least one argument for the "..." in a variadic macro"
+//  so we can have a single macro which can be called just like printf.
 
-#define LOG_INFV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_INFO,  verbosity, __VA_ARGS__)
-#define LOG_WRNV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_WARN,  verbosity, __VA_ARGS__)
-#define LOG_ERRV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, verbosity, __VA_ARGS__)
-#define LOG_DBGV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, verbosity, __VA_ARGS__)
-#define LOG_CNTV(verbosity, ...) LOG_TMPL(GGML_LOG_LEVEL_CONT,  verbosity, __VA_ARGS__)
+// Main LOG macro.
+//  behaves like printf, and supports arguments the exact same way.
+//
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
+#endif
+
+// Main TEE macro.
+//  does the same as LOG
+//  and
+//  simultaneously writes stderr.
+//
+// Secondary target can be changed just like LOG_TARGET
+//  by defining LOG_TEE_TARGET
+//
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
+#else
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
+#endif
+
+// LOG macro variants with auto endline.
+#if !defined(_MSC_VER) || defined(__clang__)
+    #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
+    #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
+#else
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+#endif
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler1_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, const std::string & filename = LOG_DEFAULT_FILE_NAME, FILE *target = nullptr)
+{
+    static bool _initialized = false;
+    static bool _append = false;
+    static bool _disabled = filename.empty() && target == nullptr;
+    static std::string log_current_filename{filename};
+    static FILE *log_current_target{target};
+    static FILE *logfile = nullptr;
+
+    if (change)
+    {
+        if (append != LogTriStateSame)
+        {
+            _append = append == LogTriStateTrue;
+            return logfile;
+        }
+
+        if (disable == LogTriStateTrue)
+        {
+            // Disable primary target
+            _disabled = true;
+        }
+        // If previously disabled, only enable, and keep previous target
+        else if (disable == LogTriStateFalse)
+        {
+            _disabled = false;
+        }
+        // Otherwise, process the arguments
+        else if (log_current_filename != filename || log_current_target != target)
+        {
+            _initialized = false;
+        }
+    }
+
+    if (_disabled)
+    {
+        // Log is disabled
+        return nullptr;
+    }
+
+    if (_initialized)
+    {
+        // with fallback in case something went wrong
+        return logfile ? logfile : stderr;
+    }
+
+    // do the (re)initialization
+    if (target != nullptr)
+    {
+        if (logfile != nullptr && logfile != stdout && logfile != stderr)
+        {
+            fclose(logfile);
+        }
+
+        log_current_filename = LOG_DEFAULT_FILE_NAME;
+        log_current_target = target;
+
+        logfile = target;
+    }
+    else
+    {
+        if (log_current_filename != filename)
+        {
+            if (logfile != nullptr && logfile != stdout && logfile != stderr)
+            {
+                fclose(logfile);
+            }
+        }
+
+        logfile = fopen(filename.c_str(), _append ? "a" : "w");
+    }
+
+    if (!logfile)
+    {
+        //  Verify whether the file was opened, otherwise fallback to stderr
+        logfile = stderr;
+
+        fprintf(stderr, "Failed to open logfile '%s' with error '%s'\n", filename.c_str(), std::strerror(errno));
+        fflush(stderr);
+
+        // At this point we let the init flag be to true below, and let the target fallback to stderr
+        //  otherwise we would repeatedly fopen() which was already unsuccessful
+    }
+
+    _initialized = true;
+
+    return logfile ? logfile : stderr;
+}
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler2_impl(bool change = false, LogTriState append = LogTriStateSame, LogTriState disable = LogTriStateSame, FILE *target = nullptr, const std::string & filename = LOG_DEFAULT_FILE_NAME)
+{
+    return log_handler1_impl(change, append, disable, filename, target);
+}
+
+// Disables logs entirely at runtime.
+//  Makes LOG() and LOG_TEE() produce no output,
+//  until enabled back.
+#define log_disable() log_disable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_disable_impl()
+{
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateTrue);
+}
+
+// Enables logs at runtime.
+#define log_enable() log_enable_impl()
+
+// INTERNAL, DO NOT USE
+inline FILE *log_enable_impl()
+{
+    return log_handler1_impl(true, LogTriStateSame, LogTriStateFalse);
+}
+
+// Sets target fir logs, either by a file name or FILE* pointer (stdout, stderr, or any valid FILE*)
+#define log_set_target(target) log_set_target_impl(target)
+
+// INTERNAL, DO NOT USE
+inline FILE *log_set_target_impl(const std::string & filename) { return log_handler1_impl(true, LogTriStateSame, LogTriStateSame, filename); }
+inline FILE *log_set_target_impl(FILE *target) { return log_handler2_impl(true, LogTriStateSame, LogTriStateSame, target); }
+
+// INTERNAL, DO NOT USE
+inline FILE *log_handler() { return log_handler1_impl(); }
+
+// Enable or disable creating separate log files for each run.
+//  can ONLY be invoked BEFORE first log use.
+#define log_multilog(enable) log_filename_generator_impl((enable) ? LogTriStateTrue : LogTriStateFalse, "", "")
+// Enable or disable append mode for log file.
+//  can ONLY be invoked BEFORE first log use.
+#define log_append(enable) log_append_impl(enable)
+// INTERNAL, DO NOT USE
+inline FILE *log_append_impl(bool enable)
+{
+    return log_handler1_impl(true, enable ? LogTriStateTrue : LogTriStateFalse, LogTriStateSame);
+}
+
+inline void log_test()
+{
+    log_disable();
+    LOG("01 Hello World to nobody, because logs are disabled!\n");
+    log_enable();
+    LOG("02 Hello World to default output, which is \"%s\" ( Yaaay, arguments! )!\n", LOG_STRINGIZE(LOG_TARGET));
+    LOG_TEE("03 Hello World to **both** default output and " LOG_TEE_TARGET_STRING "!\n");
+    log_set_target(stderr);
+    LOG("04 Hello World to stderr!\n");
+    LOG_TEE("05 Hello World TEE with double printing to stderr prevented!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("06 Hello World to default log file!\n");
+    log_set_target(stdout);
+    LOG("07 Hello World to stdout!\n");
+    log_set_target(LOG_DEFAULT_FILE_NAME);
+    LOG("08 Hello World to default log file again!\n");
+    log_disable();
+    LOG("09 Hello World _1_ into the void!\n");
+    log_enable();
+    LOG("10 Hello World back from the void ( you should not see _1_ in the log or the output )!\n");
+    log_disable();
+    log_set_target("llama.anotherlog.log");
+    LOG("11 Hello World _2_ to nobody, new target was selected but logs are still disabled!\n");
+    log_enable();
+    LOG("12 Hello World this time in a new file ( you should not see _2_ in the log or the output )?\n");
+    log_set_target("llama.yetanotherlog.log");
+    LOG("13 Hello World this time in yet new file?\n");
+    log_set_target(log_filename_generator("llama_autonamed", "log"));
+    LOG("14 Hello World in log with generated filename!\n");
+#ifdef _MSC_VER
+    LOG_TEE("15 Hello msvc TEE without arguments\n");
+    LOG_TEE("16 Hello msvc TEE with (%d)(%s) arguments\n", 1, "test");
+    LOG_TEELN("17 Hello msvc TEELN without arguments\n");
+    LOG_TEELN("18 Hello msvc TEELN with (%d)(%s) arguments\n", 1, "test");
+    LOG("19 Hello msvc LOG without arguments\n");
+    LOG("20 Hello msvc LOG with (%d)(%s) arguments\n", 1, "test");
+    LOGLN("21 Hello msvc LOGLN without arguments\n");
+    LOGLN("22 Hello msvc LOGLN with (%d)(%s) arguments\n", 1, "test");
+#endif
+}
+
+inline bool log_param_single_parse(const std::string & param)
+{
+    if ( param == "--log-test")
+    {
+        log_test();
+        return true;
+    }
+
+    if ( param == "--log-disable")
+    {
+        log_disable();
+        return true;
+    }
+
+    if ( param == "--log-enable")
+    {
+        log_enable();
+        return true;
+    }
+
+    if (param == "--log-new")
+    {
+        log_multilog(true);
+        return true;
+    }
+
+    if (param == "--log-append")
+    {
+        log_append(true);
+        return true;
+    }
+
+    return false;
+}
+
+inline bool log_param_pair_parse(bool check_but_dont_parse, const std::string & param, const std::string & next = std::string())
+{
+    if ( param == "--log-file")
+    {
+        if (!check_but_dont_parse)
+        {
+            log_set_target(log_filename_generator(next.empty() ? "unnamed" : next, "log"));
+        }
+
+        return true;
+    }
+
+    return false;
+}
+
+inline void log_print_usage()
+{
+    printf("log options:\n");
+    /* format
+    printf("  -h, --help            show this help message and exit\n");*/
+    /* spacing
+    printf("__-param----------------Description\n");*/
+    printf("  --log-test            Run simple logging test\n");
+    printf("  --log-disable         Disable trace logs\n");
+    printf("  --log-enable          Enable trace logs\n");
+    printf("  --log-file            Specify a log filename (without extension)\n");
+    printf("  --log-new             Create a separate new log file on start. "
+                                   "Each log file will have unique name: \"<name>.<ID>.log\"\n");
+    printf("  --log-append          Don't truncate the old log file.\n");
+    printf("\n");
+}
+
+#define log_dump_cmdline(argc, argv) log_dump_cmdline_impl(argc, argv)
+
+// INTERNAL, DO NOT USE
+inline void log_dump_cmdline_impl(int argc, char **argv)
+{
+    std::stringstream buf;
+    for (int i = 0; i < argc; ++i)
+    {
+        if (std::string(argv[i]).find(' ') != std::string::npos)
+        {
+            buf << " \"" << argv[i] <<"\"";
+        }
+        else
+        {
+            buf << " " << argv[i];
+        }
+    }
+    LOGLN("Cmd:%s", buf.str().c_str());
+}
+
+#define log_tostr(var) log_var_to_string_impl(var).c_str()
+
+inline std::string log_var_to_string_impl(bool var)
+{
+    return var ? "true" : "false";
+}
+
+inline std::string log_var_to_string_impl(std::string var)
+{
+    return var;
+}
+
+inline std::string log_var_to_string_impl(const std::vector<int> & var)
+{
+    std::stringstream buf;
+    buf << "[ ";
+    bool first = true;
+    for (auto e : var)
+    {
+        if (first)
+        {
+            first = false;
+        }
+        else
+        {
+            buf << ", ";
+        }
+        buf << std::to_string(e);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+template <typename C, typename T>
+inline std::string LOG_TOKENS_TOSTR_PRETTY(const C & ctx, const T & tokens)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (const auto & token : tokens)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, token);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "'" << detokenized << "'"
+            << ":" << std::to_string(token);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+template <typename C, typename B>
+inline std::string LOG_BATCH_TOSTR_PRETTY(const C & ctx, const B & batch)
+{
+    std::stringstream buf;
+    buf << "[ ";
+
+    bool first = true;
+    for (int i = 0; i < batch.n_tokens; ++i)
+    {
+        if (!first) {
+            buf << ", ";
+        } else {
+            first = false;
+        }
+
+        auto detokenized = llama_token_to_piece(ctx, batch.token[i]);
+
+        detokenized.erase(
+            std::remove_if(
+                detokenized.begin(),
+                detokenized.end(),
+                [](const unsigned char c) { return !std::isprint(c); }),
+            detokenized.end());
+
+        buf
+            << "\n" << std::to_string(i)
+            << ":token '" << detokenized << "'"
+            << ":pos " << std::to_string(batch.pos[i])
+            << ":n_seq_id  " << std::to_string(batch.n_seq_id[i])
+            << ":seq_id " << std::to_string(batch.seq_id[i][0])
+            << ":logits " << std::to_string(batch.logits[i]);
+    }
+    buf << " ]";
+
+    return buf.str();
+}
+
+#ifdef LOG_DISABLE_LOGS
+
+#undef LOG
+#define LOG(...) // dummy stub
+#undef LOGLN
+#define LOGLN(...) // dummy stub
+
+#undef LOG_TEE
+#define LOG_TEE(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_TEELN
+#define LOG_TEELN(...) fprintf(stderr, __VA_ARGS__) // convert to normal fprintf
+
+#undef LOG_DISABLE
+#define LOG_DISABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_ENABLE
+#define LOG_ENABLE() // dummy stub
+
+#undef LOG_SET_TARGET
+#define LOG_SET_TARGET(...) // dummy stub
+
+#undef LOG_DUMP_CMDLINE
+#define LOG_DUMP_CMDLINE(...) // dummy stub
+
+#endif // LOG_DISABLE_LOGS

common/ngram-cache.cpp

@@ -2,11 +2,8 @@
 #include "common.h"
 #include "log.h"
 
