server : add loading html page while model is loading (#9468)

* Adding loading page for '/' server requests

* set content when model is loading

* removed loading html file

* updated cmakelist

* updated makefile

* cleaned up whitespace

* cleanup for PR removed error

* updated server test to handle 503 HTML

* updated server test to handle 503 HTML

* ca†ch 503 before parsing json

* revert test

* account for both api and web browser requests

* precommit corrections

* eol fix

* revert changes to pre-commit

* removed print statement

* made loading message more descriptive

* also support .html files

---------

Co-authored-by: VJHack <flymyplane21@gmail.com>
Co-authored-by: Vinesh Janarthanan <36610342+VJHack@users.noreply.github.com>

.devops/full-cuda.Dockerfile

@@ -1,18 +1,16 @@
 ARG UBUNTU_VERSION=22.04
-
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
-
+ARG CUDA_VERSION=12.6.0
 # 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
 
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev libgomp1
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -24,13 +22,12 @@ WORKDIR /app
 
 COPY . .
 
-# 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)
+# 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/* .
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/llama-cli-cann.Dockerfile

@@ -0,0 +1,44 @@
+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

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
+ARG CUDA_VERSION=12.6.0
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
@@ -8,28 +8,30 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential git
+    apt-get install -y build-essential git cmake
 
 WORKDIR /app
 
 COPY . .
 
-# Set nvcc architecture
-ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV GGML_CUDA=1
-
-RUN make -j$(nproc) llama-cli
+# 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)
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libgomp1
 
-COPY --from=build /app/llama-cli /llama-cli
+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
 
 ENTRYPOINT [ "/llama-cli" ]

.devops/llama-server-cuda.Dockerfile

@@ -1,6 +1,6 @@
 ARG UBUNTU_VERSION=22.04
 # This needs to generally match the container host's environment.
-ARG CUDA_VERSION=11.7.1
+ARG CUDA_VERSION=12.6.0
 # Target the CUDA build image
 ARG BASE_CUDA_DEV_CONTAINER=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
 # Target the CUDA runtime image
@@ -8,31 +8,34 @@ ARG BASE_CUDA_RUN_CONTAINER=nvidia/cuda:${CUDA_VERSION}-runtime-ubuntu${UBUNTU_V
 
 FROM ${BASE_CUDA_DEV_CONTAINER} AS build
 
-# Unless otherwise specified, we make a fat build.
-ARG CUDA_DOCKER_ARCH=all
+# CUDA architecture to build for (defaults to all supported archs)
+ARG CUDA_DOCKER_ARCH=default
 
 RUN apt-get update && \
-    apt-get install -y build-essential git libcurl4-openssl-dev
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
-# 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) llama-server
+# 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)
 
 FROM ${BASE_CUDA_RUN_CONTAINER} AS runtime
 
 RUN apt-get update && \
     apt-get install -y libcurl4-openssl-dev libgomp1 curl
 
-COPY --from=build /app/llama-server /llama-server
+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
 
 HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
 

.devops/llama-server-intel.Dockerfile

@@ -26,6 +26,8 @@ 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

@@ -39,6 +39,8 @@ 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

@@ -23,6 +23,8 @@ 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

@@ -21,6 +21,8 @@ RUN apt-get update && \
 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/devshells.nix

@@ -1,13 +1,52 @@
+{ inputs, ... }:
+
 {
   perSystem =
-    { config, lib, ... }:
+    {
+      config,
+      lib,
+      system,
+      ...
+    }:
     {
       devShells =
-        lib.concatMapAttrs
-          (name: package: {
-            ${name} = package.passthru.shell;
-            ${name + "-extra"} = package.passthru.shell-extra;
-          })
-          config.packages;
+        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))
+        ];
     };
 }

.devops/nix/nixpkgs-instances.nix

@@ -26,16 +26,14 @@
           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

@@ -0,0 +1,36 @@
+{
+  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

@@ -3,31 +3,33 @@
   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,
-  useMpi ? false, # Increases the runtime closure size by ~700M
+  # Increases the runtime closure size by ~700M
+  useMpi ? false,
   useRocm ? config.rocmSupport,
   enableCurl ? true,
   useVulkan ? false,
@@ -37,8 +39,8 @@
   # otherwise we get libstdc++ errors downstream.
   effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
   enableStatic ? effectiveStdenv.hostPlatform.isStatic,
-  precompileMetalShaders ? false
-}@inputs:
+  precompileMetalShaders ? false,
+}:
 
 let
   inherit (lib)
@@ -46,7 +48,6 @@ let
     cmakeFeature
     optionals
     strings
-    versionOlder
     ;
 
   stdenv = throw "Use effectiveStdenv instead";
@@ -62,54 +63,11 @@ 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}";
 
-  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" {} ''
+  xcrunHost = runCommand "xcrunHost" { } ''
     mkdir -p $out/bin
     ln -s /usr/bin/xcrun $out/bin
   '';
@@ -144,181 +102,145 @@ let
   ];
 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")
-        ];
-      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
+  # 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 ../../.;
+  };
 
-    buildInputs =
-      optionals effectiveStdenv.isDarwin darwinBuildInputs
-      ++ optionals useCuda cudaBuildInputs
-      ++ optionals useMpi [ mpi ]
-      ++ optionals useRocm rocmBuildInputs
-      ++ optionals useBlas [ blas ]
-      ++ optionals useVulkan vulkanBuildInputs
-      ++ optionals enableCurl [ curl ];
+  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\";"
+  '';
 
-    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)
-          )
+  # 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 $PATH
+  # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
+  __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+
+  nativeBuildInputs =
+    [
+      cmake
+      ninja
+      pkg-config
+      git
+    ]
+    ++ optionals useCuda [
+      cudaPackages.cuda_nvcc
+
+      autoAddDriverRunpath
+    ]
+    ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [ glibc.static ]
+    ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [ xcrunHost ];
+
+  buildInputs =
+    optionals effectiveStdenv.isDarwin darwinBuildInputs
+    ++ optionals useCuda cudaBuildInputs
+    ++ optionals useMpi [ mpi ]
+    ++ optionals useRocm rocmBuildInputs
+    ++ optionals useBlas [ blas ]
+    ++ optionals useVulkan vulkanBuildInputs
+    ++ optionals enableCurl [ curl ];
+
+  cmakeFlags =
+    [
+      (cmakeBool "LLAMA_BUILD_SERVER" true)
+      (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
+      (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
+      (cmakeBool "LLAMA_CURL" enableCurl)
+      (cmakeBool "GGML_NATIVE" false)
+      (cmakeBool "GGML_BLAS" useBlas)
+      (cmakeBool "GGML_CUDA" useCuda)
+      (cmakeBool "GGML_HIPBLAS" useRocm)
+      (cmakeBool "GGML_METAL" useMetalKit)
+      (cmakeBool "GGML_VULKAN" useVulkan)
+      (cmakeBool "GGML_STATIC" enableStatic)
+    ]
+    ++ optionals useCuda [
+      (
+        with cudaPackages.flags;
+        cmakeFeature "CMAKE_CUDA_ARCHITECTURES" (
+          builtins.concatStringsSep ";" (map dropDot cudaCapabilities)
         )
-      ]
-      ++ optionals useRocm [
-        (cmakeFeature "CMAKE_HIP_COMPILER" "${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/
+  '';
 
-    # Define the shells here, but don't add in the inputsFrom to avoid recursion.
-    passthru = {
-      inherit
-        useBlas
-        useCuda
-        useMetalKit
-        useMpi
-        useRocm
-        useVulkan
-        ;
+  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;
 
-      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"
-        '';
-      };
+    # Configurations that are known to result in build failures. Can be
+    # overridden by importing Nixpkgs with `allowBroken = true`.
+    broken = (useMetalKit && !effectiveStdenv.isDarwin);
 
-      shell-extra = mkShell {
-        name = "shell-extra-${finalAttrs.finalPackage.name}";
-        description = "contains numpy, sentencepiece, torchWithoutCuda, and transformers";
-        buildInputs = [ llama-python-extra ];
-        inputsFrom = [ finalAttrs.finalPackage ];
-      };
-    };
+    description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
+    homepage = "https://github.com/ggerganov/llama.cpp/";
+    license = lib.licenses.mit;
 
-    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;
+    # Accommodates `nix run` and `lib.getExe`
+    mainProgram = "llama-cli";
 
-      # Configurations that are known to result in build failures. Can be
-      # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+    # 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.
 
-      description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
-      homepage = "https://github.com/ggerganov/llama.cpp/";
-      license = lib.licenses.mit;
+    # 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
+    ];
 
-      # 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;
-    };
-  }
-)
+    # Extend `badPlatforms` instead
+    platforms = lib.platforms.all;
+  };
+})

.devops/nix/python-scripts.nix

@@ -0,0 +1,66 @@
+{
+  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

@@ -1,19 +1,41 @@
 {
   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;
-    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;
+  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 { };
+})

.ecrc

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

.github/workflows/build.yml

@@ -375,7 +375,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -401,7 +401,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -442,7 +442,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -546,7 +546,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -576,7 +576,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -610,7 +610,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -857,7 +857,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_CUDA=ON -DBUILD_SHARED_LIBS=ON
+          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 --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
@@ -969,14 +969,14 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Install
         id: depends
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-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"

.github/workflows/docker.yml

@@ -37,9 +37,9 @@ jobs:
           - { tag: "light-cuda", dockerfile: ".devops/llama-cli-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-cuda", dockerfile: ".devops/llama-server-cuda.Dockerfile", platforms: "linux/amd64" }
           - { tag: "full-cuda", dockerfile: ".devops/full-cuda.Dockerfile", platforms: "linux/amd64" }
-          - { 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.
+          # 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: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
@@ -96,21 +96,12 @@ jobs:
         env:
           GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
 
-      - name: Build and push Docker image (versioned)
+      - name: Build and push Docker image (tagged + versioned)
         if: github.event_name == 'push'
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v6
         with:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ 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 }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/server.yml

@@ -173,6 +173,7 @@ 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

@@ -61,6 +61,7 @@ llama-batched-swift
 /rpc-server
 out/
 tmp/
+autogen-*.md
 
 # Deprecated
 
@@ -129,3 +130,6 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
+
+# Test models for lora adapters
+/lora-tests

CMakeLists.txt

@@ -139,10 +139,16 @@ set(LLAMA_BIN_INSTALL_DIR     ${CMAKE_INSTALL_BINDIR}     CACHE PATH "Location o
 # 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)
-set(GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES} ${GGML_DIR_DEFINES})
+if (GGML_TARGET_DEFINES)
+    list(APPEND GGML_TRANSIENT_DEFINES ${GGML_TARGET_DEFINES})
+endif()
 get_target_property(GGML_LINK_LIBRARIES ggml LINK_LIBRARIES)
 
 set_target_properties(llama PROPERTIES PUBLIC_HEADER ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)

CMakePresets.json

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

Makefile

@@ -39,10 +39,12 @@ BUILD_TARGETS = \
 	llama-tokenize \
 	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 \
@@ -432,7 +434,7 @@ 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
+ifndef RISCV_CROSS_COMPILE
 
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686 amd64))
 	# Use all CPU extensions that are available:
@@ -512,7 +514,12 @@ ifneq ($(filter loongarch64%,$(UNAME_M)),)
 	MK_CXXFLAGS += -mlasx
 endif
 
-else
+ifneq ($(filter riscv64%,$(UNAME_M)),)
+	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
+	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
+endif
+
+else # RISC-V CROSS COMPILATION
 	MK_CFLAGS   += -march=rv64gcv -mabi=lp64d
 	MK_CXXFLAGS += -march=rv64gcv -mabi=lp64d
 endif
@@ -923,11 +930,11 @@ OBJ_LLAMA = \
 
 OBJ_COMMON = \
 	common/common.o \
+	common/arg.o \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
 	common/train.o \
-	common/grammar-parser.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -1156,6 +1163,11 @@ common/common.o: \
 	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+common/arg.o: \
+	common/arg.cpp \
+	common/arg.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 common/sampling.o: \
 	common/sampling.cpp \
 	common/sampling.h \
@@ -1167,11 +1179,6 @@ 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
@@ -1433,6 +1440,7 @@ 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)
@@ -1448,6 +1456,11 @@ 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 \
@@ -1505,6 +1518,11 @@ 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, $<)

README.md

@@ -10,32 +10,14 @@
 
 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
 
-- [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
+- [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)
 
 ## Hot topics
 
-- **`convert.py` has been deprecated and moved to `examples/convert_legacy_llama.py`, please use `convert_hf_to_gguf.py`** https://github.com/ggerganov/llama.cpp/pull/7430
-- Initial Flash-Attention support: https://github.com/ggerganov/llama.cpp/pull/5021
-- BPE pre-tokenization support has been added: https://github.com/ggerganov/llama.cpp/pull/6920
-- MoE memory layout has been updated - reconvert models for `mmap` support and regenerate `imatrix` https://github.com/ggerganov/llama.cpp/pull/6387
-- Model sharding instructions using `gguf-split` https://github.com/ggerganov/llama.cpp/discussions/6404
-- Fix major bug in Metal batched inference https://github.com/ggerganov/llama.cpp/pull/6225
-- Multi-GPU pipeline parallelism support https://github.com/ggerganov/llama.cpp/pull/6017
-- Looking for contributions to add Deepseek support: https://github.com/ggerganov/llama.cpp/issues/5981
-- Quantization blind testing: https://github.com/ggerganov/llama.cpp/discussions/5962
-- Initial Mamba support has been added: https://github.com/ggerganov/llama.cpp/pull/5328
+- Huggingface GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
 
@@ -106,6 +88,8 @@ Typically finetunes of the base models below are supported as well.
 - [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))
 
@@ -180,6 +164,7 @@ 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`)*
 
@@ -425,6 +410,7 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [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
 

ci/run.sh

@@ -13,6 +13,9 @@
 # # 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>"
@@ -40,7 +43,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=native"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -52,6 +55,10 @@ 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
@@ -107,7 +114,7 @@ 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                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                  ) 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
 
@@ -138,7 +145,7 @@ 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                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time make -j$(nproc)                                    ) 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
@@ -266,7 +273,6 @@ function gg_sum_ctest_with_model_release {
 }
 
 # open_llama_7b_v2
-# requires: GG_BUILD_CUDA
 
 function gg_run_open_llama_7b_v2 {
     cd ${SRC}
@@ -290,8 +296,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (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
+    (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 ../examples/convert_legacy_llama.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
 
@@ -425,7 +431,7 @@ 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                                            ) 2>&1 | tee -a $OUT/${ci}-make.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
 
@@ -535,7 +541,6 @@ function gg_sum_pythia_1_4b {
 }
 
 # pythia_2_8b
-# requires: GG_BUILD_CUDA
 
 function gg_run_pythia_2_8b {
     cd ${SRC}
@@ -556,8 +561,8 @@ function gg_run_pythia_2_8b {
 
     set -e
 
-    (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
+    (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
 
@@ -692,7 +697,7 @@ 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                                            ) 2>&1 | tee -a $OUT/${ci}-make.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
 
@@ -761,7 +766,7 @@ 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} ]; then
+        if [ -z ${GG_BUILD_CUDA} ] && [ -z ${GG_BUILD_VULKAN} ]; then
             test $ret -eq 0 && gg_run pythia_1_4b
         else
             test $ret -eq 0 && gg_run pythia_2_8b

common/CMakeLists.txt

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

common/arg.h

@@ -0,0 +1,77 @@
+#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

@@ -4,18 +4,11 @@
 
 #include "llama.h"
 
-#include "sampling.h"
-
 #define LOG_NO_FILE_LINE_FUNCTION
 #include "log.h"
 
-#include <cmath>
 #include <string>
 #include <vector>
-#include <random>
-#include <thread>
-#include <unordered_map>
-#include <tuple>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -54,26 +47,103 @@ 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();
 
 //
-// CLI argument parsing
+// Common params
 //
 
+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,
 };
 
-struct gpt_params {
-    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+// sampler parameters
+struct gpt_sampler_params {
+    uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    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_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 {
     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)
@@ -100,6 +170,11 @@ 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;
 
@@ -110,26 +185,25 @@ 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
 
-    // // sampling parameters
-    struct llama_sampling_params sparams;
+    struct gpt_sampler_params sparams;
 
-    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::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::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)
@@ -173,15 +247,14 @@ struct gpt_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
+    bool no_perf           = false; // disable performance metrics
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool ignore_eos        = false; // ignore generated EOS tokens
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
@@ -191,7 +264,7 @@ 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
+    std::string mmproj = "";        // path to multimodal projector                                         // NOLINT
     std::vector<std::string> image; // path to image file(s)
 
     // embedding
@@ -204,18 +277,18 @@ struct gpt_params {
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int32_t n_threads_http = -1;           // number of threads to process HTTP requests
+    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
 
     std::string hostname      = "127.0.0.1";
-    std::string public_path   = "";
-    std::string chat_template = "";
-    std::string system_prompt = "";
+    std::string public_path   = "";                                                                         // NOLINT
+    std::string chat_template = "";                                                                         // NOLINT
+    std::string system_prompt = "";                                                                         // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
 
-    std::string ssl_file_key  = "";
-    std::string ssl_file_cert = "";
+    std::string ssl_file_key  = "";                                                                         // NOLINT
+    std::string ssl_file_cert = "";                                                                         // NOLINT
 
     bool endpoint_slots   = true;
     bool endpoint_metrics = false;
@@ -265,18 +338,18 @@ struct gpt_params {
     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;
 };
 
-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
 //
@@ -327,8 +400,9 @@ struct llama_init_result {
 
 struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
 
-struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
-struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
+struct llama_model_params     llama_model_params_from_gpt_params    (const gpt_params & params);
+struct llama_context_params   llama_context_params_from_gpt_params  (const gpt_params & params);
+struct ggml_threadpool_params ggml_threadpool_params_from_cpu_params(const cpu_params & params);
 
 struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);

common/grammar-parser.cpp

@@ -1,539 +0,0 @@
-#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) {
-                if (rule.empty()) {
-                    throw std::runtime_error("Undefined rule");
-                }
-                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

@@ -1,29 +0,0 @@
-// 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/sampling.cpp

@@ -1,460 +1,450 @@
-#define LLAMA_API_INTERNAL
 #include "sampling.h"
-#include <random>
 
-struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
-    struct llama_sampling_context * result = new llama_sampling_context();
+#include "common.h"
 
-    result->params  = params;
-    result->grammar = nullptr;
+#include <cmath>
+#include <unordered_map>
 
-    // if there is a grammar, parse it
-    if (!params.grammar.empty()) {
-        result->parsed_grammar = grammar_parser::parse(params.grammar.c_str());
+// 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) {}
 
-        // 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;
+    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};
         }
 
-        // 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::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;
+        cur_p = { cur.data(), cur.size(), -1, false };
     }
+};
 
