Support SYCL device register

Flake lock file updates:

• Updated input 'nixpkgs':
    'github:NixOS/nixpkgs/bc947f541ae55e999ffdb4013441347d83b00feb?narHash=sha256-NOiTvBbRLIOe5F6RbHaAh6%2B%2BBNjsb149fGZd1T4%2BKBg%3D' (2024-10-04)
  → 'github:NixOS/nixpkgs/5633bcff0c6162b9e4b5f1264264611e950c8ec7?narHash=sha256-9UTxR8eukdg%2BXZeHgxW5hQA9fIKHsKCdOIUycTryeVw%3D' (2024-10-09)

Co-authored-by: github-actions[bot] <github-actions[bot]@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/full-musa.Dockerfile

@@ -0,0 +1,26 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential cmake python3 python3-pip git libcurl4-openssl-dev libgomp1
+
+COPY requirements.txt   requirements.txt
+COPY requirements       requirements
+
+RUN pip install --upgrade pip setuptools wheel \
+    && pip install -r requirements.txt
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=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/full-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,7 +34,7 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang

.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-cli-musa.Dockerfile

@@ -0,0 +1,30 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=ON ${CMAKE_ARGS} -DCMAKE_EXE_LINKER_FLAGS=-Wl,--allow-shlib-undefined . && \
+    cmake --build build --config Release --target llama-cli -j$(nproc)
+
+FROM ${BASE_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libgomp1
+
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-cli /llama-cli
+
+ENTRYPOINT [ "/llama-cli" ]

.devops/llama-cli-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,7 +34,7 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 ENV GGML_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang

.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-musa.Dockerfile

@@ -0,0 +1,35 @@
+ARG UBUNTU_VERSION=22.04
+# This needs to generally match the container host's environment.
+ARG MUSA_VERSION=rc3.1.0
+# Target the MUSA build image
+ARG BASE_MUSA_DEV_CONTAINER=mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}
+# Target the MUSA runtime image
+ARG BASE_MUSA_RUN_CONTAINER=mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}
+
+FROM ${BASE_MUSA_DEV_CONTAINER} AS build
+
+RUN apt-get update && \
+    apt-get install -y build-essential git cmake libcurl4-openssl-dev
+
+WORKDIR /app
+
+COPY . .
+
+RUN cmake -B build -DGGML_MUSA=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_MUSA_RUN_CONTAINER} AS runtime
+
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev libgomp1 curl
+
+COPY --from=build /app/build/ggml/src/libggml.so /libggml.so
+COPY --from=build /app/build/src/libllama.so /libllama.so
+COPY --from=build /app/build/bin/llama-server /llama-server
+
+# Must be set to 0.0.0.0 so it can listen to requests from host machine
+ENV LLAMA_ARG_HOST=0.0.0.0
+
+HEALTHCHECK CMD [ "curl", "-f", "http://localhost:8080/health" ]
+
+ENTRYPOINT [ "/llama-server" ]

.devops/llama-server-rocm.Dockerfile

@@ -11,7 +11,7 @@ FROM ${BASE_ROCM_DEV_CONTAINER} AS build
 # Unless otherwise specified, we make a fat build.
 # List from https://github.com/ggerganov/llama.cpp/pull/1087#issuecomment-1682807878
 # This is mostly tied to rocBLAS supported archs.
-ARG ROCM_DOCKER_ARCH=\
+ARG ROCM_DOCKER_ARCH="\
     gfx803 \
     gfx900 \
     gfx906 \
@@ -21,7 +21,7 @@ ARG ROCM_DOCKER_ARCH=\
     gfx1030 \
     gfx1100 \
     gfx1101 \
-    gfx1102
+    gfx1102"
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -34,11 +34,13 @@ WORKDIR /app
 COPY . .
 
 # Set nvcc architecture
-ENV GPU_TARGETS=${ROCM_DOCKER_ARCH}
+ENV AMDGPU_TARGETS=${ROCM_DOCKER_ARCH}
 # Enable ROCm
 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 { };
+})

.dockerignore

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

.ecrc

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

.github/workflows/bench.yml.disabled

@@ -1,3 +1,6 @@
+# TODO: there have been some issues with the workflow, so disabling for now
+#       https://github.com/ggerganov/llama.cpp/issues/7893
+#
 # Benchmark
 name: Benchmark
 
@@ -24,10 +27,10 @@ on:
   push:
     branches:
       - master
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.cpp', 'ggml-quants.c', '**/*.cu', 'examples/server/*.h*', 'examples/server/*.cpp']
   schedule:
     -  cron: '04 2 * * *'
 
@@ -129,6 +132,8 @@ jobs:
 
       - name: Server bench
         id: server_bench
+        env:
+            HEAD_REF: ${{ github.head_ref || github.ref_name }}
         run: |
           set -eux
 
@@ -137,7 +142,7 @@ jobs:
           python bench.py \
               --runner-label ${{ env.RUNNER_LABEL }} \
               --name ${{ github.job }} \
-              --branch ${{ github.head_ref || github.ref_name }} \
+              --branch $HEAD_REF \
               --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
               --scenario script.js \
               --duration ${{ github.event.inputs.duration || env.DURATION }} \

.github/workflows/build.yml

@@ -19,10 +19,18 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  contents: write # for creating release
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
 
 jobs:
   macOS-latest-cmake-arm64:
@@ -47,7 +55,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -105,7 +113,7 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -222,7 +230,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DBUILD_SHARED_LIBS=OFF
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -375,7 +383,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -401,7 +409,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -442,7 +450,7 @@ jobs:
     continue-on-error: true
 
     steps:
-      - uses: actions/checkout@v2
+      - uses: actions/checkout@v4
 
       - name: add oneAPI to apt
         shell: bash
@@ -546,7 +554,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -576,7 +584,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -610,7 +618,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v1
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -696,22 +704,20 @@ jobs:
     strategy:
       matrix:
         include:
-          - build: 'rpc-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'noavx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx2-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'avx-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX2=OFF -DBUILD_SHARED_LIBS=ON'
           - build: 'avx512-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_AVX512=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'openblas-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_BLAS=ON -DBUILD_SHARED_LIBS=ON -DGGML_BLAS_VENDOR=OpenBLAS -DBLAS_INCLUDE_DIRS="$env:RUNNER_TEMP/openblas/include" -DBLAS_LIBRARIES="$env:RUNNER_TEMP/openblas/lib/openblas.lib"'
           - build: 'kompute-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan-x64'
-            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
+            defines: '-DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DGGML_RPC=ON -DGGML_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'llvm-arm64'
             defines: '-G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-windows-llvm.cmake -DGGML_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'msvc-arm64'
@@ -859,7 +865,7 @@ 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}
 
@@ -955,6 +961,7 @@ jobs:
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/sycl7.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/svml_dispmd.dll" ./build/bin
           cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libmmd.dll" ./build/bin
+          cp "${{ env.ONEAPI_ROOT }}/compiler/latest/bin/libiomp5md.dll" ./build/bin
           echo "cp oneAPI running time dll files to ./build/bin done"
           7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
@@ -966,19 +973,20 @@ jobs:
           name: llama-bin-win-sycl-x64.zip
 
   windows-latest-cmake-hip:
+    if: ${{ github.event.inputs.create_release != 'true' }}
     runs-on: windows-latest
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@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"
@@ -993,8 +1001,72 @@ jobs:
         run: |
           $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
           $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON
-          cmake --build build --config Release
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+
+  windows-latest-cmake-hip-release:
+    if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+    runs-on: windows-latest
+
+    strategy:
+      matrix:
+        gpu_target: [gfx1100, gfx1101, gfx1030]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Install
+        id: depends
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD HIP SDK Installer"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          write-host "Installing AMD HIP SDK"
+          Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
+          write-host "Completed AMD HIP SDK installation"
+
+      - name: Verify ROCm
+        id: verify
+        run: |
+          & 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' --version
+
+      - name: Build
+        id: cmake_build
+        run: |
+          $env:HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)
+          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
+          cmake -G "Unix Makefiles" -B build -S . -DCMAKE_C_COMPILER="${env:HIP_PATH}\bin\clang.exe" -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\bin\clang++.exe" -DGGML_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=${{ matrix.gpu_target }} -DGGML_RPC=ON
+          cmake --build build -j ${env:NUMBER_OF_PROCESSORS}
+          md "build\bin\rocblas\library\"
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
+            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
+          else
+            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
+            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
+          fi
+
+      - name: Pack artifacts
+        id: pack_artifacts
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip .\build\bin\*
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
+          name: llama-bin-win-hip-x64-${{ matrix.gpu_target }}.zip
 
   ios-xcode-build:
     runs-on: macos-latest
@@ -1059,6 +1131,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-cuda
+      - windows-latest-cmake-hip-release
       - macOS-latest-cmake-arm64
       - macOS-latest-cmake-x64
 

.github/workflows/close-issue.yml

@@ -3,6 +3,11 @@ on:
   schedule:
     - cron: "42 0 * * *"
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  issues: write
+
 jobs:
   close-issues:
     runs-on: ubuntu-latest

.github/workflows/docker.yml

@@ -15,11 +15,17 @@ on:
     branches:
       - master
     paths: ['.github/workflows/docker.yml', '.devops/*.Dockerfile', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.cuh', '**/*.swift', '**/*.m', '**/*.metal']
+  workflow_dispatch: # allows manual triggering, useful for debugging
 
 concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  packages: write
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -37,15 +43,20 @@ 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.
+          - { tag: "light-musa", dockerfile: ".devops/llama-cli-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "server-musa", dockerfile: ".devops/llama-server-musa.Dockerfile", platforms: "linux/amd64" }
+          - { tag: "full-musa", dockerfile: ".devops/full-musa.Dockerfile", platforms: "linux/amd64" }
+          # Note: the rocm images are failing due to a compiler error and are disabled until this is fixed to allow the workflow to complete
+          #- { tag: "light-rocm", dockerfile: ".devops/llama-cli-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
+          #- { tag: "server-rocm", dockerfile: ".devops/llama-server-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           #- { tag: "full-rocm", dockerfile: ".devops/full-rocm.Dockerfile", platforms: "linux/amd64,linux/arm64" }
           - { tag: "light-intel", dockerfile: ".devops/llama-cli-intel.Dockerfile", platforms: "linux/amd64" }
           - { tag: "server-intel", dockerfile: ".devops/llama-server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
         uses: actions/checkout@v4
+        with:
+          fetch-depth: 0 # preserve git history, so we can determine the build number
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -60,6 +71,34 @@ jobs:
           username: ${{ github.repository_owner }}
           password: ${{ secrets.GITHUB_TOKEN }}
 
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(git rev-list --count HEAD)"
+          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
+          REPO_OWNER="${GITHUB_REPOSITORY_OWNER@L}"  # to lower case
+          REPO_NAME="${{ github.event.repository.name }}"
+
+          # determine tag name postfix (build number, commit hash)
+          if [[ "${{ env.GITHUB_BRANCH_NAME }}" == "master" ]]; then
+            TAG_POSTFIX="b${BUILD_NUMBER}"
+          else
+            SAFE_NAME=$(echo "${{ env.GITHUB_BRANCH_NAME }}" | tr '/' '-')
+            TAG_POSTFIX="${SAFE_NAME}-${SHORT_HASH}"
+          fi
+
+          # list all tags possible
+          TAGS=""
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }},"
+          TAGS="${TAGS}ghcr.io/${REPO_OWNER}/${REPO_NAME}:${{ matrix.config.tag }}-${TAG_POSTFIX}"
+
+          echo "output_tags=$TAGS" >> $GITHUB_OUTPUT
+          echo "output_tags=$TAGS"  # print out for debugging
+        env:
+          GITHUB_BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       # https://github.com/jlumbroso/free-disk-space/tree/54081f138730dfa15788a46383842cd2f914a1be#example
       - name: Free Disk Space (Ubuntu)
         uses: jlumbroso/free-disk-space@main
@@ -77,40 +116,13 @@ jobs:
           docker-images: true
           swap-storage: true
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Downcase github.repository_owner
-        run: |
-          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
-        env:
-          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
-
-      - name: Build and push Docker image (versioned)
+      - 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 }}"
+          # tag list is generated from step above
+          tags: ${{ steps.tag.outputs.output_tags }}
           file: ${{ matrix.config.dockerfile }}

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

@@ -21,6 +21,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -12,6 +12,13 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
+# Fine-grant permission
+# https://docs.github.com/en/actions/security-for-github-actions/security-guides/automatic-token-authentication#modifying-the-permissions-for-the-github_token
+permissions:
+  # https://github.com/DeterminateSystems/nix-installer-action?tab=readme-ov-file#with-flakehub
+  id-token: write
+  contents: read
+
 jobs:
   nix-eval:
     strategy:

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

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

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

@@ -4,11 +4,13 @@ on:
   push:
     paths:
       - '.github/workflows/python-type-check.yml'
+      - 'pyrightconfig.json'
       - '**.py'
       - '**/requirements*.txt'
   pull_request:
     paths:
       - '.github/workflows/python-type-check.yml'
+      - 'pyrightconfig.json'
       - '**.py'
       - '**/requirements*.txt'
 
@@ -33,6 +35,6 @@ jobs:
       - name: Type-check with Pyright
         uses: jakebailey/pyright-action@v2
         with:
-          version: 1.1.370
+          version: 1.1.382
           level: warning
           warnings: true

.github/workflows/server.yml

@@ -20,6 +20,12 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
 
+env:
+  LLAMA_LOG_COLORS: 1
+  LLAMA_LOG_PREFIX: 1
+  LLAMA_LOG_TIMESTAMPS: 1
+  LLAMA_LOG_VERBOSITY: 10
+
 concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
@@ -173,6 +179,7 @@ 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

@@ -62,6 +62,9 @@ option(LLAMA_SANITIZE_THREAD    "llama: enable thread sanitizer"    OFF)
 option(LLAMA_SANITIZE_ADDRESS   "llama: enable address sanitizer"   OFF)
 option(LLAMA_SANITIZE_UNDEFINED "llama: enable undefined sanitizer" OFF)
 
+# utils
+option(LLAMA_BUILD_COMMON "llama: build common utils library" ${LLAMA_STANDALONE})
+
 # extra artifacts
 option(LLAMA_BUILD_TESTS    "llama: build tests"          ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES "llama: build examples"       ${LLAMA_STANDALONE})
@@ -82,11 +85,11 @@ set(GGML_FATAL_WARNINGS     ${LLAMA_FATAL_WARNINGS})
 
 # change the default for these ggml options
 if (NOT DEFINED GGML_LLAMAFILE)
-    set(GGML_LLAMAFILE ON)
+    set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
-if (NOT DEFINED GGML_CUDA_USE_GRAPHS)
-    set(GGML_CUDA_USE_GRAPHS ON)
+if (NOT DEFINED GGML_CUDA_GRAPHS)
+    set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()
 
 # transition helpers
@@ -139,10 +142,16 @@ 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 ${CMAKE_CURRENT_SOURCE_DIR}/include/llama.h)
@@ -185,17 +194,19 @@ install(FILES "${CMAKE_CURRENT_BINARY_DIR}/llama.pc"
         DESTINATION lib/pkgconfig)
 
 #
-# programs, examples and tests
+# utils, programs, examples and tests
 #
 
-add_subdirectory(common)
+if (LLAMA_BUILD_COMMON)
+    add_subdirectory(common)
+endif()
 
-if (LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TESTS AND NOT CMAKE_JS_VERSION)
     include(CTest)
     add_subdirectory(tests)
-endif ()
+endif()
 
-if (LLAMA_BUILD_EXAMPLES)
+if (LLAMA_BUILD_COMMON AND LLAMA_BUILD_EXAMPLES)
     add_subdirectory(examples)
     add_subdirectory(pocs)
 endif()

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" ] }
   ]
 }

CONTRIBUTING.md

@@ -1,24 +1,23 @@
 # Pull requests (for contributors)
 
 - Test your changes:
-  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the GGML library
+  - Using the commands in the [`tests`](tests) folder. For instance, running the `./tests/test-backend-ops` command tests different backend implementations of the `ggml` library
   - Execute [the full CI locally on your machine](ci/README.md) before publishing
-- Please rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs.
-  - The PR template has a series of review complexity checkboxes `[ ]` that [you can mark as](https://docs.github.com/en/get-started/writing-on-github/working-with-advanced-formatting/about-task-lists) `[X]` for your convenience
-- Consider allowing write access to your branch for faster review
+- Optionally rate the complexity of your PR (i.e. `Review Complexity : Low`, `Review Complexity : Medium`, `Review Complexity : High`). This makes it easier for maintainers to triage the PRs
+- Consider allowing write access to your branch for faster reviews, as reviewers can push commits directly
 - If your PR becomes stale, don't hesitate to ping the maintainers in the comments
 
 # Pull requests (for collaborators)
 
 - Squash-merge PRs
 - Use the following format for the squashed commit title: `<module> : <commit title> (#<issue_number>)`. For example: `utils : fix typo in utils.py (#1234)`
-- Optionally, pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
+- Optionally pick a `<module>` from here: https://github.com/ggerganov/llama.cpp/wiki/Modules
 
 # Coding guidelines
 
 - Avoid adding third-party dependencies, extra files, extra headers, etc.
 - Always consider cross-compatibility with other operating systems and architectures
-- Avoid fancy looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
+- Avoid fancy-looking modern STL constructs, use basic `for` loops, avoid templates, keep it simple
 - There are no strict rules for the code style, but try to follow the patterns in the code (indentation, spaces, etc.). Vertical alignment makes things more readable and easier to batch edit
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - Naming usually optimizes for common prefix (see https://github.com/ggerganov/ggml/pull/302#discussion_r1243240963)
@@ -27,3 +26,8 @@
 
 ![matmul](media/matmul.png)
 
+# Resources
+
+The Github issues, PRs and discussions contain a lot of information that can be useful to get familiar with the codebase. For convenience, some of the more important information is referenced from Github projects:
+
+https://github.com/ggerganov/llama.cpp/projects

Makefile

@@ -5,7 +5,6 @@ BUILD_TARGETS = \
 	llama-batched \
 	llama-batched-bench \
 	llama-bench \
-	llama-benchmark-matmult \
 	llama-cli \
 	llama-convert-llama2c-to-ggml \
 	llama-embedding \
@@ -39,10 +38,12 @@ 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 \
@@ -52,6 +53,7 @@ TEST_TARGETS = \
 	tests/test-grammar-parser \
 	tests/test-json-schema-to-grammar \
 	tests/test-llama-grammar \
+	tests/test-log \
 	tests/test-model-load-cancel \
 	tests/test-opt \
 	tests/test-quantize-fns \
@@ -65,7 +67,7 @@ TEST_TARGETS = \
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
 LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama \
-	retrieval speculative infill tokenize benchmark-matmult parallel export-lora lookahead lookup passkey gritlm
+	retrieval speculative infill tokenize parallel export-lora lookahead lookup passkey gritlm
 
 # Legacy build targets that were renamed in #7809, but we want to build binaries that for them that output a deprecation warning if people try to use them.
 #  We don't want to clutter things too much, so we only build replacements for the most commonly used binaries.
@@ -146,6 +148,14 @@ GGML_NO_METAL := 1
 DEPRECATE_WARNING := 1
 endif
 
+ifdef LLAMA_DISABLE_LOGS
+REMOVE_WARNING := 1
+endif
+
+ifdef LLAMA_SERVER_VERBOSE
+REMOVE_WARNING := 1
+endif
+
 ifndef UNAME_S
 UNAME_S := $(shell uname -s)
 endif
@@ -349,19 +359,11 @@ ifdef LLAMA_SANITIZE_UNDEFINED
 	MK_LDFLAGS  += -fsanitize=undefined -g
 endif
 
-ifdef LLAMA_SERVER_VERBOSE
-	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
-endif
-
 ifdef LLAMA_SERVER_SSL
 	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
 	MK_LDFLAGS += -lssl -lcrypto
 endif
 
-ifdef LLAMA_DISABLE_LOGS
-	MK_CPPFLAGS += -DLOG_DISABLE_LOGS
-endif # LLAMA_DISABLE_LOGS
-
 # warnings
 WARN_FLAGS = \
 	-Wall \
@@ -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
@@ -603,7 +610,7 @@ ifdef GGML_CUDA
 
 		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include
 		MK_LDFLAGS   += -lmusa -lmublas -lmusart -lpthread -ldl -lrt -L$(CUDA_PATH)/lib -L/usr/lib64
-		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_22
+		MK_NVCCFLAGS += -x musa -mtgpu --cuda-gpu-arch=mp_21 --cuda-gpu-arch=mp_22
 	else
 		ifneq ('', '$(wildcard /opt/cuda)')
 			CUDA_PATH ?= /opt/cuda
@@ -611,7 +618,7 @@ ifdef GGML_CUDA
 			CUDA_PATH ?= /usr/local/cuda
 		endif
 
-		MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include -DGGML_CUDA_USE_GRAPHS
+		MK_CPPFLAGS  += -DGGML_USE_CUDA -DGGML_CUDA_USE_GRAPHS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
 		MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L$(CUDA_PATH)/lib64 -L/usr/lib64 -L$(CUDA_PATH)/targets/$(UNAME_M)-linux/lib -L$(CUDA_PATH)/lib64/stubs -L/usr/lib/wsl/lib
 		MK_NVCCFLAGS += -use_fast_math
 	endif # GGML_MUSA
@@ -763,6 +770,10 @@ ifdef GGML_VULKAN_MEMORY_DEBUG
 	MK_CPPFLAGS  += -DGGML_VULKAN_MEMORY_DEBUG
 endif
 
+ifdef GGML_VULKAN_PERF
+	MK_CPPFLAGS  += -DGGML_VULKAN_PERF
+endif
+
 ifdef GGML_VULKAN_VALIDATE
 	MK_CPPFLAGS  += -DGGML_VULKAN_VALIDATE
 endif
@@ -919,11 +930,12 @@ OBJ_LLAMA = \
 
 OBJ_COMMON = \
 	common/common.o \
+	common/arg.o \
+	common/log.o \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
 	common/train.o \
-	common/grammar-parser.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -1016,6 +1028,14 @@ $(info   - LLAMA_NO_CCACHE)
 $(info )
 endif
 