-#include <cinttypes>
 #include <cstdint>
-#include <cstdio>
 #include <fstream>
-#include <thread>
 
 void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
                               std::vector<llama_token> & inp, int nnew, bool print_progress) {

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

@@ -1,458 +1,460 @@
+#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 {
-        if (params.n_probs > 0) {
-            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
-            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
-            //
-            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
-            // it is much faster, since we avoid sorting all tokens and should give a good approximation
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
-            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 = "logits ";
-
-    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;
+
+    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) {
+        // Get a pointer to the logits
+        float * logits = llama_get_logits_ith(ctx_main, idx);
+
+        // 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 + n_vocab};
+    }
+
+    // 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.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};
+    }
+
+    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);

common/train.cpp

@@ -1,11 +1,9 @@
 #include "train.h"
 #include "common.h"
 
-#include <algorithm>
 #include <random>
 #include <sstream>
 #include <functional>
-#include <cstring>
 
 struct random_normal_distribution {
     std::mt19937 gen;

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]
@@ -81,7 +80,6 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
-    {"name": "jina-v1-en",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
@@ -93,14 +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", },
-    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
 ]
 
 
@@ -109,8 +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")
 
 
@@ -129,27 +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:
@@ -184,7 +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}")
@@ -216,7 +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()
@@ -312,7 +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)

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/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/backend/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
 
@@ -628,26 +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.
-
-- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
-
-  Device Memory is not enough.
-
-  |Reason|Solution|
-  |-|-|
-  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
-  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
-
 ### **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.

docs/build.md

@@ -16,7 +16,7 @@ In order to build llama.cpp you have four different options.
       make
       ```
 
-  - On Windows (x86/x64 only, arm64 requires cmake):
+  - On Windows:
 
     1. Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
     2. Extract `w64devkit` on your pc.
@@ -60,17 +60,6 @@ In order to build llama.cpp you have four different options.
       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):
 
@@ -178,11 +167,7 @@ For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](ht
   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:
+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                                                                                                                                                                                                                                                                             |
 |-------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
@@ -196,19 +181,6 @@ The following compilation options are also available to tweak performance:
 | 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.
@@ -270,45 +242,6 @@ The following compilation options are also available to tweak performance (yes,
 
 ### 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.
@@ -352,37 +285,6 @@ cmake --build build --config Release
 # 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/development/HOWTO-add-model.md