-    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) {
+std::string gpt_sampler_params::print() const {
     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",
-            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);
+            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);
 
     return std::string(result);
 }
 
-std::string llama_sampling_order_print(const llama_sampling_params & params) {
-    std::string result = "CFG -> Penalties ";
-    if (params.mirostat == 0) {
-        for (auto 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 + " ";
+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");
+                }
             }
+            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 {
-        result += "-> mirostat ";
+        llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
     }
 
     return result;
 }
 
-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";
+void gpt_sampler_free(struct gpt_sampler * gsmpl) {
+    if (gsmpl) {
+        llama_sampler_free(gsmpl->grmr);
+
+        llama_sampler_free(gsmpl->chain);
+
+        delete gsmpl;
+    }
+}
+
+void gpt_sampler_accept(struct gpt_sampler * gsmpl, llama_token token, bool accept_grammar) {
+    if (accept_grammar) {
+        llama_sampler_accept(gsmpl->grmr, token);
+    }
+
+    llama_sampler_accept(gsmpl->chain, token);
+
+    gsmpl->prev.push_back(token);
+}
+
+void gpt_sampler_reset(struct gpt_sampler * gsmpl) {
+    llama_sampler_reset(gsmpl->grmr);
+
+    llama_sampler_reset(gsmpl->chain);
+}
+
+struct gpt_sampler * gpt_sampler_clone(gpt_sampler * gsmpl) {
+    return new gpt_sampler {
+        /* .params = */ gsmpl->params,
+        /* .grmr   = */ llama_sampler_clone(gsmpl->grmr),
+        /* .chain  = */ llama_sampler_clone(gsmpl->chain),
+        /* .prev   = */ gsmpl->prev,
+        /* .cur    = */ gsmpl->cur,
+        /* .cur_p  = */ gsmpl->cur_p,
+    };
+}
+
+void gpt_perf_print(const struct llama_context * ctx, const struct gpt_sampler * gsmpl) {
+    // TODO: measure grammar performance
+
+    if (gsmpl) {
+        llama_perf_sampler_print(gsmpl->chain);
+    }
+    if (ctx) {
+        llama_perf_context_print(ctx);
+    }
+}
+
+llama_token gpt_sampler_sample(struct gpt_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first) {
+    gsmpl->set_logits(ctx, idx);
+
+    auto & grmr  = gsmpl->grmr;
+    auto & chain = gsmpl->chain;
+    auto & cur_p = gsmpl->cur_p; // initialized by set_logits
+
+    if (grammar_first) {
+        llama_sampler_apply(grmr, &cur_p);
+    }
+
+    llama_sampler_apply(chain, &cur_p);
+
+    GGML_ASSERT(cur_p.selected != -1 && "no selected token during sampling - check your sampling configuration");
+
+    const llama_token id = cur_p.data[cur_p.selected].id;
+
+    if (grammar_first) {
+        return id;
+    }
+
+    // check if it the sampled token fits the grammar
+    {
+        llama_token_data       single_token_data       = { id, 1.0f, 0.0f };
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, -1, false };
+
+        llama_sampler_apply(grmr, &single_token_data_array);
+
+        const bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+        if (is_valid) {
+            return id;
+        }
+    }
+
+    // resampling:
+    // if the token is not valid, sample again, but first apply the grammar sampler and then the sampling chain
+    gsmpl->set_logits(ctx, idx);
+
+    llama_sampler_apply(grmr,  &cur_p);
+    llama_sampler_apply(chain, &cur_p);
+
+    GGML_ASSERT(cur_p.selected != -1 && "no selected token during re-sampling - check your sampling configuration");
+
+    return cur_p.data[cur_p.selected].id;
+}
+
+uint32_t gpt_sampler_get_seed(const struct gpt_sampler * gsmpl) {
+    return llama_sampler_get_seed(gsmpl->chain);
+}
+
+// helpers
+
+llama_token_data_array * gpt_sampler_get_candidates(struct gpt_sampler * gsmpl) {
+    return &gsmpl->cur_p;
+}
+
+llama_token gpt_sampler_last(const struct gpt_sampler * gsmpl) {
+    return gsmpl->prev.rat(0);
+}
+
+std::string gpt_sampler_print(const struct gpt_sampler * gsmpl) {
+    std::string result = "\tlogits ";
+
+    for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
+        const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
+        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
+    }
+
+    return result;
+}
+
+std::string gpt_sampler_prev_str(gpt_sampler * gsmpl, llama_context * ctx_main, int n) {
+    n = std::min(n, (int) gsmpl->prev.size());
+
+    if (n <= 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(8*n); // 8 is the average length of a token [citation needed], TODO: compute this from the vocab
+
+    for (int i = n - 1; i >= 0; i--) {
+        const llama_token id = gsmpl->prev.rat(i);
+
+        GGML_ASSERT(id != LLAMA_TOKEN_NULL && "null token in the sampling history - should not happen");
+
+        result += llama_token_to_piece(ctx_main, id);
+    }
+
+    return result;
+}
+
+char gpt_sampler_type_to_chr(enum gpt_sampler_type cnstr) {
+    switch (cnstr) {
+        case GPT_SAMPLER_TYPE_TOP_K:       return 'k';
+        case GPT_SAMPLER_TYPE_TFS_Z:       return 'f';
+        case GPT_SAMPLER_TYPE_TYPICAL_P:   return 'y';
+        case GPT_SAMPLER_TYPE_TOP_P:       return 'p';
+        case GPT_SAMPLER_TYPE_MIN_P:       return 'm';
+        case GPT_SAMPLER_TYPE_TEMPERATURE: return 't';
+        default : return '?';
+    }
+}
+
+std::string gpt_sampler_type_to_str(enum gpt_sampler_type cnstr) {
+    switch (cnstr) {
+        case GPT_SAMPLER_TYPE_TOP_K:       return "top_k";
+        case GPT_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
+        case GPT_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
+        case GPT_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case GPT_SAMPLER_TYPE_MIN_P:       return "min_p";
+        case GPT_SAMPLER_TYPE_TEMPERATURE: return "temperature";
         default : return "";
     }
 }
 
-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}
+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 },
     };
 
     // since samplers names are written multiple ways
     // make it ready for both system names and input names
-    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::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::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;
-}
+    std::vector<gpt_sampler_type> samplers;
+    samplers.reserve(names.size());
 
-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);
+    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);
         } else {
-            // 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_grammar_sample(ctx_sampling->grammar, ctx_main, &single_token_data_array);
-
-        // 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;
+            if (allow_alt_names) {
+                sampler = sampler_alt_name_map.find(name);
+                if (sampler != sampler_alt_name_map.end()) {
+                    samplers.push_back(sampler->second);
                 }
             }
         }
     }
 
-    // apply grammar checks before sampling logic
-    if (apply_grammar && ctx_sampling->grammar != NULL) {
-        llama_grammar_sample(ctx_sampling->grammar, ctx_main, &cur_p);
+    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);
+        }
     }
 
-    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_sampling->grammar, ctx_main, id);
-    }
+    return samplers;
 }

common/sampling.h

@@ -2,159 +2,82 @@
 
 #include "llama.h"
 
-#include "grammar-parser.h"
-
-#include <random>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-// 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);
+#include <string>
+#include <vector>
 
-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
+// gpt_sampler extends llama_sampler with additional functionality:
 //
-// required:
-//  - ctx_main:     context to use for sampling
-//  - ctx_sampling: sampling-specific context
+//  - grammar support
+//  - custom sampler logic based on the parameters
+//  - history of the last accepted tokens
+//  - performance metrics
 //
-// optional:
-//  - ctx_cfg:      context to use for classifier-free guidance
-//  - idx:          sample from llama_get_logits_ith(ctx, idx)
+// 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).
 //
-// returns:
-//  - token:      sampled token
-//  - candidates: vector of candidate tokens
+// 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
 //
-llama_token llama_sampling_sample(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = -1);
 
-// 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);
+struct gpt_sampler;
 
-void llama_sampling_accept(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        llama_token id,
-        bool apply_grammar);
+// 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);

convert_hf_to_gguf.py

@@ -3,6 +3,7 @@
 
 from __future__ import annotations
 
+import ast
 import logging
 import argparse
 import contextlib
@@ -63,6 +64,7 @@ class Model:
     model_name: str | None
     metadata_override: Path | None
     dir_model_card: Path
+    is_lora: bool
 
     # subclasses should define this!
     model_arch: gguf.MODEL_ARCH
@@ -70,7 +72,7 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False, is_lora: bool = False):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -92,6 +94,7 @@ class Model:
         self.metadata_override = metadata_override
         self.model_name = model_name
         self.dir_model_card = dir_model  # overridden in convert_lora_to_gguf.py
+        self.is_lora = is_lora  # true if model is used inside convert_lora_to_gguf.py
 
         # Apply heuristics to figure out typical tensor encoding based on first layer tensor encoding type
         if self.ftype == gguf.LlamaFileType.GUESSED:
@@ -295,12 +298,32 @@ class Model:
                             gguf.MODEL_TENSOR.FFN_GATE_INP,
                             gguf.MODEL_TENSOR.POS_EMBD,
                             gguf.MODEL_TENSOR.TOKEN_TYPES,
+                            gguf.MODEL_TENSOR.SSM_CONV1D,
+                            gguf.MODEL_TENSOR.TIME_MIX_FIRST,
+                            gguf.MODEL_TENSOR.TIME_MIX_W1,
+                            gguf.MODEL_TENSOR.TIME_MIX_W2,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
+                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
                         )
                     )
-                    or not name.endswith(".weight")
+                    or not new_name.endswith(".weight")
                 ):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
+                if data_qtype is False and any(
+                    self.match_model_tensor_name(new_name, key, bid)
+                    for key in (
+                        gguf.MODEL_TENSOR.TOKEN_EMBD,
+                        gguf.MODEL_TENSOR.OUTPUT,
+                    )
+                ):
+                    if self.ftype in (
+                        gguf.LlamaFileType.MOSTLY_TQ1_0,
+                        gguf.LlamaFileType.MOSTLY_TQ2_0,
+                    ):
+                        # TODO: use Q4_K and Q6_K
+                        data_qtype = gguf.GGMLQuantizationType.F16
+
                 # No override (data_qtype is False), or wants to be quantized (data_qtype is True)
                 if isinstance(data_qtype, bool):
                     if self.ftype == gguf.LlamaFileType.ALL_F32:
@@ -311,6 +334,10 @@ class Model:
                         data_qtype = gguf.GGMLQuantizationType.BF16
                     elif self.ftype == gguf.LlamaFileType.MOSTLY_Q8_0:
                         data_qtype = gguf.GGMLQuantizationType.Q8_0
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_TQ1_0:
+                        data_qtype = gguf.GGMLQuantizationType.TQ1_0
+                    elif self.ftype == gguf.LlamaFileType.MOSTLY_TQ2_0:
+                        data_qtype = gguf.GGMLQuantizationType.TQ2_0
                     else:
                         raise ValueError(f"Unknown file type: {self.ftype.name}")
 
@@ -599,6 +626,9 @@ class Model:
         if chkhsh == "4e2b24cc4770243d65a2c9ec19770a72f08cffc161adbb73fcbb6b7dd45a0aae":
             # ref: https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct
             res = "exaone"
+        if chkhsh == "fcace8b9cac38ce847670c970cd5892031a753a1ef381abd1d9af00f713da085":
+            # ref: https://huggingface.co/microsoft/phi-2
+            res = "phi-2"
 
         if res is None:
             logger.warning("\n")
@@ -1569,7 +1599,7 @@ class LlamaModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -1592,7 +1622,8 @@ class LlamaModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
-                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+                if not self.is_lora:
+                    self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
 
         super().prepare_tensors()
 
@@ -1615,15 +1646,16 @@ class BitnetModel(Model):
         self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
         self.gguf_writer.add_rope_scaling_factor(1.0)
 
-    def weight_quant(self, weight):
+    def weight_quant(self, weight: Tensor) -> Tensor:
         dtype = weight.dtype
         weight = weight.float()
-        s = 1 / weight.abs().mean().clamp(min=1e-5)
-        weight = (weight * s).round().clamp(-1, 1) / s
-        scale = weight.abs().max().unsqueeze(0)
-        weight = torch.where(weight.abs().less(1e-6), 0, weight).type(dtype)
-        weight = torch.sign(weight).type(dtype)
-        return weight.type(dtype), scale.type(torch.float32)
+        scale = weight.abs().mean().clamp(min=1e-5)
+        iscale = 1 / scale
+        # TODO: multiply by the scale directly instead of inverting it twice
+        # (this is also unnecessarily doubly inverted upstream)
+        # ref: https://huggingface.co/1bitLLM/bitnet_b1_58-3B/blob/af89e318d78a70802061246bf037199d2fb97020/utils_quant.py#L10
+        result = (weight * iscale).round().clamp(-1, 1) / iscale
+        return result.type(dtype)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
         new_name = self.map_tensor_name(name)
@@ -1638,11 +1670,9 @@ class BitnetModel(Model):
             gguf.MODEL_TENSOR.FFN_GATE,
         ]):
             # transform weight into 1/0/-1 (in fp32)
-            weight_torch, scale_torch = self.weight_quant(data_torch)
-            yield (new_name, weight_torch)
-            yield (new_name.removesuffix(".weight") + ".scale", scale_torch)
-        else:
-            yield (new_name, data_torch)
+            data_torch = self.weight_quant(data_torch)
+
+        yield (new_name, data_torch)
 
 
 @Model.register("GrokForCausalLM")
@@ -2139,8 +2169,9 @@ class Phi3MiniModel(Model):
         if len(long_factors) != len(short_factors) or len(long_factors) != rope_dims / 2:
             raise ValueError(f'The length of rope long and short factors must be {rope_dims / 2}')
 
-        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
-        self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
+        if not self.is_lora:
+            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_LONG]  + ".weight", np.array(long_factors, dtype=np.float32))
+            self.gguf_writer.add_tensor(gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT] + ".weight", np.array(short_factors, dtype=np.float32))
 
 
 @Model.register("PlamoForCausalLM")
@@ -2711,7 +2742,87 @@ class StarCoder2Model(Model):
     model_arch = gguf.MODEL_ARCH.STARCODER2
 
 
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
+@Model.register("Rwkv6ForCausalLM")
+class Rwkv6Model(Model):
+    model_arch = gguf.MODEL_ARCH.RWKV6
+
+    def set_vocab(self):
+        assert (self.dir_model / "rwkv_vocab_v20230424.txt").is_file()
+        vocab_size = self.hparams.get("vocab_size", 65536)
+
+        tokens: list[bytes] = ['<s>'.encode("utf-8")]
+        toktypes: list[int] = [gguf.TokenType.CONTROL]
+
+        with open(self.dir_model / "rwkv_vocab_v20230424.txt", "r", encoding="utf-8") as f:
+            lines = f.readlines()
+            for line in lines:
+                parts = line.split(' ')
+                assert len(parts) >= 3
+                token, token_len = ast.literal_eval(' '.join(parts[1:-1])), int(parts[-1])
+                token = token.encode("utf-8") if isinstance(token, str) else token
+                assert isinstance(token, bytes)
+                assert len(token) == token_len
+                token_text: str = repr(token)[2:-1]  # "b'\xff'" -> "\xff"
+                tokens.append(token_text.encode("utf-8"))
+                toktypes.append(gguf.TokenType.NORMAL)
+        remainder = vocab_size - len(tokens)
+        assert remainder >= 0
+        for i in range(len(tokens), vocab_size):
+            tokens.append(f"[PAD{i}]".encode("utf-8"))
+            toktypes.append(gguf.TokenType.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("rwkv")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_size = self.hparams["head_size"]
+        hidden_size = self.hparams["hidden_size"]
+        layer_norm_eps = self.hparams["layer_norm_epsilon"]
+        rescale_every_n_layers = self.hparams["rescale_every"]
+        intermediate_size = self.hparams["intermediate_size"] if self.hparams["intermediate_size"] is not None else int((hidden_size * 3.5) // 32 * 32)
+        time_mix_extra_dim = 64 if hidden_size == 4096 else 32
+        time_decay_extra_dim = 128 if hidden_size == 4096 else 64
+
+        # RWKV isn't context limited
+        self.gguf_writer.add_context_length(1048576)
+        self.gguf_writer.add_embedding_length(hidden_size)
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_layer_norm_eps(layer_norm_eps)
+        self.gguf_writer.add_rescale_every_n_layers(rescale_every_n_layers)
+        self.gguf_writer.add_wkv_head_size(head_size)
+        self.gguf_writer.add_time_mix_extra_dim(time_mix_extra_dim)
+        self.gguf_writer.add_time_decay_extra_dim(time_decay_extra_dim)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        # required by llama.cpp, unused
+        self.gguf_writer.add_head_count(0)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        new_name = self.map_tensor_name(name)
+
+        if not (new_name.endswith(".weight") or new_name.endswith(".bias")):
+            new_name += ".weight"
+
+        if new_name.endswith("time_mix_w1.weight") or new_name.endswith("time_mix_decay_w1.weight") or new_name.endswith("time_mix_decay_w2.weight"):
+            data_torch = data_torch.transpose(0, 1)
+
+        if new_name.endswith("time_mix_w2.weight"):
+            data_torch = data_torch.permute(0, 2, 1)
+
+        rescale_every_n_layers = self.hparams["rescale_every"]
+        if rescale_every_n_layers > 0:
+            if new_name.endswith("time_mix_output.weight") or new_name.endswith("channel_mix_value.weight"):
+                data_torch = data_torch.div_(2 ** int(bid // rescale_every_n_layers))
+
+        yield (new_name, data_torch)
+
+
+@Model.register("MambaForCausalLM", "MambaLMHeadModel", "FalconMambaForCausalLM")
 class MambaModel(Model):
     model_arch = gguf.MODEL_ARCH.MAMBA
 
@@ -2742,7 +2853,10 @@ class MambaModel(Model):
         # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
         dt_rank      = self.find_hparam(["time_step_rank",     "dt_rank"],      optional=True) or -(d_model // -16)
         rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
+        use_dt_b_c_norm = False
+        # For falconmamba we do apply RMS norm on B / DT and C layers
+        if self.find_hparam(["model_type"], optional=True) in ("falcon_mamba",):
+            use_dt_b_c_norm = True
         # Fail early for models which don't have a block expansion factor of 2
         assert d_inner == 2 * d_model
 
@@ -2750,12 +2864,13 @@ class MambaModel(Model):
         self.gguf_writer.add_embedding_length(d_model)
         self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
         self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
+        self.gguf_writer.add_block_count(self.block_count)
         self.gguf_writer.add_ssm_conv_kernel(d_conv)
         self.gguf_writer.add_ssm_inner_size(d_inner)
         self.gguf_writer.add_ssm_state_size(d_state)
         self.gguf_writer.add_ssm_time_step_rank(dt_rank)
         self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
+        self.gguf_writer.add_ssm_dt_b_c_rms(use_dt_b_c_norm) # For classic Mamba we don't apply rms norm on B / DT layers
         self.gguf_writer.add_file_type(self.ftype)
 
     _tok_embd = None
@@ -2782,23 +2897,6 @@ class MambaModel(Model):
 
         return [(new_name, data_torch)]
 
-    def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        if bid is not None and new_name in (
-            self.format_tensor_name(
-                n, bid, ".weight" if name.endswith(".weight") else ""
-            )
-            for n in [
-                gguf.MODEL_TENSOR.SSM_CONV1D,
-                gguf.MODEL_TENSOR.SSM_X,
-                gguf.MODEL_TENSOR.SSM_DT,
-                gguf.MODEL_TENSOR.SSM_A,
-                gguf.MODEL_TENSOR.SSM_D,
-            ]
-        ):
-            return gguf.GGMLQuantizationType.F32
-
-        return super().tensor_force_quant(name, new_name, bid, n_dims)
-
 
 @Model.register("CohereForCausalLM")
 class CommandR2Model(Model):
@@ -3792,7 +3890,7 @@ class ExaoneModel(Model):
     def set_gguf_parameters(self):
         hparams = self.hparams
 
-        assert(hparams["activation_function"] == "silu")
+        assert (hparams["activation_function"] == "silu")
 
         max_position_embeddings = hparams["max_position_embeddings"]
         embed_dim = hparams["hidden_size"]
@@ -3828,7 +3926,7 @@ class ExaoneModel(Model):
         if rope_scaling := self.find_hparam(["rope_scaling"], optional=True):
             if rope_scaling.get("rope_type", '').lower() == "llama3":
                 base = self.hparams.get("rope_theta", 10000.0)
-                dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]
+                dim = self.hparams.get("head_dim", self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
                 freqs = 1.0 / (base ** (torch.arange(0, dim, 2, dtype=torch.float32) / dim))
 
                 factor = rope_scaling.get("factor", 8.0)
@@ -3851,12 +3949,13 @@ class ExaoneModel(Model):
                         smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
                         rope_factors.append(1 / ((1 - smooth) / factor + smooth))
 
-                self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
+                if not self.is_lora:
+                    self.gguf_writer.add_tensor(self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), np.array(rope_factors, dtype=np.float32))
 
         super().prepare_tensors()
 
-###### CONVERSION LOGIC ######
 
+###### CONVERSION LOGIC ######
 
 # tree of lazy tensors
 class LazyTorchTensor(gguf.LazyBase):
@@ -3936,8 +4035,8 @@ def parse_args() -> argparse.Namespace:
         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",
+        "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="f16",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type depending on the first loaded tensor type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -4024,6 +4123,8 @@ def main() -> None:
         "f16": gguf.LlamaFileType.MOSTLY_F16,
         "bf16": gguf.LlamaFileType.MOSTLY_BF16,
         "q8_0": gguf.LlamaFileType.MOSTLY_Q8_0,
+        "tq1_0": gguf.LlamaFileType.MOSTLY_TQ1_0,
+        "tq2_0": gguf.LlamaFileType.MOSTLY_TQ2_0,
         "auto": gguf.LlamaFileType.GUESSED,
     }
 

convert_hf_to_gguf_update.py

@@ -31,6 +31,7 @@ import re
 import requests
 import sys
 import json
+import shutil
 
 from hashlib import sha256
 from enum import IntEnum, auto
@@ -97,6 +98,7 @@ models = [
     {'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", },
 ]
 
 
@@ -125,12 +127,27 @@ def download_model(model):
     if tokt == TOKENIZER_TYPE.UGM:
         files.append("spiece.model")
 
-    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)
+    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 model in models:

convert_lora_to_gguf.py

@@ -363,7 +363,13 @@ if __name__ == '__main__':
                     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 = super().modify_tensors(data_torch, name, bid)
+                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()
@@ -386,6 +392,7 @@ if __name__ == '__main__':
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
+            is_lora=True,
         )
 
         logger.info("Exporting model...")

docs/backend/CANN.md

@@ -0,0 +1,259 @@
+# 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 - Math Kernel Library)*.
+- **oneAPI Libraries**: A set of highly optimized libraries targeting multiple domains *(e.g. oneMKL and oneDNN)*.
 - **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.
 