+ifdef REMOVE_WARNING
+$(info !!! REMOVAL WARNING !!!)
+$(info The following LLAMA_ options have been removed and are no longer supported)
+$(info   - LLAMA_DISABLE_LOGS   (https://github.com/ggerganov/llama.cpp/pull/9418))
+$(info   - LLAMA_SERVER_VERBOSE (https://github.com/ggerganov/llama.cpp/pull/9418))
+$(info )
+endif
+
 #
 # Build libraries
 #
@@ -1034,10 +1054,11 @@ ggml/src/ggml-alloc.o: \
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
 ggml/src/ggml-backend.o: \
-	ggml/src/ggml-backend.c \
+	ggml/src/ggml-backend.cpp \
+	ggml/src/ggml-backend-impl.h \
 	ggml/include/ggml.h \
 	ggml/include/ggml-backend.h
-	$(CC)  $(CFLAGS)   -c $< -o $@
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 ggml/src/ggml-quants.o: \
 	ggml/src/ggml-quants.c \
@@ -1152,6 +1173,16 @@ common/common.o: \
 	include/llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+common/arg.o: \
+	common/arg.cpp \
+	common/arg.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+common/log.o: \
+	common/log.cpp \
+	common/log.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 common/sampling.o: \
 	common/sampling.cpp \
 	common/sampling.h \
@@ -1163,11 +1194,6 @@ 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
@@ -1335,7 +1361,7 @@ llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp \
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
 llama-convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp \
-	$(OBJ_GGML) $(OBJ_LLAMA)
+	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -1429,6 +1455,7 @@ 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)
@@ -1444,6 +1471,11 @@ 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 \
@@ -1491,21 +1523,21 @@ common/build-info.o: common/build-info.cpp
 
 tests: $(TEST_TARGETS)
 
-llama-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp \
-	$(OBJ_GGML) common/build-info.o
+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)
 
-run-benchmark-matmult: llama-benchmark-matmult
-	./$@
-
-.PHONY: run-benchmark-matmult swift
-
 tests/test-llama-grammar: tests/test-llama-grammar.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-log: tests/test-log.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-grammar-parser: tests/test-grammar-parser.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)

Package.swift

@@ -11,7 +11,7 @@ var sources = [
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
     "ggml/src/ggml-alloc.c",
-    "ggml/src/ggml-backend.c",
+    "ggml/src/ggml-backend.cpp",
     "ggml/src/ggml-quants.c",
     "ggml/src/ggml-aarch64.c",
 ]

README.md

@@ -10,32 +10,15 @@
 
 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
+- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
+- Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
 
@@ -48,7 +31,7 @@ variety of hardware - locally and in the cloud.
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
-- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
+- Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
 - Vulkan and SYCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
 
@@ -95,6 +78,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
 - [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
 - [x] [OLMo](https://allenai.org/olmo)
+- [x] [OLMoE](https://huggingface.co/allenai/OLMoE-1B-7B-0924)
 - [x] [Granite models](https://huggingface.co/collections/ibm-granite/granite-code-models-6624c5cec322e4c148c8b330)
 - [x] [GPT-NeoX](https://github.com/EleutherAI/gpt-neox) + [Pythia](https://github.com/EleutherAI/pythia)
 - [x] [Snowflake-Arctic MoE](https://huggingface.co/collections/Snowflake/arctic-66290090abe542894a5ac520)
@@ -105,6 +89,10 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Open Elm models](https://huggingface.co/collections/apple/openelm-instruct-models-6619ad295d7ae9f868b759ca)
 - [x] [ChatGLM3-6b](https://huggingface.co/THUDM/chatglm3-6b) + [ChatGLM4-9b](https://huggingface.co/THUDM/glm-4-9b)
 - [x] [SmolLM](https://huggingface.co/collections/HuggingFaceTB/smollm-6695016cad7167254ce15966)
+- [x] [EXAONE-3.0-7.8B-Instruct](https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct)
+- [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
+- [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
+- [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -126,6 +114,7 @@ Typically finetunes of the base models below are supported as well.
 - Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
+- JS/TS (Programmable Prompt Engine CLI): [offline-ai/cli](https://github.com/offline-ai/cli)
 - JavaScript/Wasm (works in browser): [tangledgroup/llama-cpp-wasm](https://github.com/tangledgroup/llama-cpp-wasm)
 - Typescript/Wasm (nicer API, available on npm): [ngxson/wllama](https://github.com/ngxson/wllama)
 - Ruby: [yoshoku/llama_cpp.rb](https://github.com/yoshoku/llama_cpp.rb)
@@ -179,17 +168,23 @@ 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)
+- [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 **Tools:**
 
 - [akx/ggify](https://github.com/akx/ggify) – download PyTorch models from HuggingFace Hub and convert them to GGML
+- [akx/ollama-dl](https://github.com/akx/ollama-dl) – download models from the Ollama library to be used directly with llama.cpp
 - [crashr/gppm](https://github.com/crashr/gppm) – launch llama.cpp instances utilizing NVIDIA Tesla P40 or P100 GPUs with reduced idle power consumption
+- [gpustack/gguf-parser](https://github.com/gpustack/gguf-parser-go/tree/main/cmd/gguf-parser) - review/check the GGUF file and estimate the memory usage
+- [Styled Lines](https://marketplace.unity.com/packages/tools/generative-ai/styled-lines-llama-cpp-model-292902) (proprietary licensed, async wrapper of inference part for game development in Unity3d with prebuild Mobile and Web platform wrappers and a model example)
 
 **Infrastructure:**
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
+- [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.
@@ -418,10 +413,11 @@ Please refer to [Build llama.cpp locally](./docs/build.md)
 | [BLAS](./docs/build.md#blas-build) | All |
 | [BLIS](./docs/backend/BLIS.md) | All |
 | [SYCL](./docs/backend/SYCL.md) | Intel and Nvidia GPU |
-| [MUSA](./docs/build.md#musa) | Moore Threads GPU |
+| [MUSA](./docs/build.md#musa) | Moore Threads MTT GPU |
 | [CUDA](./docs/build.md#cuda) | Nvidia GPU |
 | [hipBLAS](./docs/build.md#hipblas) | AMD GPU |
 | [Vulkan](./docs/build.md#vulkan) | GPU |
+| [CANN](./docs/build.md#cann) | Ascend NPU |
 
 ## Tools
 
@@ -449,7 +445,7 @@ To learn more how to measure perplexity using llama.cpp, [read this documentatio
 - Contributors can open PRs
 - Collaborators can push to branches in the `llama.cpp` repo and merge PRs into the `master` branch
 - Collaborators will be invited based on contributions
-- Any help with managing issues and PRs is very appreciated!
+- Any help with managing issues, PRs and projects is very appreciated!
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Read the [CONTRIBUTING.md](CONTRIBUTING.md) for more information
 - Make sure to read this: [Inference at the edge](https://github.com/ggerganov/llama.cpp/discussions/205)

ci/run.sh

@@ -1,4 +1,4 @@
-#/bin/bash
+#!/bin/bash
 #
 # sample usage:
 #
@@ -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
 
@@ -707,6 +712,82 @@ function gg_run_embd_bge_small {
     set +e
 }
 
+function gg_sum_embd_bge_small {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'BGE Small (BERT):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
+    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
+}
+
+# rerank_tiny
+
+function gg_run_rerank_tiny {
+    cd ${SRC}
+
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/tokenizer_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/special_tokens_map.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/resolve/main/pytorch_model.bin
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/sentence_bert_config.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/vocab.txt
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/modules.json
+    gg_wget models-mnt/rerank-tiny/ https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/config.json
+
+    gg_wget models-mnt/rerank-tiny/1_Pooling https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/raw/main/1_Pooling/config.json
+
+    path_models="../models-mnt/rerank-tiny"
+
+    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
+
+    set -e
+
+    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j$(nproc)                                    ) 2>&1 | tee -a $OUT/${ci}-make.log
+
+    python3 ../convert_hf_to_gguf.py ${path_models} --outfile ${path_models}/ggml-model-f16.gguf
+
+    model_f16="${path_models}/ggml-model-f16.gguf"
+
+    # for this model, the SEP token is "</s>"
+    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+
+    # sample output
+    # rerank score 0:    0.029
+    # rerank score 1:    0.029
+    # rerank score 2:    0.135
+
+    # check that the score is in the range [$3, $4]
+    function check_score {
+        qnt="$1"
+        score=$(echo "$2" | grep -oE "[0-9]+\.[0-9]+" | tail -n 1)
+
+        if [ $(echo "$score < $3" | bc) -eq 1 ] || [ $(echo "$score > $4" | bc) -eq 1 ]; then
+            printf '  - %s @ %s (FAIL: score not in range [%s, %s])\n' "$qnt" "$score" "$3" "$4"
+            return 20
+        fi
+
+        printf '  - %s @ %s OK\n' "$qnt" "$score"
+        return 0
+    }
+
+    check_score "rerank score 0" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 0")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 1" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 1")" "0.00" "0.05" | tee -a $OUT/${ci}-rk-f16.log
+    check_score "rerank score 2" "$(cat $OUT/${ci}-rk-f16.log | grep "rerank score 2")" "0.10" "0.30" | tee -a $OUT/${ci}-rk-f16.log
+
+    set +e
+}
+
+function gg_sum_rerank_tiny {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Rerank Tiny (Jina):\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-rk-f16.log)"
+}
+
 function gg_check_build_requirements {
     if ! command -v cmake &> /dev/null; then
         gg_printf 'cmake not found, please install'
@@ -721,17 +802,11 @@ function gg_check_build_requirements {
     fi
 }
 
-function gg_sum_embd_bge_small {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'BGE Small (BERT):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-}
-
 ## main
 
+export LLAMA_LOG_PREFIX=1
+export LLAMA_LOG_TIMESTAMPS=1
+
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     # Create symlink: ./llama.cpp/models-mnt -> $MNT/models/models-mnt
     rm -rf ${SRC}/models-mnt
@@ -754,6 +829,7 @@ test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
+    test $ret -eq 0 && gg_run rerank_tiny
 
     if [ -z ${GG_BUILD_CLOUD} ] || [ ${GG_BUILD_EXTRA_TESTS_0} ]; then
         test $ret -eq 0 && gg_run test_scripts_debug
@@ -761,7 +837,7 @@ 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

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

common/arg.h

@@ -0,0 +1,77 @@
+#pragma once
+
+#include "common.h"
+
+#include <set>
+#include <string>
+#include <vector>
+
+//
+// CLI argument parsing
+//
+
+struct common_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)   (common_params & params) = nullptr;
+    void (*handler_string) (common_params & params, const std::string &) = nullptr;
+    void (*handler_str_str)(common_params & params, const std::string &, const std::string &) = nullptr;
+    void (*handler_int)    (common_params & params, int) = nullptr;
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(common_params & params, const std::string &)
+    ) : args(args), value_hint(value_hint), help(help), handler_string(handler) {}
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const std::string & help,
+        void (*handler)(common_params & params, int)
+    ) : args(args), value_hint(value_hint), help(help), handler_int(handler) {}
+
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const std::string & help,
+        void (*handler)(common_params & params)
+    ) : args(args), help(help), handler_void(handler) {}
+
+    // support 2 values for arg
+    common_arg(
+        const std::initializer_list<const char *> & args,
+        const char * value_hint,
+        const char * value_hint_2,
+        const std::string & help,
+        void (*handler)(common_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) {}
+
+    common_arg & set_examples(std::initializer_list<enum llama_example> examples);
+    common_arg & set_env(const char * env);
+    common_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 common_params_context {
+    enum llama_example ex = LLAMA_EXAMPLE_COMMON;
+    common_params & params;
+    std::vector<common_arg> options;
+    void(*print_usage)(int, char **) = nullptr;
+    common_params_context(common_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 common_params_parse(int argc, char ** argv, common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);
+
+// function to be used by test-arg-parser
+common_params_context common_params_parser_init(common_params & params, llama_example ex, void(*print_usage)(int, char **) = nullptr);

common/common.h

@@ -4,18 +4,9 @@
 
 #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>
+#include <sstream>
 
 #ifdef _WIN32
 #define DIRECTORY_SEPARATOR '\\'
@@ -33,12 +24,12 @@
 
 #define DEFAULT_MODEL_PATH "models/7B/ggml-model-f16.gguf"
 
-struct llama_lora_adapter_info {
+struct common_lora_adapter_info {
     std::string path;
     float scale;
 };
 
-struct llama_lora_adapter_container : llama_lora_adapter_info {
+struct common_lora_adapter_container : common_lora_adapter_info {
     struct llama_lora_adapter * adapter;
 };
 
@@ -48,32 +39,109 @@ extern char const * LLAMA_COMMIT;
 extern char const * LLAMA_COMPILER;
 extern char const * LLAMA_BUILD_TARGET;
 
-struct llama_control_vector_load_info;
+struct common_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 common_sampler_type {
+    COMMON_SAMPLER_TYPE_NONE        = 0,
+    COMMON_SAMPLER_TYPE_TOP_K       = 1,
+    COMMON_SAMPLER_TYPE_TOP_P       = 2,
+    COMMON_SAMPLER_TYPE_MIN_P       = 3,
+    COMMON_SAMPLER_TYPE_TFS_Z       = 4,
+    COMMON_SAMPLER_TYPE_TYPICAL_P   = 5,
+    COMMON_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 common_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 common_sampler_type> samplers = {
+        COMMON_SAMPLER_TYPE_TOP_K,
+        COMMON_SAMPLER_TYPE_TFS_Z,
+        COMMON_SAMPLER_TYPE_TYPICAL_P,
+        COMMON_SAMPLER_TYPE_TOP_P,
+        COMMON_SAMPLER_TYPE_MIN_P,
+        COMMON_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 common_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 +168,11 @@ 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;
 
@@ -110,35 +183,34 @@ struct gpt_params {
     enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
     enum llama_attention_type    attention_type    = LLAMA_ATTENTION_TYPE_UNSPECIFIED; // attention type for embeddings
 
-    // // sampling parameters
-    struct llama_sampling_params sparams;
+    struct common_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)
     std::vector<llama_model_kv_override> kv_overrides;
 
     bool lora_init_without_apply = false; // only load lora to memory, but do not apply it to ctx (user can manually apply lora later using llama_lora_adapter_apply)
-    std::vector<llama_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
+    std::vector<common_lora_adapter_info> lora_adapters; // lora adapter path with user defined scale
 
-    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
+    std::vector<common_control_vector_load_info> control_vectors; // control vector with user defined scale
 
     int32_t verbosity                  = 0;
     int32_t control_vector_layer_start = -1; // layer range for control vector
@@ -173,15 +245,15 @@ struct gpt_params {
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
     bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
     bool flash_attn        = false; // flash attention
+    bool no_perf           = false; // disable performance metrics
+    bool ctx_shift         = true;  // context shift on inifinite text generation
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
-    bool ignore_eos        = false; // ignore generated EOS tokens
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
-    bool infill            = false; // use infill mode
     bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
     bool no_kv_offload     = false; // disable KV offloading
     bool warmup            = true;  // warmup run
@@ -191,7 +263,7 @@ 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
@@ -199,25 +271,28 @@ struct gpt_params {
     int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
     std::string embd_sep   = "\n";  // separator of embendings
+    bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    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
     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;
+    // "advanced" endpoints are disabled by default for better security
+    bool webui            = true;
+    bool endpoint_slots   = false;
+    bool endpoint_props   = false; // only control POST requests, not GET
     bool endpoint_metrics = false;
 
     bool log_json = false;
@@ -265,22 +340,39 @@ 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);
+// call once at the start of a program if it uses libcommon
+// initializes the logging system and prints info about the build
+void common_init();
 
-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 common_params_get_system_info(const common_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
 //
 
+#ifdef __GNUC__
+#ifdef __MINGW32__
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(gnu_printf, __VA_ARGS__)))
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...) __attribute__((format(printf, __VA_ARGS__)))
+#endif
+#else
+#define LLAMA_COMMON_ATTRIBUTE_FORMAT(...)
+#endif
+
+LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
+std::string string_format(const char * fmt, ...);
+
 std::vector<std::string> string_split(std::string input, char separator);
 
 std::string string_strip(const std::string & str);
@@ -305,6 +397,11 @@ static std::vector<T> string_split(const std::string & str, char delim) {
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 
+std::string string_from(bool value);
+std::string string_from(const std::vector<int> & values);
+std::string string_from(const struct llama_context * ctx, const std::vector<llama_token> & tokens);
+std::string string_from(const struct llama_context * ctx, const struct llama_batch & batch);
+
 //
 // Filesystem utils
 //
@@ -319,28 +416,29 @@ std::string fs_get_cache_file(const std::string & filename);
 // Model utils
 //
 
-struct llama_init_result {
+struct common_init_result {
     struct llama_model   * model   = nullptr;
     struct llama_context * context = nullptr;
-    std::vector<llama_lora_adapter_container> lora_adapters;
+    std::vector<common_lora_adapter_container> lora_adapters;
 };
 
-struct llama_init_result    llama_init_from_gpt_params(gpt_params & params);
+struct common_init_result     common_init_from_params(common_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     common_model_params_to_llama  (const common_params & params);
+struct llama_context_params   common_context_params_to_llama(const common_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);
+struct llama_model * common_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 * common_load_model_from_hf(const char * repo, const char * file, const char * path_model, const char * hf_token, const struct llama_model_params & params);
 
 // clear LoRA adapters from context, then apply new list of adapters
-void llama_lora_adapters_apply(struct llama_context * ctx, std::vector<llama_lora_adapter_container> & lora_adapters);
+void common_lora_adapters_apply(struct llama_context * ctx, std::vector<common_lora_adapter_container> & lora_adapters);
 
 // Batch utils
 
-void llama_batch_clear(struct llama_batch & batch);
+void common_batch_clear(struct llama_batch & batch);
 
-void llama_batch_add(
+void common_batch_add(
                  struct llama_batch & batch,
                         llama_token   id,
                           llama_pos   pos,
@@ -353,13 +451,13 @@ void llama_batch_add(
 
 // tokenizes a string into a vector of tokens
 // should work similar to Python's `tokenizer.encode`
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
                         bool   add_special,
                         bool   parse_special = false);
 
-std::vector<llama_token> llama_tokenize(
+std::vector<llama_token> common_tokenize(
     const struct llama_model * model,
            const std::string & text,
                         bool   add_special,
@@ -367,7 +465,7 @@ std::vector<llama_token> llama_tokenize(
 
 // tokenizes a token into a piece, optionally renders special/control tokens
 // should work similar to Python's `tokenizer.id_to_piece`
-std::string llama_token_to_piece(
+std::string common_token_to_piece(
         const struct llama_context * ctx,
                        llama_token   token,
                        bool          special = true);
@@ -375,45 +473,41 @@ std::string llama_token_to_piece(
 // detokenizes a vector of tokens into a string
 // should work similar to Python's `tokenizer.decode`
 // optionally renders special/control tokens
-std::string llama_detokenize(
+std::string common_detokenize(
                          llama_context * ctx,
         const std::vector<llama_token> & tokens,
                                   bool   special = true);
 
-// Uses the value from the model metadata if possible, otherwise
-// defaults to true when model type is SPM, otherwise false.
-bool llama_should_add_bos_token(const llama_model * model);
-
 //
 // Chat template utils
 //
 
 // same with llama_chat_message, but uses std::string
-struct llama_chat_msg {
+struct common_chat_msg {
     std::string role;
     std::string content;
 };
 
 // Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
-bool llama_chat_verify_template(const std::string & tmpl);
+bool common_chat_verify_template(const std::string & tmpl);
 
 // CPP wrapper for llama_chat_apply_template
 // If the built-in template is not supported, we default to chatml
 // If the custom "tmpl" is not supported, we throw an error
-std::string llama_chat_apply_template(const struct llama_model * model,
+std::string common_chat_apply_template(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & chat,
+        const std::vector<common_chat_msg> & chat,
         bool add_ass);
 
 // Format single message, while taking into account the position of that message in chat history
-std::string llama_chat_format_single(const struct llama_model * model,
+std::string common_chat_format_single(const struct llama_model * model,
         const std::string & tmpl,
-        const std::vector<llama_chat_msg> & past_msg,
-        const llama_chat_msg & new_msg,
+        const std::vector<common_chat_msg> & past_msg,
+        const common_chat_msg & new_msg,
         bool add_ass);
 
 // Returns an example of formatted chat
-std::string llama_chat_format_example(const struct llama_model * model,
+std::string common_chat_format_example(const struct llama_model * model,
         const std::string & tmpl);
 
 //
@@ -421,31 +515,31 @@ std::string llama_chat_format_example(const struct llama_model * model,
 //
 
 // Dump the KV cache view with the number of sequences per cell.
-void llama_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
+void common_kv_cache_dump_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
-void llama_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+void common_kv_cache_dump_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
 
 //
 // Embedding utils
 //
 
-void llama_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
+void common_embd_normalize(const float * inp, float * out, int n, int embd_norm = 2);
 
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
+float common_embd_similarity_cos(const float * embd1, const float * embd2, int n);
 
 //
 // Control vector utils
 //
 
-struct llama_control_vector_data {
+struct common_control_vector_data {
     int n_embd;
 
     // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
     std::vector<float> data;
 };
 
-struct llama_control_vector_load_info {
+struct common_control_vector_load_info {
     float strength;
 
     std::string fname;
@@ -453,7 +547,7 @@ struct llama_control_vector_load_info {
 
 // Load control vectors, scale each by strength, and add them together.
 // On error, returns {-1, empty}
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
+common_control_vector_data common_control_vector_load(const std::vector<common_control_vector_load_info> & load_infos);
 
 //
 // Split utils
@@ -472,5 +566,5 @@ void yaml_dump_vector_int      (FILE * stream, const char * prop_name, const std
 void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const char * data);
 
 void yaml_dump_non_result_info(
-    FILE * stream, const gpt_params & params, const llama_context * lctx,
+    FILE * stream, const common_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);

common/console.cpp

@@ -94,6 +94,9 @@ namespace console {
                 simple_io = true;
             }
         }
+        if (simple_io) {
+            _setmode(_fileno(stdin), _O_U8TEXT);
+        }
 #else
         // POSIX-specific console initialization
         if (!simple_io) {

common/grammar-parser.cpp

@@ -1,536 +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) {
-                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/log.cpp