@@ -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/development/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 "
 ```

docs/docker.md

@@ -20,7 +20,7 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 
-The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](../.devops/) and the GitHub Action defined in [.github/workflows/docker.yml](../.github/workflows/docker.yml). If you need different settings (for example, a different CUDA or ROCm library, you'll need to build the images locally for now).
+The GPU enabled images are not currently tested by CI beyond being built. They are not built with any variation from the ones in the Dockerfiles defined in [.devops/](.devops/) and the 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
 
@@ -66,8 +66,8 @@ You may want to pass in some different `ARGS`, depending on the CUDA environment
 
 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
+- `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:
 

examples/CMakeLists.txt

@@ -21,6 +21,7 @@ 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)
@@ -52,4 +53,5 @@ 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

@@ -1,28 +1,49 @@
-#include "arg.h"
 #include "common.h"
-#include "log.h"
 #include "llama.h"
 
 #include <algorithm>
+#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int, char ** argv) {
-    LOG("\nexample usage:\n");
-    LOG("\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("\n");
+// 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");
 }
 
 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;
     }
 
-    gpt_init();
-
     int is_pp_shared = params.is_pp_shared;
 
     std::vector<int> n_pp = params.n_pp;
@@ -48,7 +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);
 
@@ -79,7 +100,7 @@ int main(int argc, char ** argv) {
 
             const int ret = llama_decode(ctx, batch_view);
             if (ret != 0) {
-                LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
+                LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
 
@@ -96,18 +117,17 @@ int main(int argc, char ** argv) {
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-            LOG_ERR("%s: llama_decode() failed\n", __func__);
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
             return 1;
         }
     }
 
-    if (!params.batched_bench_output_jsonl) {
-        LOG("\n");
-        LOG("%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("\n");
-        LOG("|%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("|%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) {
@@ -136,7 +156,7 @@ int main(int argc, char ** argv) {
                 llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                    LOG_ERR("%s: llama_decode() failed\n", __func__);
+                    LOG_TEE("%s: llama_decode() failed\n", __func__);
                     return 1;
                 }
 
@@ -158,7 +178,7 @@ int main(int argc, char ** argv) {
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                        LOG_ERR("%s: llama_decode() failed\n", __func__);
+                        LOG_TEE("%s: llama_decode() failed\n", __func__);
                         return 1;
                     }
                 }
@@ -175,22 +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(
-                        "{\"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("|%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("\n");
-    llama_perf_context_print(ctx);
+    llama_print_timings(ctx);
 
     llama_batch_free(batch);
 
@@ -199,7 +209,7 @@ int main(int argc, char ** argv) {
 
     llama_backend_free();
 
-    LOG("\n\n");
+    fprintf(stderr, "\n\n");
 
     return 0;
 }

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

examples/batched/batched.cpp

@@ -1,17 +1,18 @@
-#include "arg.h"
 #include "common.h"
-#include "log.h"
 #include "llama.h"
 
 #include <algorithm>
+#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int, char ** argv) {
-    LOG("\nexample usage:\n");
-    LOG("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
-    LOG("\n");
+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");
 }
 
 int main(int argc, char ** argv) {
@@ -20,17 +21,17 @@ 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;
     }
 
-    gpt_init();
 
     // number of parallel batches
     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
 
@@ -44,7 +45,7 @@ int main(int argc, char ** argv) {
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
-        LOG_ERR("%s: error: unable to load model\n" , __func__);
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
@@ -64,39 +65,32 @@ 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) {
-        LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
     }
 
     const int n_ctx = llama_n_ctx(ctx);
 
-    LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
-        LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
-        LOG_ERR("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
     // print the prompt token-by-token
 
-    LOG("\n");
+    fprintf(stderr, "\n");
 
     for (auto id : tokens_list) {
-        LOG("%s", llama_token_to_piece(ctx, id).c_str());
+        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
     }
 
+    fflush(stderr);
+
     // create a llama_batch
     // we use this object to submit token data for decoding
     llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
@@ -114,7 +108,7 @@ int main(int argc, char ** argv) {
 
     if (llama_model_has_encoder(model)) {
         if (llama_encode(ctx, batch)) {
-            LOG_ERR("%s : failed to eval\n", __func__);
+            LOG_TEE("%s : failed to eval\n", __func__);
             return 1;
         }
 
@@ -131,7 +125,7 @@ int main(int argc, char ** argv) {
     batch.logits[batch.n_tokens - 1] = true;
 
     if (llama_decode(ctx, batch) != 0) {
-        LOG_ERR("%s: llama_decode() failed\n", __func__);
+        LOG_TEE("%s: llama_decode() failed\n", __func__);
         return 1;
     }
 
@@ -142,7 +136,7 @@ int main(int argc, char ** argv) {
     //}
 
     if (n_parallel > 1) {
-        LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
     }
 
     // main loop
@@ -170,14 +164,36 @@ 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) {
                 i_batch[i] = -1;
-                LOG("\n");
+                LOG_TEE("\n");
                 if (n_parallel > 1) {
-                    LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
                 }
 
                 continue;
@@ -185,7 +201,8 @@ int main(int argc, char ** argv) {
 
             // if there is only one stream, we print immediately to stdout
             if (n_parallel == 1) {
-                LOG("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+                fflush(stdout);
             }
 
             streams[i] += llama_token_to_piece(ctx, new_token_id);
@@ -207,33 +224,32 @@ int main(int argc, char ** argv) {
 
         // evaluate the current batch with the transformer model
         if (llama_decode(ctx, batch)) {
-            LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
     }
 
+    LOG_TEE("\n");
+
     if (n_parallel > 1) {
-        LOG("\n");
+        LOG_TEE("\n");
 
         for (int32_t i = 0; i < n_parallel; ++i) {
-            LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
         }
     }
 
     const auto t_main_end = ggml_time_us();
 
-    LOG_INF("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+    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("\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-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

@@ -9,7 +9,6 @@
 #include <climits>
 #include <cstring>
 #include <cstdarg>
-#include <cinttypes>
 #include <ctime>
 #include <random>
 #include <stdexcept>
@@ -106,43 +105,43 @@ static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_
     const int n_multiqueries = p->n_kv_heads <= 0 || p->n_kv_heads >= p->n_heads ? 1 : p->n_heads / p->n_kv_heads;
     try {
         w->token_embedding_table.resize(p->vocab_size * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
 
         w->rms_att_weight.resize(p->n_layers * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_att_weight\n",__func__,p->n_layers, p->dim, p->n_layers * p->dim);
 
         w->rms_ffn_weight.resize(p->n_layers * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
+        LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->rms_ffn_weight\n",__func__,p->n_layers , p->dim, p->n_layers * p->dim);
 
         w->wq.resize(p->n_layers * p->dim * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wq\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
         w->wk.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\n",__func__,p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
         w->wv.resize(p->n_layers * p->dim * p->dim / n_multiqueries);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim / n_multiqueries, p->n_layers * p->dim * p->dim / n_multiqueries);
 
         w->wo.resize(p->n_layers * p->dim * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->wo\n",__func__,p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
 
         w->w1.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w1\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->w2.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w2\n",__func__,p->n_layers, p->dim, p->hidden_dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->w3.resize(p->n_layers * p->hidden_dim * p->dim);
-        LOG_INF("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
+        LOG("%s: Allocating [%d] x [%d] x [%d] = [%d] float space for w->w3\n",__func__,p->n_layers, p->hidden_dim, p->dim, p->n_layers * p->hidden_dim * p->dim);
 
         w->rms_final_weight.resize(p->dim);
-        LOG_INF("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
         if (shared_weights) {
             w->wcls = {};
         } else {
             w->wcls.resize(p->vocab_size * p->dim);
-            LOG_INF("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+            LOG("%s: Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
         }
     }
     catch (std::length_error &) {
@@ -174,7 +173,7 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        LOG_ERR("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -182,26 +181,26 @@ static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FIL
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    LOG_INF("----- Quick print of first of the weight vales of all the variables\n");
-    LOG_INF("%f\n", w->token_embedding_table[0]);
-    LOG_INF("%f\n", w->rms_att_weight[0]);
-    LOG_INF("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    LOG_INF("%f\n", w->wq[0]);
-    LOG_INF("%f\n", w->wk[0]);
-    LOG_INF("%f\n", w->wv[0]);
-    LOG_INF("%f\n", w->wo[0]);
-    LOG_INF("%f\n", w->w1[0]);
-    LOG_INF("%f\n", w->w2[0]);
-    LOG_INF("%f\n", w->w3[0]);
-    LOG_INF("%f\n", w->rms_att_weight[0]);
-    if (!w->wcls.empty()) LOG_INF("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
 //////////////////////////////////////// ggml structs and functions required to load models, configs and save the model.
 
-struct my_llama_vocab {
+struct llama_vocab {
     using id    = int32_t;
     using token = std::string;
     using ttype = llama_token_type;
@@ -319,20 +318,20 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    LOG_INF("%s: n_vocab:   %u\n", __func__, params->n_vocab);
-    LOG_INF("%s: n_ctx:     %u\n", __func__, params->n_ctx);
-    LOG_INF("%s: n_embd:    %u\n", __func__, params->n_embd);
-    LOG_INF("%s: n_mult:    %u\n", __func__, params->n_mult);
-    LOG_INF("%s: n_head:    %u\n", __func__, params->n_head);
-    LOG_INF("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
-    LOG_INF("%s: n_ff:      %u\n", __func__, params->n_ff);
-    LOG_INF("%s: n_layer:   %u\n", __func__, params->n_layer);
-    LOG_INF("%s: n_rot:     %u\n", __func__, params->n_rot);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
 }
 
 static void print_tensor_info(const struct ggml_context * ctx) {
     for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
-        LOG_INF("%s: Allocating ", __func__);
+        LOG("%s: Allocating ", __func__);
         int64_t total = 1;
         int i = 0;
         for (; i < ggml_n_dims(t); ++i) {
@@ -525,9 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char * filename, const Config * config, struct my_llama_vocab * vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
-        LOG_INF("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -575,7 +574,7 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        LOG_INF("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -583,13 +582,13 @@ static void load_vocab(const char * filename, const Config * config, struct my_l
         const int  n_vocab = config->vocab_size;
         /* uint32_t max_token_length =  */ file.read_u32(); // unused
         vocab->id_to_token.resize(n_vocab);
-        for (my_llama_vocab::id id=0; id<n_vocab; ++id) {
+        for (llama_vocab::id id=0; id<n_vocab; ++id) {
             float_t score = file.read_f32();
             uint32_t len = file.read_u32();
             std::string text = file.read_string(len);
 
             unsigned char byte_val;
-            my_llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
+            llama_vocab::ttype type = LLAMA_TOKEN_TYPE_NORMAL;
             if (id == UNKNOWN_TOKEN_ID) {
                 text = "<unk>";
                 type = LLAMA_TOKEN_TYPE_UNKNOWN;
@@ -631,7 +630,7 @@ static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const floa
 }
 
 static void save_as_llama_model(
-    struct my_llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
+    struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename
 ) {
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
@@ -671,7 +670,7 @@ static void save_as_llama_model(
     std::vector<const char*> tokens;
     std::vector<float> scores;
     std::vector<llama_token_type> token_types;
-    for (const my_llama_vocab::token_data & token_data : vocab->id_to_token) {
+    for (const llama_vocab::token_data & token_data : vocab->id_to_token) {
         tokens.push_back(token_data.text.c_str());
         scores.push_back(token_data.score);
         token_types.push_back(token_data.type);
@@ -872,25 +871,23 @@ static std::string basename(const std::string &path) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_init();
-
     struct train_params params = get_default_train_params();
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
-
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        LOG_INF("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
+        LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
         FILE * file = fopen(params.fn_llama2c_model, "rb");
         if (!file) {
-            LOG_ERR("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
+            LOG("%s: Unable to open the checkpoint file %s!\n", __func__, params.fn_llama2c_model);
             return 1;
         }
         // read in the config header
         if (fread(&config, sizeof(Config), 1, file) != 1) {
-            LOG_ERR("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
             return 1;
         }
         auto shared_weights = config.vocab_size > 0;
@@ -899,13 +896,13 @@ int main(int argc, char ** argv) {
         // read in the Transformer weights
         alloc_weights(&weights, &config, shared_weights);
         if (checkpoint_init_weights(&weights, &config, file, shared_weights)) {
-            LOG_ERR("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
+            LOG("%s: Unable to initialize transformer weights from %s!",__func__,params.fn_llama2c_model);
             return 1;
         }
         fclose(file);
     }
 
-    struct my_llama_vocab vocab;
+    struct llama_vocab vocab;
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
@@ -932,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    LOG_INF("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

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
+    license: 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?)
@@ -770,7 +806,7 @@ 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) -> None:
         # Metadata About The Model And Its Provenence
         name = "LLaMA"
         if metadata is not None and metadata.name is not None:
@@ -788,73 +824,16 @@ 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.license is not None:
+                self.gguf.add_licence(metadata.license)
             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 +944,7 @@ 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,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -999,7 +978,7 @@ 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,
     ) -> None:
         check_vocab_size(params, vocab, pad_vocab=pad_vocab)
 
@@ -1042,32 +1021,35 @@ 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 +1231,34 @@ 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 +1297,8 @@ 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 +1310,32 @@ 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 +1368,7 @@ 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)
@@ -1412,21 +1399,13 @@ def main(args_in: list[str] | None = None) -> None:
 
     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"
@@ -13,15 +12,14 @@
 #include "ggml-metal.h"
 #endif
 
-#include <algorithm>
-#include <climits>
 #include <cstdio>
-#include <cstring>
-#include <fstream>
-#include <iostream>
 #include <string>
 #include <tuple>
 #include <vector>
+#include <algorithm>
+#include <iostream>
+#include <fstream>
+#include <climits>
 
 
 //////////////////////////////////////////////////
@@ -37,7 +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]);
@@ -271,7 +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());
@@ -390,7 +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;
     }
 
@@ -413,10 +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);
@@ -485,8 +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

@@ -13,6 +13,7 @@ Please update all scripts and workflows to use the new binary names.
 | server | llama-server |
 | llama-bench | llama-bench |
 | embedding | llama-embedding |
+| finetune | llama-finetune |
 | quantize | llama-quantize |
 | tokenize | llama-tokenize |
 | export-lora | llama-export-lora |
@@ -44,6 +45,7 @@ Please update all scripts and workflows to use the new binary names.
 | save-load-state | llama-save-load-state |
 | simple | llama-simple |
 | speculative | llama-speculative |
+| train-text-from-scratch | llama-train-text-from-scratch |
 | vdot | llama-vdot |
 | tests/test-c.o | tests/test-c.o |
 

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

@@ -1,6 +1,4 @@
-#include "arg.h"
 #include "common.h"
-#include "log.h"
 #include "llama.h"
 
 #include <ctime>
@@ -33,24 +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
-    LOG_INF("%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) {
-            LOG_ERR("%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) {
-            LOG_ERR("%s : failed to decode\n", __func__);
-        }
+    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_decode(ctx, batch) < 0) {
+        fprintf(stderr, "%s : failed to decode\n", __func__);
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -58,22 +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);
     }
 }
@@ -81,26 +57,35 @@ 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;
     }
 
-    gpt_init();
-
     params.embedding = true;
     // For non-causal models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
 
+    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) {
-        LOG_ERR("%s: unable to load model\n", __func__);
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
         return 1;
     }
 
@@ -108,21 +93,20 @@ 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)) {
-        LOG_ERR("%s: 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;
     }
 
     if (n_ctx > n_ctx_train) {
-        LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
-        LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     // split the prompt into lines
@@ -135,9 +119,9 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true, true);
+        auto inp = ::llama_tokenize(ctx, prompt, true, false);
         if (inp.size() > n_batch) {
-            LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
+            fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
             return 1;
         }
@@ -148,20 +132,20 @@ int main(int argc, char ** argv) {
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
         if (inp.empty() || inp.back() != llama_token_sep(model)) {
-            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
-            LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
+            fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
+            fprintf(stderr, "%s:          'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
 
     // tokenization stats
     if (params.verbose_prompt) {
         for (int i = 0; i < (int) inputs.size(); i++) {
-            LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
-            LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
             for (int j = 0; j < (int) inputs[i].size(); j++) {
-                LOG("%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
             }
-            LOG("\n\n");
+            fprintf(stderr, "\n\n");
         }
     }
 
@@ -169,23 +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
@@ -195,11 +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
@@ -208,67 +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()) {
-        LOG("\n");
+        // print the first part of the embeddings or for a single prompt, the full embedding
+        fprintf(stdout, "\n");
+        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");
+        }
 
-        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-            for (int j = 0; j < n_embd_count; j++) {
-                LOG("embedding %d: ", j);
-                for (int i = 0; i < std::min(3, n_embd); i++) {
-                    if (params.embd_normalize == 0) {
-                        LOG("%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        LOG("%9.6f ", emb[j * n_embd + i]);
-                    }
-                }
-                LOG(" ... ");
-                for (int i = n_embd - 3; i < n_embd; i++) {
-                    if (params.embd_normalize == 0) {
-                        LOG("%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        LOG("%9.6f ", emb[j * n_embd + i]);
-                    }
-                }
-                LOG("\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());
             }
-        } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
-            for (int j = 0; j < n_embd_count; j++) {
-                // NOTE: if you change this log - update the tests in ci/run.sh
-                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
-            }
-        } else {
-            // print the first part of the embeddings or for a single prompt, the full embedding
-            for (int j = 0; j < n_prompts; j++) {
-                LOG("embedding %d: ", j);
-                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                    if (params.embd_normalize == 0) {
-                        LOG("%6.0f ", emb[j * n_embd + i]);
-                    } else {
-                        LOG("%9.6f ", emb[j * n_embd + i]);
-                    }
-                }
-                LOG("\n");
-            }
-
-            // print cosine similarity matrix
-            if (n_prompts > 1) {
-                LOG("\n");
-                LOG("cosine similarity matrix:\n\n");
-                for (int i = 0; i < n_prompts; i++) {
-                    LOG("%6.6s ", prompts[i].c_str());
-                }
-                LOG("\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);
-                        LOG("%6.2f ", sim);
-                    }
-                    LOG("%1.10s", prompts[i].c_str());
-                    LOG("\n");
+            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");
             }
         }
     }
@@ -276,45 +222,43 @@ int main(int argc, char ** argv) {
     if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
         const bool notArray = params.embd_out != "array";
 
-        LOG(notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
+        fprintf(stdout, notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
         for (int j = 0;;) { // at least one iteration (one prompt)
-            if (notArray) LOG("    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
-            LOG("[");
+            if (notArray) fprintf(stdout, "    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
+            fprintf(stdout, "[");
             for (int i = 0;;) { // at least one iteration (n_embd > 0)
-                LOG(params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
+                fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
                 i++;
-                if (i < n_embd) LOG(","); else break;
+                if (i < n_embd) fprintf(stdout, ","); else break;
             }
-            LOG(notArray ? "]\n    }" : "]");
+            fprintf(stdout, notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_embd_count) LOG(notArray ? ",\n" : ","); else break;
+            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
         }
-        LOG(notArray ? "\n  ]" : "]\n");
+        fprintf(stdout, notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
-            LOG(",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
-                LOG("    [");
-                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
+            fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
+            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
+                fprintf(stdout, "    [");
+                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);
-                    LOG("%6.2f", sim);
+                    fprintf(stdout, "%6.2f", sim);
                     j++;
-                    if (j < n_embd_count) LOG(", "); else break;
+                    if (j < n_prompts) fprintf(stdout, ", "); else break;
                 }
-                LOG(" ]");
+                fprintf(stdout, " ]");
                 i++;
-                if (i < n_embd_count) LOG(",\n"); else break;
+                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
             }
-            LOG("\n  ]");
+            fprintf(stdout, "\n  ]");
         }
 
-        if (notArray) LOG("\n}\n");
+        if (notArray) fprintf(stdout, "\n}\n");
     }
 
-    LOG("\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

@@ -1,11 +1,11 @@
-#include "arg.h"
 #include "common.h"
-#include "log.h"
 #include "llama.h"
 #include "ggml.h"
 
 #include <cstdio>
+#include <random>
 #include <string>
+#include <tuple>
 #include <vector>
 
 /**
@@ -31,22 +31,22 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
     GGML_ASSERT(n > 0);
     float sum = 0;
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        LOG("                                     [\n");
+        printf("                                     [\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2*n) {
-                LOG("                                      ..., \n");
+                printf("                                      ..., \n");
                 i2 = ne[2] - n;
             }
-            LOG("                                      [\n");
+            printf("                                      [\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2*n) {
-                    LOG("                                       ..., \n");
+                    printf("                                       ..., \n");
                     i1 = ne[1] - n;
                 }
-                LOG("                                       [");
+                printf("                                       [");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2*n) {
-                        LOG("..., ");
+                        printf("..., ");
                         i0 = ne[0] - n;
                     }
                     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
@@ -62,18 +62,18 @@ 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);
                     }
-                    LOG("%12.4f", v);
+                    printf("%12.4f", v);
                     sum += v;
-                    if (i0 < ne[0] - 1) LOG(", ");
+                    if (i0 < ne[0] - 1) printf(", ");
                 }
-                LOG("],\n");
+                printf("],\n");
             }
-            LOG("                                      ],\n");
+            printf("                                      ],\n");
         }
-        LOG("                                     ]\n");
-        LOG("                                     sum = %f\n", sum);
+        printf("                                     ]\n");
+        printf("                                     sum = %f\n", sum);
     }
 }
 
@@ -99,14 +99,14 @@ 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());
     }
 
-    LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
-         t->name, ggml_type_name(t->type), ggml_op_desc(t),
-         src0->name, ggml_ne_string(src0).c_str(),
-         src1 ? src1_str : "",
-         ggml_ne_string(t).c_str());
+    printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
+           t->name, ggml_type_name(t->type), ggml_op_desc(t),
+           src0->name, ggml_ne_string(src0).c_str(),
+           src1 ? src1_str : "",
+           ggml_ne_string(t).c_str());
 
 
     // copy the data from the GPU memory if needed
@@ -127,12 +127,12 @@ 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);
 
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
-        LOG_ERR("%s : failed to eval\n", __func__);
+        fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }
 
@@ -144,11 +144,14 @@ 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;
     }
 
-    gpt_init();
+    print_build_info();
+
+    std::mt19937 rng(params.seed);
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -160,20 +163,18 @@ 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) {
-        LOG_ERR("%s : failed to init\n", __func__);
+        fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
     }
 
     // print system information
     {
-        LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
-        LOG_INF("\n");
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     bool OK = run(ctx, params);
@@ -181,8 +182,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    LOG("\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 `llama-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.prod(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 "llama-cli --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

@@ -1,83 +0,0 @@
-#include "arg.h"
-#include "common.h"
-
-#include <fstream>
-#include <string>
-
-// Export usage message (-h) to markdown format
-
-static void write_table_header(std::ofstream & file) {
-    file << "| Argument | Explanation |\n";
-    file << "| -------- | ----------- |\n";
-}
-
-static void write_table_entry(std::ofstream & file, const llama_arg & opt) {
-    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";
-}
-
-static void write_table(std::ofstream & file, std::vector<llama_arg *> & opts) {
-    write_table_header(file);
-    for (const auto & opt : opts) {
-        write_table_entry(file, *opt);
-    }
-}
-
-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);
-
-    std::vector<llama_arg *> common_options;
-    std::vector<llama_arg *> sparam_options;
-    std::vector<llama_arg *> specific_options;
-    for (auto & opt : ctx_arg.options) {
-        // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
-        if (opt.is_sparam) {
-            sparam_options.push_back(&opt);
-        } else if (opt.in_example(ctx_arg.ex)) {
-            specific_options.push_back(&opt);
-        } else {
-            common_options.push_back(&opt);
-        }
-    }
-
-    file << "**Common params**\n\n";
-    write_table(file, common_options);
-    file << "\n\n**Sampling params**\n\n";
-    write_table(file, sparam_options);
-    file << "\n\n**Example-specific params**\n\n";
-    write_table(file, specific_options);
-}
-
-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/README.md

@@ -201,6 +201,6 @@ Verification results for test.gguf.manifest - Success
 
 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/mofosyne/xxHash (From: https://github.com/Cyan4973/xxHash)
 - https://github.com/clibs/sha1/
 - https://github.com/jb55/sha256.c

examples/gguf-hash/deps/xxhash/clib.json

@@ -1,7 +1,7 @@
 {
   "name": "xxhash",
   "version": "0.8.2",
-  "repo": "Cyan4973/xxhash",
+  "repo": "mofosyne/xxhash",
   "description": "Extremely fast non-cryptographic hash algorithm",
   "keywords": ["xxhash", "hashing"],
   "license": "BSD-2-Clause",

examples/gguf-hash/gguf-hash.cpp

@@ -347,7 +347,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
                 char hex_result[17];
                 for (int  offset = 0; offset < 8; offset++) {
                     unsigned int shift_bits_by = (8 * (8 - offset - 1));
-                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
+                    sprintf( ( hex_result + (2*offset)), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
                 }
 
                 if (hash_params.manifest_is_usable) {
@@ -384,7 +384,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 
                 char hex_result[41] = {0};
                 for (int  offset = 0; offset < 20; offset++) {
-                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+                    sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff);
                 }
 
                 if (hash_params.manifest_is_usable) {
@@ -421,7 +421,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 
                 char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0};
                 for (int  offset = 0; offset < SHA256_DIGEST_SIZE; offset++) {
-                    snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+                    sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff);
                 }
 
                 if (hash_params.manifest_is_usable) {
@@ -460,7 +460,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
         char hex_result[17];
         for (int  offset = 0; offset < 8; offset++) {
             unsigned int shift_bits_by = (8 * (8 - offset - 1));
-            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
+            sprintf( ( hex_result + (2*offset)), "%02x", (unsigned char) (hash >> shift_bits_by)&0xff);
         }
 
         if (hash_params.manifest_is_usable) {
@@ -490,7 +490,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 
         char hex_result[41];
         for (int  offset = 0; offset < 20; offset++) {
-            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+            sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff);
         }
 
         if (hash_params.manifest_is_usable) {
@@ -520,7 +520,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
 
         char hex_result[SHA256_DIGEST_SIZE * 2 + 1] = {0};
         for (int  offset = 0; offset < SHA256_DIGEST_SIZE; offset++) {
-            snprintf( ( hex_result + (2*offset)), sizeof(hex_result) - (2*offset), "%02x", result[offset]&0xff);
+            sprintf( ( hex_result + (2*offset)), "%02x", result[offset]&0xff);
         }
 
         if (hash_params.manifest_is_usable) {
@@ -552,7 +552,7 @@ static hash_exit_code_t gguf_hash(const hash_params & hash_params) {
         generate_uuidv5(result, uuid);
 
         char string_buffer[37] = {0};
-        snprintf(string_buffer, sizeof(string_buffer), "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
+        sprintf(string_buffer, "%02x%02x%02x%02x-%02x%02x-%02x%02x-%02x%02x-%02x%02x%02x%02x%02x%02x",
             uuid[0], uuid[1], uuid[2], uuid[3],
             uuid[4], uuid[5], uuid[6], uuid[7],
             uuid[8], uuid[9], uuid[10], uuid[11],

examples/gguf-split/gguf-split.cpp

@@ -152,7 +152,7 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
         throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
 
-    if (argc - arg_idx != 2) {
+    if (argc - arg_idx < 2) {
         throw std::invalid_argument("error: bad arguments");
     }
 
@@ -389,17 +389,10 @@ static void gguf_merge(const split_params & split_params) {
     int n_split = 1;
     int total_tensors = 0;
 
-    // avoid overwriting existing output file
-    if (std::ifstream(split_params.output.c_str())) {
-        fprintf(stderr, "%s: output file %s already exists\n", __func__, split_params.output.c_str());
-        exit(EXIT_FAILURE);
-    }
-
+    auto * ctx_out = gguf_init_empty();
     std::ofstream fout(split_params.output.c_str(), std::ios::binary);
     fout.exceptions(std::ofstream::failbit); // fail fast on write errors
 
-    auto * ctx_out = gguf_init_empty();
-