@@ -28,10 +28,6 @@
 
 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.
@@ -45,6 +41,10 @@ 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.
@@ -196,7 +196,7 @@ 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 MKL for intel GPUs.
+Upon a successful installation, SYCL is enabled for the available intel devices, along with relevant libraries such as oneAPI oneDNN for Intel GPUs.
 
 - **Adding support to Nvidia GPUs**
 
@@ -255,8 +255,6 @@ or
 # 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
 
@@ -338,12 +336,12 @@ Choose one of following methods to run.
 - Use device 0:
 
 ```sh
-./examples/sycl/run_llama2.sh 0
+./examples/sycl/run-llama2.sh 0
 ```
 - Use multiple devices:
 
 ```sh
-./examples/sycl/run_llama2.sh
+./examples/sycl/run-llama2.sh
 ```
 
 2. Command line

docs/build.md

@@ -352,6 +352,37 @@ 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/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 `11.7.1`
-- `CUDA_DOCKER_ARCH` set to `all`
+- `CUDA_VERSION` set to `12.6.0`
+- `CUDA_DOCKER_ARCH` set to the cmake build default, which includes all the supported architectures
 
 The resulting images, are essentially the same as the non-CUDA images:
 

examples/baby-llama/baby-llama.cpp

@@ -18,7 +18,7 @@ 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);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     if (plan.work_size > 0) {
         buf.resize(plan.work_size);

examples/batched-bench/README.md

@@ -49,3 +49,12 @@ 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

@@ -1,36 +1,13 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-// 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);
-
+static void print_usage(int, char ** argv) {
     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");
@@ -39,8 +16,7 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
         return 1;
     }
 
@@ -122,12 +98,13 @@ int main(int argc, char ** argv) {
         }
     }
 
-    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", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
+    if (!params.batched_bench_output_jsonl) {
+        LOG_TEE("\n");
+        LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+        LOG_TEE("\n");
+        LOG_TEE("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
+        LOG_TEE("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
+    }
 
     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) {
@@ -195,12 +172,22 @@ int main(int argc, char ** argv) {
                 const float speed_tg = pl*tg / t_tg;
                 const float speed    = n_kv / t;
 
-                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);
+                if(params.batched_bench_output_jsonl) {
+                    LOG_TEE(
+                        "{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
+                        "\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
+                        n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
+                        pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
+                    );
+                } else {
+                    LOG_TEE("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
+                }
             }
         }
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/batched.swift/Sources/main.swift

@@ -27,7 +27,6 @@ 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)
 }
@@ -37,7 +36,6 @@ 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
@@ -48,11 +46,26 @@ 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")
@@ -125,32 +138,7 @@ while n_cur <= n_len {
             continue
         }
 
-        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);
+        let new_token_id = llama_sampler_sample(smpl, context, i_batch[i])
 
         // is it an end of stream? -> mark the stream as finished
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
@@ -210,9 +198,10 @@ 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")
+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")
 
-llama_print_timings(context)
+llama_perf_sampler_print(smpl)
+llama_perf_context_print(context)
 
 private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
     let utf8Count = text.utf8.count

examples/batched/batched.cpp

@@ -1,15 +1,13 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     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");
@@ -21,8 +19,7 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
@@ -65,6 +62,15 @@ int main(int argc, char ** argv) {
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_k(params.sparams.top_k));
+    llama_sampler_chain_add(smpl, llama_sampler_init_top_p(params.sparams.top_p, params.sparams.min_keep));
+    llama_sampler_chain_add(smpl, llama_sampler_init_temp (params.sparams.temp));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist (params.sparams.seed));
+
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
@@ -164,29 +170,7 @@ int main(int argc, char ** argv) {
                 continue;
             }
 
-            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);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, i_batch[i]);
 
             // is it an end of generation? -> mark the stream as finished
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
@@ -244,12 +228,15 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(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);
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads, nullptr);
 
     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 {
-    int32_t n_threads     = 1;
+    int     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(gf->nodes[0]);
+    TENSOR_DUMP(ggml_graph_node(gf, 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(gf->nodes[0]);
+    float sum_of_F32_reference = tensor_sum_elements(ggml_graph_node(gf, 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(gf31->nodes[0]);
+        float sum_of_Q4_result = tensor_sum_elements(ggml_graph_node(gf31, 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/cvector-generator/cvector-generator.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -35,9 +36,7 @@ static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     return ret;
 }
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     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]);
@@ -390,8 +389,7 @@ 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)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
 
@@ -486,8 +484,8 @@ int main(int argc, char ** argv) {
     if (use_pca) {
         // run PCA
         PCA::pca_params pca_params;
-        pca_params.n_threads = params.n_threads;
-        pca_params.n_batch = params.n_pca_batch;
+        pca_params.n_threads    = params.cpuparams.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

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

examples/embedding/embedding.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -79,8 +80,7 @@ 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)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
         return 1;
     }
 
@@ -90,14 +90,6 @@ int main(int argc, char ** argv) {
 
     print_build_info();
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -313,8 +305,10 @@ int main(int argc, char ** argv) {
         if (notArray) fprintf(stdout, "\n}\n");
     }
 
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
+
     // clean up
-    llama_print_timings(ctx);
     llama_batch_free(batch);
     llama_free(ctx);
     llama_free_model(model);

examples/eval-callback/eval-callback.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -144,15 +145,12 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
     print_build_info();
 
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -183,7 +181,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/export-lora/export-lora.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
@@ -369,7 +370,7 @@ struct lora_merge_ctx {
 
         // write data to output file
         {
-            auto result = gf->nodes[gf->n_nodes - 1];
+            auto * result = ggml_graph_node(gf, -1);
             size_t len = ggml_nbytes(result);
             if (read_buf.size() < len) {
                 read_buf.resize(len);
@@ -391,9 +392,7 @@ struct lora_merge_ctx {
     }
 };
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+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");
@@ -403,14 +402,13 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
         return 1;
     }
 
     g_verbose = (params.verbosity == 1);
     try {
-        lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
+        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());

examples/gbnf-validator/gbnf-validator.cpp

@@ -1,9 +1,5 @@
-#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>
@@ -12,29 +8,28 @@
 #include <string>
 #include <vector>
 
-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;
+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 & cur_stacks = llama_grammar_get_stacks(grammar);
+          llama_grammar_stacks & stacks_cur = llama_grammar_get_stacks(grammar);
 
     size_t pos = 0;
-    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
-        const llama_grammar_stacks prev_stacks = llama_grammar_get_stacks(grammar); // copy
+    for (const auto & cpt : cpts) {
+        const llama_grammar_stacks stacks_prev = llama_grammar_get_stacks(grammar); // copy
 
-        llama_grammar_accept(rules, prev_stacks, *it, cur_stacks);
+        llama_grammar_accept(rules, stacks_prev, cpt, stacks_cur);
 
-        if (cur_stacks.empty()) {
+        if (stacks_cur.empty()) {
             error_pos = pos;
-            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
-            cur_stacks = prev_stacks;
+            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(cpt) + "'";
+            stacks_cur = stacks_prev;
             return false;
         }
         ++pos;
     }
 
-    for (const auto & stack : cur_stacks) {
+    for (const auto & stack : stacks_cur) {
         if (stack.empty()) {
             return true;
         }
@@ -85,27 +80,7 @@ int main(int argc, char** argv) {
         grammar_str = buffer.str();
     }
 
-    // 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"));
+    llama_grammar * grammar = llama_grammar_init_impl(nullptr, grammar_str.c_str(), "root");
     if (grammar == nullptr) {
         throw std::runtime_error("Failed to initialize llama_grammar");
     }
@@ -122,7 +97,7 @@ 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_sample_grammar_string(grammar, input_str, error_pos, error_msg);
+    bool is_valid = llama_grammar_validate(grammar, input_str, error_pos, error_msg);
 
     if (is_valid) {
         fprintf(stdout, "Input string is valid according to the grammar.\n");
@@ -131,7 +106,7 @@ int main(int argc, char** argv) {
     }
 
     // Clean up
-    llama_grammar_free(grammar);
+    llama_grammar_free_impl(grammar);
 
     return 0;
 }

examples/gen-docs/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-gen-docs)
+add_executable(${TARGET} gen-docs.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/gen-docs/gen-docs.cpp

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

examples/gritlm/gritlm.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -9,7 +10,7 @@
 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 * mdl = llama_get_model(ctx);
+    const llama_model * model = llama_get_model(ctx);
 
     llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
@@ -18,16 +19,16 @@ 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(mdl, input_string, true, false);
+        std::vector<llama_token> inputs = llama_tokenize(model, 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(mdl));
+        // inputs.push_back(llama_token_eos(model));
 
         // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
+        const int32_t n_inst = llama_tokenize(model, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -51,7 +52,7 @@ 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(mdl);
+        uint64_t n_embd = llama_n_embd(model);
 
         // allocate embedding output
         std::vector<float> emb_unorm(n_embd, 0.0f);
@@ -92,11 +93,11 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
     return result;
 }
 
-static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
+static std::string generate(llama_context * ctx, llama_sampler * smpl, const std::string & prompt, bool stream) {
     std::string result;
 
-    const llama_model * mdl = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(mdl);
+    const llama_model * model = llama_get_model(ctx);
+    llama_token eos_token = llama_token_eos(model);
 
     llama_kv_cache_clear(ctx);
     llama_set_embeddings(ctx, false);
@@ -104,28 +105,24 @@ static std::string generate(llama_context * ctx, const std::string & prompt, boo
 
     llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
 
-    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
+    std::vector<llama_token> inputs = llama_tokenize(model, prompt, false, true);
     int32_t i_current_token = 0;
 
     while (true) {
         llama_batch_clear(bat);
-        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);
+        {
+            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);
+            }
         }
         inputs.clear();
 
         llama_decode(ctx, bat);
-        auto logits = llama_get_logits_ith(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_sampler_sample(smpl, ctx, bat.n_tokens - 1);
 
-        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
         if (token == eos_token) {
             break;
         }
@@ -157,8 +154,7 @@ 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)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -167,10 +163,18 @@ int main(int argc, char * argv[]) {
 
     llama_backend_init();
 
-    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
 
     // create generation context
-    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
+    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());
 
     // ### Embedding/Representation ###
     // samples taken from: https://github.com/ContextualAI/gritlm#basic
@@ -191,7 +195,7 @@ 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(mdl);
+        const int n_embd = llama_n_embd(model);
 
         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);
@@ -208,11 +212,12 @@ 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, prompt, true);
+        std::string response = generate(ctx, smpl, prompt, true);
     }
 
+    llama_sampler_free(smpl);
     llama_free(ctx);
-    llama_free_model(mdl);
+    llama_free_model(model);
     llama_backend_free();
 
     return 0;

examples/imatrix/imatrix.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -17,9 +18,7 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     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"
@@ -579,8 +578,7 @@ int main(int argc, char ** argv) {
     params.logits_all = true;
     params.verbosity = 1;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
         return 1;
     }
 
@@ -638,7 +636,8 @@ int main(int argc, char ** argv) {
 
     g_collector.save_imatrix();
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/infill/infill.cpp

@@ -1,8 +1,8 @@
+#include "arg.h"
 #include "common.h"
-
 #include "console.h"
+#include "sampling.h"
 #include "llama.h"
-#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -34,6 +34,7 @@
 
 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;
@@ -81,7 +82,7 @@ static void write_logfile(
     yaml_dump_string_multiline(logfile, "output", output.c_str());
     yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -93,7 +94,7 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             printf("\n");
-            llama_print_timings(*g_ctx);
+            gpt_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
             _exit(130);
         }
@@ -103,14 +104,14 @@ 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)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
         return 1;
     }
 
+    auto & sparams = params.sparams;
+
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("infill", "log"));
     LOG_TEE("Log start\n");
@@ -156,26 +157,19 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
+    print_build_info();
 
     LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    gpt_sampler  * smpl = nullptr;
 
     g_model = &model;
     g_ctx = &ctx;
+    g_smpl = &smpl;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
@@ -305,16 +299,14 @@ int main(int argc, char ** argv) {
             LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
+    smpl = gpt_sampler_init(model, sparams);
+
+    LOG_TEE("sampling seed: %u\n", gpt_sampler_get_seed(smpl));
+    LOG_TEE("sampling: \n%s\n", sparams.print().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_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) {
@@ -349,8 +341,6 @@ 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()) {
@@ -421,11 +411,11 @@ int main(int argc, char ** argv) {
         embd.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, nullptr);
+            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            // LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
 
@@ -444,7 +434,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                gpt_sampler_accept(smpl, embd_inp[n_consumed], false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -476,7 +466,7 @@ 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 ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
+            if ((gpt_sampler_last(smpl) == llama_token_eot(model) || is_interacting) && params.interactive){
                 if (is_interacting && !params.interactive_first) {
                     // print an eot token
                     printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
@@ -542,7 +532,7 @@ int main(int argc, char ** argv) {
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            else if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -615,7 +605,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    gpt_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -638,13 +628,14 @@ int main(int argc, char ** argv) {
         fflush(stdout);
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     llama_backend_free();
 
 #ifndef LOG_DISABLE_LOGS

examples/llama-bench/README.md

@@ -14,7 +14,8 @@ Performance testing tool for llama.cpp.
     1. [Markdown](#markdown)
     2. [CSV](#csv)
     3. [JSON](#json)
-    4. [SQL](#sql)
+    4. [JSONL](#jsonl)
+    5. [SQL](#sql)
 
 ## Syntax
 
@@ -23,27 +24,34 @@ 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: 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)
+  -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)
 
 Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.
 ```
@@ -238,6 +246,19 @@ $ ./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

@@ -16,6 +16,7 @@
 #include <sstream>
 #include <string>
 #include <vector>
+#include <thread>
 
 #include "ggml.h"
 #include "llama.h"
@@ -123,6 +124,9 @@ static std::string get_cpu_info() {
                         (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
@@ -170,13 +174,14 @@ static std::string get_gpu_info() {
 }
 
 // command line params
-enum output_formats {NONE, CSV, JSON, MARKDOWN, SQL};
+enum output_formats {NONE, CSV, JSON, JSONL, 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");
@@ -190,6 +195,8 @@ 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") {
@@ -225,6 +232,9 @@ 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;
@@ -236,7 +246,10 @@ 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;
 };
@@ -251,6 +264,9 @@ 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},
@@ -262,7 +278,10 @@ 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,
 };
@@ -272,29 +291,37 @@ 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("  -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("  -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("\n");
     printf("Multiple values can be given for each parameter by separating them with ',' or by specifying the parameter multiple times.\n");
 }
@@ -338,6 +365,9 @@ 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];
@@ -433,6 +463,27 @@ 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;
@@ -440,12 +491,14 @@ 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;
@@ -541,6 +594,18 @@ 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;
@@ -555,6 +620,8 @@ 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;
@@ -585,6 +652,9 @@ 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;
 }
@@ -598,6 +668,9 @@ 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;
@@ -667,7 +740,10 @@ 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 & 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 & n_prompt : params.n_prompt) {
             if (n_prompt == 0) {
                 continue;
@@ -681,6 +757,9 @@ 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,
@@ -707,6 +786,9 @@ 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,
@@ -733,6 +815,9 @@ 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,
@@ -769,6 +854,9 @@ 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;
@@ -795,6 +883,9 @@ 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;
@@ -872,13 +963,14 @@ struct test {
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
             "n_batch", "n_ubatch",
-            "n_threads", "type_k", "type_v",
+            "n_threads", "cpu_mask", "cpu_strict", "poll",
+            "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;
     }
@@ -887,7 +979,7 @@ 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 == "n_threads" || field == "poll" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -896,6 +988,7 @@ 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;
         }
@@ -928,7 +1021,8 @@ 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), ggml_type_name(type_k), ggml_type_name(type_v),
+            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_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),
@@ -996,38 +1090,39 @@ 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;
@@ -1036,7 +1131,7 @@ 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_value(fields.at(i), values.at(i)).c_str());
+            fprintf(fout, "    \"%s\": %s,\n", fields.at(i).c_str(), format_json_value(fields.at(i), values.at(i)).c_str());
         }
     }
 
@@ -1059,6 +1154,25 @@ 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;
 
@@ -1067,7 +1181,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 16;
+            return 20;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1149,6 +1263,15 @@ 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");
         }
@@ -1350,6 +1473,8 @@ 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:
@@ -1383,6 +1508,8 @@ 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);
@@ -1402,7 +1529,13 @@ 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) {
@@ -1428,12 +1561,40 @@ 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);
         }
 
@@ -1443,9 +1604,15 @@ 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);
             }
 
@@ -1463,9 +1630,11 @@ int main(int argc, char ** argv) {
             fflush(p_err->fout);
         }
 
-        llama_print_timings(ctx);
+        llama_perf_context_print(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.seed  = 1234;
-    ctx_params.n_ctx = 2048;
+
+    ctx_params.n_ctx           = 2048;
     ctx_params.n_threads       = n_threads;
     ctx_params.n_threads_batch = n_threads;
 
@@ -269,12 +269,6 @@ 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) {
@@ -311,6 +305,29 @@ 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) {
@@ -381,31 +398,21 @@ 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 batch   = reinterpret_cast<llama_batch   *>(batch_pointer);
+    const auto sampler = reinterpret_cast<llama_sampler *>(sampler_pointer);
     const auto model = llama_get_model(context);
 
     if (!la_int_var) la_int_var = env->GetObjectClass(intvar_ncur);
     if (!la_int_var_value) la_int_var_value = env->GetMethodID(la_int_var, "getValue", "()I");
     if (!la_int_var_inc) la_int_var_inc = env->GetMethodID(la_int_var, "inc", "()V");
 
-    auto n_vocab = llama_n_vocab(model);
-    auto logits = llama_get_logits_ith(context, batch->n_tokens - 1);
-
-    std::vector<llama_token_data> candidates;
-    candidates.reserve(n_vocab);
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
-    }
-
-    llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
-
     // sample the most likely token
-    const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
+    const auto new_token_id = llama_sampler_sample(sampler, context, -1);
 
     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) {

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

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

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -24,6 +24,7 @@ func llama_batch_add(_ batch: inout llama_batch, _ id: llama_token, _ pos: llama
 actor LlamaContext {
     private var model: OpaquePointer
     private var context: OpaquePointer
+    private var sampling: UnsafeMutablePointer<llama_sampler>
     private var batch: llama_batch
     private var tokens_list: [llama_token]
     var is_done: Bool = false
@@ -42,9 +43,15 @@ actor LlamaContext {
         self.tokens_list = []
         self.batch = llama_batch_init(512, 0, 1)
         self.temporary_invalid_cchars = []
+        let sparams = llama_sampler_chain_default_params()
+        self.sampling = llama_sampler_chain_init(sparams)
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
+        llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 
     deinit {
+        llama_sampler_free(sampling)
         llama_batch_free(batch)
         llama_free(context)
         llama_free_model(model)
@@ -69,10 +76,9 @@ actor LlamaContext {
         print("Using \(n_threads) threads")
 
         var ctx_params = llama_context_default_params()
-        ctx_params.seed  = 1234
         ctx_params.n_ctx = 2048
-        ctx_params.n_threads       = UInt32(n_threads)
-        ctx_params.n_threads_batch = UInt32(n_threads)
+        ctx_params.n_threads       = Int32(n_threads)
+        ctx_params.n_threads_batch = Int32(n_threads)
 
         let context = llama_new_context_with_model(model, ctx_params)
         guard let context else {
@@ -144,20 +150,7 @@ actor LlamaContext {
     func completion_loop() -> String {
         var new_token_id: llama_token = 0
 
-        let n_vocab = llama_n_vocab(model)
-        let logits = llama_get_logits_ith(context, batch.n_tokens - 1)
-
-        var candidates = Array<llama_token_data>()
-        candidates.reserveCapacity(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))
-        }
-        candidates.withUnsafeMutableBufferPointer() { buffer in
-            var candidates_p = llama_token_data_array(data: buffer.baseAddress, size: buffer.count, sorted: false)
-
-            new_token_id = llama_sample_token_greedy(context, &candidates_p)
-        }
+        new_token_id = llama_sampler_sample(sampling, context, batch.n_tokens - 1)
 
         if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             print("\n")

examples/llava/MobileVLM-README.md

@@ -39,7 +39,7 @@ python ./examples/llava/llava_surgery.py -m path/to/MobileVLM-1.7B
 3. Use `convert_image_encoder_to_gguf.py` with `--projector-type ldp` (for **V2** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B/llava.projector \
     --output-dir path/to/MobileVLM-1.7B \
@@ -47,7 +47,7 @@ python ./examples/llava/convert_image_encoder_to_gguf \
 ```
 
 ```sh
-python ./examples/llava/convert_image_encoder_to_gguf \
+python ./examples/llava/convert_image_encoder_to_gguf.py \
     -m path/to/clip-vit-large-patch14-336 \
     --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
     --output-dir path/to/MobileVLM-1.7B_V2 \
@@ -57,12 +57,12 @@ python ./examples/llava/convert_image_encoder_to_gguf \
 4. Use `examples/convert_legacy_llama.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B
+python ./examples/convert_legacy_llama.py path/to/MobileVLM-1.7B --skip-unknown
 ```
 