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

common/log.h

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

common/ngram-cache.cpp

@@ -2,10 +2,13 @@
 #include "common.h"
 #include "log.h"
 
+#include <cinttypes>
 #include <cstdint>
+#include <cstdio>
 #include <fstream>
+#include <thread>
 
-void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
+void common_ngram_cache_update(common_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
                               std::vector<llama_token> & inp, int nnew, bool print_progress) {
     const int64_t t_start_ms = ggml_time_ms();
     const int64_t inp_size = inp.size();
@@ -17,16 +20,16 @@ void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, in
         const int64_t i_start = std::max(inp_size - nnew, ngram_size);
         for (int64_t i = i_start; i < inp_size; ++i) {
             const int64_t ngram_start = i - ngram_size;
-            llama_ngram ngram(&inp[ngram_start], ngram_size);
+            common_ngram ngram(&inp[ngram_start], ngram_size);
             const llama_token token = inp[i];
 
-            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
+            common_ngram_cache::iterator part_it = ngram_cache.find(ngram);
             if (part_it == ngram_cache.end()) {
-                llama_ngram_cache_part part;
+                common_ngram_cache_part part;
                 part.emplace(token, 1);
                 ngram_cache.emplace(ngram, part);
             } else {
-                llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
+                common_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
                 if (token_count_it == part_it->second.end()) {
                     part_it->second.emplace(token, 1);
                 } else {
@@ -59,12 +62,12 @@ constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4,  3,  2,  2};
 constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
 
 // Helper function that tries to draft a token from only the static ngram cache:
-static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
-    llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+static llama_token try_draft(common_ngram_cache & nc_static, const common_ngram ngram_static) {
+    common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
     if (part_static_it == nc_static.end()) {
         return -1;
     }
-    const llama_ngram_cache_part part_static = part_static_it->second;
+    const common_ngram_cache_part part_static = part_static_it->second;
 
     int max_count_static  = 0;
     int sum_count_static  = 0;
@@ -92,19 +95,19 @@ static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ng
 
 // Try to draft a token from primary cache (context/dynamic), validate with static cache:
 static llama_token try_draft(
-    llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
+    common_ngram_cache & nc_primary, const std::vector<common_ngram> & ngrams_primary, common_ngram_cache_part & part_static,
     const int * min_sample_size, const int * min_percent) {
 
     llama_token drafted_token = -1;
 
     for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
-        const llama_ngram ngram_primary = ngrams_primary[i];
+        const common_ngram ngram_primary = ngrams_primary[i];
 
-        llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
+        common_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
         if (part_primary_it == nc_primary.end()) {
             continue;
         }
-        const llama_ngram_cache_part part_primary = part_primary_it->second;
+        const common_ngram_cache_part part_primary = part_primary_it->second;
 
         int max_count_primary = 0;
         int max_count_static  = 0;
@@ -114,7 +117,7 @@ static llama_token try_draft(
         for (std::pair<llama_token, int> token_count_primary : part_primary) {
             const llama_token token = token_count_primary.first;
 
-            llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
+            common_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
 
             const int32_t count_primary = token_count_primary.second;
             const int32_t count_static  = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
@@ -139,9 +142,9 @@ static llama_token try_draft(
     return drafted_token;
 }
 
-void llama_ngram_cache_draft(
+void common_ngram_cache_draft(
     std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
+    common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static
 ) {
     GGML_ASSERT(draft.size() == 1);
     const int inp_size = inp.size();
@@ -154,21 +157,21 @@ void llama_ngram_cache_draft(
         llama_token drafted_token = -1;
 
         const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
-        llama_ngram ngram_static;
+        common_ngram ngram_static;
         for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
             ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
         }
-        llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
-        llama_ngram_cache_part part_static;
+        common_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
+        common_ngram_cache_part part_static;
         if (part_static_it != nc_static.end()) {
             part_static = part_static_it->second;
         }
 
         // cd = context + dynamic
-        std::vector<llama_ngram> ngrams_cd;
+        std::vector<common_ngram> ngrams_cd;
         for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
             const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
-            llama_ngram ngram_cd;
+            common_ngram ngram_cd;
             for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
                 ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
             }
@@ -193,16 +196,16 @@ void llama_ngram_cache_draft(
     }
 }
 
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename) {
     std::ofstream file_out(filename, std::ios::binary);
-    for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
-        const llama_ngram      ngram        = item.first;
-        llama_ngram_cache_part token_counts = item.second;
+    for (std::pair<common_ngram, common_ngram_cache_part> item : ngram_cache) {
+        const common_ngram      ngram        = item.first;
+        common_ngram_cache_part token_counts = item.second;
         GGML_ASSERT(!token_counts.empty());
         const int32_t ntokens = token_counts.size();
         GGML_ASSERT(ntokens > 0);
 
-        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(llama_ngram));
+        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(common_ngram));
         file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
         for (std::pair<llama_token, int32_t> item2 : token_counts) {
             const llama_token token = item2.first;
@@ -216,14 +219,14 @@ void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filen
 
 }
 
-llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
+common_ngram_cache common_ngram_cache_load(std::string & filename) {
     std::ifstream hashmap_file(filename, std::ios::binary);
     if (!hashmap_file) {
         throw std::ifstream::failure("Unable to open file " + filename);
     }
-    llama_ngram_cache ngram_cache;
+    common_ngram_cache ngram_cache;
 
-    llama_ngram ngram;
+    common_ngram ngram;
     int32_t     ntokens;
     llama_token token;
     int32_t     count;
@@ -232,11 +235,11 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
     char * ntokensc = reinterpret_cast<char*>(&ntokens);
     char * tokenc   = reinterpret_cast<char*>(&token);
     char * countc   = reinterpret_cast<char*>(&count);
-    while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
+    while(hashmap_file.read(ngramc, sizeof(common_ngram))) {
         GGML_ASSERT(!hashmap_file.eof());
         GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
         GGML_ASSERT(ntokens > 0);
-        llama_ngram_cache_part token_counts;
+        common_ngram_cache_part token_counts;
 
         for (int i = 0; i < ntokens; ++i) {
             GGML_ASSERT(!hashmap_file.eof());
@@ -254,12 +257,12 @@ llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
     return ngram_cache;
 }
 
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
-    for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
-        const llama_ngram      ngram = ngram_part.first;
-        llama_ngram_cache_part  part = ngram_part.second;
+void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add) {
+    for (std::pair<common_ngram, common_ngram_cache_part> ngram_part : ngram_cache_add) {
+        const common_ngram      ngram = ngram_part.first;
+        common_ngram_cache_part  part = ngram_part.second;
 
-        llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
+        common_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
         if (part_merged_it == ngram_cache_target.end()) {
             ngram_cache_target.emplace(ngram, part);
             continue;
@@ -270,7 +273,7 @@ void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram
             const int32_t     count = token_count.second;
             GGML_ASSERT(count > 0);
 
-            llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
+            common_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
             if (token_count_merged_it == part_merged_it->second.end()) {
                 part_merged_it->second.emplace(token, count);
                 continue;

common/ngram-cache.h

@@ -12,22 +12,22 @@
 
 // Data structures to map n-grams to empirical token probabilities:
 
-struct llama_ngram {
+struct common_ngram {
     llama_token tokens[LLAMA_NGRAM_MAX];
 
-    llama_ngram() {
+    common_ngram() {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             tokens[i] = -1;
         }
     }
 
-    llama_ngram(const llama_token * input, const int ngram_size) {
+    common_ngram(const llama_token * input, const int ngram_size) {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             tokens[i] = i < ngram_size ? input[i] : -1;
         }
     }
 
-    bool operator==(const llama_ngram & other) const {
+    bool operator==(const common_ngram & other) const {
         for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
             if (tokens[i] != other.tokens[i]) {
                 return false;
@@ -37,28 +37,28 @@ struct llama_ngram {
     }
 };
 
-struct llama_token_hash_function {
+struct common_token_hash_function {
     size_t operator()(const llama_token token) const {
         // see https://probablydance.com/2018/06/16/fibonacci-hashing-the-optimization-that-the-world-forgot-or-a-better-alternative-to-integer-modulo/
         return token * 11400714819323198485llu;
     }
 };
 
-struct llama_ngram_hash_function {
-    size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = llama_token_hash_function{}(ngram.tokens[0]);
+struct common_ngram_hash_function {
+    size_t operator()(const common_ngram & ngram) const {
+        size_t hash = common_token_hash_function{}(ngram.tokens[0]);
         for (int i = 1; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= llama_token_hash_function{}(ngram.tokens[i]);
+            hash ^= common_token_hash_function{}(ngram.tokens[i]);
         }
         return hash;
     }
 };
 
 // token -> number of times token has been seen
-typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
+typedef std::unordered_map<llama_token, int32_t> common_ngram_cache_part;
 
 // n-gram -> empirical distribution of following tokens
-typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
+typedef std::unordered_map<common_ngram, common_ngram_cache_part, common_ngram_hash_function> common_ngram_cache;
 
 
 // Update an ngram cache with tokens.
@@ -70,8 +70,8 @@ typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash
 //
 // In order to get correct results inp_data can ONLY BE APPENDED TO.
 // Changes in the middle need a complete rebuild.
-void llama_ngram_cache_update(
-    llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
+void common_ngram_cache_update(
+    common_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
 
 // Try to draft tokens from ngram caches.
 // inp:                the tokens generated so far.
@@ -81,21 +81,21 @@ void llama_ngram_cache_update(
 // nc_context:         ngram cache based on current context.
 // nc_dynamic:         ngram cache based on previous user generations.
 // nc_static:          ngram cache generated from a large text corpus, used for validation.
-void llama_ngram_cache_draft(
+void common_ngram_cache_draft(
     std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
+    common_ngram_cache & nc_context, common_ngram_cache & nc_dynamic, common_ngram_cache & nc_static);
 
 // Save an ngram cache to a file.
 // ngram_cache: the ngram cache to save.
 // filename:    the path under which to save the ngram cache.
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
+void common_ngram_cache_save(common_ngram_cache & ngram_cache, std::string & filename);
 
-// Load an ngram cache saved with llama_ngram_cache_save.
+// Load an ngram cache saved with common_ngram_cache_save.
 // filename: the path from which to load the ngram cache.
 // returns:  an ngram cache containing the information saved to filename.
-llama_ngram_cache llama_ngram_cache_load(std::string & filename);
+common_ngram_cache common_ngram_cache_load(std::string & filename);
 
 // Merge two ngram caches.
 // ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
 // ngram_cache_add:    the ngram cache to add to ngram_cache_target.
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);
+void common_ngram_cache_merge(common_ngram_cache & ngram_cache_target, common_ngram_cache & ngram_cache_add);

common/sampling.cpp

@@ -1,460 +1,458 @@
-#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 common_sampler {
+    common_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 common_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 common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params) {
+    llama_sampler_chain_params lparams = llama_sampler_chain_default_params();
+
+    lparams.no_perf = params.no_perf;
+
+    auto * result = new common_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 COMMON_SAMPLER_TYPE_TOP_K:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TOP_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_MIN_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TFS_Z:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
+                        break;
+                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                        break;
+                    case COMMON_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 ";
+        if (params.n_probs > 0) {
+            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
+            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
+            //
+            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
+            // it is much faster, since we avoid sorting all tokens and should give a good approximation
+            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
+            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
+        }
+        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
     }
 
     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 common_sampler_free(struct common_sampler * gsmpl) {
+    if (gsmpl) {
+        llama_sampler_free(gsmpl->grmr);
+
+        llama_sampler_free(gsmpl->chain);
+
+        delete gsmpl;
+    }
+}
+
+void common_sampler_accept(struct common_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 common_sampler_reset(struct common_sampler * gsmpl) {
+    llama_sampler_reset(gsmpl->grmr);
+
+    llama_sampler_reset(gsmpl->chain);
+}
+
+struct common_sampler * common_sampler_clone(common_sampler * gsmpl) {
+    return new common_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 common_perf_print(const struct llama_context * ctx, const struct common_sampler * gsmpl) {
+    // TODO: measure grammar performance
+
+    if (gsmpl) {
+        llama_perf_sampler_print(gsmpl->chain);
+    }
+    if (ctx) {
+        llama_perf_context_print(ctx);
+    }
+}
+
+llama_token common_sampler_sample(struct common_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 common_sampler_get_seed(const struct common_sampler * gsmpl) {
+    return llama_sampler_get_seed(gsmpl->chain);
+}
+
+// helpers
+
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl) {
+    return &gsmpl->cur_p;
+}
+
+llama_token common_sampler_last(const struct common_sampler * gsmpl) {
+    return gsmpl->prev.rat(0);
+}
+
+std::string common_sampler_print(const struct common_sampler * gsmpl) {
+    std::string result = "logits ";
+
+    for (int i = 0; i < llama_sampler_chain_n(gsmpl->chain); i++) {
+        const auto * smpl = llama_sampler_chain_get(gsmpl->chain, i);
+        result += std::string("-> ") + llama_sampler_name(smpl) + " ";
+    }
+
+    return result;
+}
+
+std::string common_sampler_prev_str(common_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 += common_token_to_piece(ctx_main, id);
+    }
+
+    return result;
+}
+
+char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
+    switch (cnstr) {
+        case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
+        case COMMON_SAMPLER_TYPE_TFS_Z:       return 'f';
+        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
+        case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
+        case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
+        case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
+        default : return '?';
+    }
+}
+
+std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
+    switch (cnstr) {
+        case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
+        case COMMON_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
+        case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
+        case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
+        case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
+        case COMMON_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<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
+    std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
+        { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
+        { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "tfs_z",       COMMON_SAMPLER_TYPE_TFS_Z },
+        { "temperature", COMMON_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, common_sampler_type> sampler_alt_name_map {
+        { "top-k",       COMMON_SAMPLER_TYPE_TOP_K },
+        { "top-p",       COMMON_SAMPLER_TYPE_TOP_P },
+        { "nucleus",     COMMON_SAMPLER_TYPE_TOP_P },
+        { "typical-p",   COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typical",     COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
+        { "tfs-z",       COMMON_SAMPLER_TYPE_TFS_Z },
+        { "tfs",         COMMON_SAMPLER_TYPE_TFS_Z },
+        { "temp",        COMMON_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<common_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<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
+    std::unordered_map<char, common_sampler_type> sampler_name_map = {
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z),       COMMON_SAMPLER_TYPE_TFS_Z },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }
+    };
+
+    std::vector<common_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
+// common_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 common_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 common_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 common_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 common_sampler * common_sampler_init(const struct llama_model * model, const struct common_sampler_params & params);
+
+void common_sampler_free(struct common_sampler * gsmpl);
+
+// if accept_grammar is true, the token is accepted both by the sampling chain and the grammar
+void                    common_sampler_accept(struct common_sampler * gsmpl, llama_token token, bool accept_grammar);
+void                    common_sampler_reset (struct common_sampler * gsmpl);
+struct common_sampler * common_sampler_clone (struct common_sampler * gsmpl);
+
+// arguments can be nullptr to skip printing
+void common_perf_print(const struct llama_context * ctx, const struct common_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 common_sampler_sample(struct common_sampler * gsmpl, struct llama_context * ctx, int idx, bool grammar_first = false);
+
+uint32_t common_sampler_get_seed(const struct common_sampler * gsmpl);
+
+// helpers
+
+// access the internal list of current candidate tokens
+llama_token_data_array * common_sampler_get_candidates(struct common_sampler * gsmpl);
+
+// get the last accepted token
+llama_token common_sampler_last(const struct common_sampler * gsmpl);
+
+// print the sampler chain into a string
+std::string common_sampler_print(const struct common_sampler * gsmpl);
+
+// get a string representation of the last accepted tokens
+std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx, int n);
+
+char        common_sampler_type_to_chr(enum common_sampler_type cnstr);
+std::string common_sampler_type_to_str(enum common_sampler_type cnstr);
+
+std::vector<enum common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
+std::vector<enum common_sampler_type> common_sampler_types_from_chars(const std::string & chars);

common/train.cpp

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

convert_hf_to_gguf.py

@@ -3,6 +3,7 @@
 
 from __future__ import annotations
 
+import ast
 import logging
 import argparse
 import contextlib
@@ -14,6 +15,7 @@ from enum import IntEnum
 from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
+from itertools import chain
 
 import math
 import numpy as np
@@ -129,12 +131,14 @@ class Model:
     def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
         tensor_names_from_parts: set[str] = set()
 
-        if len(self.part_names) > 1:
+        index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
+        index_name += ".index.json"
+        index_file = self.dir_model / index_name
+
+        if index_file.is_file():
             self.tensor_names = set()
-            index_name = "model.safetensors" if self.is_safetensors else "pytorch_model.bin"
-            index_name += ".index.json"
             logger.info(f"gguf: loading model weight map from '{index_name}'")
-            with open(self.dir_model / index_name, "r", encoding="utf-8") as f:
+            with open(index_file, "r", encoding="utf-8") as f:
                 index: dict[str, Any] = json.load(f)
                 weight_map = index.get("weight_map")
                 if weight_map is None or not isinstance(weight_map, dict):
@@ -142,6 +146,7 @@ class Model:
                 self.tensor_names.update(weight_map.keys())
         else:
             self.tensor_names = tensor_names_from_parts
+            weight_map = {}
 
         for part_name in self.part_names:
             logger.info(f"gguf: loading model part '{part_name}'")
@@ -168,9 +173,17 @@ class Model:
                             data = LazyTorchTensor.from_eager(data)
                     yield name, data
 
-        # only verify tensor name presence; it doesn't matter if they are not in the right files
-        if len(sym_diff := tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
-            raise ValueError(f"Mismatch between weight map and model parts for tensor names: {sym_diff}")
+        # verify tensor name presence and identify potentially missing files
+        if len(tensor_names_from_parts.symmetric_difference(self.tensor_names)) > 0:
+            missing = sorted(self.tensor_names.difference(tensor_names_from_parts))
+            extra = sorted(tensor_names_from_parts.difference(self.tensor_names))
+            missing_files = sorted(set(weight_map[n] for n in missing if n in weight_map))
+            if len(extra) == 0 and len(missing_files) > 0:
+                raise ValueError(f"Missing or incomplete model files: {missing_files}")
+            else:
+                raise ValueError("Mismatch between weight map and model parts for tensor names:\n"
+                                 f"Missing tensors: {missing}\n"
+                                 f"Extra tensors: {extra}")
 
     def format_tensor_name(self, key: gguf.MODEL_TENSOR, bid: int | None = None, suffix: str = ".weight") -> str:
         if key not in gguf.MODEL_TENSORS[self.model_arch]:
@@ -256,10 +269,14 @@ class Model:
 
         return False
 
+    # some models need extra generated tensors (like rope_freqs)
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        return ()
+
     def prepare_tensors(self):
         max_name_len = max(len(s) for _, s in self.tensor_map.mapping.values()) + len(".weight,")
 
-        for name, data_torch in self.get_tensors():
+        for name, data_torch in chain(self.generate_extra_tensors(), self.get_tensors()):
             # we don't need these
             if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
                 continue
@@ -277,8 +294,13 @@ class Model:
                     bid = int(part)
                     break
 
-            for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
-                data: np.ndarray  # type hint
+            for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
+                data = data_torch.squeeze().numpy()
+
+                # if data ends up empty, it means data_torch was a scalar tensor -> restore
+                if len(data.shape) == 0:
+                    data = data_torch.numpy()
+
                 n_dims = len(data.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
@@ -295,12 +317,32 @@ 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 +353,10 @@ 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}")
 
@@ -554,6 +600,9 @@ class Model:
         if chkhsh == "a8594e3edff7c29c003940395316294b2c623e09894deebbc65f33f1515df79e":
             # ref: https://huggingface.co/databricks/dbrx-base
             res = "dbrx"
+        if chkhsh == "c7699093ba4255a91e702aa38a596aa81669f3525dae06c2953267dde580f448":
+            # ref: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+            res = "jina-v1-en"
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/jinaai/jina-embeddings-v2-base-en
             res = "jina-v2-en"
@@ -590,6 +639,21 @@ class Model:
         if chkhsh == "855059429035d75a914d1eda9f10a876752e281a054a7a3d421ef0533e5b6249":
             # ref: https://huggingface.co/HuggingFaceTB/SmolLM-135M
             res = "smollm"
+        if chkhsh == "3c30d3ad1d6b64202cd222813e7736c2db6e1bd6d67197090fc1211fbc612ae7":
+            # ref: https://huggingface.co/bigscience/bloom
+            res = "bloom"
+        if chkhsh == "bc01ce58980e1db43859146dc51b1758b3b88729b217a74792e9f8d43e479d21":
+            # ref: https://huggingface.co/TurkuNLP/gpt3-finnish-small
+            res = "gpt3-finnish"
+        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 chkhsh == "60824e3c0d9401f89943cbb2fff727f0e2d4c545ba4df2d6e4f09a6db0f5b450":
+            # ref: https://huggingface.co/facebook/chameleon-7b
+            res = "chameleon"
 
         if res is None:
             logger.warning("\n")
@@ -893,7 +957,7 @@ class GPTNeoXModel(Model):
         return tensors
 
 
-@Model.register("BloomForCausalLM")
+@Model.register("BloomForCausalLM", "BloomModel")
 class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
@@ -1448,7 +1512,7 @@ class StableLMModel(Model):
                 raise ValueError(f"Unprocessed norms: {norms}")
 
 
-@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+@Model.register("LLaMAForCausalLM", "LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
 class LlamaModel(Model):
     model_arch = gguf.MODEL_ARCH.LLAMA
 
@@ -1556,11 +1620,11 @@ class LlamaModel(Model):
 
         return [(self.map_tensor_name(name), data_torch)]
 
-    def prepare_tensors(self):
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
         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)
@@ -1583,8 +1647,9 @@ 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))
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
 