     std::vector<uint8_t> read_data;
     std::vector<ggml_context *> ctx_metas;
     std::vector<gguf_context *> ctx_ggufs;

examples/gguf/gguf.cpp

@@ -92,11 +92,6 @@ static bool gguf_ex_read_0(const std::string & fname) {
 
     struct gguf_context * ctx = gguf_init_from_file(fname.c_str(), params);
 
-    if (!ctx) {
-        fprintf(stderr, "%s: failed to load '%s'\n", __func__, fname.c_str());
-        return false;
-    }
-
     printf("%s: version:      %d\n", __func__, gguf_get_version(ctx));
     printf("%s: alignment:   %zu\n", __func__, gguf_get_alignment(ctx));
     printf("%s: data offset: %zu\n", __func__, gguf_get_data_offset(ctx));

examples/gritlm/gritlm.cpp

@@ -1,4 +1,3 @@
-#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -10,7 +9,7 @@
 static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
     std::vector<std::vector<float>> result;
 
-    const llama_model * model = llama_get_model(ctx);
+    const llama_model * mdl = llama_get_model(ctx);
 
     llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
@@ -19,16 +18,16 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         const std::string input_string = instruction + sentences[i];
 
-        std::vector<llama_token> inputs = llama_tokenize(model, input_string, true, false);
+        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
 
         const int32_t n_toks = inputs.size();
 
         // GritLM seems to have EOS = ""
         // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_token_eos(model));
+        // inputs.push_back(llama_token_eos(mdl));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(model, instruction, true, false).size();
+        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -52,7 +51,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         llama_decode(ctx, batch);
 
         // get embedding dimensions
-        uint64_t n_embd = llama_n_embd(model);
+        uint64_t n_embd = llama_n_embd(mdl);
 
         // allocate embedding output
         std::vector<float> emb_unorm(n_embd, 0.0f);
@@ -93,11 +92,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     return result;
 }
 
-static std::string generate(llama_context * ctx, llama_sampler * smpl, const std::string & prompt, bool stream) {
+static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
     std::string result;
 
-    const llama_model * model = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(model);
+    const llama_model * mdl = llama_get_model(ctx);
+    llama_token eos_token = llama_token_eos(mdl);
 
     llama_kv_cache_clear(ctx);
     llama_set_embeddings(ctx, false);
@@ -105,24 +104,28 @@ static std::string generate(llama_context * ctx, llama_sampler * smpl, const std
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
-    std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true);
+    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
         llama_batch_clear(bat);
-        {
-            const int32_t n_inputs = inputs.size();
-
-            for (int32_t i = 0; i < n_inputs; i++) {
-                llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
-            }
+        auto n_inputs = (int32_t)inputs.size();
+        for (int32_t i = 0; i < n_inputs; i++) {
+            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
         }
         inputs.clear();
 
         llama_decode(ctx, bat);
+        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
 
-        llama_token token = llama_sampler_sample(smpl, ctx, bat.n_tokens - 1);
+        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
+        auto n_candidates = (int32_t)candidates.size();
+        for (int32_t token = 0; token < n_candidates; token++) {
+            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
+        }
+        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
 
+        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
         if (token == eos_token) {
             break;
         }
@@ -154,29 +157,20 @@ static std::string gritlm_instruction(const std::string & instruction) {
 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;
     }
 
-    gpt_init();
-
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
     llama_context_params cparams = llama_context_params_from_gpt_params(params);
 
     llama_backend_init();
 
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
+    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
 
     // create generation context
-    llama_context * ctx = llama_new_context_with_model(model, cparams);
-
-    auto sparams = llama_sampler_chain_default_params();
-
-    sparams.no_perf = false;
-
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
-
-    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
 
     // ### Embedding/Representation ###
     // samples taken from: https://github.com/ContextualAI/gritlm#basic
@@ -197,7 +191,7 @@ int main(int argc, char * argv[]) {
         const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
         const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
 
-        const int n_embd = llama_n_embd(model);
+        const int n_embd = llama_n_embd(mdl);
 
         const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
         const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
@@ -214,12 +208,11 @@ int main(int argc, char * argv[]) {
     // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
     {
         const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, smpl, prompt, true);
+        std::string response = generate(ctx, prompt, true);
     }
 
-    llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_free_model(model);
+    llama_free_model(mdl);
     llama_backend_free();
 
     return 0;

examples/imatrix/README.md

@@ -1,6 +1,6 @@
 # llama.cpp/examples/imatrix
 
-Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantized models.
+Compute an importance matrix for a model and given text dataset. Can be used during quantization to enchance the quality of the quantum models.
 More information is available here: https://github.com/ggerganov/llama.cpp/pull/4861
 
 ## Usage

examples/imatrix/imatrix.cpp

@@ -1,6 +1,4 @@
-#include "arg.h"
 #include "common.h"
-#include "log.h"
 #include "llama.h"
 
 #include <cmath>
@@ -19,13 +17,15 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static void print_usage(int, char ** argv) {
-    LOG("\nexample usage:\n");
-    LOG("\n    %s \\\n"
-            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n"
+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 \\\n"
+            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] [--verbosity 1] \\\n"
             "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
             "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
-    LOG("\n");
+    LOG_TEE("\n");
 }
 
 struct Stats {
@@ -126,10 +126,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             e.counts.resize(src1->ne[0]*n_as, 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
-            LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
-            exit(1); //GGML_ABORT("fatal error");
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
+            exit(1); //GGML_ASSERT(false);
+        }
+        if (m_params.verbosity > 1) {
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
         }
-        LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
         // loop over all possible experts, regardless if they are used or not in the batch
         for (int ex = 0; ex < n_as; ++ex) {
             size_t e_start = ex*src1->ne[0];
@@ -150,8 +152,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                         e.values[e_start + j] += x[j]*x[j];
                         e.counts[e_start + j]++;
                         if (!std::isfinite(e.values[e_start + j])) {
-                            LOG("\n");
-                            LOG_ERR("%f detected in %s\n", e.values[e_start + j], wname.c_str());
+                            fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
                             exit(1);
                         }
                     }
@@ -174,18 +175,20 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             e.counts.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
-            LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
-            exit(1); //GGML_ABORT("fatal error");
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
+            exit(1); //GGML_ASSERT(false);
         }
         ++e.ncall;
-        LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+        if (m_params.verbosity > 1) {
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
+        }
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];
             for (int j = 0; j < (int)src1->ne[0]; ++j) {
                 e.values[j] += x[j]*x[j];
                 e.counts[j]++;
                 if (!std::isfinite(e.values[j])) {
-                    LOG_ERR("%f detected in %s\n", e.values[j], wname.c_str());
+                    fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
                     exit(1);
                 }
             }
@@ -237,17 +240,17 @@ void IMatrixCollector::save_imatrix(int ncall) const {
         }
 
         if (n_zeros != 0 && is_first) {
-            LOG_INF("\n");
+            fprintf(stderr, "\n");
             is_first = false;
         }
 
         if (n_zeros == n_all) {
-            LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
+            fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
             continue;
         }
 
         if (n_zeros > 0) {
-            LOG_WRN("%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
+            fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
             continue;
         }
 
@@ -256,7 +259,7 @@ void IMatrixCollector::save_imatrix(int ncall) const {
     }
 
     if (to_store.size() < m_stats.size()) {
-        LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
+        fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
     }
 
     std::ofstream out(fname, std::ios::binary);
@@ -288,20 +291,21 @@ void IMatrixCollector::save_imatrix(int ncall) const {
         out.write(m_params.prompt_file.c_str(), len);
     }
 
-    LOGV(1, "\n");
-    LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
+    if (m_params.verbosity > 0) {
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
+    }
 }
 
 bool IMatrixCollector::load_imatrix(const char * fname) {
     std::ifstream in(fname, std::ios::binary);
     if (!in) {
-        LOG_ERR("%s: failed to open %s\n",__func__, fname);
+        printf("%s: failed to open %s\n",__func__, fname);
         return false;
     }
     int n_entries;
     in.read((char*)&n_entries, sizeof(n_entries));
     if (in.fail() || n_entries < 1) {
-        LOG_ERR("%s: no data in file %s\n", __func__, fname);
+        printf("%s: no data in file %s\n", __func__, fname);
         return false;
     }
     for (int i = 0; i < n_entries; ++i) {
@@ -309,7 +313,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         std::vector<char> name_as_vec(len+1);
         in.read((char *)name_as_vec.data(), len);
         if (in.fail()) {
-            LOG_ERR("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
+            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
             return false;
         }
         name_as_vec[len] = 0;
@@ -320,7 +324,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         int nval;
         in.read((char *)&nval, sizeof(nval));
         if (in.fail() || nval < 1) {
-            LOG_ERR("%s: failed reading number of values for entry %d\n",__func__,i);
+            printf("%s: failed reading number of values for entry %d\n",__func__,i);
             m_stats = {};
             return false;
         }
@@ -333,7 +337,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         std::vector<float> tmp(nval);
         in.read((char*)tmp.data(), nval*sizeof(float));
         if (in.fail()) {
-            LOG_ERR("%s: failed reading data for entry %d\n",__func__,i);
+            printf("%s: failed reading data for entry %d\n",__func__,i);
             m_stats = {};
             return false;
         }
@@ -429,30 +433,31 @@ static void process_logits(
 }
 
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
-    LOG_INF("%s: tokenizing the input ..\n", __func__);
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
     std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
 
     auto tim2 = std::chrono::high_resolution_clock::now();
-    LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
     if (params.i_chunk > 0) {
         if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
-            LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
+            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
             return false;
         }
-        LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
+        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
         tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
     }
 
     if (int(tokens.size()) < 2*n_ctx) {
-        LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);
-        LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
         return false;
     }
 
@@ -474,7 +479,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
     double nll = 0.0;
     double nll2 = 0.0;
 
-    LOG_INF("%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -510,7 +515,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 
             // TODO: use batch.logits to save computations instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
-                LOG_ERR("%s : failed to eval\n", __func__);
+                fprintf(stderr, "%s : failed to eval\n", __func__);
                 return false;
             }
 
@@ -527,29 +532,29 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 
         if (i == 0) {
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
-            LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
             int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
-                LOG("%d hours ", total_seconds / (60*60));
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
-            LOG("%.2f minutes\n", total_seconds / 60.0);
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
         }
 
         if (params.compute_ppl) {
             const int first = n_ctx/2;
-            const auto * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+            const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
             process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
                     workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
             count += n_ctx - first - 1;
 
-            LOG("[%d]%.4lf,", i + 1, std::exp(nll / count));
+            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
             fflush(stdout);
 
             logits.clear();
         }
     }
-    LOG("\n");
+    printf("\n");
 
     if (params.compute_ppl) {
         nll2 /= count;
@@ -558,9 +563,9 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
         nll2 -= nll * nll;
         if (nll2 > 0) {
             nll2 = sqrt(nll2/(count-1));
-            LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+            printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
         } else {
-            LOG("Unexpected negative standard deviation of log(prob)\n");
+            printf("Unexpected negative standard deviation of log(prob)\n");
         }
     }
 
@@ -572,28 +577,27 @@ int main(int argc, char ** argv) {
 
     params.n_ctx = 512;
     params.logits_all = true;
-    params.escape = false;
+    params.verbosity = 1;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        print_usage(argc, argv, params);
         return 1;
     }
 
-    gpt_init();
-
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     g_collector.set_params(params);
 
     for (const auto & in_file : params.in_files) {
-        LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
+        printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
         if (!g_collector.load_imatrix(in_file.c_str())) {
-            LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());
+            fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
             return 1;
         }
     }
 
     if (params.in_files.size() > 1) {
-        LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
+        printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
         g_collector.save_imatrix();
     }
 
@@ -607,25 +611,25 @@ 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_context * ctx;
 
-    llama_model * model = llama_init.model;
-    llama_context * ctx = llama_init.context;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
     if (model == nullptr || ctx == nullptr) {
-        LOG_ERR("%s : failed to init\n", __func__);
+        fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
     if (params.n_ctx > n_ctx_train) {
-        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, params.n_ctx);
     }
 
     // print system information
     {
-        LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
     }
 
     if (!compute_imatrix(ctx, params)) {
@@ -634,8 +638,7 @@ int main(int argc, char ** argv) {
 
     g_collector.save_imatrix();
 
-    LOG("\n");
-    llama_perf_context_print(ctx);
+    llama_print_timings(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/infill/infill.cpp

@@ -1,9 +1,8 @@
-#include "arg.h"
 #include "common.h"
+
 #include "console.h"
-#include "sampling.h"
-#include "log.h"
 #include "llama.h"
+#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -35,7 +34,6 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
-static gpt_sampler             ** g_smpl;
 static gpt_params               * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
@@ -56,7 +54,7 @@ static void write_logfile(
 
     const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
-        LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
+        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
         return;
     }
@@ -65,7 +63,7 @@ static void write_logfile(
     FILE * logfile = fopen(logfile_path.c_str(), "w");
 
     if (logfile == NULL) {
-        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
         return;
     }
 
@@ -83,7 +81,7 @@ static void write_logfile(
     yaml_dump_string_multiline(logfile, "output", output.c_str());
     yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
-    llama_perf_dump_yaml(logfile, ctx);
+    llama_dump_timing_info_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -94,14 +92,9 @@ static void sigint_handler(int signo) {
             is_interacting = true;
         } else {
             console::cleanup();
-            LOG("\n");
-            gpt_perf_print(*g_ctx, *g_smpl);
+            printf("\n");
+            llama_print_timings(*g_ctx);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
-
-            // make sure all logs are flushed
-            LOG("Interrupted by user\n");
-            gpt_log_pause(gpt_log_main());
-
             _exit(130);
         }
     }
@@ -110,95 +103,106 @@ static void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
+    llama_sampling_params & sparams = params.sparams;
     g_params = &params;
 
-    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
+    if (!gpt_params_parse(argc, argv, params)) {
+        gpt_params_print_usage(argc, argv, params);
         return 1;
     }
 
-    gpt_init();
-
-    auto & sparams = params.sparams;
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("infill", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+#endif // LOG_DISABLE_LOGS
 
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
 
     if (params.logits_all) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
-        LOG_ERR("************\n\n");
+        printf("\n************\n");
+        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        printf("************\n\n");
 
         return 0;
     }
 
     if (params.embedding) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
-        LOG_ERR("************\n\n");
+        printf("\n************\n");
+        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        printf("************\n\n");
 
         return 0;
     }
 
     if (params.n_ctx != 0 && params.n_ctx < 8) {
-        LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
+        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
-
     if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
-        LOG_ERR("\n************\n");
-        LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
-        LOG_ERR("************\n\n");
+        printf("\n************\n");
+        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        printf("************\n\n");
 
         return 0;
     }
 
     if (params.rope_freq_base != 0.0) {
-        LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
     }
 
     if (params.rope_freq_scale != 0.0) {
-        LOG_WRN("%s: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_INF("%s: llama backend init\n", __func__);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+
+    LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model = nullptr;
-    llama_context * ctx = nullptr;
-    gpt_sampler  * smpl = nullptr;
+    llama_model * model;
+    llama_context * ctx;
 
     g_model = &model;
     g_ctx = &ctx;
-    g_smpl = &smpl;
 
     // load the model and apply lora adapter, if any
-    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
-
-    model = llama_init.model;
-    ctx = llama_init.context;
+    LOG("%s: load the model and apply lora adapter, if any\n", __func__);
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     if (model == NULL) {
-        LOG_ERR("%s: unable to load model\n", __func__);
+        LOG_TEE("%s: error: unable to load model\n", __func__);
         return 1;
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
-    LOG_DBG("n_ctx: %d\n", n_ctx);
+    LOG("n_ctx: %d\n", n_ctx);
 
     if (n_ctx > n_ctx_train) {
-        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
+        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
-        LOG_INF("\n");
-        LOG_INF("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_TEE("\n");
+        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_add_bos_token(model);
-    GGML_ASSERT(!llama_add_eos_token(model));
+    const bool add_bos = llama_should_add_bos_token(model);
+    GGML_ASSERT(llama_add_eos_token(model) != 1);
+    LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;
     std::vector<llama_token> embd_end;
@@ -223,19 +227,18 @@ int main(int argc, char ** argv) {
         embd_inp.push_back(middle_token);
     }
 
-    LOG_DBG("add_bos: %d\n", add_bos);
-    LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
-    LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
-    LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
+    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
+    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
+    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(model));
-        LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
+        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
     }
 
     if ((int) embd_inp.size() > n_ctx - 4) {
-        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
+        LOG_TEE("%s: error: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
     }
 
@@ -244,8 +247,9 @@ int main(int argc, char ** argv) {
         params.n_keep = (int)embd_inp.size();
     }
 
-    LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
-    LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
+    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
+    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
+
 
     // enable interactive mode if interactive start is specified
     if (params.interactive_first) {
@@ -253,21 +257,21 @@ int main(int argc, char ** argv) {
     }
 
     if (params.verbose_prompt) {
-        LOG_INF("\n");
-        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        LOG_TEE("\n");
+        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            LOG_INF("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (params.n_keep > 0) {
-        LOG_INF("%s: static prompt based on n_keep: '", __func__);
+        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                LOG_CNT("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            LOG_CNT("'\n");
+            LOG_TEE("'\n");
         }
-        LOG_INF("\n");
+        LOG_TEE("\n");
     }
 
     if (params.interactive) {
@@ -284,30 +288,30 @@ int main(int argc, char ** argv) {
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-        LOG_INF("%s: interactive mode on.\n", __func__);
+        LOG_TEE("%s: interactive mode on.\n", __func__);
 
         if (params.input_prefix_bos) {
-            LOG_INF("Input prefix with BOS\n");
+            LOG_TEE("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
         }
 
         if (!params.input_suffix.empty()) {
-            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    smpl = gpt_sampler_init(model, sparams);
+    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+    LOG_TEE("\n\n");
 
-    LOG_INF("sampler seed: %u\n",     gpt_sampler_get_seed(smpl));
-    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
-    LOG_INF("sampler chain: %s\n",    gpt_sampler_print(smpl).c_str());
-
-    LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-
-    LOG_INF("\n");
-    LOG_INF("\n#####  Infill mode  #####\n\n");
+    LOG_TEE("\n#####  Infill mode  #####\n\n");
+    if (params.infill) {
+        printf("\n************\n");
+        printf("no need to specify '--infill', always running infill\n");
+        printf("************\n\n");
+    }
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
@@ -318,11 +322,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        LOG_INF("== Running in interactive mode. ==\n");
+        LOG_TEE("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
+        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-        LOG_INF(       "%s\n", control_message);
+        LOG_TEE(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -342,6 +346,8 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
+    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
+
     while (n_remain != 0 || params.interactive) {
         // predict
         if (!embd.empty()) {
@@ -355,8 +361,9 @@ int main(int argc, char ** argv) {
                 embd.resize(max_embd_size);
 
                 console::set_display(console::error);
-                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console::set_display(console::reset);
+                fflush(stdout);
             }
 
             // infinite text generation via context swapping
@@ -365,14 +372,14 @@ int main(int argc, char ** argv) {
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() > n_ctx) {
                 if (params.n_predict == -2) {
-                    LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
                 }
 
                 const int n_left    = n_past - params.n_keep - 1;
                 const int n_discard = n_left/2;
 
-                LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
                 llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
@@ -380,9 +387,9 @@ int main(int argc, char ** argv) {
 
                 n_past -= n_discard;
 
-                LOG_DBG("after swap: n_past = %d\n", n_past);
+                LOG("after swap: n_past = %d\n", n_past);
 
-                LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
             }
 
@@ -394,16 +401,16 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
+                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
-                    LOG_ERR("%s : failed to eval\n", __func__);
+                    LOG_TEE("%s : failed to eval\n", __func__);
                     return 1;
                 }
 
                 n_past += n_eval;
 
-                LOG_DBG("n_past = %d\n", n_past);
+                LOG("n_past = %d\n", n_past);
             }
 
         }
@@ -411,11 +418,11 @@ int main(int argc, char ** argv) {
         embd.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
+            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, nullptr);
 
-            gpt_sampler_accept(smpl, id, true);
+            llama_sampling_accept(ctx_sampling, ctx, id, true);
 
-            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
+            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
 
             embd.push_back(id);
 
@@ -425,16 +432,16 @@ int main(int argc, char ** argv) {
             // decrement remaining sampling budget
             --n_remain;
 
-            LOG_DBG("n_remain: %d\n", n_remain);
+            LOG("n_remain: %d\n", n_remain);
         } else {
             // some user input remains from prompt or interaction, forward it to processing
-            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                gpt_sampler_accept(smpl, embd_inp[n_consumed], false);
+                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -447,7 +454,7 @@ int main(int argc, char ** argv) {
         if (input_echo) {
             for (auto id : embd) {
                 const std::string token_str = llama_token_to_piece(ctx, id);
-                LOG("%s", token_str.c_str());
+                printf("%s", token_str.c_str());
 
                 if (embd.size() > 1) {
                     input_tokens.push_back(id);
@@ -456,6 +463,7 @@ int main(int argc, char ** argv) {
                     output_ss << token_str;
                 }
             }
+            fflush(stdout);
         }
         // reset color to default if we there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
@@ -465,12 +473,13 @@ int main(int argc, char ** argv) {
         // if not currently processing queued inputs;
         if ((int) embd_inp.size() <= n_consumed) {
             // deal with eot token in infill mode
-            if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
+            if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
                     // print an eot token
-                    LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+                    printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
-                LOG("\n");
+                fflush(stdout);
+                printf("\n");
                 console::set_display(console::user_input);
                 std::string buffer;
                 std::string line;
@@ -526,33 +535,35 @@ int main(int argc, char ** argv) {
                 n_remain = params.n_predict;
                 n_past = 0;
                 n_consumed = 0;
+                // LOG_TEE("took new input\n");
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
-                LOG_DBG("found EOS token\n");
+            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+                LOG("found EOS token\n");
 
                 if (params.interactive) {
 
                     is_interacting = true;
-                    LOG("\n");
+                    printf("\n");
                     console::set_display(console::user_input);
+                    fflush(stdout);
                }
             }
 
             if (n_past > 0 && is_interacting && !params.interactive) {
-                LOG_DBG("waiting for user input\n");
+                LOG("waiting for user input\n");
 
                 if (params.input_prefix_bos) {
-                    LOG_DBG("adding input prefix BOS token\n");
+                    LOG("adding input prefix BOS token\n");
                     embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
-                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     buffer += params.input_prefix;
-                    LOG("%s", buffer.c_str());
+                    printf("%s", buffer.c_str());
                 }
 
                 std::string line;
@@ -570,17 +581,17 @@ int main(int argc, char ** argv) {
                 if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
-                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         buffer += params.input_suffix;
-                        LOG("%s", params.input_suffix.c_str());
+                        printf("%s", params.input_suffix.c_str());
                     }
 
-                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
+                    LOG("buffer: '%s'\n", buffer.c_str());
 
                     const size_t original_size = embd_inp.size();
 
                     const auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
+                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
@@ -591,9 +602,9 @@ int main(int argc, char ** argv) {
                     }
 
                     n_remain -= line_inp.size();
-                    LOG_DBG("n_remain: %d\n", n_remain);
+                    LOG("n_remain: %d\n", n_remain);
                 } else {
-                    LOG_DBG("empty line, passing control back\n");
+                    LOG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again
@@ -601,7 +612,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    gpt_sampler_reset(smpl);
+                    llama_sampling_reset(ctx_sampling);
                 }
                 is_interacting = false;
             }
@@ -620,18 +631,22 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        LOG("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
+        fflush(stdout);
     }
 
-    LOG("\n");
-    gpt_perf_print(ctx, smpl);
+    llama_print_timings(ctx);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);
 
-    gpt_sampler_free(smpl);
+    llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
+#ifndef LOG_DISABLE_LOGS
+    LOG_TEE("Log end\n");
+#endif // LOG_DISABLE_LOGS
+
     return 0;
 }

examples/llama-bench/README.md

@@ -14,8 +14,7 @@ Performance testing tool for llama.cpp.
     1. [Markdown](#markdown)
     2. [CSV](#csv)
     3. [JSON](#json)
-    4. [JSONL](#jsonl)
-    5. [SQL](#sql)
+    4. [SQL](#sql)
 
 ## Syntax
 
@@ -24,34 +23,27 @@ usage: ./llama-bench [options]
 
 options:
   -h, --help
-  -m, --model <filename>                    (default: models/7B/ggml-model-q4_0.gguf)
-  -p, --n-prompt <n>                        (default: 512)
-  -n, --n-gen <n>                           (default: 128)
-  -pg <pp,tg>                               (default: )
-  -b, --batch-size <n>                      (default: 2048)
-  -ub, --ubatch-size <n>                    (default: 512)
-  -ctk, --cache-type-k <t>                  (default: f16)
-  -ctv, --cache-type-v <t>                  (default: f16)
-  -t, --threads <n>                         (default: 8)
-  -C, --cpu-mask <hex,hex>                  (default: 0x0)
-  --cpu-strict <0|1>                        (default: 0)
-  --poll <0...100>                          (default: 50)
-  -ngl, --n-gpu-layers <n>                  (default: 99)
-  -rpc, --rpc <rpc_servers>                 (default: )
-  -sm, --split-mode <none|layer|row>        (default: layer)
-  -mg, --main-gpu <i>                       (default: 0)
-  -nkvo, --no-kv-offload <0|1>              (default: 0)
-  -fa, --flash-attn <0|1>                   (default: 0)
-  -mmp, --mmap <0|1>                        (default: 1)
-  --numa <distribute|isolate|numactl>       (default: disabled)
-  -embd, --embeddings <0|1>                 (default: 0)
-  -ts, --tensor-split <ts0/ts1/..>          (default: 0)
-  -r, --repetitions <n>                     (default: 5)
-  --prio <0|1|2|3>                          (default: 0)
-  --delay <0...N> (seconds)                 (default: 0)
-  -o, --output <csv|json|jsonl|md|sql>      (default: md)
-  -oe, --output-err <csv|json|jsonl|md|sql> (default: none)
-  -v, --verbose                             (default: 0)
+  -m, --model <filename>              (default: models/7B/ggml-model-q4_0.gguf)
+  -p, --n-prompt <n>                  (default: 512)
+  -n, --n-gen <n>                     (default: 128)
+  -pg <pp,tg>                         (default: 512,128)
+  -b, --batch-size <n>                (default: 2048)
+  -ub, --ubatch-size <n>              (default: 512)
+  -ctk, --cache-type-k <t>            (default: f16)
+  -ctv, --cache-type-v <t>            (default: f16)
+  -t, --threads <n>                   (default: 16)
+  -ngl, --n-gpu-layers <n>            (default: 99)
+  -sm, --split-mode <none|layer|row>  (default: layer)
+  -mg, --main-gpu <i>                 (default: 0)
+  -nkvo, --no-kv-offload <0|1>        (default: 0)
+  -fa, --flash-attn <0|1>             (default: 0)
+  -mmp, --mmap <0|1>                  (default: 1)
+  --numa <distribute|isolate|numactl> (default: disabled)
+  -embd, --embeddings <0|1>           (default: 0)
+  -ts, --tensor-split <ts0/ts1/..>    (default: 0)
+  -r, --repetitions <n>               (default: 5)
+  -o, --output <csv|json|md|sql>      (default: md)
+  -v, --verbose                       (default: 0)
 
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```
@@ -246,19 +238,6 @@ $ ./llama-bench -o json
 ]
 ```
 