-5. Use `quantize` to convert LLaMA part's DataType from `fp16` to `q4_k`
+5. Use `quantize` to convert LLaMA part's DataType from `fp32` to `q4_k`
 ```sh
-./llama-quantize path/to/MobileVLM-1.7B/ggml-model-f16.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
+./llama-quantize path/to/MobileVLM-1.7B/ggml-model-F32.gguf path/to/MobileVLM-1.7B/ggml-model-q4_k.gguf q4_k_s
 ```
 
 Now both the LLaMA part and the image encoder is in the `MobileVLM-1.7B` directory.

examples/llava/README-minicpmv2.5.md

@@ -15,8 +15,8 @@ cd llama.cpp
 Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
 
 ```bash
-python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
-python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 2
 python ./convert_hf_to_gguf.py ../MiniCPM-Llama3-V-2_5/model
 
 # quantize int4 version

examples/llava/clip.cpp

@@ -20,6 +20,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -212,13 +216,19 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
 
 static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
     if (search.empty()) {
-        return; // Avoid infinite loop if 'search' is an empty string
+        return;
     }
+    std::string builder;
+    builder.reserve(s.length());
     size_t pos = 0;
-    while ((pos = s.find(search, pos)) != std::string::npos) {
-        s.replace(pos, search.length(), replace);
-        pos += replace.length();
+    size_t last_pos = 0;
+    while ((pos = s.find(search, last_pos)) != std::string::npos) {
+        builder.append(s, last_pos, pos - last_pos);
+        builder.append(replace);
+        last_pos = pos + search.length();
     }
+    builder.append(s, last_pos, std::string::npos);
+    s = std::move(builder);
 }
 
 static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@@ -1108,7 +1118,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx;
+    clip_ctx * new_clip = new clip_ctx{};
 
     // update projector type
     {
@@ -1142,6 +1152,10 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     LOG_TEE("%s: CLIP using CANN backend\n", __func__);
 #endif
 
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_TEE("%s: CLIP using Vulkan backend\n", __func__);
+#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
@@ -1609,7 +1623,7 @@ static void normalize_image_u8_to_f32(const clip_image_u8* src, clip_image_f32*
     }
 }
 
-inline float clip(float x, float lower, float upper) {
+inline int clip(int x, int lower, int upper) {
     return std::max(lower, std::min(x, upper));
 }
 
@@ -1813,10 +1827,6 @@ static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size
     return refine_size;
 }
 
-inline int clip(int x, int lower, int upper) {
-    return std::max(lower, std::min(x, upper));
-}
-
 static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
     std::vector<int> candidate_split_grids_nums;
     for (int i : {multiple - 1, multiple, multiple + 1}) {
@@ -2439,7 +2449,7 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor
-    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = ggml_graph_node(gf, -1);
 
     // copy the embeddings to the location passed by the user
     ggml_backend_tensor_get(embeddings, vec, 0, ggml_nbytes(embeddings));

examples/llava/llava-cli.cpp

@@ -1,11 +1,12 @@
-#include "ggml.h"
+#include "arg.h"
+#include "base64.hpp"
 #include "log.h"
 #include "common.h"
+#include "sampling.h"
 #include "clip.h"
 #include "llava.h"
 #include "llama.h"
-
-#include "base64.hpp"
+#include "ggml.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -40,11 +41,11 @@ static bool eval_string(struct llama_context * ctx_llama, const char* str, int n
     return true;
 }
 
-static const char * sample(struct llama_sampling_context * ctx_sampling,
+static const char * sample(struct gpt_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
-    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
+    gpt_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
@@ -112,9 +113,7 @@ struct llava_context {
     struct llama_model * model = NULL;
 };
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\n example usage:\n");
     LOG_TEE("\n     %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
     LOG_TEE("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
@@ -129,14 +128,14 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
         if (!params->image.empty()) {
             LOG_TEE("using base64 encoded image instead of command line image path\n");
         }
-        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
+        embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
         if (!embed) {
             LOG_TEE("%s: can't load image from prompt\n", __func__);
             return NULL;
         }
         params->prompt = remove_image_from_prompt(prompt);
     } else {
-        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, fname.c_str());
+        embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
         if (!embed) {
             fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
             return NULL;
@@ -191,15 +190,15 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     LOG_TEE("\n");
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    if (!ctx_sampling) {
+    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    if (!smpl) {
         fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
         exit(1);
     }
 
     std::string response = "";
     for (int i = 0; i < max_tgt_len; i++) {
-        const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+        const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
         response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
@@ -211,7 +210,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         fflush(stdout);
     }
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     printf("\n");
 }
 
@@ -280,8 +279,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
         return 1;
     }
 
@@ -293,7 +291,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
-        print_usage(argc, argv, {});
+        print_usage(argc, argv);
         return 1;
     }
     auto model = llava_init(&params);
@@ -310,7 +308,7 @@ int main(int argc, char ** argv) {
         // process the prompt
         process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-        llama_print_timings(ctx_llava->ctx_llama);
+        llama_perf_context_print(ctx_llava->ctx_llama);
         llava_image_embed_free(image_embed);
         ctx_llava->model = NULL;
         llava_free(ctx_llava);
@@ -327,7 +325,7 @@ int main(int argc, char ** argv) {
             // process the prompt
             process_prompt(ctx_llava, image_embed, &params, params.prompt);
 
-            llama_print_timings(ctx_llava->ctx_llama);
+            llama_perf_context_print(ctx_llava->ctx_llama);
             llava_image_embed_free(image_embed);
             ctx_llava->model = NULL;
             llava_free(ctx_llava);

examples/llava/llava.cpp

@@ -184,7 +184,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
     ggml_build_forward_expand(gf, flatten);
     ggml_graph_compute_with_ctx(model.ctx, gf, 1);
-    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor* result = ggml_graph_node(gf, -1);
 
     memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
     // append without newline tokens (default behavior in llava_arch when not using unpad ):

examples/llava/minicpmv-cli.cpp

@@ -1,9 +1,11 @@
-#include "ggml.h"
+#include "arg.h"
 #include "log.h"
 #include "common.h"
+#include "sampling.h"
 #include "clip.h"
 #include "llava.h"
 #include "llama.h"
+#include "ggml.h"
 
 #include <cstdio>
 #include <cstdlib>
@@ -16,8 +18,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    LOG_TEE("\nexample usage:\n\n%s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
@@ -163,11 +165,11 @@ static void process_image(struct llava_context * ctx_llava, struct llava_image_e
     LOG_TEE("%s: image token past: %d\n", __func__, n_past);
 }
 
-static const char * sample(struct llama_sampling_context * ctx_sampling,
+static const char * sample(struct gpt_sampler * smpl,
                            struct llama_context * ctx_llama,
                            int * n_past) {
-    const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
-    llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
+    const llama_token id = gpt_sampler_sample(smpl, ctx_llama, -1);
+    gpt_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
@@ -180,7 +182,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
 
 static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
     auto ctx_clip = clip_init_context(params);
-    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
+    auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->cpuparams.n_threads, fname.c_str());
     if (!embeds) {
         std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
         return NULL;
@@ -214,7 +216,7 @@ static struct llava_context * minicpmv_init(gpt_params * params, const std::stri
     return ctx_llava;
 }
 
-static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
+static struct gpt_sampler * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
     std::string user_prompt = prompt;
     int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
     if (!is_first) {
@@ -238,13 +240,13 @@ static struct llama_sampling_context * llama_init(struct llava_context * ctx_lla
 
     LOG_TEE("\n");
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    return ctx_sampling;
+    struct gpt_sampler * smpl = gpt_sampler_init(ctx_llava->model, params->sparams);
+    return smpl;
 }
 
-static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
+static const char * llama_loop(struct llava_context * ctx_llava,struct gpt_sampler * smpl, int &n_past){
 
-    const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
+    const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
     return tmp;
 }
 
@@ -253,8 +255,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        show_additional_info(argc, argv);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
         return 1;
     }
 
@@ -266,7 +267,6 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     if (params.mmproj.empty() || (params.image.empty())) {
-        gpt_params_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
         return 1;
     }
@@ -278,12 +278,12 @@ int main(int argc, char ** argv) {
         if (!params.prompt.empty()) {
             LOG_TEE("<user>%s\n", params.prompt.c_str());
             LOG_TEE("<assistant>");
-            auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
+            auto smpl = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
             const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
             std::string response = "";
             bool have_tmp = false;
             for (int i = 0; i < max_tgt_len; i++) {
-                auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                auto tmp = llama_loop(ctx_llava, smpl, n_past);
                 response += tmp;
                 if (strcmp(tmp, "</s>") == 0){
                     if(!have_tmp)continue;
@@ -296,18 +296,18 @@ int main(int argc, char ** argv) {
 
                 fflush(stdout);
             }
-            llama_sampling_free(ctx_sampling);
+            gpt_sampler_free(smpl);
         }else {
             while (true) {
                 LOG_TEE("<user>");
                 std::string prompt;
                 std::getline(std::cin, prompt);
                 LOG_TEE("<assistant>");
-                auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
+                auto smpl = llama_init(ctx_llava, &params, prompt, n_past, true);
                 const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
                 std::string response = "";
                 for (int i = 0; i < max_tgt_len; i++) {
-                    auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
+                    auto tmp = llama_loop(ctx_llava, smpl, n_past);
                     response += tmp;
                     if (strcmp(tmp, "</s>") == 0) break;
                     if (strstr(tmp, "###")) break; // Yi-VL behavior
@@ -315,11 +315,11 @@ int main(int argc, char ** argv) {
                     if (strstr(response.c_str(), "<user>")) break; // minicpm-v
                     fflush(stdout);
                 }
-                llama_sampling_free(ctx_sampling);
+                gpt_sampler_free(smpl);
             }
         }
         printf("\n");
-        llama_print_timings(ctx_llava->ctx_llama);
+        llama_perf_context_print(ctx_llava->ctx_llama);
 
         ctx_llava->model = NULL;
         llava_free(ctx_llava);

examples/lookahead/lookahead.cpp

@@ -1,7 +1,8 @@
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
@@ -37,8 +38,7 @@ struct ngram_container {
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -118,7 +118,7 @@ int main(int argc, char ** argv) {
     llama_batch batch = llama_batch_init(params.n_ctx, 0, W + G + 1);
 
     // target model sampling context
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
 
     // verification n-grams
     std::vector<ngram_data> ngrams_cur(G);
@@ -159,9 +159,9 @@ int main(int argc, char ** argv) {
 
     // sample first token
     {
-        id = llama_sampling_sample(ctx_sampling, ctx, NULL, 0);
+        id = gpt_sampler_sample(smpl, ctx, 0);
 
-        llama_sampling_accept(ctx_sampling, ctx, id, true);
+        gpt_sampler_accept(smpl, id, true);
 
         {
             const std::string token_str = llama_token_to_piece(ctx, id);
@@ -284,9 +284,9 @@ int main(int argc, char ** argv) {
             }
 
             // sample the next token
-            id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_batch);
+            id = gpt_sampler_sample(smpl, ctx, i_batch);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
             // print
             {
@@ -361,7 +361,7 @@ int main(int argc, char ** argv) {
                 if (v == 0) {
                     // sample from the last level
                     for (int i = 0; i < W; i++) {
-                        tokens_j[N - 2][i] = llama_sampling_sample(ctx_sampling, ctx, NULL, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
+                        tokens_j[N - 2][i] = gpt_sampler_sample(smpl, ctx, ngrams_cur.size()*(N-1) + W*(N - 2) + i);
                     }
                 } else {
                     for (int i = 0; i < W; i++) {
@@ -468,10 +468,12 @@ int main(int argc, char ** argv) {
     LOG_TEE("n_predict = %d\n", n_predict);
     LOG_TEE("n_accept  = %d\n", n_accept);
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
+
+    gpt_sampler_free(smpl);
 
     llama_kv_cache_view_free(&kvc_view);
-    llama_sampling_free(ctx_sampling);
 
     llama_batch_free(batch);
 

examples/lookup/lookup-create.cpp

@@ -1,7 +1,8 @@
-#include "ggml.h"
-#include "llama.h"
+#include "arg.h"
 #include "common.h"
 #include "ngram-cache.h"
+#include "ggml.h"
+#include "llama.h"
 
 #include <cstdint>
 #include <fstream>
@@ -13,8 +14,7 @@
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
@@ -40,4 +40,6 @@ int main(int argc, char ** argv){
     fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
 
     llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+
+    return 0;
 }

examples/lookup/lookup-stats.cpp

@@ -1,8 +1,9 @@
-#include "ggml.h"
+#include "arg.h"
 #include "common.h"
-#include "llama.h"
 #include "log.h"
 #include "ngram-cache.h"
+#include "llama.h"
+#include "ggml.h"
 
 #include <cmath>
 #include <cstdint>
@@ -15,8 +16,7 @@
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 

examples/lookup/lookup.cpp

@@ -1,21 +1,20 @@
+#include "arg.h"
 #include "ggml.h"
-#include "llama.h"
 #include "common.h"
 #include "ngram-cache.h"
+#include "sampling.h"
+#include "llama.h"
 
-#include <cmath>
 #include <cstdint>
 #include <cstdio>
 #include <fstream>
 #include <string>
 #include <vector>
-#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_LOOKUP)) {
         return 1;
     }
 
@@ -106,7 +105,7 @@ int main(int argc, char ** argv){
 
     bool has_eos = false;
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    struct gpt_sampler * smpl = gpt_sampler_init(model, params.sparams);
 
     std::vector<llama_token> draft;
 
@@ -130,9 +129,9 @@ int main(int argc, char ** argv){
         int i_dft = 0;
         while (true) {
             // sample from the target model
-            llama_token id = llama_sampling_sample(ctx_sampling, ctx, NULL, i_dft);
+            llama_token id = gpt_sampler_sample(smpl, ctx, i_dft);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            gpt_sampler_accept(smpl, id, true);
 
             const std::string token_str = llama_token_to_piece(ctx, id);
 
@@ -240,10 +239,11 @@ int main(int argc, char ** argv){
     LOG_TEE("n_accept     = %d\n", n_accept);
     LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
-    LOG_TEE("\ntarget:\n");
-    llama_print_timings(ctx);
+    LOG_TEE("\ntarget:\n\n");
+    gpt_perf_print(ctx, smpl);
+
+    gpt_sampler_free(smpl);
 
-    llama_sampling_free(ctx_sampling);
     llama_batch_free(batch_tgt);
 
     llama_free(ctx);

examples/main/main.cpp

@@ -1,6 +1,7 @@
+#include "arg.h"
 #include "common.h"
-
 #include "console.h"
+#include "sampling.h"
 #include "llama.h"
 
 #include <cassert>
@@ -33,6 +34,7 @@
 
 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;
@@ -40,6 +42,13 @@ static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting  = false;
 static bool need_insert_eot = false;
 
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n  text generation:     %s -m your_model.gguf -p \"I believe the meaning of life is\" -n 128\n", argv[0]);
+    printf("\n  chat (conversation): %s -m your_model.gguf -p \"You are a helpful assistant\" -cnv\n", argv[0]);
+    printf("\n");
+}
+
 static bool file_exists(const std::string & path) {
     std::ifstream f(path.c_str());
     return f.good();
@@ -92,7 +101,7 @@ static void write_logfile(
     yaml_dump_string_multiline(logfile, "output", output.c_str());
     yaml_dump_vector_int(logfile, "output_tokens", output_tokens);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -105,7 +114,7 @@ static void sigint_handler(int signo) {
         } else {
             console::cleanup();
             printf("\n");
-            llama_print_timings(*g_ctx);
+            gpt_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
             _exit(130);
         }
@@ -121,8 +130,7 @@ static void llama_log_callback_logTee(ggml_log_level level, const char * text, v
 
 static std::string chat_add_and_format(struct llama_model * model, std::vector<llama_chat_msg> & chat_msgs, std::string role, std::string content) {
     llama_chat_msg new_msg{role, content};
-    auto formatted = llama_chat_format_single(
-        model, g_params->chat_template, chat_msgs, new_msg, role == "user");
+    auto formatted = llama_chat_format_single(model, g_params->chat_template, chat_msgs, new_msg, role == "user");
     chat_msgs.push_back({role, content});
     LOG("formatted: %s\n", formatted.c_str());
     return formatted;
@@ -131,13 +139,11 @@ static std::string chat_add_and_format(struct llama_model * model, std::vector<l
 int main(int argc, char ** argv) {
     gpt_params params;
     g_params = &params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_MAIN, print_usage)) {
         return 1;
     }
 
-    llama_sampling_params & sparams = params.sparams;
+    auto & sparams = params.sparams;
 
 #ifndef LOG_DISABLE_LOGS
     log_set_target(log_filename_generator("main", "log"));
@@ -183,27 +189,21 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
     }
 
-    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    LOG_TEE("%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
+    print_build_info();
 
     LOG("%s: llama backend init\n", __func__);
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model * model;
-    llama_context * ctx;
-    llama_context * ctx_guidance = NULL;
+    llama_model * model = nullptr;
+    llama_context * ctx = nullptr;
+    gpt_sampler * smpl = nullptr;
+
     std::vector<llama_chat_msg> chat_msgs;
+
     g_model = &model;
     g_ctx = &ctx;
+    g_smpl = &smpl;
 
     // load the model and apply lora adapter, if any
     LOG("%s: load the model and apply lora adapter, if any\n", __func__);
@@ -211,16 +211,43 @@ int main(int argc, char ** argv) {
 
     model = llama_init.model;
     ctx = llama_init.context;
-    if (sparams.cfg_scale > 1.f) {
-        struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
-        ctx_guidance = llama_new_context_with_model(model, lparams);
-    }
 
     if (model == NULL) {
         LOG_TEE("%s: error: unable to load model\n", __func__);
         return 1;
     }
 
+    LOG("%s: llama threadpool init = n_threads = %d\n",
+        __func__,
+        (int) params.cpuparams.n_threads
+    );
+    struct ggml_threadpool_params tpp_batch =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams_batch);
+    struct ggml_threadpool_params tpp =
+            ggml_threadpool_params_from_cpu_params(params.cpuparams);
+
+    set_process_priority(params.cpuparams.priority);
+
+    struct ggml_threadpool * threadpool_batch = NULL;
+    if (!ggml_threadpool_params_match(&tpp, &tpp_batch)) {
+        threadpool_batch = ggml_threadpool_new(&tpp_batch);
+        if (!threadpool_batch) {
+            LOG_TEE("%s: batch threadpool create failed : n_threads %d\n", __func__, tpp_batch.n_threads);
+            exit(1);
+        }
+
+        // Start the non-batch threadpool in the paused state
+        tpp.paused = true;
+    }
+
+    struct ggml_threadpool * threadpool = ggml_threadpool_new(&tpp);
+    if (!threadpool) {
+        LOG_TEE("%s: threadpool create failed : n_threads %d\n", __func__, tpp.n_threads);
+        exit(1);
+    }
+
+    llama_attach_threadpool(ctx, threadpool, threadpool_batch);
+
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
     LOG("n_ctx: %d\n", n_ctx);
@@ -303,24 +330,6 @@ int main(int argc, char ** argv) {
     }
 
     // Tokenize negative prompt
-    std::vector<llama_token> guidance_inp;
-    int guidance_offset = 0;
-    int original_prompt_len = 0;
-    if (ctx_guidance) {
-        LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
-
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true, true);
-        LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
-
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true, true);
-        LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
-
-        original_prompt_len = original_inp.size();
-        guidance_offset = (int)guidance_inp.size() - original_prompt_len;
-        LOG("original_prompt_len: %s", log_tostr(original_prompt_len));
-        LOG("guidance_offset:     %s", log_tostr(guidance_offset));
-    }
-
     if ((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;
@@ -352,8 +361,8 @@ int main(int argc, char ** argv) {
     }
 
     LOGLN(
-            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu, embd_inp.size() %zu",
-            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size(), embd_inp.size());
+            "recalculate the cached logits (check): embd_inp.empty() %s, n_matching_session_tokens %zu, embd_inp.size() %zu, session_tokens.size() %zu",
+            log_tostr(embd_inp.empty()), n_matching_session_tokens, embd_inp.size(), session_tokens.size());
 
     // if we will use the cache for the full prompt without reaching the end of the cache, force
     // reevaluation of the last token to recalculate the cached logits
@@ -387,15 +396,6 @@ int main(int argc, char ** argv) {
             LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
-        if (ctx_guidance) {
-            LOG_TEE("\n");
-            LOG_TEE("%s: negative prompt: '%s'\n", __func__, sparams.cfg_negative_prompt.c_str());
-            LOG_TEE("%s: number of tokens in negative prompt = %zu\n", __func__, guidance_inp.size());
-            for (int i = 0; i < (int) guidance_inp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", guidance_inp[i], llama_token_to_piece(ctx, guidance_inp[i]).c_str());
-            }
-        }
-
         if (params.n_keep > add_bos) {
             LOG_TEE("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
@@ -461,8 +461,17 @@ int main(int argc, char ** argv) {
             }
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
-    LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
+
+    smpl = gpt_sampler_init(model, sparams);
+    if (!smpl) {
+        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+        exit(1);
+    }
+
+    LOG_TEE("sampling seed: %u\n", gpt_sampler_get_seed(smpl));
+    LOG_TEE("sampling params: \n%s\n", sparams.print().c_str());
+    LOG_TEE("sampler constr: \n%s\n", gpt_sampler_print(smpl).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);
 
     // group-attention state
@@ -509,7 +518,6 @@ int main(int argc, char ** argv) {
     int n_remain           = params.n_predict;
     int n_consumed         = 0;
     int n_session_consumed = 0;
-    int n_past_guidance    = 0;
 
     std::vector<int>   input_tokens;  g_input_tokens  = &input_tokens;
     std::vector<int>   output_tokens; g_output_tokens = &output_tokens;
@@ -521,7 +529,6 @@ int main(int argc, char ** argv) {
     display = params.display_prompt;
 
     std::vector<llama_token> embd;
-    std::vector<llama_token> embd_guidance;
 
     // tokenized antiprompts
     std::vector<std::vector<llama_token>> antiprompt_ids;
@@ -531,12 +538,6 @@ int main(int argc, char ** argv) {
         antiprompt_ids.emplace_back(::llama_tokenize(ctx, antiprompt, false, true));
     }
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(sparams);
-    if (!ctx_sampling) {
-        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
-        exit(1);
-    }
-
     if (llama_model_has_encoder(model)) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
@@ -578,7 +579,7 @@ int main(int argc, char ** argv) {
                 // if we run out of context:
                 // - take the n_keep first tokens from the original prompt (via n_past)
                 // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) >= n_ctx) {
+                if (n_past + (int) embd.size() >= n_ctx) {
                     if (params.n_predict == -2) {
                         LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                         break;
@@ -595,11 +596,7 @@ int main(int argc, char ** argv) {
 
                     n_past -= n_discard;
 
-                    if (ctx_guidance) {
-                        n_past_guidance -= n_discard;
-                    }
-
-                    LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+                    LOG("after swap: n_past = %d\n", n_past);
 
                     LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
 
@@ -652,46 +649,6 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            // evaluate tokens in batches
-            // embd is typically prepared beforehand to fit within a batch, but not always
-            if (ctx_guidance) {
-                int input_size = 0;
-                llama_token * input_buf = NULL;
-
-                if (n_past_guidance < (int) guidance_inp.size()) {
-                    // Guidance context should have the same data with these modifications:
-                    //
-                    // * Replace the initial prompt
-                    // * Shift everything by guidance_offset
-                    embd_guidance = guidance_inp;
-                    if (embd.begin() + original_prompt_len < embd.end()) {
-                        embd_guidance.insert(
-                            embd_guidance.end(),
-                            embd.begin() + original_prompt_len,
-                            embd.end()
-                        );
-                    }
-
-                    input_buf  = embd_guidance.data();
-                    input_size = embd_guidance.size();
-
-                    LOG("guidance context: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_guidance).c_str());
-                } else {
-                    input_buf  = embd.data();
-                    input_size = embd.size();
-                }
-
-                for (int i = 0; i < input_size; i += params.n_batch) {
-                    int n_eval = std::min(input_size - i, params.n_batch);
-                    if (llama_decode(ctx_guidance, llama_batch_get_one(input_buf + i, n_eval, n_past_guidance, 0))) {
-                        LOG_TEE("%s : failed to eval\n", __func__);
-                        return 1;
-                    }
-
-                    n_past_guidance += n_eval;
-                }
-            }
-
             for (int i = 0; i < (int) embd.size(); i += params.n_batch) {
                 int n_eval = (int) embd.size() - i;
                 if (n_eval > params.n_batch) {
@@ -721,7 +678,6 @@ int main(int argc, char ** argv) {
         }
 
         embd.clear();
-        embd_guidance.clear();
 
         if ((int) embd_inp.size() <= n_consumed && !is_interacting) {
             // optionally save the session on first sample (for faster prompt loading next time)
@@ -732,11 +688,11 @@ int main(int argc, char ** argv) {
                 LOG("saved session to %s\n", path_session.c_str());
             }
 
-            const llama_token id = llama_sampling_sample(ctx_sampling, ctx, ctx_guidance);
+            const llama_token id = gpt_sampler_sample(smpl, ctx, -1);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, /* apply_grammar= */ true);
+            gpt_sampler_accept(smpl, id, /* apply_grammar= */ true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            // LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
 
@@ -755,7 +711,7 @@ int main(int argc, char ** argv) {
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], /* apply_grammar= */ false);
+                gpt_sampler_accept(smpl, embd_inp[n_consumed], /* apply_grammar= */ false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -798,7 +754,7 @@ int main(int argc, char ** argv) {
             // check for reverse prompt in the last n_prev tokens
             if (!params.antiprompt.empty()) {
                 const int n_prev = 32;
-                const std::string last_output = llama_sampling_prev_str(ctx_sampling, ctx, n_prev);
+                const std::string last_output = gpt_sampler_prev_str(smpl, ctx, n_prev);
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
@@ -820,7 +776,7 @@ int main(int argc, char ** argv) {
                 }
 
                 // check for reverse prompt using special tokens
-                llama_token last_token = llama_sampling_last(ctx_sampling);
+                llama_token last_token = gpt_sampler_last(smpl);
                 for (std::vector<llama_token> ids : antiprompt_ids) {
                     if (ids.size() == 1 && last_token == ids[0]) {
                         if (params.interactive) {
@@ -837,7 +793,7 @@ int main(int argc, char ** argv) {
             }
 
             // deal with end of generation tokens in interactive mode
-            if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
+            if (llama_token_is_eog(model, gpt_sampler_last(smpl))) {
                 LOG("found an EOG token\n");
 
                 if (params.interactive) {
@@ -858,7 +814,7 @@ int main(int argc, char ** argv) {
 
             // if current token is not EOG, we add it to current assistant message
             if (params.conversation) {
-                auto id = llama_sampling_last(ctx_sampling);
+                const auto id = gpt_sampler_last(smpl);
                 assistant_ss << llama_token_to_piece(ctx, id, false);
             }
 
@@ -954,7 +910,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    gpt_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -979,16 +935,20 @@ int main(int argc, char ** argv) {
         llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    gpt_perf_print(ctx, smpl);
     write_logfile(ctx, params, model, input_tokens, output_ss.str(), output_tokens);
 
-    if (ctx_guidance) { llama_free(ctx_guidance); }
+    gpt_sampler_free(smpl);
+
     llama_free(ctx);
     llama_free_model(model);
 
-    llama_sampling_free(ctx_sampling);
     llama_backend_free();
 
+    ggml_threadpool_free(threadpool);
+    ggml_threadpool_free(threadpool_batch);
+
 #ifndef LOG_DISABLE_LOGS
     LOG_TEE("Log end\n");
 #endif // LOG_DISABLE_LOGS

examples/parallel/parallel.cpp

@@ -1,7 +1,9 @@
 // A basic application simulating a server with multiple clients.
 // The clients submit requests to the server and they are processed in parallel.
 
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
 #include <cmath>
@@ -50,8 +52,8 @@ static std::vector<std::string> k_prompts = {
 
 struct client {
     ~client() {
-        if (ctx_sampling) {
-            llama_sampling_free(ctx_sampling);
+        if (smpl) {
+            gpt_sampler_free(smpl);
         }
     }
 
@@ -72,7 +74,7 @@ struct client {
     std::string prompt;
     std::string response;
 
-    struct llama_sampling_context * ctx_sampling = nullptr;
+    struct gpt_sampler * smpl = nullptr;
 };
 
 static void print_date_time() {
@@ -100,8 +102,7 @@ int main(int argc, char ** argv) {
 
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PARALLEL)) {
         return 1;
     }
 
@@ -161,7 +162,7 @@ int main(int argc, char ** argv) {
     for (size_t i = 0; i < clients.size(); ++i) {
         auto & client = clients[i];
         client.id = i;
-        client.ctx_sampling = llama_sampling_init(params.sparams);
+        client.smpl = gpt_sampler_init(model, params.sparams);
     }
 
     std::vector<llama_token> tokens_system;
@@ -253,7 +254,7 @@ int main(int argc, char ** argv) {
                     client.prompt   = client.input + "\nAssistant:";
                     client.response = "";
 
-                    llama_sampling_reset(client.ctx_sampling);
+                    gpt_sampler_reset(client.smpl);
 
                     // do not prepend BOS because we have a system prompt!
                     std::vector<llama_token> tokens_prompt;
@@ -341,9 +342,9 @@ int main(int argc, char ** argv) {
                 //printf("client %d, seq %d, token %d, pos %d, batch %d\n",
                 //        client.id, client.seq_id, client.sampled, client.n_decoded, client.i_batch);
 
-                const llama_token id = llama_sampling_sample(client.ctx_sampling, ctx, NULL, client.i_batch - i);
+                const llama_token id = gpt_sampler_sample(client.smpl, ctx, client.i_batch - i);
 
-                llama_sampling_accept(client.ctx_sampling, ctx, id, true);
+                gpt_sampler_accept(client.smpl, id, true);
 
                 if (client.n_decoded == 1) {
                     // start measuring generation time after the first token to make sure all concurrent clients
@@ -371,7 +372,7 @@ int main(int argc, char ** argv) {
                     }
 
                     // delete only the generated part of the sequence, i.e. keep the system prompt in the cache
-                    llama_kv_cache_seq_rm(ctx, client.id + 1, -1, -1);
+                    llama_kv_cache_seq_rm(ctx,    client.id + 1, -1, -1);
                     llama_kv_cache_seq_cp(ctx, 0, client.id + 1, -1, -1);
 
                     const auto t_main_end = ggml_time_us();
@@ -413,7 +414,8 @@ int main(int argc, char ** argv) {
 
     LOG_TEE("\n");
 
-    llama_print_timings(ctx);
+    // TODO: print sampling/grammar timings for all clients
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 

examples/passkey/passkey.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -6,9 +7,7 @@
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf --junk 250 --pos 90 --keep 32 --grp-attn-n 2 [--seed 1234]\n", argv[0]);
     LOG_TEE("\n");
@@ -21,13 +20,10 @@ int main(int argc, char ** argv) {
     params.n_keep = 32;
     params.i_pos  = -1;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PASSKEY, print_usage)) {
         return 1;
     }
 
-    srand(params.seed == LLAMA_DEFAULT_SEED ? time(NULL) : params.seed);
-
     int n_junk = params.n_junk;
     int n_keep = params.n_keep;
     int n_grp  = params.grp_attn_n;
@@ -80,12 +76,17 @@ int main(int argc, char ** argv) {
     GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
 
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
-
     if (ctx == NULL) {
         fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
         return 1;
     }
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_greedy());
+
     // tokenize the prompt
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(ctx, params.prompt, true);
@@ -217,20 +218,7 @@ int main(int argc, char ** argv) {
     while (n_cur <= n_len) {
         // sample the next token
         {
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, 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 llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
 
             // is it an end of generation?
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
@@ -267,10 +255,13 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
+    llama_sampler_free(smpl);
+
     llama_batch_free(batch);
 
     llama_free(ctx);

examples/perplexity/perplexity.cpp

@@ -1,18 +1,19 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
+#include <array>
+#include <atomic>
 #include <cmath>
 #include <cstdio>
 #include <cstring>
 #include <ctime>
+#include <fstream>
+#include <mutex>
+#include <random>
 #include <sstream>
 #include <thread>
-#include <mutex>
-#include <atomic>
 #include <vector>
-#include <array>
-#include <fstream>
-#include <sstream>
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
@@ -76,7 +77,7 @@ static void write_logfile(
     fprintf(logfile, "ppl_value: %f\n", results.ppl_value);
     yaml_dump_vector_float(logfile, "probs", results.probs);
 
-    llama_dump_timing_info_yaml(logfile, ctx);
+    llama_perf_dump_yaml(logfile, ctx);
     fclose(logfile);
 }
 
@@ -1967,8 +1968,7 @@ int main(int argc, char ** argv) {
     params.n_ctx = 512;
     params.logits_all = true;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_PERPLEXITY)) {
         return 1;
     }
 
@@ -2007,14 +2007,6 @@ int main(int argc, char ** argv) {
 
     print_build_info();
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
@@ -2054,7 +2046,8 @@ int main(int argc, char ** argv) {
         results = perplexity(ctx, params, n_ctx);
     }
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
     write_logfile(ctx, params, model, results);
 
     llama_free(ctx);

examples/quantize-stats/quantize-stats.cpp

@@ -1,7 +1,7 @@
-#define LLAMA_API_INTERNAL
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
+#include "llama-impl.h"
 
 #include <algorithm>
 #include <cassert>
@@ -319,8 +319,7 @@ int main(int argc, char ** argv) {
         }
 
         auto cparams = llama_context_default_params();
-        cparams.n_ctx      = 256;
-        cparams.seed       = 1;
+        cparams.n_ctx = 256;
 
         ctx = llama_new_context_with_model(model, cparams);
 

examples/quantize/CMakeLists.txt

@@ -1,6 +1,6 @@
 set(TARGET llama-quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama common ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/quantize/README.md

@@ -54,6 +54,8 @@ As the models are currently fully loaded into memory, you will need adequate dis
 
 Several quantization methods are supported. They differ in the resulting model disk size and inference speed.
 
+The quantization formats `Q4_0_4_4`, `Q4_0_4_8` and `Q4_0_8_8` are block interleaved variants of the `Q4_0` format, providing a data layout that is better suited for specific implementations of optimized mulmat kernels. Since these formats differ only in data layout, they have the same quantized size as the `Q4_0` format.
+
 *(outdated)*
 
 | Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |

examples/quantize/quantize.cpp

@@ -26,6 +26,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "IQ2_M",    LLAMA_FTYPE_MOSTLY_IQ2_M,    " 2.7  bpw quantization",            },
     { "IQ1_S",    LLAMA_FTYPE_MOSTLY_IQ1_S,    " 1.56 bpw quantization",            },
     { "IQ1_M",    LLAMA_FTYPE_MOSTLY_IQ1_M,    " 1.75 bpw quantization",            },
+    { "TQ1_0",    LLAMA_FTYPE_MOSTLY_TQ1_0,    " 1.69 bpw ternarization",           },
+    { "TQ2_0",    LLAMA_FTYPE_MOSTLY_TQ2_0,    " 2.06 bpw ternarization",           },
     { "Q2_K",     LLAMA_FTYPE_MOSTLY_Q2_K,     " 2.96G, +3.5199 ppl @ Llama-3-8B",  },
     { "Q2_K_S",   LLAMA_FTYPE_MOSTLY_Q2_K_S,   " 2.96G, +3.1836 ppl @ Llama-3-8B",  },
     { "IQ3_XXS",  LLAMA_FTYPE_MOSTLY_IQ3_XXS,  " 3.06 bpw quantization",            },
@@ -104,7 +106,7 @@ static void usage(const char * executable) {
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
     printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
-    printf("  --keep-split: will generate quatized model in the same shards as input");
+    printf("  --keep-split: will generate quantized model in the same shards as input\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");

examples/retrieval/retrieval.cpp

@@ -1,12 +1,11 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <algorithm>
 #include <fstream>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .\n", argv[0]);
     LOG_TEE("\n");
@@ -113,8 +112,7 @@ 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)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_RETRIEVAL, print_usage)) {
         return 1;
     }
 
@@ -293,9 +291,11 @@ int main(int argc, char ** argv) {
         }
     }
 
+    LOG_TEE("\n");
+    llama_perf_context_print(ctx);
+
     // clean up
     llama_batch_free(query_batch);
-    llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
     llama_backend_free();

examples/rpc/README.md

@@ -10,20 +10,21 @@ This can be used for distributed LLM inference with `llama.cpp` in the following
 