+    def prepare_tensors(self):
         super().prepare_tensors()
 
         if self._experts is not None:
@@ -1606,15 +1671,16 @@ 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)
@@ -1629,11 +1695,9 @@ 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")
@@ -1802,6 +1866,59 @@ class MiniCPMModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("MiniCPM3ForCausalLM")
+class MiniCPM3Model(Model):
+    model_arch = gguf.MODEL_ARCH.MINICPM3
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
+        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count_kv(hparams["num_key_value_heads"])
+        self.gguf_writer.add_layer_norm_rms_eps(hparams["rms_norm_eps"])
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        if "q_lora_rank" in hparams and hparams["q_lora_rank"] is not None:
+            self.gguf_writer.add_q_lora_rank(hparams["q_lora_rank"])
+        self.gguf_writer.add_kv_lora_rank(hparams["kv_lora_rank"])
+        self.gguf_writer.add_key_length(hparams["qk_nope_head_dim"] + hparams["qk_rope_head_dim"])
+        self.gguf_writer.add_rope_dimension_count(hparams["qk_rope_head_dim"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        rope_scaling = self.find_hparam(['rope_scaling'], True)
+        if rope_scaling is not None:
+            rope_dims = self.hparams["qk_rope_head_dim"]
+
+            long_factors = rope_scaling.get('long_factor', None)
+            short_factors = rope_scaling.get('short_factor', None)
+
+            if long_factors is None or short_factors is None:
+                raise KeyError('Missing the required key rope_scaling.long_factor or rope_scaling_short_factor')
+
+            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}')
+
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+
+    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
+        if n_kv_head is not None and n_head != n_kv_head:
+            n_head //= n_kv_head
+
+        return (
+            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
+            .swapaxes(1, 2)
+            .reshape(weights.shape)
+        )
+
+
 @Model.register("QWenLMHeadModel")
 class QwenModel(Model):
     model_arch = gguf.MODEL_ARCH.QWEN
@@ -2101,6 +2218,13 @@ class Phi3MiniModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_sliding_window(self.find_hparam(["sliding_window"]))
 
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        max_pos_embds = self.find_hparam(["n_positions", "max_position_embeddings"])
+        orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
+        rope_dims = n_embd // n_head
+
         # write rope scaling for long context (128k) model
         rope_scaling = self.find_hparam(['rope_scaling'], True)
         if rope_scaling is None:
@@ -2130,8 +2254,8 @@ 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))
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_LONG), torch.tensor(long_factors, dtype=torch.float32))
+        yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FACTORS_SHORT), torch.tensor(short_factors, dtype=torch.float32))
 
 
 @Model.register("PlamoForCausalLM")
@@ -2493,7 +2617,7 @@ class NomicBertModel(BertModel):
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
 
-@Model.register("XLMRobertaModel")
+@Model.register("XLMRobertaModel", "XLMRobertaForSequenceClassification")
 class XLMRobertaModel(BertModel):
     model_arch = gguf.MODEL_ARCH.BERT
 
@@ -2591,6 +2715,11 @@ class XLMRobertaModel(BertModel):
         self.gguf_writer.add_add_eos_token(True)
 
     def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "roberta.", remove the prefix
+        # e.g. https://huggingface.co/BAAI/bge-reranker-v2-m3/tree/main
+        if name.startswith("roberta."):
+            name = name[8:]
+
         # position embeddings start at pad_token_id + 1, so just chop down the weight tensor
         if name == "embeddings.position_embeddings.weight":
             if self._position_offset is not None:
@@ -2702,7 +2831,87 @@ 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
 
@@ -2733,7 +2942,10 @@ 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
 
@@ -2741,12 +2953,13 @@ 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
@@ -2773,23 +2986,6 @@ 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):
@@ -2837,6 +3033,66 @@ class OlmoModel(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("OlmoeForCausalLM")
+class OlmoeModel(Model):
+    model_arch = gguf.MODEL_ARCH.OLMOE
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_layer_norm_rms_eps(1e-5)
+        if (n_experts := self.hparams.get("num_experts")) is not None:
+            self.gguf_writer.add_expert_count(n_experts)
+
+    _experts: list[dict[str, Tensor]] | None = None
+
+    # Copied from: Qwen2MoeModel
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # process the experts separately
+        if name.find("experts") != -1:
+            n_experts = self.hparams["num_experts"]
+            assert bid is not None
+
+            if self._experts is None:
+                self._experts = [{} for _ in range(self.block_count)]
+
+            self._experts[bid][name] = data_torch
+
+            if len(self._experts[bid]) >= n_experts * 3:
+                tensors: list[tuple[str, Tensor]] = []
+
+                # merge the experts into a single 3d tensor
+                for w_name in ["down_proj", "gate_proj", "up_proj"]:
+                    datas: list[Tensor] = []
+
+                    for xid in range(n_experts):
+                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
+                        datas.append(self._experts[bid][ename])
+                        del self._experts[bid][ename]
+
+                    data_torch = torch.stack(datas, dim=0)
+
+                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
+
+                    new_name = self.map_tensor_name(merged_name)
+
+                    tensors.append((new_name, data_torch))
+                return tensors
+            else:
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # Copied from: Qwen2MoeModel
+    def prepare_tensors(self):
+        super().prepare_tensors()
+
+        if self._experts is not None:
+            # flatten `list[dict[str, Tensor]]` into `list[str]`
+            experts = [k for d in self._experts for k in d.keys()]
+            if len(experts) > 0:
+                raise ValueError(f"Unprocessed experts: {experts}")
+
+
 @Model.register("JinaBertModel", "JinaBertForMaskedLM")
 class JinaBertV2Model(BertModel):
     model_arch = gguf.MODEL_ARCH.JINA_BERT_V2
@@ -2875,6 +3131,14 @@ class JinaBertV2Model(BertModel):
         self.gguf_writer.add_add_bos_token(True)
         self.gguf_writer.add_add_eos_token(True)
 
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # if name starts with "bert.", remove the prefix
+        # e.g. https://huggingface.co/jinaai/jina-reranker-v1-tiny-en
+        if name.startswith("bert."):
+            name = name[5:]
+
+        return super().modify_tensors(data_torch, name, bid)
+
 
 @Model.register("OpenELMForCausalLM")
 class OpenELMModel(Model):
@@ -3324,6 +3588,145 @@ class T5Model(Model):
         return [(self.map_tensor_name(name), data_torch)]
 
 
+@Model.register("T5EncoderModel")
+class T5EncoderModel(Model):
+    model_arch = gguf.MODEL_ARCH.T5ENCODER
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.shared_token_embeddings_found = False
+
+    def set_vocab(self):
+        # to avoid TypeError: Descriptors cannot be created directly
+        # exception when importing sentencepiece_model_pb2
+        os.environ["PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION"] = "python"
+        from sentencepiece import SentencePieceProcessor
+        from sentencepiece import sentencepiece_model_pb2 as model
+
+        tokenizer_path = self.dir_model / 'tokenizer.model'
+
+        # many older models use spiece.model tokenizer model filename
+        if not tokenizer_path.is_file():
+            tokenizer_path = self.dir_model / 'spiece.model'
+
+        if not tokenizer_path.is_file():
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
+
+        sentencepiece_model = model.ModelProto()  # pyright: ignore[reportAttributeAccessIssue]
+        sentencepiece_model.ParseFromString(open(tokenizer_path, "rb").read())
+
+        # some models like Pile-T5 family use BPE tokenizer instead of Unigram
+        if sentencepiece_model.trainer_spec.model_type == 2:  # BPE
+            # assure the tokenizer model file name is correct
+            assert tokenizer_path.name == 'tokenizer.model'
+            return self._set_vocab_sentencepiece()
+        else:
+            assert sentencepiece_model.trainer_spec.model_type == 1  # UNIGRAM
+
+        add_prefix = sentencepiece_model.normalizer_spec.add_dummy_prefix
+        remove_whitespaces = sentencepiece_model.normalizer_spec.remove_extra_whitespaces
+        precompiled_charsmap = sentencepiece_model.normalizer_spec.precompiled_charsmap
+
+        tokenizer = SentencePieceProcessor()
+        tokenizer.LoadFromFile(str(tokenizer_path))
+
+        vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
+
+        tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
+        scores: list[float] = [-10000.0] * vocab_size
+        toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
+
+        for token_id in range(tokenizer.vocab_size()):
+            piece = tokenizer.IdToPiece(token_id)
+            text = piece.encode("utf-8")
+            score = tokenizer.GetScore(token_id)
+
+            toktype = SentencePieceTokenTypes.NORMAL
+            if tokenizer.IsUnknown(token_id):
+                toktype = SentencePieceTokenTypes.UNKNOWN
+            elif tokenizer.IsControl(token_id):
+                toktype = SentencePieceTokenTypes.CONTROL
+            elif tokenizer.IsUnused(token_id):
+                toktype = SentencePieceTokenTypes.UNUSED
+            elif tokenizer.IsByte(token_id):
+                toktype = SentencePieceTokenTypes.BYTE
+
+            tokens[token_id] = text
+            scores[token_id] = score
+            toktypes[token_id] = toktype
+
+        added_tokens_file = self.dir_model / 'added_tokens.json'
+        if added_tokens_file.is_file():
+            with open(added_tokens_file, "r", encoding="utf-8") as f:
+                added_tokens_json = json.load(f)
+                for key in added_tokens_json:
+                    token_id = added_tokens_json[key]
+                    if token_id >= vocab_size:
+                        logger.warning(f'ignore token {token_id}: id is out of range, max={vocab_size - 1}')
+                        continue
+
+                    tokens[token_id] = key.encode("utf-8")
+                    scores[token_id] = -1000.0
+                    toktypes[token_id] = SentencePieceTokenTypes.USER_DEFINED
+
+        if vocab_size > len(tokens):
+            pad_count = vocab_size - len(tokens)
+            logger.debug(f"Padding vocab with {pad_count} token(s) - [PAD1] through [PAD{pad_count}]")
+            for i in range(1, pad_count + 1):
+                tokens.append(bytes(f"[PAD{i}]", encoding="utf-8"))
+                scores.append(-1000.0)
+                toktypes.append(SentencePieceTokenTypes.UNUSED)
+
+        self.gguf_writer.add_tokenizer_model("t5")
+        self.gguf_writer.add_tokenizer_pre("default")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_scores(scores)
+        self.gguf_writer.add_token_types(toktypes)
+        self.gguf_writer.add_add_space_prefix(add_prefix)
+        self.gguf_writer.add_remove_extra_whitespaces(remove_whitespaces)
+        if precompiled_charsmap:
+            self.gguf_writer.add_precompiled_charsmap(precompiled_charsmap)
+
+        special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+        self.gguf_writer.add_add_bos_token(False)
+        self.gguf_writer.add_add_eos_token(True)
+
+    def set_gguf_parameters(self):
+        if (n_ctx := self.find_hparam(["n_positions"], optional=True)) is None:
+            logger.warning("Couldn't find context length in config.json, assuming default value of 512")
+            n_ctx = 512
+        self.gguf_writer.add_context_length(n_ctx)
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["d_ff"])
+        self.gguf_writer.add_block_count(self.hparams["num_layers"])
+        self.gguf_writer.add_head_count(self.hparams["num_heads"])
+        self.gguf_writer.add_key_length(self.hparams["d_kv"])
+        self.gguf_writer.add_value_length(self.hparams["d_kv"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_relative_attn_buckets_count(self.hparams["relative_attention_num_buckets"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        del bid  # unused
+
+        # T5 based models contain shared token embeddings tensors saved randomly as either "encoder.embed_tokens.weight",
+        # "decoder.embed_tokens.weight" or "shared.weight" tensor. In some models there are even multiple of them stored
+        # in the safetensors files. We use the first tensor from these three as the token embeddings for both encoder
+        # and decoder and ignore the remaining ones.
+        if name in ["decoder.embed_tokens.weight", "encoder.embed_tokens.weight", "shared.weight"]:
+            if not self.shared_token_embeddings_found:
+                name = "shared.weight"
+                self.shared_token_embeddings_found = True
+            else:
+                logger.debug(f"Skipping shared tensor {name!r} in safetensors so that convert can end normally.")
+                return []
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
 @Model.register("JAISLMHeadModel")
 class JaisModel(Model):
     model_arch = gguf.MODEL_ARCH.JAIS
@@ -3595,6 +3998,217 @@ class ChatGLMModel(Model):
         name = name.removeprefix("transformer.")
         return [(self.map_tensor_name(name), data_torch)]
 
+
+@Model.register("NemotronForCausalLM")
+class NemotronModel(Model):
+    model_arch = gguf.MODEL_ARCH.NEMOTRON
+
+    def set_vocab(self):
+        self._set_vocab_sentencepiece()
+        self.gguf_writer.add_pad_token_id(0)
+        self.gguf_writer.add_unk_token_id(1)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+
+        f_norm_eps = self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon", "norm_eps"])
+        self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+
+        # * Partial RoPE
+        rot_pct = self.find_hparam(["partial_rotary_factor", "rope_pct", "rope_percent"])
+        n_embd = self.find_hparam(["hidden_size", "n_embd"])
+        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
+
+        # * RopeScaling for Nemotron
+        if "rope_scaling" not in self.hparams or self.hparams["rope_scaling"] is None:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
+        else:
+            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+            self.gguf_writer.add_rope_scaling_factor(self.hparams["factor"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # * Adding +1 to LayerNorm's weights here to implement layernorm1p w/o changing anything on the GGML engine side
+        #   model.layers.{l}.input_layernorm.weight
+        #   model.layers.{l}.post_attention_layernorm.weight
+        #   model.norm.weight
+        if name.endswith("norm.weight"):
+            data_torch = data_torch + 1
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+
+@Model.register("ExaoneForCausalLM")
+class ExaoneModel(Model):
+    model_arch = gguf.MODEL_ARCH.EXAONE
+
+    def set_gguf_parameters(self):
+        hparams = self.hparams
+
+        assert (hparams["activation_function"] == "silu")
+
+        max_position_embeddings = hparams["max_position_embeddings"]
+        embed_dim = hparams["hidden_size"]
+        num_heads = hparams["num_attention_heads"]
+        num_kv_heads = hparams.get("num_key_value_heads", num_heads)
+        layer_norm_eps = hparams["layer_norm_epsilon"]
+        intermediate_size = hparams["intermediate_size"] if "intermediate_size" in hparams else 4 * embed_dim
+        num_layers = hparams["num_layers"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct attentino_dropout is 0.0
+        # attention_dropout_rate = hparams["attention_dropout"]
+        # ignore for now as EXAONE-3.0-7.8B-Instruct embed_dropout is 0.0
+        # embed_dropout_rate = hparams["embed_dropout"]
+        self.gguf_writer.add_embedding_length(embed_dim)
+        self.gguf_writer.add_head_count(num_heads)
+        self.gguf_writer.add_head_count_kv(num_kv_heads)
+        self.gguf_writer.add_context_length(max_position_embeddings)
+        self.gguf_writer.add_layer_norm_rms_eps(layer_norm_eps)
+        self.gguf_writer.add_feed_forward_length(intermediate_size)
+        self.gguf_writer.add_block_count(num_layers)
+        self.gguf_writer.add_file_type(self.ftype)
+
+        if (rope_theta := self.hparams.get("rope_theta")) is not None:
+            self.gguf_writer.add_rope_freq_base(rope_theta)
+        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"], optional=True)
+        rotary_factor = rotary_factor if rotary_factor is not None else 1.0
+        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        if hparams.get("rope_scaling") is not None and "factor" in hparams["rope_scaling"]:
+            if hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(hparams["rope_scaling"]["factor"])
+
+    def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+        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.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)
+                low_freq_factor = rope_scaling.get("low_freq_factor", 1.0)
+                high_freq_factor = rope_scaling.get("high_freq_factor", 4.0)
+                old_context_len = self.hparams.get("original_max_position_embeddings", 8192)
+
+                low_freq_wavelen = old_context_len / low_freq_factor
+                high_freq_wavelen = old_context_len / high_freq_factor
+                assert low_freq_wavelen != high_freq_wavelen
+
+                rope_factors = []
+                for freq in freqs:
+                    wavelen = 2 * math.pi / freq
+                    if wavelen < high_freq_wavelen:
+                        rope_factors.append(1)
+                    elif wavelen > low_freq_wavelen:
+                        rope_factors.append(factor)
+                    else:
+                        smooth = (old_context_len / wavelen - low_freq_factor) / (high_freq_factor - low_freq_factor)
+                        rope_factors.append(1 / ((1 - smooth) / factor + smooth))
+
+                yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.tensor(rope_factors, dtype=torch.float32))
+
+
+@Model.register("GraniteForCausalLM")
+class GraniteModel(LlamaModel):
+    """Conversion for IBM's GraniteForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE
+
+    def set_gguf_parameters(self):
+        """Granite uses standard llama parameters with the following differences:
+
+        - No head_dim support
+        - New multiplier params:
+            - attention_scale
+            - embedding_scale
+            - residual_scale
+        - logits_scaling
+        """
+        if head_dim := self.hparams.pop("head_dim", None):
+            logger.warning("Ignoring head_dim (%s) from config for Granite", head_dim)
+        super().set_gguf_parameters()
+        # NOTE: Convert _multiplier params to _scale params for naming
+        #   consistency
+        if attention_scale := self.hparams.get("attention_multiplier"):
+            self.gguf_writer.add_attention_scale(attention_scale)
+            logger.info("gguf: (granite) attention_scale = %s", attention_scale)
+        if embedding_scale := self.hparams.get("embedding_multiplier"):
+            self.gguf_writer.add_embedding_scale(embedding_scale)
+            logger.info("gguf: (granite) embedding_scale = %s", embedding_scale)
+        if residual_scale := self.hparams.get("residual_multiplier"):
+            self.gguf_writer.add_residual_scale(residual_scale)
+            logger.info("gguf: (granite) residual_scale = %s", residual_scale)
+        if logits_scale := self.hparams.get("logits_scaling"):
+            self.gguf_writer.add_logit_scale(logits_scale)
+            logger.info("gguf: (granite) logits_scale = %s", logits_scale)
+
+
+@Model.register("GraniteMoeForCausalLM")
+class GraniteMoeModel(GraniteModel):
+    """Conversion for IBM's GraniteMoeForCausalLM"""
+    model_arch = gguf.MODEL_ARCH.GRANITE_MOE
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        """In modeling_granitemoe, the JetMoe implementation of parallel experts
+        is used. This essentially merges w1 and w3 into a single tensor with 2x
+        the hidden size that is then split during forward. To keep compatibility
+        with existing mixtral support, we pull them apart here.
+        """
+
+        if name.endswith("block_sparse_moe.input_linear.weight"):
+            ffn_dim = self.hparams["intermediate_size"]
+            assert data_torch.shape[-2] == 2 * ffn_dim, "Merged FFN tensor size must be 2 * intermediate_size"
+            gate, up = data_torch[..., :ffn_dim, :], data_torch[..., ffn_dim:, :]
+            return [
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid), gate),
+                (self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid), up),
+            ]
+
+        return super().modify_tensors(data_torch, name, bid)
+
+
+@Model.register("ChameleonForConditionalGeneration")
+@Model.register("ChameleonForCausalLM")  # obsolete
+class ChameleonModel(Model):
+    model_arch = gguf.MODEL_ARCH.CHAMELEON
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # ignore image tokenizer for now
+        # TODO: remove this once image support is implemented for Chameleon
+        if name.startswith("model.vqmodel"):
+            return []
+
+        n_head = self.hparams["num_attention_heads"]
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        hidden_dim = self.hparams.get("hidden_size")
+
+        if name.endswith(("q_proj.weight", "q_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
+        if name.endswith(("k_proj.weight", "k_proj.bias")):
+            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
+        if name.endswith(("q_norm.weight", "q_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
+        if name.endswith(("k_norm.weight", "k_norm.bias")):
+            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
+
+        return [(self.map_tensor_name(name), data_torch)]
+
+    # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
+    @staticmethod
+    def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
+        head_dim = hidden_dim // n_heads
+        data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
+        data_torch = data_torch.repeat_interleave(n_heads, 0)
+        return data_torch
+
+
 ###### CONVERSION LOGIC ######
 
 
@@ -3676,8 +4290,8 @@ 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",
@@ -3764,6 +4378,8 @@ 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
@@ -80,6 +81,7 @@ models = [
     {"name": "qwen2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/Qwen/Qwen1.5-7B", },
     {"name": "olmo",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/allenai/OLMo-1.7-7B-hf", },
     {"name": "dbrx",           "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/databricks/dbrx-base", },
+    {"name": "jina-v1-en",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-reranker-v1-tiny-en", },
     {"name": "jina-v2-en",     "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-en", }, # WPM!
     {"name": "jina-v2-es",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-es", },
     {"name": "jina-v2-de",     "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/jinaai/jina-embeddings-v2-base-de", },
@@ -94,6 +96,11 @@ models = [
     {"name": "codeshell",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/WisdomShell/CodeShell-7B", },
     {"name": "tekken",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mistralai/Mistral-Nemo-Base-2407", },
     {"name": "smollm",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/HuggingFaceTB/SmolLM-135M", },
+    {'name': "bloom",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigscience/bloom", },
+    {'name': "gpt3-finnish",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/TurkuNLP/gpt3-finnish-small", },
+    {"name": "exaone",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/LGAI-EXAONE/EXAONE-3.0-7.8B-Instruct", },
+    {"name": "phi-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/microsoft/phi-2", },
+    {"name": "chameleon",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/facebook/chameleon-7b", },
 ]
 
 
@@ -122,12 +129,27 @@ 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_llama_ggml_to_gguf.py

@@ -116,7 +116,7 @@ class Tensor:
         assert quant is not None, 'Unknown tensor type'
         (blksize, tysize) = quant
         offset += 12
-        self.dtype= dtype
+        self.dtype= gguf.GGMLQuantizationType(dtype)
         self.dims = struct.unpack(f'<{n_dims}I', data[offset:offset + (4 * n_dims)])
         offset += 4 * n_dims
         self.name = bytes(data[offset:offset + name_len])

convert_lora_to_gguf.py

@@ -331,6 +331,10 @@ if __name__ == '__main__':
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
                 super().set_gguf_parameters()
 
+            def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
+                # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
+                return ()
+
             def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                 tensor_map: dict[str, PartialLoraTensor] = {}
 
@@ -363,7 +367,13 @@ 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()

docs/android.md

@@ -2,55 +2,82 @@
 # Android
 
 ## Build on Android using Termux
-[Termux](https://github.com/termux/termux-app#installation) is a method to execute `llama.cpp` on an Android device (no root required).
-```
-apt update && apt upgrade -y
-apt install git make cmake
-```
 
-It's recommended to move your model inside the `~/` directory for best performance:
-```
-cd storage/downloads
-mv model.gguf ~/
-```
+[Termux](https://termux.dev/en/) is an Android terminal emulator and Linux environment app (no root required). As of writing, Termux is available experimentally in the Google Play Store; otherwise, it may be obtained directly from the project repo or on F-Droid.
 