-
-### JSONL
-
-```sh
-$ ./llama-bench -o jsonl
-```
-
-```json lines
-{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":512,"n_gen":0,"test_time":"2023-09-23T12:09:57Z","avg_ns":212365953,"stddev_ns":985423,"avg_ts":2410.974041,"stddev_ts":11.163766,"samples_ns":[213837238,211635853,212328053,211329715,212698907],"samples_ts":[2394.34,2419.25,2411.36,2422.75,2407.16]}
-{"build_commit":"3469684","build_number":1275,"cuda":true,"metal":false,"gpu_blas":true,"blas":true,"cpu_info":"13th Gen Intel(R) Core(TM) i9-13900K","gpu_info":"NVIDIA GeForce RTX 3090 Ti","model_filename":"models/7B/ggml-model-q4_0.gguf","model_type":"llama 7B mostly Q4_0","model_size":3825065984,"model_n_params":6738415616,"n_batch":512,"n_threads":16,"f16_kv":true,"n_gpu_layers":99,"main_gpu":0,"mul_mat_q":true,"tensor_split":"0.00","n_prompt":0,"n_gen":128,"test_time":"2023-09-23T12:09:59Z","avg_ns":977425219,"stddev_ns":9268593,"avg_ts":130.965708,"stddev_ts":1.238924,"samples_ns":[984472709,974901233,989474741,970729355,967548060],"samples_ts":[130.019,131.295,129.362,131.86,132.293]}
-```
-
-
 ### SQL
 