 ```mermaid
 flowchart TD
-    rpcb---|TCP|srva
-    rpcb---|TCP|srvb
-    rpcb-.-|TCP|srvn
+    rpcb<-->|TCP|srva
+    rpcb<-->|TCP|srvb
+    rpcb<-.->|TCP|srvn
     subgraph hostn[Host N]
-    srvn[rpc-server]-.-backend3["Backend (CUDA,Metal,etc.)"]
+    srvn[rpc-server]<-.->backend3["Backend (CUDA,Metal,etc.)"]
     end
     subgraph hostb[Host B]
-    srvb[rpc-server]---backend2["Backend (CUDA,Metal,etc.)"]
+    srvb[rpc-server]<-->backend2["Backend (CUDA,Metal,etc.)"]
     end
     subgraph hosta[Host A]
-    srva[rpc-server]---backend["Backend (CUDA,Metal,etc.)"]
+    srva[rpc-server]<-->backend["Backend (CUDA,Metal,etc.)"]
     end
     subgraph host[Main Host]
-    ggml[llama.cpp]---rpcb[RPC backend]
+    local["Backend (CUDA,Metal,etc.)"]<-->ggml[llama-cli]
+    ggml[llama-cli]<-->rpcb[RPC backend]
     end
     style hostn stroke:#66,stroke-width:2px,stroke-dasharray: 5 5
 ```
@@ -62,17 +63,12 @@ $ CUDA_VISIBLE_DEVICES=0 bin/rpc-server -p 50052
 This way you can run multiple `rpc-server` instances on the same host, each with a different CUDA device.
 
 
-On the main host build `llama.cpp` only with `-DGGML_RPC=ON`:
-
-```bash
-mkdir build-rpc
-cd build-rpc
-cmake .. -DGGML_RPC=ON
-cmake --build . --config Release
-```
-
-Finally, use the `--rpc` option to specify the host and port of each `rpc-server`:
+On the main host build `llama.cpp` for the local backend and add `-DGGML_RPC=ON` to the build options.
+Finally, when running `llama-cli`, use the `--rpc` option to specify the host and port of each `rpc-server`:
 