-[Get the code](https://github.com/ggerganov/llama.cpp#get-the-code) & [follow the Linux build instructions](https://github.com/ggerganov/llama.cpp#build) to build `llama.cpp`.
-
-## Building the Project using Android NDK
-Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
-
-Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
-```
-$ mkdir build-android
-$ cd build-android
-$ export NDK=<your_ndk_directory>
-$ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
-$ make
-```
-
-Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
-
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+With Termux, you can install and run `llama.cpp` as if the environment were Linux. Once in the Termux shell:
 
 ```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$ apt update && apt upgrade -y
+$ apt install git cmake
 ```
 
-Now, you can start chatting:
+Then, follow the [build instructions](https://github.com/ggerganov/llama.cpp/blob/master/docs/build.md), specifically for CMake.
+
+Once the binaries are built, download your model of choice (e.g., from Hugging Face). It's recommended to place it in the `~/` directory for best performance:
+
 ```
-$cd /data/data/com.termux/files/home/bin
-$./llama-cli -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
+$ curl -L {model-url} -o ~/{model}.gguf
 ```
 
-Here's a demo of an interactive session running on Pixel 5 phone:
+Then, if you are not already in the repo directory, `cd` into `llama.cpp` and:
+
+```
+$ ./build/bin/llama-simple -m ~/{model}.gguf -c {context-size} -p "{your-prompt}"
+```
+
+Here, we show `llama-simple`, but any of the executables under `examples` should work, in theory. Be sure to set `context-size` to a reasonable number (say, 4096) to start with; otherwise, memory could spike and kill your terminal.
+
+To see what it might look like visually, here's an old demo of an interactive session running on a Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
+
+## Cross-compile using Android NDK
+It's possible to build `llama.cpp` for Android on your host system via CMake and the Android NDK. If you are interested in this path, ensure you already have an environment prepared to cross-compile programs for Android (i.e., install the Android SDK). Note that, unlike desktop environments, the Android environment ships with a limited set of native libraries, and so only those libraries are available to CMake when building with the Android NDK (see: https://developer.android.com/ndk/guides/stable_apis.)
+
+Once you're ready and have cloned `llama.cpp`, invoke the following in the project directory:
+
+```
+$ cmake \
+  -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DCMAKE_C_FLAGS="-march=armv8.7a" \
+  -DCMAKE_CXX_FLAGS="-march=armv8.7a" \
+  -DGGML_OPENMP=OFF \
+  -DGGML_LLAMAFILE=OFF \
+  -B build-android
+```
+
+Notes:
+  - While later versions of Android NDK ship with OpenMP, it must still be installed by CMake as a dependency, which is not supported at this time
+  - `llamafile` does not appear to support Android devices (see: https://github.com/Mozilla-Ocho/llamafile/issues/325)
+
+The above command should configure `llama.cpp` with the most performant options for modern devices. Even if your device is not running `armv8.7a`, `llama.cpp` includes runtime checks for available CPU features it can use.
+
+Feel free to adjust the Android ABI for your target. Once the project is configured:
+
+```
+$ cmake --build build-android --config Release -j{n}
+$ cmake --install build-android --prefix {install-dir} --config Release
+```
+
+After installing, go ahead and download the model of your choice to your host system. Then:
+
+```
+$ adb shell "mkdir /data/local/tmp/llama.cpp"
+$ adb push {install-dir} /data/local/tmp/llama.cpp/
+$ adb push {model}.gguf /data/local/tmp/llama.cpp/
+$ adb shell
+```
+
+In the `adb shell`:
+
+```
+$ cd /data/local/tmp/llama.cpp
+$ LD_LIBRARY_PATH=lib ./bin/llama-simple -m {model}.gguf -c {context-size} -p "{your-prompt}"
+```
+
+That's it!
+
+Be aware that Android will not find the library path `lib` on its own, so we must specify `LD_LIBRARY_PATH` in order to run the installed executables. Android does support `RPATH` in later API levels, so this could change in the future. Refer to the previous section for information about `context-size` (very important!) and running other `examples`.

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

@@ -20,17 +20,13 @@
 **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.
 
 ### Llama.cpp + SYCL
 
-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.
+The llama.cpp SYCL backend is designed to support **Intel GPU** firstly. Based on the cross-platform feature of SYCL, it also supports other vendor GPUs: Nvidia and AMD.
 
 ## Recommended Release
 
@@ -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.
@@ -80,7 +80,14 @@ The following release is verified with good quality:
 
 ### Intel GPU
 
-**Verified devices**
+SYCL backend supports Intel GPU Family:
+
+- Intel Data Center Max Series
+- Intel Flex Series, Arc Series
+- Intel Built-in Arc GPU
+- Intel iGPU in Core CPU (11th Generation Core CPU and newer, refer to [oneAPI supported GPU](https://www.intel.com/content/www/us/en/developer/articles/system-requirements/intel-oneapi-base-toolkit-system-requirements.html#inpage-nav-1-1)).
+
+#### Verified devices
 
 | Intel GPU                     | Status  | Verified Model                        |
 |-------------------------------|---------|---------------------------------------|
@@ -88,7 +95,7 @@ The following release is verified with good quality:
 | Intel Data Center Flex Series | Support | Flex 170                              |
 | Intel Arc Series              | Support | Arc 770, 730M, Arc A750               |
 | Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
-| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
+| Intel iGPU                    | Support | iGPU in 13700k, i5-1250P, i7-1260P, i7-1165G7 |
 
 *Notes:*
 
@@ -104,10 +111,18 @@ The following release is verified with good quality:
 
 **Verified devices**
 
-| Nvidia GPU               | Status  | Verified Model |
-|--------------------------|---------|----------------|
-| Ampere Series            | Support | A100, A4000    |
-| Ampere Series *(Mobile)* | Support | RTX 40 Series  |
+| Nvidia GPU               | Status    | Verified Model |
+|--------------------------|-----------|----------------|
+| Ampere Series            | Supported | A100, A4000    |
+| Ampere Series *(Mobile)* | Supported | RTX 40 Series  |
+
+| AMD GPU                  | Status       | Verified Model |
+|--------------------------|--------------|----------------|
+| Radeon Pro               | Experimental | W6800          |
+| Radeon RX                | Experimental | 6700 XT        |
+
+Note: AMD GPU support is highly experimental and is incompatible with F16.
+Additionally, it only supports GPUs with a sub_group_size (warp size) of 32.
 
 ## Docker
 The docker build option is currently limited to *intel GPU* targets.
@@ -179,6 +194,10 @@ Platform #0: Intel(R) OpenCL HD Graphics
 
 In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
 
+- **AMD GPU**
+
+To target AMD GPUs with SYCL, the ROCm stack must be installed first.
+
 2. **Install Intel® oneAPI Base toolkit**
 
 - **For Intel GPU**
@@ -189,7 +208,7 @@ 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**
 
@@ -205,6 +224,19 @@ cmake -B buildWithCublas -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENAB
 cmake --build buildWithCublas --config Release
 ```
 
+- **Adding support to AMD GPUs**
+
+**oneAPI Plugin**: In order to enable SYCL support on AMD GPUs, please install the [Codeplay oneAPI Plugin for AMD GPUs](https://developer.codeplay.com/products/oneapi/amd/download). As with Nvidia GPUs, the user should also make sure the plugin version matches the installed base toolkit.
+
+**oneMKL for rocBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* doesn't contain the rocBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *rocBLAS* backend enabled is thus required to run it on AMD GPUs.
+
+```sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+# Find your HIPTARGET with rocminfo, under the key 'Name:'
+cmake -B buildWithrocBLAS -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_ROCBLAS_BACKEND=ON -DHIPTARGETS=${HIPTARGET} -DTARGET_DOMAINS=blas
+cmake --build buildWithrocBLAS --config Release
+```
 
 3. **Verify installation and environment**
 
@@ -216,33 +248,48 @@ sycl-ls
 
 - **Intel GPU**
 
-When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`ext_oneapi_level_zero:gpu:0`] in the sample output below:
+When targeting an intel GPU, the user should expect one or more level-zero devices among the available SYCL devices. Please make sure that at least one GPU is present, for instance [`level_zero:gpu`] in the sample output below:
 
 ```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
-[opencl:cpu:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
-[opencl:gpu:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
-[ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
+[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.10.0.17_160000]
+[opencl:cpu][opencl:1] Intel(R) OpenCL, 13th Gen Intel(R) Core(TM) i7-13700K OpenCL 3.0 (Build 0) [2023.16.10.0.17_160000]
+[opencl:gpu][opencl:2] Intel(R) OpenCL Graphics, Intel(R) Arc(TM) A770 Graphics OpenCL 3.0 NEO  [23.30.26918.50]
+[level_zero:gpu][level_zero:0] Intel(R) Level-Zero, Intel(R) Arc(TM) A770 Graphics 1.3 [1.3.26918]
 ```
 
 - **Nvidia GPU**
 
-Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`cuda:gpu`] as below:
+
 ```
-[opencl:acc:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
-[opencl:cpu:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
-[ext_oneapi_cuda:gpu:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.2]
+[opencl:acc][opencl:0] Intel(R) FPGA Emulation Platform for OpenCL(TM), Intel(R) FPGA Emulation Device OpenCL 1.2  [2023.16.12.0.12_195853.xmain-hotfix]
+[opencl:cpu][opencl:1] Intel(R) OpenCL, Intel(R) Xeon(R) Gold 6326 CPU @ 2.90GHz OpenCL 3.0 (Build 0) [2023.16.12.0.12_195853.xmain-hotfix]
+[cuda:gpu][cuda:0] NVIDIA CUDA BACKEND, NVIDIA A100-PCIE-40GB 8.0 [CUDA 12.5]
+```
+
+- **AMD GPU**
+
+For AMD GPUs we should expect at least one SYCL-HIP device [`hip:gpu`]:
+
+```
+[opencl:cpu][opencl:0] Intel(R) OpenCL, 12th Gen Intel(R) Core(TM) i9-12900K OpenCL 3.0 (Build 0) [2024.18.6.0.02_160000]
+[hip:gpu][hip:0] AMD HIP BACKEND, AMD Radeon PRO W6800 gfx1030 [HIP 60140.9]
 ```
 
 ### II. Build llama.cpp
 
 #### Intel GPU
+
+```
+./examples/sycl/build.sh
+```
+
+or
+
 ```sh
 # Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
-# Build LLAMA with MKL BLAS acceleration for intel GPU
-
 # Option 1: Use FP32 (recommended for better performance in most cases)
 cmake -B build -DGGML_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
@@ -254,6 +301,7 @@ cmake --build build --config Release -j -v
 ```
 
 #### Nvidia GPU
+
 ```sh
 # Export relevant ENV variables
 export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
@@ -271,28 +319,49 @@ cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -
 
 # build all binary
 cmake --build build --config Release -j -v
+```
 
+#### AMD GPU
+
+```sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithrocBLAS/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithrocBLAS/include:$CPLUS_INCLUDE_DIR
+
+# Build LLAMA with rocBLAS acceleration through SYCL
+
+## AMD
+# Use FP32, FP16 is not supported
+# Find your GGML_SYCL_HIP_TARGET with rocminfo, under the key 'Name:'
+cmake -B build -DGGML_SYCL=ON -DGGML_SYCL_TARGET=AMD -DGGML_SYCL_HIP_TARGET=${GGML_SYCL_HIP_TARGET} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# build all binary
+cmake --build build --config Release -j -v
 ```
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
 You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 ```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/llama-ls-sycl-device
 ```
+
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
 ```
 found 2 SYCL devices:
@@ -304,12 +373,37 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 ```
 
+#### Choose level-zero devices
 
-4. Launch inference
+|Chosen Device ID|Setting|
+|-|-|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
+
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+- Use device 0:
+
+```sh
+./examples/sycl/run-llama2.sh 0
+```
+- Use multiple devices:
+
+```sh
+./examples/sycl/run-llama2.sh
+```
+
+2. Command line
+Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device target specified by the user.
+- Single device: Use one device assigned by user. Default device id is 0.
 - Multiple devices: Automatically choose the devices with the same backend.
 
 In two device selection modes, the default SYCL backend is level_zero, you can choose other backend supported by SYCL by setting environment variable ONEAPI_DEVICE_SELECTOR.
@@ -326,11 +420,6 @@ Examples:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm none -mg 0
 ```
-or run by script:
-
-```sh
-./examples/sycl/run_llama2.sh 0
-```
 
 - Use multiple devices:
 
@@ -338,12 +427,6 @@ or run by script:
 ZES_ENABLE_SYSMAN=1 ./build/bin/llama-cli -m models/llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33 -sm layer
 ```
 
-Otherwise, you can run the script:
-
-```sh
-./examples/sycl/run_llama2.sh
-```
-
 *Notes:*
 
 - Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
@@ -390,7 +473,7 @@ c. Verify installation
 In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
-sycl-ls
+sycl-ls.exe
 ```
 
 There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
@@ -411,6 +494,18 @@ b. The new Visual Studio will install Ninja as default. (If not, please install
 
 ### II. Build llama.cpp
 
+You could download the release package for Windows directly, which including binary files and depended oneAPI dll files.
+
+Choose one of following methods to build from source code.
+
+1. Script
+
+```sh
+.\examples\sycl\win-build-sycl.bat
+```
+
+2. CMake
+
 On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
@@ -425,12 +520,8 @@ cmake -B build -G "Ninja" -DGGML_SYCL=ON -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPI
 cmake --build build --config Release -j
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
-.\examples\sycl\win-build-sycl.bat
-```
-
 Or, use CMake presets to build:
+
 ```sh
 cmake --preset x64-windows-sycl-release
 cmake --build build-x64-windows-sycl-release -j --target llama-cli
@@ -442,7 +533,9 @@ cmake --preset x64-windows-sycl-debug
 cmake --build build-x64-windows-sycl-debug -j --target llama-cli
 ```
 
-Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
+3. Visual Studio
+
+You can use Visual Studio to open llama.cpp folder as a CMake project. Choose the sycl CMake presets (`x64-windows-sycl-release` or `x64-windows-sycl-debug`) before you compile the project.
 
 *Notes:*
 
@@ -450,23 +543,25 @@ Or, you can use Visual Studio to open llama.cpp folder as a CMake project. Choos
 
 ### III. Run the inference
 
-1. Retrieve and prepare model
+#### Retrieve and prepare model
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You can refer to the general [*Prepare and Quantize*](README.md#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
-2. Enable oneAPI running environment
+##### Check device
+
+1. Enable oneAPI running environment
 
 On the oneAPI command line window, run the following and step into the llama.cpp directory:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+2. List devices information
 
 Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
-build\bin\ls-sycl-device.exe
+build\bin\llama-ls-sycl-device.exe
 ```
 
 This command will only display the selected backend that is supported by SYCL. The default backend is level_zero. For example, in a system with 2 *intel GPU* it would look like the following:
@@ -479,9 +574,27 @@ found 2 SYCL devices:
 | 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
 
 ```
+#### Choose level-zero devices
 
+|Chosen Device ID|Setting|
+|-|-|
+|0|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|1|`set ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
+|0 & 1|`set ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 
-4. Launch inference
+#### Execute
+
+Choose one of following methods to run.
+
+1. Script
+
+```
+examples\sycl\win-run-llama2.bat
+```
+
+2. Command line
+
+Launch inference
 
 There are two device selection modes:
 
@@ -508,11 +621,7 @@ build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website ca
 ```
 build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be done in 10 simple steps:\nStep 1:" -n 400 -e -ngl 33 -s 0 -sm layer
 ```
-Otherwise, run the following wrapper script:
 
-```
-.\examples\sycl\win-run-llama2.bat
-```
 
 Note:
 
@@ -526,17 +635,18 @@ Or
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
+
 ## Environment Variable
 
 #### Build
 
-| Name               | Value                             | Function                                    |
-|--------------------|-----------------------------------|---------------------------------------------|
-| GGML_SYCL          | ON (mandatory)                    | Enable build with SYCL code path.           |
-| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
-| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)* | Enable FP16 build with SYCL code path.      |
-| CMAKE_C_COMPILER   | icx                               | Set *icx* compiler for SYCL code path.      |
-| CMAKE_CXX_COMPILER | icpx *(Linux)*, icx *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
+| Name               | Value                                 | Function                                    |
+|--------------------|---------------------------------------|---------------------------------------------|
+| GGML_SYCL          | ON (mandatory)                        | Enable build with SYCL code path.<br>FP32 path - recommended for better perforemance than FP16 on quantized model|
+| GGML_SYCL_TARGET   | INTEL *(default)* \| NVIDIA \| AMD    | Set the SYCL target device type.            |
+| GGML_SYCL_F16      | OFF *(default)* \|ON *(optional)*     | Enable FP16 build with SYCL code path.      |
+| CMAKE_C_COMPILER   | `icx` *(Linux)*, `icx/cl` *(Windows)* | Set `icx` compiler for SYCL code path.      |
+| CMAKE_CXX_COMPILER | `icpx` *(Linux)*, `icx` *(Windows)*   | Set `icpx/icx` compiler for SYCL code path. |
 
 #### Runtime
 
@@ -572,9 +682,26 @@ use 1 SYCL GPUs: [0] with Max compute units:512
   ```
   Otherwise, please double-check the GPU driver installation steps.
 
+- Can I report Ollama issue on Intel GPU to llama.cpp SYCL backend?
+
+  No. We can't support Ollama issue directly, because we aren't familiar with Ollama.
+
+  Sugguest reproducing on llama.cpp and report similar issue to llama.cpp. We will surpport it.
+
+  It's same for other projects including llama.cpp SYCL backend.
+
+- Meet issue: `Native API failed. Native API returns: -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -6 (PI_ERROR_OUT_OF_HOST_MEMORY) -999 (UNKNOWN PI error)` or `failed to allocate SYCL0 buffer`
+
+  Device Memory is not enough.
+
+  |Reason|Solution|
+  |-|-|
+  |Default Context is too big. It leads to more memory usage.|Set `-c 8192` or smaller value.|
+  |Model is big and require more memory than device's.|Choose smaller quantized model, like Q5 -> Q4;<br>Use more than one devices to load model.|
+
 ### **GitHub contribution**:
 Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
 ## TODO
 
-- Support row layer split for multiple card runs.
+- NA

docs/build.md

@@ -198,6 +198,8 @@ The following compilation options are also available to tweak performance:
 
 ### MUSA
 
+This provides GPU acceleration using the MUSA cores of your Moore Threads MTT GPU. Make sure to have the MUSA SDK installed. You can download it from here: [MUSA SDK](https://developer.mthreads.com/sdk/download/musa).
+
 - Using `make`:
   ```bash
   make GGML_MUSA=1
@@ -209,6 +211,12 @@ The following compilation options are also available to tweak performance:
   cmake --build build --config Release
   ```
 
+The environment variable [`MUSA_VISIBLE_DEVICES`](https://docs.mthreads.com/musa-sdk/musa-sdk-doc-online/programming_guide/Z%E9%99%84%E5%BD%95/) can be used to specify which GPU(s) will be used.
+
+The environment variable `GGML_CUDA_ENABLE_UNIFIED_MEMORY=1` can be used to enable unified memory in Linux. This allows swapping to system RAM instead of crashing when the GPU VRAM is exhausted.
+
+Most of the compilation options available for CUDA should also be available for MUSA, though they haven't been thoroughly tested yet.
+
 ### hipBLAS
 
 This provides BLAS acceleration on HIP-supported AMD GPUs.
@@ -352,6 +360,37 @@ 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

@@ -19,8 +19,11 @@ Additionally, there the following images, similar to the above:
 - `ghcr.io/ggerganov/llama.cpp:full-rocm`: Same as `full` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:light-rocm`: Same as `light` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
 - `ghcr.io/ggerganov/llama.cpp:server-rocm`: Same as `server` but compiled with ROCm support. (platforms: `linux/amd64`, `linux/arm64`)
+- `ghcr.io/ggerganov/llama.cpp:full-musa`: Same as `full` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:light-musa`: Same as `light` but compiled with MUSA support. (platforms: `linux/amd64`)
+- `ghcr.io/ggerganov/llama.cpp:server-musa`: Same as `server` but compiled with MUSA support. (platforms: `linux/amd64`)
 
-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, ROCm or MUSA library, you'll need to build the images locally for now).
 
 ## Usage
 
@@ -66,8 +69,8 @@ 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:
 
@@ -84,3 +87,37 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:full-cuda --run
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:light-cuda -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
 docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
 ```
+
+## Docker With MUSA
+
+Assuming one has the [mt-container-toolkit](https://developer.mthreads.com/musa/native) properly installed on Linux, `muBLAS` should be accessible inside the container.
+
+## Building Docker locally
+
+```bash
+docker build -t local/llama.cpp:full-musa -f .devops/full-musa.Dockerfile .
+docker build -t local/llama.cpp:light-musa -f .devops/llama-cli-musa.Dockerfile .
+docker build -t local/llama.cpp:server-musa -f .devops/llama-server-musa.Dockerfile .
+```
+
+You may want to pass in some different `ARGS`, depending on the MUSA environment supported by your container host, as well as the GPU architecture.
+
+The defaults are:
+
+- `MUSA_VERSION` set to `rc3.1.0`
+
+The resulting images, are essentially the same as the non-MUSA images:
+
+1. `local/llama.cpp:full-musa`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization.
+2. `local/llama.cpp:light-musa`: This image only includes the main executable file.
+3. `local/llama.cpp:server-musa`: This image only includes the server executable file.
+
+## Usage
+
+After building locally, Usage is similar to the non-MUSA examples, but you'll need to set `mthreads` as default Docker runtime. This can be done by executing `(cd /usr/bin/musa && sudo ./docker setup $PWD)` and verifying the changes by executing `docker info | grep mthreads` on the host machine. You will also want to use the `--n-gpu-layers` flag.
+
+```bash
+docker run -v /path/to/models:/models local/llama.cpp:full-musa --run -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run -v /path/to/models:/models local/llama.cpp:light-musa -m /models/7B/ggml-model-q4_0.gguf -p "Building a website can be done in 10 simple steps:" -n 512 --n-gpu-layers 1
+docker run -v /path/to/models:/models local/llama.cpp:server-musa -m /models/7B/ggml-model-q4_0.gguf --port 8000 --host 0.0.0.0 -n 512 --n-gpu-layers 1
+```

examples/CMakeLists.txt

@@ -16,7 +16,6 @@ else()
     add_subdirectory(baby-llama)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)
-    add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(eval-callback)