 SQL output is suitable for importing into a SQLite database. The output can be piped into the `sqlite3` command line tool to add the results to a database.

examples/llama-bench/llama-bench.cpp

@@ -16,7 +16,6 @@
 #include <sstream>
 #include <string>
 #include <vector>
-#include <thread>
 
 #include "ggml.h"
 #include "llama.h"
@@ -24,18 +23,6 @@
 #include "ggml-cuda.h"
 #include "ggml-sycl.h"
 
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
-
-#ifdef _WIN32
-#define WIN32_LEAN_AND_MEAN
-#ifndef NOMINMAX
-#   define NOMINMAX
-#endif
-#include <windows.h>
-#endif
-
 // utils
 static uint64_t get_time_ns() {
     using clock = std::chrono::high_resolution_clock;
@@ -105,30 +92,6 @@ static std::string get_cpu_info() {
         }
         fclose(f);
     }
-#elif defined(_WIN32)
-    HKEY hKey;
-    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                     0,
-                     KEY_READ,
-                     &hKey) != ERROR_SUCCESS) {
-        // fail to open registry key
-        return "";
-    }
-    char cpu_brand[256];
-    DWORD cpu_brand_size = sizeof(cpu_brand);
-    if (RegQueryValueExA(hKey,
-                        TEXT("ProcessorNameString"),
-                        NULL,
-                        NULL,
-                        (LPBYTE)cpu_brand,
-                        &cpu_brand_size) == ERROR_SUCCESS) {
-        id.assign(cpu_brand, cpu_brand_size);
-        if (id.find('\0') != std::string::npos) {
-            id.resize(id.find('\0'));
-        }
-    }
-    RegCloseKey(hKey);
 #endif
     // TODO: other platforms
     return id;
@@ -157,34 +120,22 @@ static std::string get_gpu_info() {
             id += "/";
         }
     }
-#endif
-#ifdef GGML_USE_CANN
-    uint32_t count = ggml_backend_cann_get_device_count();
-    for (uint32_t i = 0; i < count; i++) {
-        char buf[128];
-        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
-        }
-    }
 #endif
     // TODO: other backends
     return id;
 }
 