 ```bash
 $ bin/llama-cli -m ../models/tinyllama-1b/ggml-model-f16.gguf -p "Hello, my name is" --repeat-penalty 1.0 -n 64 --rpc 192.168.88.10:50052,192.168.88.11:50052 -ngl 99
 ```
+
+This way you can offload model layers to both local and remote devices.
+

examples/save-load-state/save-load-state.cpp

@@ -1,17 +1,17 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
 #include <vector>
 #include <cstdio>
-#include <chrono>
 
 int main(int argc, char ** argv) {
     gpt_params params;
 
     params.prompt = "The quick brown fox";
+    params.sparams.seed = 1234;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
@@ -38,6 +38,13 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    auto sparams = llama_sampler_chain_default_params();
+
+    llama_sampler * smpl = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
+
     // tokenize prompt
     auto tokens = llama_tokenize(ctx, params.prompt, true);
 
@@ -64,16 +71,7 @@ int main(int argc, char ** argv) {
     printf("\nfirst run: %s", params.prompt.c_str());
 
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx);
-        auto n_vocab = llama_n_vocab(model);
-
-        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 };
-        auto next_token = llama_sample_token(ctx, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl, ctx, -1);
         auto next_token_str = llama_token_to_piece(ctx, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -96,6 +94,11 @@ int main(int argc, char ** argv) {
     // make new context
     auto * ctx2 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
 
+    llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
+
     printf("\nsecond run: %s", params.prompt.c_str());
 
     // load state (rng, logits, embedding and kv_cache) from file
@@ -124,15 +127,7 @@ int main(int argc, char ** argv) {
 
     // second run
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx2);
-        auto n_vocab = llama_n_vocab(model);
-        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 };
-        auto next_token = llama_sample_token(ctx2, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl2, ctx2, -1);
         auto next_token_str = llama_token_to_piece(ctx2, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -157,7 +152,12 @@ int main(int argc, char ** argv) {
     }
 
     // make new context
-    auto* ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+    auto * ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+
+    llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
+
+    llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
+    llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
 
@@ -215,15 +215,7 @@ int main(int argc, char ** argv) {
 
     // third run with seq 1 instead of 0
     for (auto i = 0; i < params.n_predict; i++) {
-        auto * logits = llama_get_logits(ctx3);
-        auto n_vocab = llama_n_vocab(model);
-        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 };
-        auto next_token = llama_sample_token(ctx3, &candidates_p);
+        auto next_token     = llama_sampler_sample(smpl3, ctx3, -1);
         auto next_token_str = llama_token_to_piece(ctx3, next_token);
 
         printf("%s", next_token_str.c_str());
@@ -240,6 +232,10 @@ int main(int argc, char ** argv) {
 
     printf("\n");
 
+    llama_sampler_free(smpl);
+    llama_sampler_free(smpl2);
+    llama_sampler_free(smpl3);
+
     llama_free(ctx3);
     llama_free_model(model);
 

examples/server/CMakeLists.txt

@@ -30,6 +30,7 @@ set(PUBLIC_ASSETS
     system-prompts.js
     prompt-formats.js
     json-schema-to-grammar.mjs
+    loading.html
 )
 
 foreach(asset ${PUBLIC_ASSETS})

examples/server/README.md

@@ -17,236 +17,145 @@ The project is under active development, and we are [looking for feedback and co
 
 ## Usage
 
+| Argument | Explanation |
+| -------- | ----------- |
+| `-h, --help, --usage` | print usage and exit |
+| `--version` | show version and build info |
+| `-v, --verbose` | print verbose information |
+| `--verbosity N` | set specific verbosity level (default: 0) |
+| `-t, --threads N` | number of threads to use during generation (default: -1)<br/>(env: LLAMA_ARG_THREADS) |
+| `-tb, --threads-batch N` | number of threads to use during batch and prompt processing (default: same as --threads) |
+| `-C, --cpu-mask M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range (default: "") |
+| `-Cr, --cpu-range lo-hi` | range of CPUs for affinity. Complements --cpu-mask |
+| `--cpu-strict <0\|1>` | use strict CPU placement (default: 0)<br/> |
+| `--prio N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
+| `--poll <0...100>` | use polling level to wait for work (0 - no polling, default: 50)<br/> |
+| `-Cb, --cpu-mask-batch M` | CPU affinity mask: arbitrarily long hex. Complements cpu-range-batch (default: same as --cpu-mask) |
+| `-Crb, --cpu-range-batch lo-hi` | ranges of CPUs for affinity. Complements --cpu-mask-batch |
+| `--cpu-strict-batch <0\|1>` | use strict CPU placement (default: same as --cpu-strict) |
+| `--prio-batch N` | set process/thread priority : 0-normal, 1-medium, 2-high, 3-realtime (default: 0)<br/> |
+| `--poll-batch <0\|1>` | use polling to wait for work (default: same as --poll) |
+| `-c, --ctx-size N` | size of the prompt context (default: 0, 0 = loaded from model)<br/>(env: LLAMA_ARG_CTX_SIZE) |
+| `-n, --predict, --n-predict N` | number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)<br/>(env: LLAMA_ARG_N_PREDICT) |
+| `-b, --batch-size N` | logical maximum batch size (default: 2048)<br/>(env: LLAMA_ARG_BATCH) |
+| `-ub, --ubatch-size N` | physical maximum batch size (default: 512)<br/>(env: LLAMA_ARG_UBATCH) |
+| `--keep N` | number of tokens to keep from the initial prompt (default: 0, -1 = all) |
+| `-fa, --flash-attn` | enable Flash Attention (default: disabled)<br/>(env: LLAMA_ARG_FLASH_ATTN) |
+| `-p, --prompt PROMPT` | prompt to start generation with |
+| `-f, --file FNAME` | a file containing the prompt (default: none) |
+| `-bf, --binary-file FNAME` | binary file containing the prompt (default: none) |
+| `-e, --escape` | process escapes sequences (\n, \r, \t, \', \", \\) (default: true) |
+| `--no-escape` | do not process escape sequences |
+| `--spm-infill` | use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
+| `-s, --seed SEED` | RNG seed (default: -1, use random seed for < 0) |
+| `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
+| `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
+| `--penalize-nl` | penalize newline tokens (default: false) |
+| `--temp N` | temperature (default: 0.8) |
+| `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
+| `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
+| `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
+| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) |
+| `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
+| `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
+| `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
+| `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
+| `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
+| `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
+| `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
+| `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
+| `--grammar GRAMMAR` | BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '') |
+| `--grammar-file FNAME` | file to read grammar from |
+| `-j, --json-schema SCHEMA` | JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object<br/>For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead |
+| `--rope-scaling {none,linear,yarn}` | RoPE frequency scaling method, defaults to linear unless specified by the model |
+| `--rope-scale N` | RoPE context scaling factor, expands context by a factor of N |
+| `--rope-freq-base N` | RoPE base frequency, used by NTK-aware scaling (default: loaded from model) |
+| `--rope-freq-scale N` | RoPE frequency scaling factor, expands context by a factor of 1/N |
+| `--yarn-orig-ctx N` | YaRN: original context size of model (default: 0 = model training context size) |
+| `--yarn-ext-factor N` | YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation) |
+| `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0) |
+| `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0) |
+| `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0) |
+| `-gan, --grp-attn-n N` | group-attention factor (default: 1) |
+| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0) |
+| `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
+| `-nkvo, --no-kv-offload` | disable KV offload |
+| `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16) |
+| `-ctv, --cache-type-v TYPE` | KV cache data type for V (default: f16) |
+| `-dt, --defrag-thold N` | KV cache defragmentation threshold (default: -1.0, < 0 - disabled)<br/>(env: LLAMA_ARG_DEFRAG_THOLD) |
+| `-np, --parallel N` | number of parallel sequences to decode (default: 1) |
+| `-cb, --cont-batching` | enable continuous batching (a.k.a dynamic batching) (default: enabled)<br/>(env: LLAMA_ARG_CONT_BATCHING) |
+| `-nocb, --no-cont-batching` | disable continuous batching<br/>(env: LLAMA_ARG_NO_CONT_BATCHING) |
+| `--mlock` | force system to keep model in RAM rather than swapping or compressing |
+| `--no-mmap` | do not memory-map model (slower load but may reduce pageouts if not using mlock) |
+| `--numa TYPE` | attempt optimizations that help on some NUMA systems<br/>- distribute: spread execution evenly over all nodes<br/>- isolate: only spawn threads on CPUs on the node that execution started on<br/>- numactl: use the CPU map provided by numactl<br/>if run without this previously, it is recommended to drop the system page cache before using this<br/>see https://github.com/ggerganov/llama.cpp/issues/1437 |
+| `-ngl, --gpu-layers, --n-gpu-layers N` | number of layers to store in VRAM<br/>(env: LLAMA_ARG_N_GPU_LAYERS) |
+| `-sm, --split-mode {none,layer,row}` | how to split the model across multiple GPUs, one of:<br/>- none: use one GPU only<br/>- layer (default): split layers and KV across GPUs<br/>- row: split rows across GPUs |
+| `-ts, --tensor-split N0,N1,N2,...` | fraction of the model to offload to each GPU, comma-separated list of proportions, e.g. 3,1 |
+| `-mg, --main-gpu INDEX` | the GPU to use for the model (with split-mode = none), or for intermediate results and KV (with split-mode = row) (default: 0) |
+| `--check-tensors` | check model tensor data for invalid values (default: false) |
+| `--override-kv KEY=TYPE:VALUE` | advanced option to override model metadata by key. may be specified multiple times.<br/>types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false |
+| `--lora FNAME` | path to LoRA adapter (can be repeated to use multiple adapters) |
+| `--lora-scaled FNAME SCALE` | path to LoRA adapter with user defined scaling (can be repeated to use multiple adapters) |
+| `--control-vector FNAME` | add a control vector<br/>note: this argument can be repeated to add multiple control vectors |
+| `--control-vector-scaled FNAME SCALE` | add a control vector with user defined scaling SCALE<br/>note: this argument can be repeated to add multiple scaled control vectors |
+| `--control-vector-layer-range START END` | layer range to apply the control vector(s) to, start and end inclusive |
+| `-a, --alias STRING` | set alias for model name (to be used by REST API) |
+| `-m, --model FNAME` | model path (default: `models/$filename` with filename from `--hf-file` or `--model-url` if set, otherwise models/7B/ggml-model-f16.gguf)<br/>(env: LLAMA_ARG_MODEL) |
+| `-mu, --model-url MODEL_URL` | model download url (default: unused)<br/>(env: LLAMA_ARG_MODEL_URL) |
+| `-hfr, --hf-repo REPO` | Hugging Face model repository (default: unused)<br/>(env: LLAMA_ARG_HF_REPO) |
+| `-hff, --hf-file FILE` | Hugging Face model file (default: unused)<br/>(env: LLAMA_ARG_HF_FILE) |
+| `-hft, --hf-token TOKEN` | Hugging Face access token (default: value from HF_TOKEN environment variable)<br/>(env: HF_TOKEN) |
+| `--host HOST` | ip address to listen (default: 127.0.0.1)<br/>(env: LLAMA_ARG_HOST) |
+| `--port PORT` | port to listen (default: 8080)<br/>(env: LLAMA_ARG_PORT) |
+| `--path PATH` | path to serve static files from (default: ) |
+| `--embedding, --embeddings` | restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)<br/>(env: LLAMA_ARG_EMBEDDINGS) |
+| `--api-key KEY` | API key to use for authentication (default: none)<br/>(env: LLAMA_API_KEY) |
+| `--api-key-file FNAME` | path to file containing API keys (default: none) |
+| `--ssl-key-file FNAME` | path to file a PEM-encoded SSL private key |
+| `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate |
+| `-to, --timeout N` | server read/write timeout in seconds (default: 600) |
+| `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
+| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
+| `--log-format {text, json}` | log output format: json or text (default: json) |
+| `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
+| `--no-slots` | disables slots monitoring endpoint (default: enabled)<br/>(env: LLAMA_ARG_NO_ENDPOINT_SLOTS) |
+| `--slot-save-path PATH` | path to save slot kv cache (default: disabled) |
+| `--chat-template JINJA_TEMPLATE` | set custom jinja chat template (default: template taken from model's metadata)<br/>if suffix/prefix are specified, template will be disabled<br/>only commonly used templates are accepted:<br/>https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template<br/>(env: LLAMA_ARG_CHAT_TEMPLATE) |
+| `-sps, --slot-prompt-similarity SIMILARITY` | how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)<br/> |
+| `--lora-init-without-apply` | load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled) |
+| `-ld, --logdir LOGDIR` | path under which to save YAML logs (no logging if unset) |
+| `--log-test` | Log test |
+| `--log-disable` | Log disable |
+| `--log-enable` | Log enable |
+| `--log-new` | Log new |
+| `--log-append` | Log append |
+| `--log-file FNAME` | Log file |
+
+Note: If both command line argument and environment variable are both set for the same param, the argument will take precedence over env var.
+
+Example usage of docker compose with environment variables:
+
+```yml
+services:
+  llamacpp-server:
+    image: ghcr.io/ggerganov/llama.cpp:server
+    ports:
+      - 8080:8080
+    volumes:
+      - ./models:/models
+    environment:
+      # alternatively, you can use "LLAMA_ARG_MODEL_URL" to download the model
+      LLAMA_ARG_MODEL: /models/my_model.gguf
+      LLAMA_ARG_CTX_SIZE: 4096
+      LLAMA_ARG_N_PARALLEL: 2
+      LLAMA_ARG_ENDPOINT_METRICS: 1
+      LLAMA_ARG_PORT: 8080
 ```
-usage: ./llama-server [options]
-
-general:
-
-  -h,    --help, --usage          print usage and exit
-         --version                show version and build info
-  -v,    --verbose                print verbose information
-         --verbosity N            set specific verbosity level (default: 0)
-         --verbose-prompt         print a verbose prompt before generation (default: false)
-         --no-display-prompt      don't print prompt at generation (default: false)
-  -co,   --color                  colorise output to distinguish prompt and user input from generations (default: false)
-  -s,    --seed SEED              RNG seed (default: -1, use random seed for < 0)
-  -t,    --threads N              number of threads to use during generation (default: 8)
-  -tb,   --threads-batch N        number of threads to use during batch and prompt processing (default: same as --threads)
-  -td,   --threads-draft N        number of threads to use during generation (default: same as --threads)
-  -tbd,  --threads-batch-draft N  number of threads to use during batch and prompt processing (default: same as --threads-draft)
-         --draft N                number of tokens to draft for speculative decoding (default: 5)
-  -ps,   --p-split N              speculative decoding split probability (default: 0.1)
-  -lcs,  --lookup-cache-static FNAME
-                                  path to static lookup cache to use for lookup decoding (not updated by generation)
-  -lcd,  --lookup-cache-dynamic FNAME
-                                  path to dynamic lookup cache to use for lookup decoding (updated by generation)
-  -c,    --ctx-size N             size of the prompt context (default: 0, 0 = loaded from model)
-  -n,    --predict N              number of tokens to predict (default: -1, -1 = infinity, -2 = until context filled)
-  -b,    --batch-size N           logical maximum batch size (default: 2048)
-  -ub,   --ubatch-size N          physical maximum batch size (default: 512)
-         --keep N                 number of tokens to keep from the initial prompt (default: 0, -1 = all)
-         --chunks N               max number of chunks to process (default: -1, -1 = all)
-  -fa,   --flash-attn             enable Flash Attention (default: disabled)
-  -p,    --prompt PROMPT          prompt to start generation with
-                                  in conversation mode, this will be used as system prompt
-                                  (default: '')
-  -f,    --file FNAME             a file containing the prompt (default: none)
-         --in-file FNAME          an input file (repeat to specify multiple files)
-  -bf,   --binary-file FNAME      binary file containing the prompt (default: none)
-  -e,    --escape                 process escapes sequences (\n, \r, \t, \', \", \\) (default: true)
-         --no-escape              do not process escape sequences
-  -ptc,  --print-token-count N    print token count every N tokens (default: -1)
-         --prompt-cache FNAME     file to cache prompt state for faster startup (default: none)
-         --prompt-cache-all       if specified, saves user input and generations to cache as well
-                                  not supported with --interactive or other interactive options
-         --prompt-cache-ro        if specified, uses the prompt cache but does not update it
-  -r,    --reverse-prompt PROMPT  halt generation at PROMPT, return control in interactive mode
-                                  can be specified more than once for multiple prompts
-  -sp,   --special                special tokens output enabled (default: false)
-  -cnv,  --conversation           run in conversation mode, does not print special tokens and suffix/prefix
-                                  if suffix/prefix are not specified, default chat template will be used
-                                  (default: false)
-  -i,    --interactive            run in interactive mode (default: false)
-  -if,   --interactive-first      run in interactive mode and wait for input right away (default: false)
-  -mli,  --multiline-input        allows you to write or paste multiple lines without ending each in '\'
-         --in-prefix-bos          prefix BOS to user inputs, preceding the `--in-prefix` string
-         --in-prefix STRING       string to prefix user inputs with (default: empty)
-         --in-suffix STRING       string to suffix after user inputs with (default: empty)
-         --spm-infill             use Suffix/Prefix/Middle pattern for infill (instead of Prefix/Suffix/Middle) as some models prefer this. (default: disabled)
-
-sampling:
-
-         --samplers SAMPLERS      samplers that will be used for generation in the order, separated by ';'
-                                  (default: top_k;tfs_z;typical_p;top_p;min_p;temperature)
-         --sampling-seq SEQUENCE  simplified sequence for samplers that will be used (default: kfypmt)
-         --ignore-eos             ignore end of stream token and continue generating (implies --logit-bias EOS-inf)
-         --penalize-nl            penalize newline tokens (default: false)
-         --temp N                 temperature (default: 0.8)
-         --top-k N                top-k sampling (default: 40, 0 = disabled)
-         --top-p N                top-p sampling (default: 0.9, 1.0 = disabled)
-         --min-p N                min-p sampling (default: 0.1, 0.0 = disabled)
-         --tfs N                  tail free sampling, parameter z (default: 1.0, 1.0 = disabled)
-         --typical N              locally typical sampling, parameter p (default: 1.0, 1.0 = disabled)
-         --repeat-last-n N        last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size)
-         --repeat-penalty N       penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled)
-         --presence-penalty N     repeat alpha presence penalty (default: 0.0, 0.0 = disabled)
-         --frequency-penalty N    repeat alpha frequency penalty (default: 0.0, 0.0 = disabled)
-         --dynatemp-range N       dynamic temperature range (default: 0.0, 0.0 = disabled)
-         --dynatemp-exp N         dynamic temperature exponent (default: 1.0)
-         --mirostat N             use Mirostat sampling.
-                                  Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.
-                                  (default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)
-         --mirostat-lr N          Mirostat learning rate, parameter eta (default: 0.1)
-         --mirostat-ent N         Mirostat target entropy, parameter tau (default: 5.0)
-         -l TOKEN_ID(+/-)BIAS     modifies the likelihood of token appearing in the completion,
-                                  i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',
-                                  or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'
-         --cfg-negative-prompt PROMPT
-                                  negative prompt to use for guidance (default: '')
-         --cfg-negative-prompt-file FNAME
-                                  negative prompt file to use for guidance
-         --cfg-scale N            strength of guidance (default: 1.0, 1.0 = disable)
-         --chat-template JINJA_TEMPLATE
-                                  set custom jinja chat template (default: template taken from model's metadata)
-                                  if suffix/prefix are specified, template will be disabled
-                                  only commonly used templates are accepted:
-                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-
-grammar:
-
-         --grammar GRAMMAR        BNF-like grammar to constrain generations (see samples in grammars/ dir) (default: '')
-         --grammar-file FNAME     file to read grammar from
-  -j,    --json-schema SCHEMA     JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object
-                                  For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead
-
-embedding:
-
-         --pooling {none,mean,cls,last}
-                                  pooling type for embeddings, use model default if unspecified
-         --attention {causal,non-causal}
-                                  attention type for embeddings, use model default if unspecified
-
-context hacking:
-
-         --rope-scaling {none,linear,yarn}
-                                  RoPE frequency scaling method, defaults to linear unless specified by the model
-         --rope-scale N           RoPE context scaling factor, expands context by a factor of N
-         --rope-freq-base N       RoPE base frequency, used by NTK-aware scaling (default: loaded from model)
-         --rope-freq-scale N      RoPE frequency scaling factor, expands context by a factor of 1/N
-         --yarn-orig-ctx N        YaRN: original context size of model (default: 0 = model training context size)
-         --yarn-ext-factor N      YaRN: extrapolation mix factor (default: -1.0, 0.0 = full interpolation)
-         --yarn-attn-factor N     YaRN: scale sqrt(t) or attention magnitude (default: 1.0)
-         --yarn-beta-slow N       YaRN: high correction dim or alpha (default: 1.0)
-         --yarn-beta-fast N       YaRN: low correction dim or beta (default: 32.0)
-  -gan,  --grp-attn-n N           group-attention factor (default: 1)
-  -gaw,  --grp-attn-w N           group-attention width (default: 512.0)
-  -dkvc, --dump-kv-cache          verbose print of the KV cache
-  -nkvo, --no-kv-offload          disable KV offload
-  -ctk,  --cache-type-k TYPE      KV cache data type for K (default: f16)
-  -ctv,  --cache-type-v TYPE      KV cache data type for V (default: f16)
-
-perplexity:
-
-         --all-logits             return logits for all tokens in the batch (default: false)
-         --hellaswag              compute HellaSwag score over random tasks from datafile supplied with -f
-         --hellaswag-tasks N      number of tasks to use when computing the HellaSwag score (default: 400)
-         --winogrande             compute Winogrande score over random tasks from datafile supplied with -f
-         --winogrande-tasks N     number of tasks to use when computing the Winogrande score (default: 0)
-         --multiple-choice        compute multiple choice score over random tasks from datafile supplied with -f
-         --multiple-choice-tasks N
-                                  number of tasks to use when computing the multiple choice score (default: 0)
-         --kl-divergence          computes KL-divergence to logits provided via --kl-divergence-base
-         --ppl-stride N           stride for perplexity calculation (default: 0)
-         --ppl-output-type {0,1}  output type for perplexity calculation (default: 0)
-
-parallel:
-
-  -dt,   --defrag-thold N         KV cache defragmentation threshold (default: -1.0, < 0 - disabled)
-  -np,   --parallel N             number of parallel sequences to decode (default: 1)
-  -ns,   --sequences N            number of sequences to decode (default: 1)
-  -cb,   --cont-batching          enable continuous batching (a.k.a dynamic batching) (default: enabled)
-
-multi-modality:
-
-         --mmproj FILE            path to a multimodal projector file for LLaVA. see examples/llava/README.md
-         --image FILE             path to an image file. use with multimodal models. Specify multiple times for batching
-
-backend:
-
-         --rpc SERVERS            comma separated list of RPC servers
-         --mlock                  force system to keep model in RAM rather than swapping or compressing
-         --no-mmap                do not memory-map model (slower load but may reduce pageouts if not using mlock)
-         --numa TYPE              attempt optimizations that help on some NUMA systems
-                                    - distribute: spread execution evenly over all nodes
-                                    - isolate: only spawn threads on CPUs on the node that execution started on
-                                    - numactl: use the CPU map provided by numactl
-                                  if run without this previously, it is recommended to drop the system page cache before using this
-                                  see https://github.com/ggerganov/llama.cpp/issues/1437
-
-model:
-
-         --check-tensors          check model tensor data for invalid values (default: false)
-         --override-kv KEY=TYPE:VALUE
-                                  advanced option to override model metadata by key. may be specified multiple times.
-                                  types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false
-         --lora FNAME             apply LoRA adapter (implies --no-mmap)
-         --lora-scaled FNAME S    apply LoRA adapter with user defined scaling S (implies --no-mmap)
-         --lora-base FNAME        optional model to use as a base for the layers modified by the LoRA adapter
-         --control-vector FNAME   add a control vector
-                                  note: this argument can be repeated to add multiple control vectors
-         --control-vector-scaled FNAME SCALE
-                                  add a control vector with user defined scaling SCALE
-                                  note: this argument can be repeated to add multiple scaled control vectors
-         --control-vector-layer-range START END
-                                  layer range to apply the control vector(s) to, start and end inclusive
-  -m,    --model FNAME            model path (default: models/$filename with filename from --hf-file
-                                  or --model-url if set, otherwise models/7B/ggml-model-f16.gguf)
-  -md,   --model-draft FNAME      draft model for speculative decoding (default: unused)
-  -mu,   --model-url MODEL_URL    model download url (default: unused)
-  -hfr,  --hf-repo REPO           Hugging Face model repository (default: unused)
-  -hff,  --hf-file FILE           Hugging Face model file (default: unused)
-  -hft,  --hf-token TOKEN         Hugging Face access token (default: value from HF_TOKEN environment variable)
-
-server:
-
-         --host HOST              ip address to listen (default: 127.0.0.1)
-         --port PORT              port to listen (default: 8080)
-         --path PATH              path to serve static files from (default: )
-         --embedding(s)           restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
-         --api-key KEY            API key to use for authentication (default: none)
-         --api-key-file FNAME     path to file containing API keys (default: none)
-         --ssl-key-file FNAME     path to file a PEM-encoded SSL private key
-         --ssl-cert-file FNAME    path to file a PEM-encoded SSL certificate
-         --timeout N              server read/write timeout in seconds (default: 600)
-         --threads-http N         number of threads used to process HTTP requests (default: -1)
-         --system-prompt-file FNAME
-                                  set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications
-         --log-format {text,json}
-                                  log output format: json or text (default: json)
-         --metrics                enable prometheus compatible metrics endpoint (default: disabled)
-         --no-slots               disables slots monitoring endpoint (default: enabled)
-         --slot-save-path PATH    path to save slot kv cache (default: disabled)
-         --chat-template JINJA_TEMPLATE
-                                  set custom jinja chat template (default: template taken from model's metadata)
-                                  only commonly used templates are accepted:
-                                  https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-  -sps,  --slot-prompt-similarity SIMILARITY
-                                  how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
-         --lora-init-without-apply
-                                  load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
-
-logging:
-
-         --simple-io              use basic IO for better compatibility in subprocesses and limited consoles
-  -ld,   --logdir LOGDIR          path under which to save YAML logs (no logging if unset)
-         --log-test               Run simple logging test
-         --log-disable            Disable trace logs
-         --log-enable             Enable trace logs
-         --log-file FNAME         Specify a log filename (without extension)
-         --log-new                Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
-         --log-append             Don't truncate the old log file.
-```
-
 