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

@@ -1,49 +1,28 @@
+#include "arg.h"
 #include "common.h"
+#include "log.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);
-
-    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");
+static void print_usage(int, char ** argv) {
+    LOG("\nexample usage:\n");
+    LOG("\n    %s -m model.gguf -c 2048 -b 2048 -ub 512 -npp 128,256,512 -ntg 128,256 -npl 1,2,4,8,16,32 [-pps]\n", argv[0]);
+    LOG("\n");
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_BENCH, print_usage)) {
         return 1;
     }
 
+    common_init();
+
     int is_pp_shared = params.is_pp_shared;
 
     std::vector<int> n_pp = params.n_pp;
@@ -57,7 +36,7 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = common_model_params_to_llama(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
@@ -66,7 +45,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     // ensure enough sequences are available
     ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
@@ -100,7 +79,7 @@ int main(int argc, char ** argv) {
 
             const int ret = llama_decode(ctx, batch_view);
             if (ret != 0) {
-                LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
+                LOG_ERR("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
 
@@ -113,21 +92,22 @@ int main(int argc, char ** argv) {
     // warm up
     {
         for (int i = 0; i < 16; ++i) {
-            llama_batch_add(batch, 0, i, { 0 }, false);
+            common_batch_add(batch, 0, i, { 0 }, false);
         }
 
         if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            LOG_ERR("%s: llama_decode() failed\n", __func__);
             return 1;
         }
     }
 
-    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("\n");
+        LOG("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, flash_attn = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+        LOG("\n");
+        LOG("|%6s | %6s | %4s | %6s | %8s | %8s | %8s | %8s | %8s | %8s |\n", "PP", "TG", "B", "N_KV", "T_PP s", "S_PP t/s", "T_TG s", "S_TG t/s", "T s", "S t/s");
+        LOG("|%6s-|-%6s-|-%4s-|-%6s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|-%8s-|\n", "------", "------", "----", "------", "--------", "--------", "--------", "--------", "--------", "--------");
+    }
 
     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) {
@@ -142,11 +122,11 @@ int main(int argc, char ** argv) {
                     continue;
                 }
 
-                llama_batch_clear(batch);
+                common_batch_clear(batch);
 
                 for (int i = 0; i < pp; ++i) {
                     for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        llama_batch_add(batch, 0, i, { j }, false);
+                        common_batch_add(batch, 0, i, { j }, false);
                     }
                 }
                 batch.logits[batch.n_tokens - 1] = true;
@@ -156,7 +136,7 @@ int main(int argc, char ** argv) {
                 llama_kv_cache_clear(ctx);
 
                 if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                    LOG_TEE("%s: llama_decode() failed\n", __func__);
+                    LOG_ERR("%s: llama_decode() failed\n", __func__);
                     return 1;
                 }
 
@@ -171,14 +151,14 @@ int main(int argc, char ** argv) {
                 const auto t_tg_start = ggml_time_us();
 
                 for (int i = 0; i < tg; ++i) {
-                    llama_batch_clear(batch);
+                    common_batch_clear(batch);
 
                     for (int j = 0; j < pl; ++j) {
-                        llama_batch_add(batch, 0, pp + i, { j }, true);
+                        common_batch_add(batch, 0, pp + i, { j }, true);
                     }
 
                     if (!decode_helper(ctx, batch, ctx_params.n_batch)) {
-                        LOG_TEE("%s: llama_decode() failed\n", __func__);
+                        LOG_ERR("%s: llama_decode() failed\n", __func__);
                         return 1;
                     }
                 }
@@ -195,12 +175,22 @@ 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(
+                        "{\"n_kv_max\": %d, \"n_batch\": %d, \"n_ubatch\": %d, \"flash_attn\": %d, \"is_pp_shared\": %d, \"n_gpu_layers\": %d, \"n_threads\": %u, \"n_threads_batch\": %u, "
+                        "\"pp\": %d, \"tg\": %d, \"pl\": %d, \"n_kv\": %d, \"t_pp\": %f, \"speed_pp\": %f, \"t_tg\": %f, \"speed_tg\": %f, \"t\": %f, \"speed\": %f}\n",
+                        n_kv_max, params.n_batch, params.n_ubatch, params.flash_attn, params.is_pp_shared, params.n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch,
+                        pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed
+                    );
+                } else {
+                    LOG("|%6d | %6d | %4d | %6d | %8.3f | %8.2f | %8.3f | %8.2f | %8.3f | %8.2f |\n", pp, tg, pl, n_kv, t_pp, speed_pp, t_tg, speed_tg, t, speed);
+                }
             }
         }
     }
 
-    llama_print_timings(ctx);
+    LOG("\n");
+    llama_perf_context_print(ctx);
 
     llama_batch_free(batch);
 
@@ -209,7 +199,7 @@ int main(int argc, char ** argv) {
 
     llama_backend_free();
 
-    fprintf(stderr, "\n\n");
+    LOG("\n\n");
 
     return 0;
 }

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

@@ -1,31 +1,30 @@
+#include "arg.h"
 #include "common.h"
+#include "log.h"
 #include "llama.h"
 
 #include <algorithm>
-#include <cmath>
 #include <cstdio>
 #include <string>
 #include <vector>
 
-static void print_usage(int argc, char ** argv, const gpt_params & params) {
-    gpt_params_print_usage(argc, argv, params);
-
-    LOG_TEE("\nexample usage:\n");
-    LOG_TEE("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
-    LOG_TEE("\n");
+static void print_usage(int, char ** argv) {
+    LOG("\nexample usage:\n");
+    LOG("\n    %s -m model.gguf -p \"Hello my name is\" -n 32 -np 4\n", argv[0]);
+    LOG("\n");
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.prompt = "Hello my name is";
     params.n_predict = 32;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON, print_usage)) {
         return 1;
     }
 
+    common_init();
 
     // number of parallel batches
     int n_parallel = params.n_parallel;
@@ -40,57 +39,64 @@ int main(int argc, char ** argv) {
 
     // initialize the model
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(params);
+    llama_model_params model_params = common_model_params_to_llama(params);
 
     llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
 
     if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        LOG_ERR("%s: error: unable to load model\n" , __func__);
         return 1;
     }
 
     // tokenize the prompt
 
     std::vector<llama_token> tokens_list;
-    tokens_list = ::llama_tokenize(model, params.prompt, true);
+    tokens_list = common_tokenize(model, params.prompt, true);
 
     const int n_kv_req = tokens_list.size() + (n_predict - tokens_list.size())*n_parallel;
 
     // initialize the context
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
+    llama_context_params ctx_params = common_context_params_to_llama(params);
 
     ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_predict, n_parallel);
 
     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__);
+        LOG_ERR("%s: error: failed to create the llama_context\n" , __func__);
         return 1;
     }
 
     const int n_ctx = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_INF("\n%s: n_predict = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_predict, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
-        LOG_TEE("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
-        LOG_TEE("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
+        LOG_ERR("%s: error: n_kv_req (%d) > n_ctx, the required KV cache size is not big enough\n", __func__,  n_kv_req);
+        LOG_ERR("%s:        either reduce n_parallel or increase n_ctx\n", __func__);
         return 1;
     }
 
     // print the prompt token-by-token
 
-    fprintf(stderr, "\n");
+    LOG("\n");
 
     for (auto id : tokens_list) {
-        fprintf(stderr, "%s", llama_token_to_piece(ctx, id).c_str());
+        LOG("%s", common_token_to_piece(ctx, id).c_str());
     }
 
-    fflush(stderr);
-
     // create a llama_batch
     // we use this object to submit token data for decoding
     llama_batch batch = llama_batch_init(std::max(tokens_list.size(), (size_t) n_parallel), 0, n_parallel);
@@ -102,13 +108,13 @@ int main(int argc, char ** argv) {
 
     // evaluate the initial prompt
     for (size_t i = 0; i < tokens_list.size(); ++i) {
-        llama_batch_add(batch, tokens_list[i], i, seq_ids, false);
+        common_batch_add(batch, tokens_list[i], i, seq_ids, false);
     }
     GGML_ASSERT(batch.n_tokens == (int) tokens_list.size());
 
     if (llama_model_has_encoder(model)) {
         if (llama_encode(ctx, batch)) {
-            LOG_TEE("%s : failed to eval\n", __func__);
+            LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
 
@@ -117,15 +123,15 @@ int main(int argc, char ** argv) {
             decoder_start_token_id = llama_token_bos(model);
         }
 
-        llama_batch_clear(batch);
-        llama_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
+        common_batch_clear(batch);
+        common_batch_add(batch, decoder_start_token_id, 0, seq_ids, false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
     batch.logits[batch.n_tokens - 1] = true;
 
     if (llama_decode(ctx, batch) != 0) {
-        LOG_TEE("%s: llama_decode() failed\n", __func__);
+        LOG_ERR("%s: llama_decode() failed\n", __func__);
         return 1;
     }
 
@@ -136,7 +142,7 @@ int main(int argc, char ** argv) {
     //}
 
     if (n_parallel > 1) {
-        LOG_TEE("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
+        LOG("\n\n%s: generating %d sequences ...\n", __func__, n_parallel);
     }
 
     // main loop
@@ -155,7 +161,7 @@ int main(int argc, char ** argv) {
 
     while (n_cur <= n_predict) {
         // prepare the next batch
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         // sample the next token for each parallel sequence / stream
         for (int32_t i = 0; i < n_parallel; ++i) {
@@ -164,36 +170,14 @@ 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) {
                 i_batch[i] = -1;
-                LOG_TEE("\n");
+                LOG("\n");
                 if (n_parallel > 1) {
-                    LOG_TEE("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
+                    LOG_INF("%s: stream %d finished at n_cur = %d", __func__, i, n_cur);
                 }
 
                 continue;
@@ -201,16 +185,15 @@ int main(int argc, char ** argv) {
 
             // if there is only one stream, we print immediately to stdout
             if (n_parallel == 1) {
-                LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
-                fflush(stdout);
+                LOG("%s", common_token_to_piece(ctx, new_token_id).c_str());
             }
 
-            streams[i] += llama_token_to_piece(ctx, new_token_id);
+            streams[i] += common_token_to_piece(ctx, new_token_id);
 
             i_batch[i] = batch.n_tokens;
 
             // push this new token for next evaluation
-            llama_batch_add(batch, new_token_id, n_cur, { i }, true);
+            common_batch_add(batch, new_token_id, n_cur, { i }, true);
 
             n_decode += 1;
         }
@@ -224,32 +207,33 @@ int main(int argc, char ** argv) {
 
         // evaluate the current batch with the transformer model
         if (llama_decode(ctx, batch)) {
-            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            LOG_ERR("%s : failed to eval, return code %d\n", __func__, 1);
             return 1;
         }
     }
 
-    LOG_TEE("\n");
-
     if (n_parallel > 1) {
-        LOG_TEE("\n");
+        LOG("\n");
 
         for (int32_t i = 0; i < n_parallel; ++i) {
-            LOG_TEE("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
+            LOG("sequence %d:\n\n%s%s\n\n", i, params.prompt.c_str(), streams[i].c_str());
         }
     }
 
     const auto t_main_end = ggml_time_us();
 
-    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+    LOG_INF("%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("\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/CMakeLists.txt

@@ -1,6 +0,0 @@
-set(TARGET llama-bench-matmult)
-add_executable(${TARGET} benchmark-matmult.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
-target_include_directories(${TARGET} PRIVATE ../../common)
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/benchmark/benchmark-matmult.cpp

@@ -1,275 +0,0 @@
-#include "common.h"
-#include "ggml.h"
-
-#include <locale.h>
-#include <assert.h>
-#include <math.h>
-#include <cstring>
-#include <cstdio>
-#include <cinttypes>
-#include <unordered_map>
-#include <queue>
-#include <string.h>
-#include <cassert>
-#include <fstream>
-#include <string>
-#include <iterator>
-#include <algorithm>
-
-#if defined(_MSC_VER)
-#pragma warning(disable: 4244 4267) // possible loss of data
-#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);
-
-    if (plan.work_size > 0) {
-        buf.resize(plan.work_size);
-        plan.work_data = buf.data();
-    }
-
-    ggml_graph_compute(graph, &plan);
-}
-
-static float tensor_sum_elements(const ggml_tensor * tensor) {
-    double sum = 0;
-    if (tensor->type == GGML_TYPE_F32) {
-        for (int j = 0; j < tensor->ne[1]; j++) {
-            for (int k = 0; k < tensor->ne[0]; k++) {
-                sum += ((float *) tensor->data)[j*tensor->ne[0] + k];
-            }
-        }
-    }
-    return sum;
-}
-
-static void tensor_dump(const ggml_tensor * tensor, const char * name) {
-    printf("%15s: type = %i (%5s) ne = %5" PRIi64 " x %5" PRIi64 " x %5" PRIi64 ", nb = (%5zi, %5zi, %5zi) - ", name,
-        tensor->type, ggml_type_name(tensor->type),
-        tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
-    float sum = tensor_sum_elements(tensor);
-    printf("Sum of tensor %s is %6.2f\n", name, sum);
-}
-
-#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
-
-struct benchmark_params_struct {
-    int32_t n_threads     = 1;
-    int32_t n_iterations  = 10;
-};
-
-static void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct params) {
-    fprintf(stderr, "usage: %s [options]\n", argv[0]);
-    fprintf(stderr, "\n");
-    fprintf(stderr, "options:\n");
-    fprintf(stderr, "  -h, --help            show this help message and exit\n");
-    fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
-    fprintf(stderr, "  -i N, --iter N     number of iterations to use during computation (default: %d)\n", params.n_iterations);
-    fprintf(stderr, "\n");
-}
-
-int main(int argc, char ** argv)  {
-    struct benchmark_params_struct benchmark_params;
-
-    bool invalid_param = false;
-    std::string arg;
-    for (int i = 1; i < argc; i++) {
-        arg = argv[i];
-
-        if (arg == "-t" || arg == "--threads") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_threads = std::stoi(argv[i]);
-        } else if (arg == "-i" || arg == "--iter") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            benchmark_params.n_iterations = std::stoi(argv[i]);
-        }  else if (arg == "-h" || arg == "--help") {
-            print_usage(argc, argv, benchmark_params);
-            exit(0);
-        }
-    }
-    if (invalid_param) {
-        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        print_usage(argc, argv, benchmark_params);
-        exit(1);
-    }
-
-    print_build_info();
-    printf("Starting Test\n");
-
-    // create the ggml context
-    struct ggml_context * ctx;
-    //const int sizex = 4096;
-    //const int sizey = 11008;
-
-#undef VERBOSE_DEBUGGING
-#ifndef VERBOSE_DEBUGGING
-    const int sizey = 4096;
-    const int sizex = 11008;
-    const int sizez = 128;
-#else
-    /* Working - let's increase size */
-    const int sizey = 1;
-    const int sizex = (8*32);
-    const int sizez = 1;
-
-    /*const int sizey = 1;
-    const int sizex = 3*(8*32);
-    const int sizez = 1;*/
-#endif
-
-    //printf("Memsize required = %i\n", sizex*sizex);
-
-    // TODO: perform the bench for all types or for a user specified type
-    const ggml_type qtype = GGML_TYPE_Q4_1;
-
-    size_t ctx_size = 0;
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizez);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(qtype,         sizex*sizey);
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += ggml_row_size(GGML_TYPE_F32, sizex*sizey); // BLAS
-    ctx_size += 1024*1024*16;
-
-    printf("Allocating Memory of size %zi bytes, %zi MB\n",ctx_size, (ctx_size/1024/1024));
-
-    struct ggml_init_params params = {
-        /*.mem_size   =*/ ctx_size,
-        /*.mem_buffer =*/ NULL,
-        /* no_alloc   =*/ 0
-    };
-
-    ctx = ggml_init(params);
-    if (!ctx) {
-        fprintf(stderr, "%s: ggml_init() failed\n", __func__);
-        return 1;
-    }
-
-
-    printf("Creating new tensors\n");
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m11 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m11, 1.0f);
-
-    // printf("Creating new tensor m1\n");
-    struct ggml_tensor * m12 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizey);
-    ggml_set_f32(m12, 1.5f);
-
-    // printf("Creating new tensor m2\n");
-    struct ggml_tensor * m2 = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, sizex, sizez);
-    ggml_set_f32(m2, 2.0f);
-
-    printf("\n------ Test 1 - Matrix Mult via F32 code\n");
-    // printf("Creating new tensor m11xm2\n");
-    struct ggml_tensor * m11xm2 = ggml_mul_mat(ctx, m11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf, m11xm2);
-
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    TENSOR_DUMP(m11);
-    TENSOR_DUMP(m2);
-
-    std::vector<uint8_t> work_buffer;
-
-    ggml_graph_compute_helper(work_buffer, gf, benchmark_params.n_threads);
-
-    TENSOR_DUMP(gf->nodes[0]);
-
-    printf("\n------ Test 2 - Matrix Mult via %s code\n", ggml_type_name(qtype));
-
-    int32_t nelements = sizex*sizey;
-
-    // Set up a the benchmark matrices
-    // printf("Creating new tensor q11 & Running quantize\n");
-    struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
-
-    // Set up a the compute graph
-    // printf("Creating new tensor q31\n");
-    struct ggml_tensor * q31 = ggml_mul_mat(ctx, q11, m2);
-
-    // printf("Creating compute graph\n");
-    struct ggml_cgraph * gf31 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf31, q31);
-
-    // Set up a second graph computation to make sure we override the CPU cache lines
-    // printf("Creating new tensor q12 & Running quantize\n");
-    struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
-
-    // printf("Creating new tensor q32\n");
-    struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);
-
-    //printf("Creating compute graph\n");
-    struct ggml_cgraph * gf32 = ggml_new_graph(ctx);
-    ggml_build_forward_expand(gf32, q32);
-    printf("n_threads=%i\n", benchmark_params.n_threads);
-
-    const int dimx = sizex;
-    const int dimy = sizey;
-    const int dimz = sizez;
-    long long int flops_per_dot_product = dimy + dimy;
-    long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
-    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
-
-
-    // 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]);
-
-    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
-    printf("=====================================================================================\n");
-
-    double  gflops_sum = 0;
-    for (int i=0;i<benchmark_params.n_iterations ;i++) {
-
-        long long int start = ggml_time_us();
-        //printf("Running ggml_graph_compute\n");
-        ggml_graph_compute_helper(work_buffer, gf31, benchmark_params.n_threads);
-
-        long long int stop = ggml_time_us();
-        long long int usec = stop-start;
-        double gflops = (double)(flops_per_matrix)/usec/1000.0;
-        gflops_sum += gflops;
-        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
-            i,
-            benchmark_params.n_threads,
-            sizex, sizey, sizez, flops_per_matrix,
-            usec,gflops);
-
-#ifdef VERBOSE_DEBUGGING
-        TENSOR_DUMP("res",gf31.nodes[0])
-#endif
-
-        // 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 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
-
-        if (delta > allowed_delta)  {
-            printf("\nABORT - ERROR in Matrix Multiplication result - expected %6.2f, got %6.2f (delta %6.2f > allowed_delta %6.2f)\n",
-                sum_of_F32_reference,
-                sum_of_Q4_result,
-                delta,
-                allowed_delta
-            );
-            exit(0);
-        }
-
-        // Running a different graph computation to make sure we override the CPU cache lines
-        ggml_graph_compute_helper(work_buffer, gf32, benchmark_params.n_threads);
-    }
-    printf("\n");
-    printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
-    printf("=====================================================================================\n");
-}

examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp

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

examples/cvector-generator/cvector-generator.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 #include "ggml.h"
@@ -12,14 +13,15 @@
 #include "ggml-metal.h"
 #endif
 
+#include <algorithm>
+#include <climits>
 #include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <iostream>
 #include <string>
 #include <tuple>
 #include <vector>
-#include <algorithm>
-#include <iostream>
-#include <fstream>
-#include <climits>
 
 
 //////////////////////////////////////////////////
@@ -29,15 +31,13 @@ template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
     std::string ret;
     for (; begin != end; ++begin) {
-        ret += llama_token_to_piece(ctx, *begin);
+        ret += common_token_to_piece(ctx, *begin);
     }
 
     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]);
@@ -271,9 +271,9 @@ struct tokenized_prompt {
     size_t max_seq_len;
 
     tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) {
-        const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-        tokens_pos = ::llama_tokenize(ctx, pos, add_bos, true);
-        tokens_neg = ::llama_tokenize(ctx, neg, add_bos, true);
+        const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+        tokens_pos = common_tokenize(ctx, pos, add_bos, true);
+        tokens_neg = common_tokenize(ctx, neg, add_bos, true);
         max_seq_len = std::max(tokens_pos.size(), tokens_neg.size());
         padding_seq(ctx, tokens_pos, max_seq_len);
         padding_seq(ctx, tokens_neg, max_seq_len);
@@ -281,7 +281,7 @@ struct tokenized_prompt {
 
     void padding_seq(llama_context * ctx, std::vector<llama_token> & tokens, size_t len) {
         // TODO: customize padding token
-        std::vector<llama_token> pad_tokens = ::llama_tokenize(ctx, " ", false);
+        std::vector<llama_token> pad_tokens = common_tokenize(ctx, " ", false);
         llama_token pad_tok = pad_tokens.back();
         while (tokens.size() < len) {
             tokens.push_back(pad_tok);
@@ -370,7 +370,7 @@ static void export_gguf(const std::vector<struct ggml_tensor *> & v_ctrl, const
  * Load prompt files and completion file.
  * Then format each pair of prompt + completion to make an entry.
  */
-static int prepare_entries(gpt_params & params, train_context & ctx_train) {
+static int prepare_entries(common_params & params, train_context & ctx_train) {
     // load prompts
     std::vector<std::string> positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true);
     std::vector<std::string> negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true);
@@ -388,10 +388,9 @@ static int prepare_entries(gpt_params & params, train_context & ctx_train) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_CVECTOR_GENERATOR, print_usage)) {
         return 1;
     }
 