 // command line params
-enum output_formats {NONE, CSV, JSON, JSONL, MARKDOWN, SQL};
+enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
 
 static const char * output_format_str(output_formats format) {
     switch (format) {
         case NONE:     return "none";
         case CSV:      return "csv";
         case JSON:     return "json";
-        case JSONL:    return "jsonl";
         case MARKDOWN: return "md";
         case SQL:      return "sql";
-        default: GGML_ABORT("invalid output format");
+        default: GGML_ASSERT(!"invalid output format");
     }
 }
 
@@ -195,8 +146,6 @@ static bool output_format_from_str(const std::string & s, output_formats & forma
         format = CSV;
     } else if (s == "json") {
         format = JSON;
-    } else if (s == "jsonl") {
-        format = JSONL;
     } else if (s == "md") {
         format = MARKDOWN;
     } else if (s == "sql") {
@@ -212,7 +161,7 @@ static const char * split_mode_str(llama_split_mode mode) {
         case LLAMA_SPLIT_MODE_NONE:  return "none";
         case LLAMA_SPLIT_MODE_LAYER: return "layer";
         case LLAMA_SPLIT_MODE_ROW:   return "row";
-        default: GGML_ABORT("invalid split mode");
+        default: GGML_ASSERT(!"invalid split mode");
     }
 }
 
@@ -232,9 +181,6 @@ struct cmd_params {
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
-    std::vector<std::string> cpu_mask;
-    std::vector<bool> cpu_strict;
-    std::vector<int> poll;
     std::vector<int> n_gpu_layers;
     std::vector<std::string> rpc_servers;
     std::vector<llama_split_mode> split_mode;
@@ -246,10 +192,7 @@ struct cmd_params {
     std::vector<bool> embeddings;
     ggml_numa_strategy numa;
     int reps;
-    ggml_sched_priority prio;
-    int delay;
     bool verbose;
-    bool progress;
     output_formats output_format;
     output_formats output_format_stderr;
 };
@@ -264,9 +207,6 @@ static const cmd_params cmd_params_defaults = {
     /* type_k               */ {GGML_TYPE_F16},
     /* type_v               */ {GGML_TYPE_F16},
     /* n_threads            */ {cpu_get_num_math()},
-    /* cpu_mask             */ {"0x0"},
-    /* cpu_strict           */ {false},
-    /* poll                 */ {50},
     /* n_gpu_layers         */ {99},
     /* rpc_servers          */ {""},
     /* split_mode           */ {LLAMA_SPLIT_MODE_LAYER},
@@ -278,10 +218,7 @@ static const cmd_params cmd_params_defaults = {
     /* embeddings           */ {false},
     /* numa                 */ GGML_NUMA_STRATEGY_DISABLED,
     /* reps                 */ 5,
-    /* prio                 */ GGML_SCHED_PRIO_NORMAL,
-    /* delay                */ 0,
     /* verbose              */ false,
-    /* progress             */ false,
     /* output_format        */ MARKDOWN,
     /* output_format_stderr */ NONE,
 };
@@ -291,37 +228,29 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help\n");
-    printf("  -m, --model <filename>                    (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
-    printf("  -p, --n-prompt <n>                        (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
-    printf("  -n, --n-gen <n>                           (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
-    printf("  -pg <pp,tg>                               (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
-    printf("  -b, --batch-size <n>                      (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ub, --ubatch-size <n>                    (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
-    printf("  -ctk, --cache-type-k <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
-    printf("  -ctv, --cache-type-v <t>                  (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
-    printf("  -t, --threads <n>                         (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
-    printf("  -C, --cpu-mask <hex,hex>                  (default: %s)\n", join(cmd_params_defaults.cpu_mask, ",").c_str());
-    printf("  --cpu-strict <0|1>                        (default: %s)\n", join(cmd_params_defaults.cpu_strict, ",").c_str());
-    printf("  --poll <0...100>                          (default: %s)\n", join(cmd_params_defaults.poll, ",").c_str());
-    printf("  -ngl, --n-gpu-layers <n>                  (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
-#ifdef GGML_USE_RPC
-    printf("  -rpc, --rpc <rpc_servers>                 (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
-#endif
-    printf("  -sm, --split-mode <none|layer|row>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
-    printf("  -mg, --main-gpu <i>                       (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
-    printf("  -nkvo, --no-kv-offload <0|1>              (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
-    printf("  -fa, --flash-attn <0|1>                   (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
-    printf("  -mmp, --mmap <0|1>                        (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  --numa <distribute|isolate|numactl>       (default: disabled)\n");
-    printf("  -embd, --embeddings <0|1>                 (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
-    printf("  -ts, --tensor-split <ts0/ts1/..>          (default: 0)\n");
-    printf("  -r, --repetitions <n>                     (default: %d)\n", cmd_params_defaults.reps);
-    printf("  --prio <0|1|2|3>                          (default: %d)\n", cmd_params_defaults.prio);
-    printf("  --delay <0...N> (seconds)                 (default: %d)\n", cmd_params_defaults.delay);
-    printf("  -o, --output <csv|json|jsonl|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
-    printf("  -oe, --output-err <csv|json|jsonl|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
-    printf("  -v, --verbose                             (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
-    printf("  --progress                                (default: %s)\n", cmd_params_defaults.progress ? "1" : "0");
+    printf("  -m, --model <filename>              (default: %s)\n", join(cmd_params_defaults.model, ",").c_str());
+    printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
+    printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
+    printf("  -pg <pp,tg>                         (default: %s)\n", join(transform_to_str(cmd_params_defaults.n_pg, pair_str), ",").c_str());
+    printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
+    printf("  -ub, --ubatch-size <n>              (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
+    printf("  -ctk, --cache-type-k <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
+    printf("  -ctv, --cache-type-v <t>            (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
+    printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
+    printf("  -ngl, --n-gpu-layers <n>            (default: %s)\n", join(cmd_params_defaults.n_gpu_layers, ",").c_str());
+    printf("  -rpc, --rpc <rpc_servers>           (default: %s)\n", join(cmd_params_defaults.rpc_servers, ",").c_str());
+    printf("  -sm, --split-mode <none|layer|row>  (default: %s)\n", join(transform_to_str(cmd_params_defaults.split_mode, split_mode_str), ",").c_str());
+    printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
+    printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
+    printf("  -fa, --flash-attn <0|1>             (default: %s)\n", join(cmd_params_defaults.flash_attn, ",").c_str());
+    printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
+    printf("  --numa <distribute|isolate|numactl> (default: disabled)\n");
+    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
+    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
+    printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
+    printf("  -oe, --output-err <csv|json|md|sql> (default: %s)\n", output_format_str(cmd_params_defaults.output_format_stderr));
+    printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
     printf("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 }
@@ -365,9 +294,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     params.output_format_stderr = cmd_params_defaults.output_format_stderr;
     params.reps = cmd_params_defaults.reps;
     params.numa = cmd_params_defaults.numa;
-    params.prio = cmd_params_defaults.prio;
-    params.delay = cmd_params_defaults.delay;
-    params.progress = cmd_params_defaults.progress;
 
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
@@ -439,9 +365,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
-            if (invalid_param) {
-                break;
-            }
             params.type_k.insert(params.type_k.end(), types.begin(), types.end());
         } else if (arg == "-ctv" || arg == "--cache-type-v") {
             if (++i >= argc) {
@@ -458,9 +381,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
-            if (invalid_param) {
-                break;
-            }
             params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
@@ -469,27 +389,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_threads.insert(params.n_threads.end(), p.begin(), p.end());
-        } else if (arg == "-C" || arg == "--cpu-mask") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<std::string>(argv[i], split_delim);
-            params.cpu_mask.insert(params.cpu_mask.end(), p.begin(), p.end());
-        } else if (arg == "--cpu-strict") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<bool>(argv[i], split_delim);
-            params.cpu_strict.insert(params.cpu_strict.end(), p.begin(), p.end());
-        } else if (arg == "--poll") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = string_split<int>(argv[i], split_delim);
-            params.poll.insert(params.poll.end(), p.begin(), p.end());
         } else if (arg == "-ngl" || arg == "--n-gpu-layers") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -497,14 +396,12 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = string_split<int>(argv[i], split_delim);
             params.n_gpu_layers.insert(params.n_gpu_layers.end(), p.begin(), p.end());
-#ifdef GGML_USE_RPC
         } else if (arg == "-rpc" || arg == "--rpc") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
             params.rpc_servers.push_back(argv[i]);
-#endif
         } else if (arg == "-sm" || arg == "--split-mode") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -526,9 +423,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 modes.push_back(mode);
             }
-            if (invalid_param) {
-                break;
-            }
             params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
@@ -603,18 +497,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 break;
             }
             params.reps = std::stoi(argv[i]);
-        } else if (arg == "--prio") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.prio = (enum ggml_sched_priority) std::stoi(argv[i]);
-        } else if (arg == "--delay") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.delay = std::stoi(argv[i]);
         } else if (arg == "-o" || arg == "--output") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -629,8 +511,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             invalid_param = !output_format_from_str(argv[i], params.output_format_stderr);
         } else if (arg == "-v" || arg == "--verbose") {
             params.verbose = true;
-        } else if (arg == "--progress") {
-            params.progress = true;
         } else {
             invalid_param = true;
             break;
@@ -661,9 +541,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
     if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
-    if (params.cpu_mask.empty())     { params.cpu_mask  = cmd_params_defaults.cpu_mask;  }
-    if (params.cpu_strict.empty())   { params.cpu_strict = cmd_params_defaults.cpu_strict; }
-    if (params.poll.empty())         { params.poll = cmd_params_defaults.poll; }
 
     return params;
 }
@@ -677,9 +554,6 @@ struct cmd_params_instance {
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
-    std::string cpu_mask;
-    bool cpu_strict;
-    int poll;
     int n_gpu_layers;
     std::string rpc_servers;
     llama_split_mode split_mode;
@@ -749,10 +623,7 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & tv : params.type_v)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & fa : params.flash_attn)
-    for (const auto & nt : params.n_threads)
-    for (const auto & cm : params.cpu_mask)
-    for (const auto & cs : params.cpu_strict)
-    for (const auto & pl : params.poll) {
+    for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
                 continue;
@@ -766,9 +637,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
-                /* .cpu_mask     = */ cm,
-                /* .cpu_strict   = */ cs,
-                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -795,9 +663,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
-                /* .cpu_mask     = */ cm,
-                /* .cpu_strict   = */ cs,
-                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -824,9 +689,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
-                /* .cpu_mask     = */ cm,
-                /* .cpu_strict   = */ cs,
-                /* .poll         = */ pl,
                 /* .n_gpu_layers = */ nl,
                 /* .rpc_servers  = */ rpc,
                 /* .split_mode   = */ sm,
@@ -863,9 +725,6 @@ struct test {
     int n_batch;
     int n_ubatch;
     int n_threads;
-    std::string cpu_mask;
-    bool cpu_strict;
-    int poll;
     bool has_rpc;
     ggml_type type_k;
     ggml_type type_v;
@@ -892,9 +751,6 @@ struct test {
         n_batch = inst.n_batch;
         n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
-        cpu_mask = inst.cpu_mask;
-        cpu_strict = inst.cpu_strict;
-        poll = inst.poll;
         has_rpc = !inst.rpc_servers.empty();
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -972,14 +828,13 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
-            "n_threads", "cpu_mask", "cpu_strict", "poll",
-            "type_k", "type_v",
+            "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload", "flash_attn",
             "tensor_split", "use_mmap", "embeddings",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
-            "avg_ts", "stddev_ts",
+            "avg_ts", "stddev_ts"
         };
         return fields;
     }
@@ -988,7 +843,7 @@ struct test {
 
     static field_type get_field_type(const std::string & field) {
         if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" || field == "poll" ||
+            field == "n_threads" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -997,7 +852,6 @@ struct test {
         }
         if (field == "cuda" || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "cpu_strict" ||
             field == "flash_attn" || field == "use_mmap" || field == "embeddings") {
             return BOOL;
         }
@@ -1030,8 +884,7 @@ struct test {
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
             std::to_string(n_batch), std::to_string(n_ubatch),
-            std::to_string(n_threads), cpu_mask, std::to_string(cpu_strict), std::to_string(poll),
-            ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload), std::to_string(flash_attn),
             tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
@@ -1099,39 +952,38 @@ struct csv_printer : public printer {
     }
 };
 