 ## Build
 
@@ -368,15 +277,16 @@ node index.js
 
 ## API Endpoints
 
-### GET `/health`: Returns the current state of the server
+### GET `/health`: Returns heath check result
 
-  - 503 -> `{"status": "loading model"}` if the model is still being loaded.
-  - 500 -> `{"status": "error"}` if the model failed to load.
-  - 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
-  - 200 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if no slots are currently available.
-  - 503 -> `{"status": "no slot available", "slots_idle": 0, "slots_processing": 32}` if the query parameter `fail_on_no_slot` is provided and no slots are currently available.
+**Response format**
 
-  If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
+- HTTP status code 503
+  - Body: `{"error": {"code": 503, "message": "Loading model", "type": "unavailable_error"}}`
+  - Explanation: the model is still being loaded.
+- HTTP status code 200
+  - Body: `{"status": "ok" }`
+  - Explanation: the model is successfully loaded and the server is ready.
 
 ### POST `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -424,8 +334,6 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
-    `penalty_prompt`: This will replace the `prompt` for the purpose of the penalty evaluation. Can be either `null`, a string or an array of numbers representing tokens. Default: `null`, which is to use the original `prompt`.
-
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -499,9 +407,44 @@ Notice that each `probs` is an array of length `n_probs`.
 
     *Options:*
 
-    `content`: Set the text to tokenize.
+    `content`: (Required) The text to tokenize.
 
-    `add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
+    `add_special`: (Optional) Boolean indicating if special tokens, i.e. `BOS`, should be inserted.  Default: `false`
+
+    `with_pieces`: (Optional) Boolean indicating whether to return token pieces along with IDs.  Default: `false`
+
+**Response:**
+
+Returns a JSON object with a `tokens` field containing the tokenization result. The `tokens` array contains either just token IDs or objects with `id` and `piece` fields, depending on the `with_pieces` parameter. The piece field is a string if the piece is valid unicode or a list of bytes otherwise.
+
+
+If `with_pieces` is `false`:
+```json
+{
+  "tokens": [123, 456, 789]
+}
+```
+
+If `with_pieces` is `true`:
+```json
+{
+  "tokens": [
+    {"id": 123, "piece": "Hello"},
+    {"id": 456, "piece": " world"},
+    {"id": 789, "piece": "!"}
+  ]
+}
+```
+
+With input 'á' (utf8 hex: C3 A1) on tinyllama/stories260k
+```json
+{
+  "tokens": [
+    {"id": 198, "piece": [195]}, // hex C3
+    {"id": 164, "piece": [161]} // hex A1
+  ]
+}
+```
 
 ### POST `/detokenize`: Convert tokens to text
 
@@ -639,10 +582,16 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
     }'
     ```
 
-### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
+### GET `/slots`: Returns the current slots processing state
+
+This endpoint can be disabled with `--no-slots`
+
+If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
 
 **Response format**
 
+Example:
+
 ```json
 [
     {
@@ -672,7 +621,6 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
             "stopping_word": ""
         },
         "penalize_nl": true,
-        "penalty_prompt_tokens": [],
         "presence_penalty": 0.0,
         "prompt": "Say hello to llama.cpp",
         "repeat_last_n": 64,
@@ -696,13 +644,18 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
         "tfs_z": 1.0,
         "top_k": 40,
         "top_p": 0.949999988079071,
-        "typical_p": 1.0,
-        "use_penalty_prompt_tokens": false
+        "typical_p": 1.0
     }
 ]
 ```
 
-### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
+Possible values for `slot[i].state` are:
+- `0`: SLOT_STATE_IDLE
+- `1`: SLOT_STATE_PROCESSING
+
+### GET `/metrics`: Prometheus compatible metrics exporter
+
+This endpoint is only accessible if `--metrics` is set.
 
 Available metrics:
 - `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@@ -767,6 +720,10 @@ Available metrics:
 
 ### GET `/lora-adapters`: Get list of all LoRA adapters
 
+This endpoint returns the loaded LoRA adapters. You can add adapters using `--lora` when starting the server, for example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...`
+
+By default, all adapters will be loaded with scale set to 1. To initialize all adapters scale to 0, add `--lora-init-without-apply`
+
 If an adapter is disabled, the scale will be set to 0.
 
 **Response format**

examples/server/public/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

examples/server/tests/features/embeddings.feature

@@ -9,8 +9,11 @@ Feature: llama.cpp server
     And   a model alias bert-bge-small
     And   42 as server seed
     And   2 slots
-    And   1024 as batch size
-    And   1024 as ubatch size
+    # the bert-bge-small model has context size of 512
+    # since the generated prompts are as big as the batch size, we need to set the batch size to 512
+    # ref: https://huggingface.co/BAAI/bge-small-en-v1.5/blob/5c38ec7c405ec4b44b94cc5a9bb96e735b38267a/config.json#L20
+    And   512 as batch size
+    And   512 as ubatch size
     And   2048 KV cache size
     And   embeddings extraction
     Then  the server is starting

examples/server/tests/features/parallel.feature

@@ -77,6 +77,35 @@ Feature: Parallel
       | disabled  | 128       |
       | enabled   | 64        |
 
+  Scenario Outline: Multi users with number of prompts exceeding number of slots
+    Given a system prompt You are a writer.
+    And   a model tinyllama-2
+    Given a prompt:
+      """
+      Write a very long book.
+      """
+    And a prompt:
+      """
+      Write another a poem.
+      """
+    And a prompt:
+      """
+      What is LLM?
+      """
+    And a prompt:
+      """
+      The sky is blue and I love it.
+      """
+    And <n_predict> max tokens to predict
+    And streaming is <streaming>
+    Given concurrent OAI completions requests
+    Then the server is busy
+    Then the server is idle
+    Then all prompts are predicted with <n_predict> tokens
+    Examples:
+      | streaming | n_predict |
+      | disabled  | 128       |
+      | enabled   | 64        |
 
   Scenario:  Multi users with total number of tokens to predict exceeds the KV Cache size #3969
     Given a prompt:

examples/server/tests/features/passkey.feature

@@ -15,6 +15,7 @@ Feature: Passkey / Self-extend with context shift
     And   <n_junk> as number of junk
     And   <n_predicted> server max tokens to predict
     And   42 as seed
+    And   0.0 temperature
     And   <n_ctx> KV cache size
     And   1 slots
     And   <n_ga> group attention factor to extend context size through self-extend
@@ -22,7 +23,8 @@ Feature: Passkey / Self-extend with context shift
     # Can be override with N_GPU_LAYERS
     And   <ngl> GPU offloaded layers
     Then  the server is starting
-    Then  the server is healthy
+    # Higher timeout because the model may need to be downloaded from the internet
+    Then  the server is healthy with timeout 120 seconds
     Given available models
     Then  model 0 is trained on <n_ctx_train> tokens context
     Given a prefix prompt:

examples/server/tests/features/server.feature

@@ -105,6 +105,14 @@ Feature: llama.cpp server
     Given first token is removed
     Then  tokens can be detokenized
 
+  Scenario: Tokenize with pieces
+    When  tokenizing with pieces:
+    """
+    What is the capital of Germany?
+    媽
+    """
+    Then  tokens are given with pieces
+
   Scenario: Models available
     Given available models
     Then  1 models are supported

examples/server/tests/features/steps/steps.py

@@ -1,3 +1,6 @@
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
 import asyncio
 import json
 import os
@@ -23,6 +26,8 @@ from prometheus_client import parser
 
 # pyright: reportRedeclaration=false
 
+DEFAULT_TIMEOUT_SECONDS = aiohttp.ClientTimeout(total=600)
+
 @step("a server listening on {server_fqdn}:{server_port}")
 def step_server_config(context, server_fqdn: str, server_port: str):
     context.server_fqdn = server_fqdn
@@ -200,36 +205,39 @@ def step_start_server(context):
             time.sleep(0.1)
 
 
-@step("the server is {expecting_status}")
-@async_run_until_complete
-async def step_wait_for_the_server_to_be_started(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
+async def wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
     match expecting_status:
         case 'healthy':
-            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=30)
+            await wait_for_slots_status(context, context.base_url, 200,
+                                        timeout=timeout)
 
         case 'ready' | 'idle':
-            await wait_for_health_status(context, context.base_url, 200, 'ok',
-                                         timeout=30,
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=context.n_slots,
-                                         slots_processing=0,
-                                         expected_slots=[{'id': slot_id, 'state': 0}
-                                                         for slot_id in
-                                                         range(context.n_slots if context.n_slots else 1)])
+            await wait_for_slots_status(context, context.base_url, 200,
+                                        timeout=timeout,
+                                        params={'fail_on_no_slot': 1},
+                                        slots_idle=context.n_slots,
+                                        slots_processing=0)
         case 'busy':
-            await wait_for_health_status(context, context.base_url, 503,
-                                         'no slot available',
-                                         params={'fail_on_no_slot': 0, 'include_slots': 0},
-                                         slots_idle=0,
-                                         slots_processing=context.n_slots,
-                                         expected_slots=[{'id': slot_id, 'state': 1}
-                                                         for slot_id in
-                                                         range(context.n_slots if context.n_slots else 1)])
+            await wait_for_slots_status(context, context.base_url, 503,
+                                        params={'fail_on_no_slot': 1},
+                                        slots_idle=0,
+                                        slots_processing=context.n_slots)
         case _:
             assert False, "unknown status"
 
 
+@step("the server is {expecting_status} with timeout {timeout:d} seconds")
+@async_run_until_complete
+async def step_wait_for_server_status_with_timeout(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str, timeout: int):
+    await wait_for_server_status_with_timeout(context, expecting_status, timeout)
+
+
+@step("the server is {expecting_status}")
+@async_run_until_complete
+async def step_wait_for_server_status(context, expecting_status: Literal['healthy', 'ready', 'idle', 'busy'] | str):
+    await wait_for_server_status_with_timeout(context, expecting_status, 30)
+
+
 @step('all slots are {expected_slot_status_string}')
 @async_run_until_complete
 async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
@@ -692,11 +700,37 @@ def step_tokenize_set_add_special(context):
     context.tokenize_add_special = True
 
 
+@step("tokenizing with pieces")
+@async_run_until_complete
+async def step_tokenize_with_pieces(context):
+    context.tokenized_text = context_text(context)
+    async with aiohttp.ClientSession() as session:
+        tokenize_args = {"content": context.tokenized_text, "with_pieces": True}
+        if getattr(context, "tokenize_add_special", None) is not None:
+            tokenize_args["add_special"] = context.tokenize_add_special
+
+        async with session.post(
+            f"{context.base_url}/tokenize", json=tokenize_args
+        ) as response:
+            assert response.status == 200
+            tokenize_json = await response.json()
+            context.tokens_with_pieces = tokenize_json["tokens"]
+
+
+@step("tokens are given with pieces")
+@async_run_until_complete
+async def step_tokenize_with_pieces(context):
+    # Verify that the response contains both token IDs and pieces
+    assert all(
+        "id" in token and "piece" in token for token in context.tokens_with_pieces
+    )
+
+
 @step('tokenizing')
 @async_run_until_complete
 async def step_tokenize(context):
     context.tokenized_text = context_text(context)
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         tokenize_args = {
             "content": context.tokenized_text,
         }
@@ -713,7 +747,7 @@ async def step_tokenize(context):
 @async_run_until_complete
 async def step_detokenize(context):
     assert len(context.tokens) > 0
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/detokenize',
                                 json={
                                     "tokens": context.tokens,
@@ -742,7 +776,7 @@ def step_strings_for_tokenization(context):
 @step('an OPTIONS request is sent from {origin}')
 @async_run_until_complete
 async def step_options_request(context, origin):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         headers = {'Authorization': f'Bearer {context.user_api_key}', 'Origin': origin}
         async with session.options(f'{context.base_url}/v1/chat/completions',
                                     headers=headers) as response:
@@ -758,7 +792,7 @@ def step_check_options_header_value(context, cors_header, cors_header_value):
 @step('prometheus metrics are exposed')
 @async_run_until_complete
 async def step_prometheus_metrics_exported(context):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with await session.get(f'{context.base_url}/metrics') as metrics_response:
             assert metrics_response.status == 200
             assert metrics_response.headers['Content-Type'] == "text/plain; version=0.0.4"
@@ -825,13 +859,13 @@ async def concurrent_requests(context, f_completion, *args, **kwargs):
     for prompt_no in range(context.n_prompts):
         shifted_args = [context.prompts.pop(), seeds[prompt_no], *args]
         context.concurrent_tasks.append(asyncio.create_task(f_completion(*shifted_args, **kwargs)))
-    await asyncio.sleep(0.1)
+    await asyncio.sleep(0.01)
 
 
 @step('the slot {slot_id:d} is saved with filename "{filename}"')
 @async_run_until_complete
 async def step_save_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=save',
                                 json={"filename": filename},
                                 headers={"Content-Type": "application/json"}) as response:
@@ -841,7 +875,7 @@ async def step_save_slot(context, slot_id, filename):
 @step('the slot {slot_id:d} is restored with filename "{filename}"')
 @async_run_until_complete
 async def step_restore_slot(context, slot_id, filename):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=restore',
                                 json={"filename": filename},
                                 headers={"Content-Type": "application/json"}) as response:
@@ -851,7 +885,7 @@ async def step_restore_slot(context, slot_id, filename):
 @step('the slot {slot_id:d} is erased')
 @async_run_until_complete
 async def step_erase_slot(context, slot_id):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/slots/{slot_id}?action=erase',
                                 headers={"Content-Type": "application/json"}) as response:
             context.response = response
@@ -860,7 +894,7 @@ async def step_erase_slot(context, slot_id):
 @step('switch {on_or_off} lora adapter {lora_id:d}')
 @async_run_until_complete
 async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{context.base_url}/lora-adapters',
                                 json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
                                 headers={"Content-Type": "application/json"}) as response:
@@ -896,7 +930,7 @@ async def request_completion(prompt,
             print(f"Set user_api_key: {user_api_key}")
         headers['Authorization'] = f'Bearer {user_api_key}'
 
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{base_url}/completion',
                                 json={
                                     "input_prefix": prompt_prefix,
@@ -909,8 +943,7 @@ async def request_completion(prompt,
                                     "temperature": temperature if temperature is not None else 0.8,
                                     "n_probs": 2,
                                 },
-                                headers=headers,
-                                timeout=3600) as response:
+                                headers=headers) as response:
             if expect_api_error is None or not expect_api_error:
                 assert response.status == 200
                 assert response.headers['Access-Control-Allow-Origin'] == origin
@@ -968,7 +1001,7 @@ async def oai_chat_completions(user_prompt,
     if async_client:
         origin = 'llama.cpp'
         headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
+        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
             async with session.post(f'{base_url}{base_path}',
                                     json=payload,
                                     headers=headers) as response:
@@ -1055,7 +1088,7 @@ async def oai_chat_completions(user_prompt,
 
 
 async def request_embedding(content, seed, base_url=None) -> list[list[float]]:
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with session.post(f'{base_url}/embedding',
                                 json={
                                     "content": content,
@@ -1075,14 +1108,13 @@ async def request_oai_embeddings(input, seed,
         headers=[]
         if user_api_key is not None:
             headers = {'Authorization': f'Bearer {user_api_key}', 'Origin': origin}
-        async with aiohttp.ClientSession() as session:
+        async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
             async with session.post(f'{base_url}/v1/embeddings',
                                     json={
                                         "input": input,
                                         "model": model,
                                     },
-                                    headers=headers,
-                                    timeout=3600) as response:
+                                    headers=headers) as response:
                 assert response.status == 200, f"received status code not expected: {response.status}"
                 assert response.headers['Access-Control-Allow-Origin'] == origin
                 assert response.headers['Content-Type'] == "application/json; charset=utf-8"
@@ -1187,44 +1219,35 @@ async def gather_tasks_results(context):
     return n_completions
 
 
-async def wait_for_health_status(context,
-                                 base_url,
-                                 expected_http_status_code,
-                                 expected_health_status,
-                                 timeout=3,
-                                 params=None,
-                                 slots_idle=None,
-                                 slots_processing=None,
-                                 expected_slots=None):
+async def wait_for_slots_status(context,
+                                base_url,
+                                expected_http_status_code,
+                                timeout=3,
+                                params=None,
+                                slots_idle=None,
+                                slots_processing=None):
     if context.debug:
-        print(f"Starting checking for health for expected_health_status={expected_health_status}")
+        print(f"Starting checking for health for expected_http_status_code={expected_http_status_code}")
     interval = 0.5
     counter = 0
     if 'GITHUB_ACTIONS' in os.environ:
         timeout *= 2
 
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         while True:
-            async with await session.get(f'{base_url}/health', params=params) as health_response:
-                status_code = health_response.status
-                health = await health_response.json()
+            async with await session.get(f'{base_url}/slots', params=params) as slots_response:
+                status_code = slots_response.status
+                slots = await slots_response.json()
                 if context.debug:
-                    print(f"HEALTH - response for expected health status='{expected_health_status}' on "
-                          f"'{base_url}/health'?{params} is {health}\n")
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
-                    return
-                if (status_code == expected_http_status_code
-                        and health['status'] == expected_health_status
-                        and (slots_idle is None or health['slots_idle'] == slots_idle)
-                        and (slots_processing is None or health['slots_processing'] == slots_processing)):
-                    if expected_slots is not None:
-                        assert_slots_status(health['slots'], expected_slots)
+                    print(f"slots responses {slots}\n")
+                if status_code == 503 and status_code == expected_http_status_code:
                     return
+                if status_code == 200 and status_code == expected_http_status_code:
+                    n_slots_idle = sum(1 if slot["state"] == 0 else 0 for slot in slots)
+                    n_slots_processing = sum(1 if slot["state"] != 0 else 0 for slot in slots)
+                    if ((slots_idle is None or slots_idle == n_slots_idle)
+                        and (slots_processing is None or slots_processing == n_slots_processing)):
+                        return
             await asyncio.sleep(interval)
 
             counter += interval
@@ -1238,7 +1261,7 @@ async def wait_for_health_status(context,
                         if n_completions > 0:
                             return
 