@@ -414,7 +413,7 @@ int main(int argc, char ** argv) {
     llama_numa_init(params.numa);
 
     // load the model to get hparams
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
@@ -486,8 +485,8 @@ 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
 
@@ -207,13 +204,6 @@ static ggml_status compute_piter(
         ggml_backend_cpu_set_n_threads(model.backend, params.n_threads);
     }
 
-// TODO: enable GPU support when support for GGML_OP_SQRT is added
-//#ifdef GGML_USE_METAL
-//    if (ggml_backend_is_metal(model.backend)) {
-//        ggml_backend_metal_set_n_cb(model.backend, params.n_threads);
-//    }
-//#endif
-
     ggml_status res = ggml_backend_graph_compute(model.backend, gf);
     if (res == GGML_STATUS_SUCCESS) {
         auto extract_i = [](std::string prefix, std::string str) -> int {
@@ -229,8 +219,8 @@ 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

@@ -1,4 +1,6 @@
+#include "arg.h"
 #include "common.h"
+#include "log.h"
 #include "llama.h"
 
 #include <ctime>
@@ -26,18 +28,29 @@ static std::vector<std::string> split_lines(const std::string & s, const std::st
 static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, llama_seq_id seq_id) {
     size_t n_tokens = tokens.size();
     for (size_t i = 0; i < n_tokens; i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, true);
+        common_batch_add(batch, tokens[i], i, { seq_id }, true);
     }
 }
 
 static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
+    const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
+    const struct llama_model * model = llama_get_model(ctx);
+
     // clear previous kv_cache values (irrelevant for embeddings)
     llama_kv_cache_clear(ctx);
 
     // run model
-    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
+    LOG_INF("%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
+    if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
+        // encoder-only model
+        if (llama_encode(ctx, batch) < 0) {
+            LOG_ERR("%s : failed to encode\n", __func__);
+        }
+    } else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        // decoder-only model
+        if (llama_decode(ctx, batch) < 0) {
+            LOG_ERR("%s : failed to decode\n", __func__);
+        }
     }
 
     for (int i = 0; i < batch.n_tokens; i++) {
@@ -45,47 +58,49 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
             continue;
         }
 
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        const float * embd = nullptr;
+        int embd_pos = 0;
 
-        float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd, embd_norm);
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            // try to get token embeddings
+            embd = llama_get_embeddings_ith(ctx, i);
+            embd_pos = i;
+            GGML_ASSERT(embd != NULL && "failed to get token embeddings");
+        } else {
+            // try to get sequence embeddings - supported only when pooling_type is not NONE
+            embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
+            embd_pos = batch.seq_id[i][0];
+            GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
+        }
+
+        float * out = output + embd_pos * n_embd;
+        common_embd_normalize(embd, out, n_embd, embd_norm);
     }
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EMBEDDING)) {
         return 1;
     }
 
+    common_init();
+
     params.embedding = true;
     // For non-causal models, batch size must be equal to ubatch size
     params.n_ubatch = params.n_batch;
 
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-
     llama_backend_init();
     llama_numa_init(params.numa);
 
     // load the model
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
     if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
@@ -93,20 +108,21 @@ int main(int argc, char ** argv) {
     const int n_ctx = llama_n_ctx(ctx);
 
     const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
-    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
-        fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
+
+    if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
+        LOG_ERR("%s: computing embeddings in encoder-decoder models is not supported\n", __func__);
         return 1;
     }
 
     if (n_ctx > n_ctx_train) {
-        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+        LOG_WRN("%s: warning: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     // split the prompt into lines
@@ -119,9 +135,9 @@ int main(int argc, char ** argv) {
     // tokenize the prompts and trim
     std::vector<std::vector<int32_t>> inputs;
     for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true, false);
+        auto inp = common_tokenize(ctx, prompt, true, true);
         if (inp.size() > n_batch) {
-            fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
+            LOG_ERR("%s: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
                     __func__, (long long int) inp.size(), (long long int) n_batch);
             return 1;
         }
@@ -132,20 +148,20 @@ int main(int argc, char ** argv) {
     // it should be automatically added by the tokenizer when 'tokenizer.ggml.add_eos_token' is set to 'true'
     for (auto & inp : inputs) {
         if (inp.empty() || inp.back() != llama_token_sep(model)) {
-            fprintf(stderr, "%s: warning: last token in the prompt is not SEP\n", __func__);
-            fprintf(stderr, "%s:          'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
+            LOG_WRN("%s: last token in the prompt is not SEP\n", __func__);
+            LOG_WRN("%s: 'tokenizer.ggml.add_eos_token' should be set to 'true' in the GGUF header\n", __func__);
         }
     }
 
     // tokenization stats
     if (params.verbose_prompt) {
         for (int i = 0; i < (int) inputs.size(); i++) {
-            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
-            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
+            LOG_INF("%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
+            LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
             for (int j = 0; j < (int) inputs[i].size(); j++) {
-                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
+                LOG("%6d -> '%s'\n", inputs[i][j], common_token_to_piece(ctx, inputs[i][j]).c_str());
             }
-            fprintf(stderr, "\n\n");
+            LOG("\n\n");
         }
     }
 
@@ -153,13 +169,23 @@ int main(int argc, char ** argv) {
     const int n_prompts = prompts.size();
     struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
 
+    // count number of embeddings
+    int n_embd_count = 0;
+    if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+        for (int k = 0; k < n_prompts; k++) {
+            n_embd_count += inputs[k].size();
+        }
+    } else {
+        n_embd_count = n_prompts;
+    }
+
     // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
+    std::vector<float> embeddings(n_embd_count * n_embd, 0);
     float * emb = embeddings.data();
 
     // break into batches
-    int p = 0; // number of prompts processed already
+    int e = 0; // number of embeddings already stored
     int s = 0; // number of prompts in current batch
     for (int k = 0; k < n_prompts; k++) {
         // clamp to n_batch tokens
@@ -169,11 +195,11 @@ int main(int argc, char ** argv) {
 
         // encode if at capacity
         if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
+            float * out = emb + e * n_embd;
             batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
-            llama_batch_clear(batch);
-            p += s;
+            e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
             s = 0;
+            common_batch_clear(batch);
         }
 
         // add to batch
@@ -182,39 +208,67 @@ int main(int argc, char ** argv) {
     }
 
     // final batch
-    float * out = emb + p * n_embd;
+    float * out = emb + e * n_embd;
     batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
 
     if (params.embd_out.empty()) {
-        // print the first part of the embeddings or for a single prompt, the full embedding
-        fprintf(stdout, "\n");
-        for (int j = 0; j < n_prompts; j++) {
-            fprintf(stdout, "embedding %d: ", j);
-            for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-                if (params.embd_normalize == 0) {
-                    fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
-                } else {
-                    fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-                }
-            }
-            fprintf(stdout, "\n");
-        }
+        LOG("\n");
 
-        // print cosine similarity matrix
-        if (n_prompts > 1) {
-            fprintf(stdout, "\n");
-            printf("cosine similarity matrix:\n\n");
-            for (int i = 0; i < n_prompts; i++) {
-                fprintf(stdout, "%6.6s ", prompts[i].c_str());
-            }
-            fprintf(stdout, "\n");
-            for (int i = 0; i < n_prompts; i++) {
-                for (int j = 0; j < n_prompts; j++) {
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f ", sim);
+        if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
+            for (int j = 0; j < n_embd_count; j++) {
+                LOG("embedding %d: ", j);
+                for (int i = 0; i < std::min(3, n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        LOG("%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        LOG("%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                LOG(" ... ");
+                for (int i = n_embd - 3; i < n_embd; i++) {
+                    if (params.embd_normalize == 0) {
+                        LOG("%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        LOG("%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                LOG("\n");
+            }
+        } else if (pooling_type == LLAMA_POOLING_TYPE_RANK) {
+            for (int j = 0; j < n_embd_count; j++) {
+                // NOTE: if you change this log - update the tests in ci/run.sh
+                LOG("rerank score %d: %8.3f\n", j, emb[j * n_embd]);
+            }
+        } else {
+            // print the first part of the embeddings or for a single prompt, the full embedding
+            for (int j = 0; j < n_prompts; j++) {
+                LOG("embedding %d: ", j);
+                for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
+                    if (params.embd_normalize == 0) {
+                        LOG("%6.0f ", emb[j * n_embd + i]);
+                    } else {
+                        LOG("%9.6f ", emb[j * n_embd + i]);
+                    }
+                }
+                LOG("\n");
+            }
+
+            // print cosine similarity matrix
+            if (n_prompts > 1) {
+                LOG("\n");
+                LOG("cosine similarity matrix:\n\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    LOG("%6.6s ", prompts[i].c_str());
+                }
+                LOG("\n");
+                for (int i = 0; i < n_prompts; i++) {
+                    for (int j = 0; j < n_prompts; j++) {
+                        float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                        LOG("%6.2f ", sim);
+                    }
+                    LOG("%1.10s", prompts[i].c_str());
+                    LOG("\n");
                 }
-                fprintf(stdout, "%1.10s", prompts[i].c_str());
-                fprintf(stdout, "\n");
             }
         }
     }
@@ -222,43 +276,45 @@ int main(int argc, char ** argv) {
     if (params.embd_out == "json" || params.embd_out == "json+" || params.embd_out == "array") {
         const bool notArray = params.embd_out != "array";
 
-        fprintf(stdout, notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
+        LOG(notArray ? "{\n  \"object\": \"list\",\n  \"data\": [\n" : "[");
         for (int j = 0;;) { // at least one iteration (one prompt)
-            if (notArray) fprintf(stdout, "    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
-            fprintf(stdout, "[");
+            if (notArray) LOG("    {\n      \"object\": \"embedding\",\n      \"index\": %d,\n      \"embedding\": ",j);
+            LOG("[");
             for (int i = 0;;) { // at least one iteration (n_embd > 0)
-                fprintf(stdout, params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
+                LOG(params.embd_normalize == 0 ? "%1.0f" : "%1.7f", emb[j * n_embd + i]);
                 i++;
-                if (i < n_embd) fprintf(stdout, ","); else break;
+                if (i < n_embd) LOG(","); else break;
             }
-            fprintf(stdout, notArray ? "]\n    }" : "]");
+            LOG(notArray ? "]\n    }" : "]");
             j++;
-            if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
+            if (j < n_embd_count) LOG(notArray ? ",\n" : ","); else break;
         }
-        fprintf(stdout, notArray ? "\n  ]" : "]\n");
+        LOG(notArray ? "\n  ]" : "]\n");
 
         if (params.embd_out == "json+" && n_prompts > 1) {
-            fprintf(stdout, ",\n  \"cosineSimilarity\": [\n");
-            for (int i = 0;;) { // at least two iteration (n_prompts > 1)
-                fprintf(stdout, "    [");
-                for (int j = 0;;) { // at least two iteration (n_prompts > 1)
-                    float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
-                    fprintf(stdout, "%6.2f", sim);
+            LOG(",\n  \"cosineSimilarity\": [\n");
+            for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
+                LOG("    [");
+                for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
+                    float sim = common_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                    LOG("%6.2f", sim);
                     j++;
-                    if (j < n_prompts) fprintf(stdout, ", "); else break;
+                    if (j < n_embd_count) LOG(", "); else break;
                 }
-                fprintf(stdout, " ]");
+                LOG(" ]");
                 i++;
-                if (i < n_prompts) fprintf(stdout, ",\n"); else break;
+                if (i < n_embd_count) LOG(",\n"); else break;
             }
-            fprintf(stdout, "\n  ]");
+            LOG("\n  ]");
         }
 
-        if (notArray) fprintf(stdout, "\n}\n");
+        if (notArray) LOG("\n}\n");
     }
 
+    LOG("\n");
+    llama_perf_context_print(ctx);
+
     // clean up
-    llama_print_timings(ctx);
     llama_batch_free(batch);
     llama_free(ctx);
     llama_free_model(model);

examples/eval-callback/eval-callback.cpp

@@ -1,11 +1,11 @@
+#include "arg.h"
 #include "common.h"
+#include "log.h"
 #include "llama.h"
 #include "ggml.h"
 
 #include <cstdio>
-#include <random>
 #include <string>
-#include <tuple>
 #include <vector>
 
 /**
@@ -31,22 +31,22 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
     GGML_ASSERT(n > 0);
     float sum = 0;
     for (int64_t i3 = 0; i3 < ne[3]; i3++) {
-        printf("                                     [\n");
+        LOG("                                     [\n");
         for (int64_t i2 = 0; i2 < ne[2]; i2++) {
             if (i2 == n && ne[2] > 2*n) {
-                printf("                                      ..., \n");
+                LOG("                                      ..., \n");
                 i2 = ne[2] - n;
             }
-            printf("                                      [\n");
+            LOG("                                      [\n");
             for (int64_t i1 = 0; i1 < ne[1]; i1++) {
                 if (i1 == n && ne[1] > 2*n) {
-                    printf("                                       ..., \n");
+                    LOG("                                       ..., \n");
                     i1 = ne[1] - n;
                 }
-                printf("                                       [");
+                LOG("                                       [");
                 for (int64_t i0 = 0; i0 < ne[0]; i0++) {
                     if (i0 == n && ne[0] > 2*n) {
-                        printf("..., ");
+                        LOG("..., ");
                         i0 = ne[0] - n;
                     }
                     size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
@@ -64,16 +64,16 @@ static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne
                     } else {
                         GGML_ABORT("fatal error");
                     }
-                    printf("%12.4f", v);
+                    LOG("%12.4f", v);
                     sum += v;
-                    if (i0 < ne[0] - 1) printf(", ");
+                    if (i0 < ne[0] - 1) LOG(", ");
                 }
-                printf("],\n");
+                LOG("],\n");
             }
-            printf("                                      ],\n");
+            LOG("                                      ],\n");
         }
-        printf("                                     ]\n");
-        printf("                                     sum = %f\n", sum);
+        LOG("                                     ]\n");
+        LOG("                                     sum = %f\n", sum);
     }
 }
 
@@ -102,11 +102,11 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
         snprintf(src1_str, sizeof(src1_str), "%s{%s}", src1->name, ggml_ne_string(src1).c_str());
     }
 
-    printf("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
-           t->name, ggml_type_name(t->type), ggml_op_desc(t),
-           src0->name, ggml_ne_string(src0).c_str(),
-           src1 ? src1_str : "",
-           ggml_ne_string(t).c_str());
+    LOG("%s: %24s = (%s) %10s(%s{%s}, %s}) = {%s}\n", __func__,
+         t->name, ggml_type_name(t->type), ggml_op_desc(t),
+         src0->name, ggml_ne_string(src0).c_str(),
+         src1 ? src1_str : "",
+         ggml_ne_string(t).c_str());
 
 
     // copy the data from the GPU memory if needed
@@ -126,13 +126,13 @@ static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
     return true;
 }
 
-static bool run(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+static bool run(llama_context * ctx, const common_params & params) {
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
     if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
-        fprintf(stderr, "%s : failed to eval\n", __func__);
+        LOG_ERR("%s : failed to eval\n", __func__);
         return false;
     }
 
@@ -142,16 +142,13 @@ static bool run(llama_context * ctx, const gpt_params & params) {
 int main(int argc, char ** argv) {
     callback_data cb_data;
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    print_build_info();
-
-    std::mt19937 rng(params.seed);
+    common_init();
 
     llama_backend_init();
     llama_numa_init(params.numa);
@@ -163,19 +160,20 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
-        fprintf(stderr, "%s : failed to init\n", __func__);
+        LOG_ERR("%s : failed to init\n", __func__);
         return 1;
     }
 
     // print system information
     {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
+        LOG_INF("\n");
     }
 
     bool OK = run(ctx, params);
@@ -183,7 +181,8 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    llama_print_timings(ctx);
+    LOG("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/export-lora/README.md

@@ -17,9 +17,9 @@ For example:
 
 ```bash
 ./bin/llama-export-lora \
-    -m open-llama-3b-v2-q8_0.gguf \
-    -o open-llama-3b-v2-q8_0-english2tokipona-chat.gguf \
-    --lora lora-open-llama-3b-v2-q8_0-english2tokipona-chat-LATEST.gguf
+    -m open-llama-3b-v2.gguf \
+    -o open-llama-3b-v2-english2tokipona-chat.gguf \
+    --lora lora-open-llama-3b-v2-english2tokipona-chat-LATEST.gguf
 ```
 
 Multiple LORA adapters can be applied by passing multiple `--lora FNAME` or `--lora-scaled FNAME S` command line parameters:

examples/export-lora/export-lora.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
@@ -10,6 +11,12 @@
 
 static bool g_verbose = false;
 
+struct tensor_transformation {
+    struct ggml_tensor * in;
+    struct ggml_tensor * out;
+    bool is_copy;
+};
+
 static std::string get_kv_str(struct gguf_context * ctx_gguf, const std::string & key){
     int id = gguf_find_key(ctx_gguf, key.c_str());
     return id < 0 ? "" : std::string(gguf_get_val_str(ctx_gguf, id));
@@ -121,7 +128,7 @@ struct lora_merge_ctx {
 
     lora_merge_ctx(
             std::string & base_fname,
-            std::vector<llama_lora_adapter_info> & lora_files,
+            std::vector<common_lora_adapter_info> & lora_files,
             std::string & outfile,
             int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
         fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@@ -198,8 +205,7 @@ struct lora_merge_ctx {
         }
 
         // mapping base tensor to out tensor (same shape with base, but different type)
-        // if out_tensor == nullptr, we only copy it
-        std::vector<std::pair<struct ggml_tensor *, struct ggml_tensor *>> base_to_out_tensors;
+        std::vector<tensor_transformation> trans;
         for (auto & it : base_model.tensors) {
             bool t_a = true;
             bool t_b = true;
@@ -212,14 +218,22 @@ struct lora_merge_ctx {
                 // only copy
                 struct ggml_tensor * cpy_tensor = ggml_dup_tensor(ctx_out_ggml, base_tensor);
                 ggml_set_name(cpy_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(cpy_tensor, nullptr));
+                trans.push_back({
+                    cpy_tensor,
+                    cpy_tensor,
+                    true,
+                });
                 gguf_add_tensor(ctx_out, cpy_tensor);
             } else if (t_a && t_b) {
                 // need merging
                 struct ggml_tensor * out_tensor = ggml_new_tensor(
                     ctx_out_ggml, get_out_tensor_type(base_tensor), GGML_MAX_DIMS, base_tensor->ne);
                 ggml_set_name(out_tensor, base_tensor->name);
-                base_to_out_tensors.push_back(std::make_pair(base_tensor, out_tensor));
+                trans.push_back({
+                    base_tensor,
+                    out_tensor,
+                    false,
+                });
                 gguf_add_tensor(ctx_out, out_tensor);
             } else {
                 throw std::runtime_error("tensor " + it.first + " missing either lora_a or lora_b");
@@ -234,12 +248,12 @@ struct lora_merge_ctx {
 
         // process base model tensors
         size_t n_merged = 0;
-        for (auto & it : base_to_out_tensors) {
-            if (it.second != nullptr) {
-                merge_tensor(it.first, it.second);
+        for (auto & it : trans) {
+            if (!it.is_copy) {
+                merge_tensor(it.in, it.out);
                 n_merged++;
             } else {
-                copy_tensor(it.first);
+                copy_tensor(it.in);
             }
         }
 
@@ -252,7 +266,7 @@ struct lora_merge_ctx {
         }
 
         printf("%s : merged %ld tensors with lora adapters\n", __func__, n_merged);
-        printf("%s : wrote %ld tensors to output file\n", __func__, base_to_out_tensors.size());
+        printf("%s : wrote %ld tensors to output file\n", __func__, trans.size());
     }
 
     void copy_tensor(struct ggml_tensor * base) {
@@ -285,6 +299,10 @@ struct lora_merge_ctx {
         for (size_t i = 0; i < adapters.size(); ++i) {
             auto t_a = adapters[i]->get_tensor(name_lora_a);
             auto t_b = adapters[i]->get_tensor(name_lora_b);
+            // TODO: add support for quantized lora
+            if (ggml_is_quantized(t_a->type) || ggml_is_quantized(t_b->type)) {
+                throw std::runtime_error("quantized LoRA adapters is not supported, please retry with f16 or f32");
+            }
             inp_a[i] = ggml_dup_tensor(ctx, t_a);
             inp_b[i] = ggml_dup_tensor(ctx, t_b);
         }
@@ -296,9 +314,9 @@ struct lora_merge_ctx {
             // optionally dequantize it
             printf("%s :   + dequantize base tensor from %s to F32\n", __func__, ggml_type_name(base->type));
             auto nels = ggml_nelements(inp_base);
-            ggml_type_traits_t qtype = ggml_internal_get_type_traits(base->type);
+            const auto * qtype = ggml_get_type_traits(base->type);
             std::vector<uint8_t> dequant_buf(nels * sizeof(float));
-            qtype.to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
+            qtype->to_float(read_buf.data(), (float *)dequant_buf.data(), nels);
             ggml_backend_tensor_set(inp_base, dequant_buf.data(), 0, dequant_buf.size());
         } else {
             ggml_backend_tensor_set(inp_base, read_buf.data(), 0, ggml_nbytes(inp_base));
@@ -352,7 +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);
@@ -374,9 +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");
@@ -384,16 +400,15 @@ static void print_usage(int argc, char ** argv, const gpt_params & params) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_EXPORT_LORA, print_usage)) {
         return 1;
     }
 