-
-static std::string escape_json(const std::string & value) {
-    std::string escaped;
-    for (auto c : value) {
-        if (c == '"') {
-            escaped += "\\\"";
-        } else if (c == '\\') {
-            escaped += "\\\\";
-        } else  if (c <= 0x1f) {
-            char buf[8];
-            snprintf(buf, sizeof(buf), "\\u%04x", c);
-            escaped += buf;
-        } else {
-            escaped += c;
-        }
-    }
-    return escaped;
-}
-
-static std::string format_json_value(const std::string & field, const std::string & value) {
-    switch (test::get_field_type(field)) {
-        case test::STRING:
-            return "\"" + escape_json(value) + "\"";
-        case test::BOOL:
-            return value == "0" ? "false" : "true";
-        default:
-            return value;
-    }
-}
-
 struct json_printer : public printer {
     bool first = true;
 
+    static std::string escape_json(const std::string & value) {
+        std::string escaped;
+        for (auto c : value) {
+            if (c == '"') {
+                escaped += "\\\"";
+            } else if (c == '\\') {
+                escaped += "\\\\";
+            } else  if (c <= 0x1f) {
+                char buf[8];
+                snprintf(buf, sizeof(buf), "\\u%04x", c);
+                escaped += buf;
+            } else {
+                escaped += c;
+            }
+        }
+        return escaped;
+    }
+
+    static std::string format_value(const std::string & field, const std::string & value) {
+        switch (test::get_field_type(field)) {
+            case test::STRING:
+                return "\"" + escape_json(value) + "\"";
+            case test::BOOL:
+                return value == "0" ? "false" : "true";
+            default:
+                return value;
+        }
+    }
+
     void print_header(const cmd_params & params) override {
         fprintf(fout, "[\n");
         (void) params;
@@ -1140,7 +992,7 @@ struct json_printer : public printer {
     void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
         assert(fields.size() == values.size());
         for (size_t i = 0; i < fields.size(); i++) {
-            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_value(fields.at(i), values.at(i)).c_str());
         }
     }
 
@@ -1163,25 +1015,6 @@ struct json_printer : public printer {
     }
 };
 
-
-struct jsonl_printer : public printer {
-    void print_fields(const std::vector<std::string> & fields, const std::vector<std::string> & values) {
-        assert(fields.size() == values.size());
-        for (size_t i = 0; i < fields.size(); i++) {
-            fprintf(fout, "\"%s\": %s, ", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
-        }
-    }
-
-    void print_test(const test & t) override {
-        fprintf(fout, "{");
-        print_fields(test::get_fields(), t.get_values());
-        fprintf(fout, "\"samples_ns\": [ %s ],", join(t.samples_ns, ", ").c_str());
-        fprintf(fout, "\"samples_ts\": [ %s ]", join(t.get_ts(), ", ").c_str());
-        fprintf(fout, "}\n");
-        fflush(fout);
-    }
-};
-
 struct markdown_printer : public printer {
     std::vector<std::string> fields;
 
@@ -1190,7 +1023,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 20;
+            return 16;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1272,15 +1105,6 @@ struct markdown_printer : public printer {
         if (params.n_threads.size() > 1 || params.n_threads != cmd_params_defaults.n_threads || is_cpu_backend) {
             fields.emplace_back("n_threads");
         }
-        if (params.cpu_mask.size() > 1 || params.cpu_mask != cmd_params_defaults.cpu_mask) {
-            fields.emplace_back("cpu_mask");
-        }
-        if (params.cpu_strict.size() > 1 || params.cpu_strict != cmd_params_defaults.cpu_strict) {
-            fields.emplace_back("cpu_strict");
-        }
-        if (params.poll.size() > 1 || params.poll != cmd_params_defaults.poll) {
-            fields.emplace_back("poll");
-        }
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
@@ -1482,14 +1306,12 @@ static std::unique_ptr<printer> create_printer(output_formats format) {
             return std::unique_ptr<printer>(new csv_printer());
         case JSON:
             return std::unique_ptr<printer>(new json_printer());
-        case JSONL:
-            return std::unique_ptr<printer>(new jsonl_printer());
         case MARKDOWN:
             return std::unique_ptr<printer>(new markdown_printer());
         case SQL:
             return std::unique_ptr<printer>(new sql_printer());
     }
-    GGML_ABORT("fatal error");
+    GGML_ASSERT(false);
 }
 
 int main(int argc, char ** argv) {
@@ -1517,8 +1339,6 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    set_process_priority(params.prio);
-
     // initialize printer
     std::unique_ptr<printer> p = create_printer(params.output_format);
     std::unique_ptr<printer> p_err = create_printer(params.output_format_stderr);
@@ -1538,13 +1358,7 @@ int main(int argc, char ** argv) {
     llama_model * lmodel = nullptr;
     const cmd_params_instance * prev_inst = nullptr;
 
-    int params_idx = 0;
-    auto params_count = params_instances.size();
     for (const auto & inst : params_instances) {
-        params_idx ++;
-        if (params.progress) {
-            fprintf(stderr, "llama-bench: benchmark %d/%ld: starting\n", params_idx, params_count);
-        }
         // keep the same model between tests when possible
         if (!lmodel || !prev_inst || !inst.equal_mparams(*prev_inst)) {
             if (lmodel) {
@@ -1570,40 +1384,12 @@ int main(int argc, char ** argv) {
 
         llama_kv_cache_clear(ctx);
 
-        // cool off before the test
-        if (params.delay) {
-            std::this_thread::sleep_for(std::chrono::seconds(params.delay));
-        }
-
-        struct ggml_threadpool_params tpp = ggml_threadpool_params_default(t.n_threads);
-        if (!parse_cpu_mask(t.cpu_mask, tpp.cpumask)) {
-            fprintf(stderr, "%s: failed to parse cpu-mask: %s\n", __func__, t.cpu_mask.c_str());
-            exit(1);
-        }
-        tpp.strict_cpu = t.cpu_strict;
-        tpp.poll       = t.poll;
-        tpp.prio       = params.prio;
-
-        struct ggml_threadpool* threadpool = ggml_threadpool_new(&tpp);
-        if (!threadpool) {
-            fprintf(stderr, "%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
-            exit(1);
-        }
-
-        llama_attach_threadpool(ctx, threadpool, NULL);
-
         // warmup run
         if (t.n_prompt > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
-            }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
             test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
-            if (params.progress) {
-                fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
-            }
             test_gen(ctx, 1, 0, t.n_threads);
         }
 
@@ -1613,15 +1399,9 @@ int main(int argc, char ** argv) {
             uint64_t t_start = get_time_ns();
 
             if (t.n_prompt > 0) {
-                if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
-                }
                 test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
-                if (params.progress) {
-                    fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
-                }
                 test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
 
@@ -1639,11 +1419,9 @@ int main(int argc, char ** argv) {
             fflush(p_err->fout);
         }
 
-        llama_perf_context_print(ctx);
+        llama_print_timings(ctx);
 
         llama_free(ctx);
-
-        ggml_threadpool_free(threadpool);
     }
 
     llama_free_model(lmodel);

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

@@ -120,8 +120,8 @@ Java_android_llama_cpp_LLamaAndroid_new_1context(JNIEnv *env, jobject, jlong jmo
     LOGi("Using %d threads", n_threads);
 
     llama_context_params ctx_params = llama_context_default_params();
-
-    ctx_params.n_ctx           = 2048;
+    ctx_params.seed  = 1234;
+    ctx_params.n_ctx = 2048;
     ctx_params.n_threads       = n_threads;
     ctx_params.n_threads_batch = n_threads;
 
@@ -269,6 +269,12 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
     return env->NewStringUTF(result.str().c_str());
 }
 
+extern "C"
+JNIEXPORT void JNICALL
+Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
+    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
+}
+
 extern "C"
 JNIEXPORT jlong JNICALL
 Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens, jint embd, jint n_seq_max) {
@@ -305,29 +311,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
     return reinterpret_cast<jlong>(batch);
 }
 
-extern "C"
-JNIEXPORT void JNICALL
-Java_android_llama_cpp_LLamaAndroid_free_1batch(JNIEnv *, jobject, jlong batch_pointer) {
-    llama_batch_free(*reinterpret_cast<llama_batch *>(batch_pointer));
-}
-
-extern "C"
-JNIEXPORT jlong JNICALL
-Java_android_llama_cpp_LLamaAndroid_new_1sampler(JNIEnv *, jobject) {
-    auto sparams = llama_sampler_chain_default_params();
-    sparams.no_perf = true;
-    llama_sampler * smpl = llama_sampler_chain_init(sparams);
-    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
-
-    return reinterpret_cast<jlong>(smpl);
-}
-
-extern "C"
-JNIEXPORT void JNICALL
-Java_android_llama_cpp_LLamaAndroid_free_1sampler(JNIEnv *, jobject, jlong sampler_pointer) {
-    llama_sampler_free(reinterpret_cast<llama_sampler *>(sampler_pointer));
-}
-
 extern "C"
 JNIEXPORT void JNICALL
 Java_android_llama_cpp_LLamaAndroid_backend_1init(JNIEnv *, jobject) {
@@ -398,25 +381,35 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         jobject,
         jlong context_pointer,
         jlong batch_pointer,
-        jlong sampler_pointer,
         jint n_len,
         jobject intvar_ncur
 ) {
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
-    const auto batch   = reinterpret_cast<llama_batch   *>(batch_pointer);
-    const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
+    const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
     const auto model = llama_get_model(context);
 
     if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
     if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
     if (!la_int_var_inc) la_int_var_inc = env->GetMethodID(la_int_var, "inc", "()V");
 
+    auto n_vocab = llama_n_vocab(model);
+    auto logits = llama_get_logits_ith(context, batch->n_tokens - 1);
+
+    std::vector<llama_token_data> candidates;
+    candidates.reserve(n_vocab);
+
+    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+        candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+    }
+
+    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
     // sample the most likely token
-    const auto new_token_id = llama_sampler_sample(sampler, context, -1);
+    const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
     if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
-        return nullptr;
+        return env->NewStringUTF("");
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);

examples/llama.android/llama/src/main/java/android/llama/cpp/LLamaAndroid.kt

@@ -45,10 +45,8 @@ class LLamaAndroid {
     private external fun free_context(context: Long)
     private external fun backend_init(numa: Boolean)
     private external fun backend_free()
-    private external fun new_batch(nTokens: Int, embd: Int, nSeqMax: Int): Long
     private external fun free_batch(batch: Long)
-    private external fun new_sampler(): Long
-    private external fun free_sampler(sampler: Long)
+    private external fun new_batch(nTokens: Int, embd: Int, nSeqMax: Int): Long
     private external fun bench_model(
         context: Long,
         model: Long,
@@ -71,7 +69,6 @@ class LLamaAndroid {
     private external fun completion_loop(
         context: Long,
         batch: Long,
-        sampler: Long,
         nLen: Int,
         ncur: IntVar
     ): String?
@@ -104,11 +101,8 @@ class LLamaAndroid {
                     val batch = new_batch(512, 0, 1)
                     if (batch == 0L) throw IllegalStateException("new_batch() failed")
 
-                    val sampler = new_sampler()
-                    if (sampler == 0L) throw IllegalStateException("new_sampler() failed")
-
                     Log.i(tag, "Loaded model $pathToModel")
-                    threadLocalState.set(State.Loaded(model, context, batch, sampler))
+                    threadLocalState.set(State.Loaded(model, context, batch))
                 }
                 else -> throw IllegalStateException("Model already loaded")
             }
@@ -120,7 +114,7 @@ class LLamaAndroid {
             is State.Loaded -> {
                 val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
-                    val str = completion_loop(state.context, state.batch, state.sampler, nlen, ncur)
+                    val str = completion_loop(state.context, state.batch, nlen, ncur)
                     if (str == null) {
                         break
                     }
@@ -144,7 +138,6 @@ class LLamaAndroid {
                     free_context(state.context)
                     free_model(state.model)
                     free_batch(state.batch)
-                    free_sampler(state.sampler);
 
                     threadLocalState.set(State.Idle)
                 }
@@ -168,7 +161,7 @@ class LLamaAndroid {
 
         private sealed interface State {
             data object Idle: State
-            data class Loaded(val model: Long, val context: Long, val batch: Long, val sampler: Long): State
+            data class Loaded(val model: Long, val context: Long, val batch: Long): State
         }
 
         // Enforce only one instance of Llm.