-                assert False, f'{expected_health_status} timeout exceeded {counter}s>={timeout}'
+                assert False, f'slots check timeout exceeded {counter}s>={timeout}'
 
 
 def assert_embeddings(embeddings):
@@ -1253,7 +1276,7 @@ def assert_embeddings(embeddings):
 
 
 async def request_slots_status(context, expected_slots):
-    async with aiohttp.ClientSession() as session:
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
         async with await session.get(f'{context.base_url}/slots') as slots_response:
             assert slots_response.status == 200
             slots = await slots_response.json()

examples/server/tests/features/wrong_usages.feature

@@ -8,9 +8,12 @@ Feature: Wrong usage of llama.cpp server
   Scenario: Infinite loop
     Given a server listening on localhost:8080
     And   a model file tinyllamas/stories260K.gguf from HF repo ggml-org/models
+    And   42 as server seed
+    And   2048 KV cache size
     # Uncomment below to fix the issue
     #And   64 server max tokens to predict
     Then  the server is starting
+    Then  the server is healthy
     Given a prompt:
       """
       Go to: infinite loop

examples/server/utils.hpp

@@ -3,6 +3,14 @@
 #include "llama.h"
 #include "common.h"
 
+#ifndef NDEBUG
+// crash the server in debug mode, otherwise send an http 500 error
+#define CPPHTTPLIB_NO_EXCEPTIONS 1
+#endif
+// increase max payload length to allow use of larger context size
+#define CPPHTTPLIB_FORM_URL_ENCODED_PAYLOAD_MAX_LENGTH 1048576
+#include "httplib.h"
+
 // Change JSON_ASSERT from assert() to GGML_ASSERT:
 #define JSON_ASSERT GGML_ASSERT
 #include "json.hpp"
@@ -279,6 +287,18 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
     return std::string::npos;
 }
 
+static bool json_is_array_of_numbers(json data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (!e.is_number()) {
+                return false;
+            }
+        }
+        return true;
+    }
+    return false;
+}
+
 // TODO: reuse llama_detokenize
 template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
@@ -343,6 +363,19 @@ static json probs_vector_to_json(const llama_context * ctx, const std::vector<co
     return out;
 }
 
+static bool server_sent_event(httplib::DataSink & sink, const char * event, json & data) {
+    const std::string str =
+        std::string(event) + ": " +
+        data.dump(-1, ' ', false, json::error_handler_t::replace) +
+        "\n\n";
+
+    LOG_VERBOSE("data stream", {
+        { "to_send", str }
+    });
+
+    return sink.write(str.c_str(), str.size());
+}
+
 //
 // OAI utils
 //
@@ -583,7 +616,40 @@ static json format_embeddings_response_oaicompat(const json & request, const jso
     return res;
 }
 
-static json format_tokenizer_response(const std::vector<llama_token> & tokens) {
+static bool is_valid_utf8(const std::string & str) {
+    const unsigned char* bytes = reinterpret_cast<const unsigned char*>(str.data());
+    const unsigned char* end = bytes + str.length();
+
+    while (bytes < end) {
+        if (*bytes <= 0x7F) {
+            // 1-byte sequence (0xxxxxxx)
+            bytes++;
+        } else if ((*bytes & 0xE0) == 0xC0) {
+            // 2-byte sequence (110xxxxx 10xxxxxx)
+            if (end - bytes < 2 || (bytes[1] & 0xC0) != 0x80)
+                return false;
+            bytes += 2;
+        } else if ((*bytes & 0xF0) == 0xE0) {
+            // 3-byte sequence (1110xxxx 10xxxxxx 10xxxxxx)
+            if (end - bytes < 3 || (bytes[1] & 0xC0) != 0x80 || (bytes[2] & 0xC0) != 0x80)
+                return false;
+            bytes += 3;
+        } else if ((*bytes & 0xF8) == 0xF0) {
+            // 4-byte sequence (11110xxx 10xxxxxx 10xxxxxx 10xxxxxx)
+            if (end - bytes < 4 || (bytes[1] & 0xC0) != 0x80 ||
+                (bytes[2] & 0xC0) != 0x80 || (bytes[3] & 0xC0) != 0x80)
+                return false;
+            bytes += 4;
+        } else {
+            // Invalid UTF-8 lead byte
+            return false;
+        }
+    }
+
+    return true;
+}
+
+static json format_tokenizer_response(const json & tokens) {
     return json {
         {"tokens", tokens}
     };

examples/simple/simple.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -6,9 +7,7 @@
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
+static void print_usage(int, char ** argv) {
     LOG_TEE("\nexample usage:\n");
     LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32\n", argv[0]);
     LOG_TEE("\n");
@@ -20,8 +19,7 @@ int main(int argc, char ** argv) {
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
@@ -55,6 +53,14 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    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());
+
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
@@ -110,20 +116,7 @@ int main(int argc, char ** argv) {
     while (n_cur <= n_predict) {
         // sample the next token
         {
-            auto   n_vocab = llama_n_vocab(model);
-            auto * logits  = llama_get_logits_ith(ctx, 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 llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+            const llama_token new_token_id = llama_sampler_sample(smpl, ctx, batch.n_tokens - 1);
 
             // is it an end of generation?
             if (llama_token_is_eog(model, new_token_id) || n_cur == n_predict) {
@@ -160,12 +153,14 @@ int main(int argc, char ** argv) {
     LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
             __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
 
-    llama_print_timings(ctx);
+    LOG_TEE("\n");
+    llama_perf_sampler_print(smpl);
+    llama_perf_context_print(ctx);
 
     fprintf(stderr, "\n");
 
     llama_batch_free(batch);
-
+    llama_sampler_free(smpl);
     llama_free(ctx);
     llama_free_model(model);
 

examples/speculative/speculative.cpp

@@ -1,11 +1,13 @@
+#include "arg.h"
 #include "common.h"
+#include "sampling.h"
 #include "llama.h"
 
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 #include <set>
+#include <random>
 
 #define SPEC_VOCAB_MAX_SIZE_DIFFERENCE  100
 #define SPEC_VOCAB_CHECK_START_TOKEN_ID 5
@@ -21,14 +23,13 @@ struct seq_draft {
     std::vector<llama_token> tokens;
     std::vector<std::vector<llama_token_data>> dists;
 
-    struct llama_sampling_context * ctx_sampling;
+    struct gpt_sampler * smpl = nullptr;
 };
 
 int main(int argc, char ** argv) {
     gpt_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!gpt_params_parse(argc, argv, params, LLAMA_EXAMPLE_SPECULATIVE)) {
         return 1;
     }
 
@@ -43,10 +44,7 @@ int main(int argc, char ** argv) {
     // probability threshold for splitting a draft branch (only for n_seq_dft > 1)
     const float p_split  = params.p_split;
 
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-    std::default_random_engine rng(params.seed);
+    std::default_random_engine rng(params.sparams.seed);
     std::uniform_real_distribution<> u_dist;
 
 #ifndef LOG_DISABLE_LOGS
@@ -73,10 +71,11 @@ int main(int argc, char ** argv) {
     // load the draft model
     params.model = params.model_draft;
     params.n_gpu_layers = params.n_gpu_layers_draft;
-    if (params.n_threads_draft > 0) {
-        params.n_threads = params.n_threads_draft;
+    if (params.draft_cpuparams.n_threads > 0) {
+        params.cpuparams.n_threads = params.draft_cpuparams.n_threads;
     }
-    params.n_threads_batch = params.n_threads_batch_draft;
+
+    params.cpuparams_batch.n_threads = params.draft_cpuparams_batch.n_threads;
     llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
     model_dft = llama_init_dft.model;
     ctx_dft = llama_init_dft.context;
@@ -178,19 +177,17 @@ int main(int argc, char ** argv) {
     // used to determine end of generation
     bool has_eos = false;
 
-    // target model sampling context
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params.sparams);
+    // target model sampling context (reuse the llama_context's sampling instance)
+    struct gpt_sampler * smpl = gpt_sampler_init(model_tgt, params.sparams);
+
+    struct llama_sampler * softmax = llama_sampler_init_softmax();
 
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
-    params.sparams.grammar.clear(); // the draft samplers will copy the target sampler's grammar
-    if (params.sparams.temp == 0) {
-        params.sparams.temp = -1.0f; // force greedy sampling with probs for the draft model
-    }
-
     for (int s = 0; s < n_seq_dft; ++s) {
-        drafts[s].ctx_sampling = llama_sampling_init(params.sparams);
+        // allocate gpt_sampler for each draft sequence
+        drafts[s].smpl = gpt_sampler_init(model_dft, params.sparams);
     }
 
     llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
@@ -232,12 +229,12 @@ int main(int argc, char ** argv) {
                 bool accept = false;
                 if (params.sparams.temp > 0) {
                     // stochastic verification
+                    gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft], true);
 
-                    llama_token_data_array dist_tgt = llama_sampling_prepare(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft], true, NULL);
-                    llama_sample_softmax(ctx_tgt, &dist_tgt);
-                    float p_tgt = 0, p_dft = 0;
+                    auto & dist_tgt = *gpt_sampler_get_candidates(smpl);
 
-                    // GGML_ASSERT(dist_tgt.size() == dist_dft.size());
+                    float p_tgt = 0.0f;
+                    float p_dft = 0.0f;
 
                     while (active_seqs.size() > 0) {
                         // randomly select a sequence to verify from active sequences
@@ -256,9 +253,13 @@ int main(int argc, char ** argv) {
                             }
                             continue;
                         }
+
                         LOG("verifying sequence #%d at pos #%d from %d active sequence(s)\n", s, i_dft, (int) active_seqs.size());
                         float r = u_dist(rng);
-                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), true };
+                        llama_token_data_array dist_dft = { drafts[s].dists[i_dft].data() , drafts[s].dists[i_dft].size(), LLAMA_TOKEN_NULL, true };
+
+                        //GGML_ASSERT(dist_tgt.size <= dist_dft.size);
+
                         // acquire the token probabilities assigned by the draft and target models
                         for (size_t i = 0; i < dist_tgt.size; i++) {
                             if (dist_tgt.data[i].id == drafts[s].tokens[i_dft]) {
@@ -277,7 +278,7 @@ int main(int argc, char ** argv) {
                             accept = true;
                             token_id = drafts[s].tokens[i_dft];
                             token_str = llama_token_to_piece(ctx_tgt, token_id);
-                            llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                            gpt_sampler_accept(smpl, token_id, true);
 
                             LOG("draft token %d of sequence %d (%d, '%s') accepted\n", i_dft, s, token_id, token_str.c_str());
                             break;
@@ -288,7 +289,6 @@ int main(int argc, char ** argv) {
                             // calculate residual probability
                             GGML_ASSERT(dist_tgt.sorted);
                             GGML_ASSERT(dist_dft.sorted);
-                            float sum_probs = 0.0f;
 
                             // sort dist by id
                             std::sort(dist_tgt.data, dist_tgt.data + dist_tgt.size, [](const llama_token_data &a, const llama_token_data &b) {
@@ -298,10 +298,18 @@ int main(int argc, char ** argv) {
                                 return a.id < b.id;
                             });
 
+                            float sum_probs = 0.0f;
+
                             for (size_t i = 0; i < dist_tgt.size; i++) {
-                                dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
+                                if (i < dist_dft.size) {
+                                    dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p - dist_dft.data[i].p);
+                                } else {
+                                    dist_tgt.data[i].p = std::max(0.0f, dist_tgt.data[i].p);
+                                }
+
                                 sum_probs += dist_tgt.data[i].p;
                             }
+
                             for (size_t i = 0; i < dist_tgt.size; i++) {
                                 dist_tgt.data[i].p /= sum_probs;
                             }
@@ -331,21 +339,29 @@ int main(int argc, char ** argv) {
                         // all drafted tokens were rejected
                         // sample from the target model
                         LOG("all drafted tokens were rejected, sampling from residual distribution\n");
-                        token_id = llama_sample_token(ctx_tgt, &dist_tgt);
-                        llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                        std::vector<float> probs(dist_tgt.size);
+                        for (size_t i = 0; i < dist_tgt.size; ++i) {
+                            probs[i] = dist_tgt.data[i].p;
+                        }
+
+                        std::discrete_distribution<> dist(probs.begin(), probs.end());
+
+                        const int idx = dist(rng);
+
+                        token_id = dist_tgt.data[idx].id;
+                        gpt_sampler_accept(smpl, token_id, true);
                         token_str = llama_token_to_piece(ctx_tgt, token_id);
                     }
-
                 } else {
                     // greedy verification
 
                     // sample from the target model
                     LOG("sampling target: s_keep = %3d, i_dft = %3d, i_batch_tgt = %3d\n", s_keep, i_dft, drafts[s_keep].i_batch_tgt[i_dft]);
-                    token_id = llama_sampling_sample(ctx_sampling, ctx_tgt, NULL, drafts[s_keep].i_batch_tgt[i_dft]);
+                    token_id = gpt_sampler_sample(smpl, ctx_tgt, drafts[s_keep].i_batch_tgt[i_dft]);
 
-                    llama_sampling_accept(ctx_sampling, ctx_tgt, token_id, true);
+                    gpt_sampler_accept(smpl, token_id, true);
 
-                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, ctx_sampling->prev).c_str());
+                    //LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_tgt, smpl->prev).c_str());
 
                     token_str = llama_token_to_piece(ctx_tgt, token_id);
 
@@ -433,7 +449,10 @@ int main(int argc, char ** argv) {
             break;
         }
 
-        llama_sampling_cp(ctx_sampling, drafts[0].ctx_sampling);
+        if (drafts[0].smpl) {
+            gpt_sampler_free(drafts[0].smpl);
+        }
+        drafts[0].smpl = gpt_sampler_clone(smpl);
 
         int n_seq_cur  = 1;
         int n_past_cur = n_past_dft;
@@ -462,20 +481,20 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                llama_sampling_sample(drafts[s].ctx_sampling, ctx_dft, NULL, drafts[s].i_batch_dft);
+                gpt_sampler_sample(drafts[s].smpl, ctx_dft, drafts[s].i_batch_dft, true);
 
-                const auto & cur_p = drafts[s].ctx_sampling->cur;
+                const auto * cur_p = gpt_sampler_get_candidates(drafts[s].smpl);
 
-                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p.size()); ++k) {
+                for (int k = 0; k < std::min(n_seq_dft + 3, (int) cur_p->size); ++k) {
                     LOG(" - draft candidate %3d for seq %3d, pos %3d: %6d (%8.3f) '%s'\n",
-                            k, s, i, cur_p[k].id, cur_p[k].p, llama_token_to_piece(ctx_dft, cur_p[k].id).c_str());
+                            k, s, i, cur_p->data[k].id, cur_p->data[k].p, llama_token_to_piece(ctx_dft, cur_p->data[k].id).c_str());
                 }
 
                 std::vector<int> sa(1, s);
 
                 // attempt to split the branch if the probability is high enough
                 for (int f = 1; f < 8; ++f) {
-                    if (n_seq_cur < n_seq_dft && cur_p[f].p > p_split) {
+                    if (n_seq_cur < n_seq_dft && cur_p->data[f].p > p_split) {
                         LOG("splitting seq %3d into %3d\n", s, n_seq_cur);
 
                         llama_kv_cache_seq_rm(ctx_dft,    n_seq_cur, -1, -1);
@@ -502,7 +521,10 @@ int main(int argc, char ** argv) {
                         drafts[n_seq_cur].i_batch_dft = drafts[s].i_batch_dft;
                         drafts[n_seq_cur].i_batch_tgt = drafts[s].i_batch_tgt;
 
-                        llama_sampling_cp(drafts[s].ctx_sampling, drafts[n_seq_cur].ctx_sampling);
+                        if (drafts[n_seq_cur].smpl) {
+                            gpt_sampler_free(drafts[n_seq_cur].smpl);
+                        }
+                        drafts[n_seq_cur].smpl = gpt_sampler_clone(drafts[s].smpl);
 
                         sa.push_back(n_seq_cur);
 
@@ -514,15 +536,15 @@ int main(int argc, char ** argv) {
 
                 // add drafted token for each sequence
                 for (int is = 0; is < (int) sa.size(); ++is) {
-                    const llama_token id = cur_p[is].id;
+                    const llama_token id = cur_p->data[is].id;
 
                     const int s = sa[is];
 
-                    llama_sampling_accept(drafts[s].ctx_sampling, ctx_dft, id, true);
+                    gpt_sampler_accept(drafts[s].smpl, id, true);
 
                     drafts[s].tokens.push_back(id);
                     // save cur_p.data into drafts[s].dists
-                    drafts[s].dists.push_back(cur_p);
+                    drafts[s].dists.push_back({cur_p->data, cur_p->data + cur_p->size});
 
                     // add unique drafted tokens to the target batch
                     drafts[s].i_batch_tgt.push_back(batch_tgt.n_tokens);
@@ -592,17 +614,19 @@ int main(int argc, char ** argv) {
     LOG_TEE("n_accept  = %d\n", n_accept);
     LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
-    LOG_TEE("\ndraft:\n");
-    llama_print_timings(ctx_dft);
+    LOG_TEE("\ndraft:\n\n");
+    // TODO: print sampling/grammar timings for all drafts
+    llama_perf_context_print(ctx_dft);
 
-    LOG_TEE("\ntarget:\n");
-    llama_print_timings(ctx_tgt);
+    LOG_TEE("\ntarget:\n\n");
+    gpt_perf_print(ctx_tgt, smpl);
 
-    llama_sampling_free(ctx_sampling);
+    gpt_sampler_free(smpl);
     for (int s = 0; s < n_seq_dft; ++s) {
-        llama_sampling_free(drafts[s].ctx_sampling);
+        gpt_sampler_free(drafts[s].smpl);
     }
 
+    llama_sampler_free(softmax);
     llama_batch_free(batch_dft);
 
     llama_free(ctx_tgt);

examples/sycl/run-llama2.sh

@@ -4,33 +4,23 @@
 #  Copyright (C) 2024 Intel Corporation
 #  SPDX-License-Identifier: MIT
 
-INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
 source /opt/intel/oneapi/setvars.sh
 
-if [ $# -gt 0 ]; then
-    GGML_SYCL_DEVICE=$1
-    GGML_SYCL_SINGLE_GPU=1
-else
-    GGML_SYCL_DEVICE=0
-    GGML_SYCL_SINGLE_GPU=0
-fi
-
 #export GGML_SYCL_DEBUG=1
 
-
 #ZES_ENABLE_SYSMAN=1, Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory. Recommended to use when --split-mode = layer.
 
-if [ $GGML_SYCL_SINGLE_GPU -eq 1 ]; then
+INPUT_PROMPT="Building a website can be done in 10 simple steps:\nStep 1:"
+MODEL_FILE=llama-2-7b.Q4_0.gguf
+NGL=33
+
+if [ $# -gt 0 ]; then
+    GGML_SYCL_DEVICE=$1
     echo "use $GGML_SYCL_DEVICE as main GPU"
     #use signle GPU only
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0 -mg $GGML_SYCL_DEVICE -sm none
+
 else
     #use multiple GPUs with same max compute units
-    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0
+    ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/${MODEL_FILE} -p "${INPUT_PROMPT}" -n 400 -e -ngl ${NGL} -s 0
 fi
-
-#use main GPU only
-#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0 -mg $GGML_SYCL_DEVICE -sm none
-
-#use multiple GPUs with same max compute units
-#ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "${INPUT2}" -n 400 -e -ngl 33 -s 0

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1722555600,
-        "narHash": "sha256-XOQkdLafnb/p9ij77byFQjDf5m5QYl9b2REiVClC+x4=",
+        "lastModified": 1725234343,
+        "narHash": "sha256-+ebgonl3NbiKD2UD0x4BszCZQ6sTfL4xioaM49o5B3Y=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "8471fe90ad337a8074e957b69ca4d0089218391d",
+        "rev": "567b938d64d4b4112ee253b9274472dc3a346eb6",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1723175592,
-        "narHash": "sha256-M0xJ3FbDUc4fRZ84dPGx5VvgFsOzds77KiBMW/mMTnI=",
+        "lastModified": 1725634671,
+        "narHash": "sha256-v3rIhsJBOMLR8e/RNWxr828tB+WywYIoajrZKFM+0Gg=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "5e0ca22929f3342b19569b21b2f3462f053e497b",
+        "rev": "574d1eac1c200690e27b8eb4e24887f8df7ac27c",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1722555339,
-        "narHash": "sha256-uFf2QeW7eAHlYXuDktm9c25OxOyCoUOQmh5SZ9amE5Q=",
+        "lastModified": 1725233747,
+        "narHash": "sha256-Ss8QWLXdr2JCBPcYChJhz4xJm+h/xjl4G0c0XlP6a74=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/a5d394176e64ab29c852d03346c1fc9b0b7d33eb.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/356624c12086a18f2ea2825fed34523d60ccc4e3.tar.gz"
       }
     },
     "root": {