-    g_verbose = (params.verbosity == 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 +1,83 @@
+#include "arg.h"
+#include "common.h"
+
+#include <fstream>
+#include <string>
+
+// Export usage message (-h) to markdown format
+
+static void write_table_header(std::ofstream & file) {
+    file << "| Argument | Explanation |\n";
+    file << "| -------- | ----------- |\n";
+}
+
+static void write_table_entry(std::ofstream & file, const common_arg & opt) {
+    file << "| `";
+    // args
+    for (const auto & arg : opt.args) {
+    if (arg == opt.args.front()) {
+            file << arg;
+            if (opt.args.size() > 1) file << ", ";
+        } else {
+            file << arg << (arg != opt.args.back() ? ", " : "");
+        }
+    }
+    // value hint
+    if (opt.value_hint) {
+        std::string md_value_hint(opt.value_hint);
+        string_replace_all(md_value_hint, "|", "\\|");
+        file << " " << md_value_hint;
+    }
+    if (opt.value_hint_2) {
+        std::string md_value_hint_2(opt.value_hint_2);
+        string_replace_all(md_value_hint_2, "|", "\\|");
+        file << " " << md_value_hint_2;
+    }
+    // help text
+    std::string md_help(opt.help);
+    string_replace_all(md_help, "\n", "<br/>");
+    string_replace_all(md_help, "|", "\\|");
+    file << "` | " << md_help << " |\n";
+}
+
+static void write_table(std::ofstream & file, std::vector<common_arg *> & opts) {
+    write_table_header(file);
+    for (const auto & opt : opts) {
+        write_table_entry(file, *opt);
+    }
+}
+
+static void export_md(std::string fname, llama_example ex) {
+    std::ofstream file(fname, std::ofstream::out | std::ofstream::trunc);
+
+    common_params params;
+    auto ctx_arg = common_params_parser_init(params, ex);
+
+    std::vector<common_arg *> common_options;
+    std::vector<common_arg *> sparam_options;
+    std::vector<common_arg *> specific_options;
+    for (auto & opt : ctx_arg.options) {
+        // in case multiple LLAMA_EXAMPLE_* are set, we prioritize the LLAMA_EXAMPLE_* matching current example
+        if (opt.is_sparam) {
+            sparam_options.push_back(&opt);
+        } else if (opt.in_example(ctx_arg.ex)) {
+            specific_options.push_back(&opt);
+        } else {
+            common_options.push_back(&opt);
+        }
+    }
+
+    file << "**Common params**\n\n";
+    write_table(file, common_options);
+    file << "\n\n**Sampling params**\n\n";
+    write_table(file, sparam_options);
+    file << "\n\n**Example-specific params**\n\n";
+    write_table(file, specific_options);
+}
+
+int main(int, char **) {
+    export_md("autogen-main.md", LLAMA_EXAMPLE_MAIN);
+    export_md("autogen-server.md", LLAMA_EXAMPLE_SERVER);
+
+    return 0;
+}

examples/gguf-split/gguf-split.cpp

@@ -22,12 +22,20 @@
 #endif
 
 enum split_operation : uint8_t {
-    SPLIT_OP_SPLIT,
-    SPLIT_OP_MERGE,
+    OP_NONE,
+    OP_SPLIT,
+    OP_MERGE,
+};
+
+enum split_mode : uint8_t {
+    MODE_NONE,
+    MODE_TENSOR,
+    MODE_SIZE,
 };
 
 struct split_params {
-    split_operation operation = SPLIT_OP_SPLIT;
+    split_operation operation = OP_NONE;
+    split_mode mode = MODE_NONE;
     size_t n_bytes_split = 0;
     int n_split_tensors = 128;
     std::string input;
@@ -87,59 +95,52 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
         }
 
         bool arg_found = false;
-        bool is_op_set = false;
-        bool is_mode_set = false;
         if (arg == "-h" || arg == "--help") {
             split_print_usage(argv[0]);
             exit(0);
-        }
-        if (arg == "--version") {
+        } else if (arg == "--version") {
             fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
             fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
             exit(0);
-        }
-        if (arg == "--dry-run") {
+        } else if (arg == "--dry-run") {
             arg_found = true;
             params.dry_run = true;
-        }
-        if (arg == "--no-tensor-first-split") {
+        } else if (arg == "--no-tensor-first-split") {
             arg_found = true;
             params.no_tensor_first_split = true;
-        }
-
-        if (is_op_set) {
-            throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
-        }
-        if (arg == "--merge") {
+        } else if (arg == "--merge") {
             arg_found = true;
-            is_op_set = true;
-            params.operation = SPLIT_OP_MERGE;
-        }
-        if (arg == "--split") {
+            if (params.operation != OP_NONE && params.operation != OP_MERGE) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_MERGE;
+        } else if (arg == "--split") {
             arg_found = true;
-            is_op_set = true;
-            params.operation = SPLIT_OP_SPLIT;
-        }
-
-        if (is_mode_set) {
-            throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
-        }
-        if (arg == "--split-max-tensors") {
+            if (params.operation != OP_NONE && params.operation != OP_SPLIT) {
+                throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+            }
+            params.operation = OP_SPLIT;
+        } else if (arg == "--split-max-tensors") {
             if (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
-            is_mode_set = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_TENSOR) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_TENSOR;
             params.n_split_tensors = atoi(argv[arg_idx]);
-        }
-        if (arg == "--split-max-size") {
+        } else if (arg == "--split-max-size") {
             if (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
-            is_mode_set = true;
+            if (params.mode != MODE_NONE && params.mode != MODE_SIZE) {
+                throw std::invalid_argument("error: either --split-max-tensors or --split-max-size can be specified, but not both");
+            }
+            params.mode = MODE_SIZE;
             params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
         }
 
@@ -148,11 +149,20 @@ static void split_params_parse_ex(int argc, const char ** argv, split_params & p
         }
     }
 
+    // the operation is split if not specified
+    if (params.operation == OP_NONE) {
+        params.operation = OP_SPLIT;
+    }
+    // the split mode is by tensor if not specified
+    if (params.mode == MODE_NONE) {
+        params.mode = MODE_TENSOR;
+    }
+
     if (invalid_param) {
         throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
 
-    if (argc - arg_idx < 2) {
+    if (argc - arg_idx != 2) {
         throw std::invalid_argument("error: bad arguments");
     }
 
@@ -265,13 +275,15 @@ struct split_strategy {
     }
 
     bool should_split(int i_tensor, size_t next_size) {
-        if (params.n_bytes_split > 0) {
+        if (params.mode == MODE_SIZE) {
             // split by max size per file
             return next_size > params.n_bytes_split;
-        } else {
+        } else if (params.mode == MODE_TENSOR) {
             // split by number of tensors per file
             return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
         }
+        // should never happen
+        GGML_ABORT("invalid mode");
     }
 
     void print_info() {
@@ -389,10 +401,17 @@ static void gguf_merge(const split_params & split_params) {
     int n_split = 1;
     int total_tensors = 0;
 
-    auto * ctx_out = gguf_init_empty();
+    // avoid overwriting existing output file
+    if (std::ifstream(split_params.output.c_str())) {
+        fprintf(stderr, "%s: output file %s already exists\n", __func__, split_params.output.c_str());
+        exit(EXIT_FAILURE);
+    }
+
     std::ofstream fout(split_params.output.c_str(), std::ios::binary);
     fout.exceptions(std::ofstream::failbit); // fail fast on write errors
 
+    auto * ctx_out = gguf_init_empty();
+
     std::vector<uint8_t> read_data;
     std::vector<ggml_context *> ctx_metas;
     std::vector<gguf_context *> ctx_ggufs;
@@ -552,9 +571,9 @@ int main(int argc, const char ** argv) {
     split_params_parse(argc, argv, params);
 
     switch (params.operation) {
-        case SPLIT_OP_SPLIT: gguf_split(params);
+        case OP_SPLIT: gguf_split(params);
             break;
-        case SPLIT_OP_MERGE: gguf_merge(params);
+        case OP_MERGE: gguf_merge(params);
             break;
         default: split_print_usage(argv[0]);
             exit(EXIT_FAILURE);

examples/gritlm/gritlm.cpp

@@ -1,3 +1,4 @@
+#include "arg.h"
 #include "common.h"
 #include "llama.h"
 
@@ -9,25 +10,25 @@
 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);
 
     for (uint64_t i = 0; i < sentences.size(); i++) {
-        llama_batch_clear(batch);
+        common_batch_clear(batch);
 
         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 = common_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 = common_tokenize(model, instruction, true, false).size();
 
 #ifdef GRIT_DEBUG
         // debug tokens - should be matching as referenced in the GritLM sample
@@ -39,7 +40,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
 
         // add input to batch (this increments n_tokens)
         for (int32_t j = 0; j < n_toks; j++) {
-            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
+            common_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
         }
 
         // clear previous kv_cache values (irrelevant for embeddings)
@@ -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);
@@ -74,7 +75,7 @@ static std::vector<std::vector<float>> encode(llama_context * ctx, const std::ve
         }
 
         std::vector<float> emb_norm(emb_unorm.size());
-        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
+        common_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
         result.push_back(emb_norm);
 
 #ifdef GRIT_DEBUG
@@ -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);
@@ -104,33 +105,29 @@ 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 = common_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);
+        common_batch_clear(bat);
+        {
+            const int32_t n_inputs = inputs.size();
+
+            for (int32_t i = 0; i < n_inputs; i++) {
+                common_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;
         }
 
-        std::string piece = llama_token_to_piece(ctx, token);
+        std::string piece = common_token_to_piece(ctx, token);
         if (stream) {
             std::printf("%s", piece.c_str());
             std::fflush(stdout);
@@ -155,22 +152,31 @@ static std::string gritlm_instruction(const std::string & instruction) {
 }
 
 int main(int argc, char * argv[]) {
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_COMMON)) {
         return 1;
     }
 
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
+    common_init();
+
+    llama_model_params mparams = common_model_params_to_llama(params);
+    llama_context_params cparams = common_context_params_to_llama(params);
 
     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
@@ -191,12 +197,12 @@ 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);
-        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q0_d0 = common_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q0_d1 = common_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
+        const float cosine_sim_q1_d0 = common_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
+        const float cosine_sim_q1_d1 = common_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
 
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
         std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
@@ -208,11 +214,12 @@ 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

@@ -1,4 +1,6 @@
+#include "arg.h"
 #include "common.h"
+#include "log.h"
 #include "llama.h"
 
 #include <cmath>
@@ -17,15 +19,13 @@
 #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);
-
-    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"
+static void print_usage(int, char ** argv) {
+    LOG("\nexample usage:\n");
+    LOG("\n    %s \\\n"
+            "       -m model.gguf -f some-text.txt [-o imatrix.dat] [--process-output] \\\n"
             "       [--no-ppl] [--chunk 123] [--output-frequency 10] [--save-frequency 0] \\\n"
             "       [--in-file imatrix-prev-0.dat --in-file imatrix-prev-1.dat ...]\n" , argv[0]);
-    LOG_TEE("\n");
+    LOG("\n");
 }
 
 struct Stats {
@@ -37,13 +37,13 @@ struct Stats {
 class IMatrixCollector {
 public:
     IMatrixCollector() = default;
-    void set_params(gpt_params params) { m_params = std::move(params); }
+    void set_params(common_params params) { m_params = std::move(params); }
     bool collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data);
     void save_imatrix(int ncall = -1) const;
     bool load_imatrix(const char * file_name);
 private:
     std::unordered_map<std::string, Stats> m_stats;
-    gpt_params                             m_params;
+    common_params                          m_params;
     std::mutex                             m_mutex;
     int                                    m_last_call = 0;
     std::vector<float>                     m_src1_data;
@@ -126,12 +126,10 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             e.counts.resize(src1->ne[0]*n_as, 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
+            LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
             exit(1); //GGML_ABORT("fatal error");
         }
-        if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
-        }
+        LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
         // loop over all possible experts, regardless if they are used or not in the batch
         for (int ex = 0; ex < n_as; ++ex) {
             size_t e_start = ex*src1->ne[0];
@@ -152,7 +150,8 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                         e.values[e_start + j] += x[j]*x[j];
                         e.counts[e_start + j]++;
                         if (!std::isfinite(e.values[e_start + j])) {
-                            fprintf(stderr, "%f detected in %s\n", e.values[e_start + j], wname.c_str());
+                            LOG("\n");
+                            LOG_ERR("%f detected in %s\n", e.values[e_start + j], wname.c_str());
                             exit(1);
                         }
                     }
@@ -175,20 +174,18 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             e.counts.resize(src1->ne[0], 0);
         }
         else if (e.values.size() != (size_t)src1->ne[0]) {
-            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
+            LOG_ERR("%s: inconsistent size for %s (%d vs %d)\n", __func__, wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
             exit(1); //GGML_ABORT("fatal error");
         }
         ++e.ncall;
-        if (m_params.verbosity > 1) {
-            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
-        }
+        LOG_DBGV(2, "%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];
             for (int j = 0; j < (int)src1->ne[0]; ++j) {
                 e.values[j] += x[j]*x[j];
                 e.counts[j]++;
                 if (!std::isfinite(e.values[j])) {
-                    fprintf(stderr, "%f detected in %s\n", e.values[j], wname.c_str());
+                    LOG_ERR("%f detected in %s\n", e.values[j], wname.c_str());
                     exit(1);
                 }
             }
@@ -240,17 +237,17 @@ void IMatrixCollector::save_imatrix(int ncall) const {
         }
 
         if (n_zeros != 0 && is_first) {
-            fprintf(stderr, "\n");
+            LOG_INF("\n");
             is_first = false;
         }
 
         if (n_zeros == n_all) {
-            fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
+            LOG_WRN("%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
             continue;
         }
 
         if (n_zeros > 0) {
-            fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
+            LOG_WRN("%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
             continue;
         }
 
@@ -259,7 +256,7 @@ void IMatrixCollector::save_imatrix(int ncall) const {
     }
 
     if (to_store.size() < m_stats.size()) {
-        fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
+        LOG_WRN("%s: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
     }
 
     std::ofstream out(fname, std::ios::binary);
@@ -291,21 +288,20 @@ void IMatrixCollector::save_imatrix(int ncall) const {
         out.write(m_params.prompt_file.c_str(), len);
     }
 
-    if (m_params.verbosity > 0) {
-        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
-    }
+    LOGV(1, "\n");
+    LOG_DBGV(1, "%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname.c_str());
 }
 
 bool IMatrixCollector::load_imatrix(const char * fname) {
     std::ifstream in(fname, std::ios::binary);
     if (!in) {
-        printf("%s: failed to open %s\n",__func__, fname);
+        LOG_ERR("%s: failed to open %s\n",__func__, fname);
         return false;
     }
     int n_entries;
     in.read((char*)&n_entries, sizeof(n_entries));
     if (in.fail() || n_entries < 1) {
-        printf("%s: no data in file %s\n", __func__, fname);
+        LOG_ERR("%s: no data in file %s\n", __func__, fname);
         return false;
     }
     for (int i = 0; i < n_entries; ++i) {
@@ -313,7 +309,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         std::vector<char> name_as_vec(len+1);
         in.read((char *)name_as_vec.data(), len);
         if (in.fail()) {
-            printf("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
+            LOG_ERR("%s: failed reading name for entry %d from %s\n",__func__,i+1, fname);
             return false;
         }
         name_as_vec[len] = 0;
@@ -324,7 +320,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         int nval;
         in.read((char *)&nval, sizeof(nval));
         if (in.fail() || nval < 1) {
-            printf("%s: failed reading number of values for entry %d\n",__func__,i);
+            LOG_ERR("%s: failed reading number of values for entry %d\n",__func__,i);
             m_stats = {};
             return false;
         }
@@ -337,7 +333,7 @@ bool IMatrixCollector::load_imatrix(const char * fname) {
         std::vector<float> tmp(nval);
         in.read((char*)tmp.data(), nval*sizeof(float));
         if (in.fail()) {
-            printf("%s: failed reading data for entry %d\n",__func__,i);
+            LOG_ERR("%s: failed reading data for entry %d\n",__func__,i);
             m_stats = {};
             return false;
         }
@@ -432,32 +428,31 @@ static void process_logits(
     }
 }
 
-static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
+static bool compute_imatrix(llama_context * ctx, const common_params & params) {
+    const bool add_bos = llama_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(!llama_add_eos_token(llama_get_model(ctx)));
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
-    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+    LOG_INF("%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
+    std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, true);
 
     auto tim2 = std::chrono::high_resolution_clock::now();
-    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+    LOG_INF("%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
 
     if (params.i_chunk > 0) {
         if (size_t((params.i_chunk + 2)*n_ctx) >= tokens.size()) {
-            fprintf(stderr, "%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
+            LOG_ERR("%s: there will be not enough tokens left after removing %d chunks\n", __func__, params.i_chunk);
             return false;
         }
-        fprintf(stderr, "%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
+        LOG_INF("%s: removing initial %d chunks (%d tokens)\n", __func__, params.i_chunk, params.i_chunk*n_ctx);
         tokens.erase(tokens.begin(), tokens.begin() + params.i_chunk*n_ctx);
     }
 
     if (int(tokens.size()) < 2*n_ctx) {
-        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
-                n_ctx);
-        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        LOG_ERR("%s: you need at least %d tokens for a context of %d tokens\n", __func__, 2*n_ctx, n_ctx);
+        LOG_ERR("%s: the data file you provided tokenizes to only %zu tokens\n", __func__, tokens.size());
         return false;
     }
 
@@ -479,7 +474,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
     double nll = 0.0;
     double nll2 = 0.0;
 
-    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+    LOG_INF("%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
 
     std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
 
@@ -515,7 +510,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 
             // TODO: use batch.logits to save computations instead of relying on logits_all == true
             if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
-                fprintf(stderr, "%s : failed to eval\n", __func__);
+                LOG_ERR("%s : failed to eval\n", __func__);
                 return false;
             }
 
@@ -532,29 +527,29 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 
         if (i == 0) {
             const float t_total = std::chrono::duration<float>(t_end - t_start).count();
-            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            LOG_INF("%s: %.2f seconds per pass - ETA ", __func__, t_total);
             int total_seconds = (int)(t_total * n_chunk);
             if (total_seconds >= 60*60) {
-                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                LOG("%d hours ", total_seconds / (60*60));
                 total_seconds = total_seconds % (60*60);
             }
-            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+            LOG("%.2f minutes\n", total_seconds / 60.0);
         }
 
         if (params.compute_ppl) {
             const int first = n_ctx/2;
-            const auto all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
+            const auto * all_logits = num_batches > 1 ? logits.data() : llama_get_logits(ctx);
             process_logits(n_vocab, all_logits + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
                     workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
             count += n_ctx - first - 1;
 
-            printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+            LOG("[%d]%.4lf,", i + 1, std::exp(nll / count));
             fflush(stdout);
 
             logits.clear();
         }
     }
-    printf("\n");
+    LOG("\n");
 
     if (params.compute_ppl) {
         nll2 /= count;
@@ -563,9 +558,9 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
         nll2 -= nll * nll;
         if (nll2 > 0) {
             nll2 = sqrt(nll2/(count-1));
-            printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+            LOG("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
         } else {
-            printf("Unexpected negative standard deviation of log(prob)\n");
+            LOG("Unexpected negative standard deviation of log(prob)\n");
         }
     }
 
@@ -573,31 +568,32 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
 }
 
 int main(int argc, char ** argv) {
-    gpt_params params;
+    common_params params;
 
     params.n_ctx = 512;
     params.logits_all = true;
-    params.verbosity = 1;
+    params.escape = false;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_IMATRIX, print_usage)) {
         return 1;
     }
 
+    common_init();
+
     params.n_batch = std::min(params.n_batch, params.n_ctx);
 
     g_collector.set_params(params);
 
     for (const auto & in_file : params.in_files) {
-        printf("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
+        LOG_INF("%s : loading imatrix from '%s'\n", __func__, in_file.c_str());
         if (!g_collector.load_imatrix(in_file.c_str())) {
-            fprintf(stderr, "%s : failed to load %s\n", __func__, in_file.c_str());
+            LOG_ERR("%s : failed to load %s\n", __func__, in_file.c_str());
             return 1;
         }
     }
 
     if (params.in_files.size() > 1) {
-        printf("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
+        LOG_INF("%s : saving combined imatrix to '%s'\n", __func__, params.out_file.c_str());
         g_collector.save_imatrix();
     }
 
@@ -611,25 +607,25 @@ int main(int argc, char ** argv) {
     params.warmup = false;
 
     // init
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    common_init_result llama_init = common_init_from_params(params);
 
     llama_model * model = llama_init.model;
     llama_context * ctx = llama_init.context;
     if (model == nullptr || ctx == nullptr) {
-        fprintf(stderr, "%s : failed to init\n", __func__);
+        LOG_ERR("%s : failed to init\n", __func__);
         return 1;
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
     if (params.n_ctx > n_ctx_train) {
-        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n",
                 __func__, n_ctx_train, params.n_ctx);
     }
 
     // print system information
     {
-        fprintf(stderr, "\n");
-        fprintf(stderr, "%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
 
     if (!compute_imatrix(ctx, params)) {
@@ -638,7 +634,8 @@ int main(int argc, char ** argv) {
 
     g_collector.save_imatrix();
 
-    llama_print_timings(ctx);
+    LOG("\n");
+    llama_perf_context_print(ctx);
 
     llama_free(ctx);
     llama_free_model(model);

examples/infill/infill.cpp

@@ -1,8 +1,9 @@
+#include "arg.h"
 #include "common.h"
-
 #include "console.h"
+#include "sampling.h"
+#include "log.h"
 #include "llama.h"
-#include "grammar-parser.h"
 
 #include <cassert>
 #include <cinttypes>
@@ -34,7 +35,8 @@
 
 static llama_context           ** g_ctx;
 static llama_model             ** g_model;
-static gpt_params               * g_params;
+static common_sampler          ** g_smpl;
+static common_params            * g_params;
 static std::vector<llama_token> * g_input_tokens;
 static std::ostringstream       * g_output_ss;
 static std::vector<llama_token> * g_output_tokens;
@@ -42,7 +44,7 @@ static std::vector<llama_token> * g_output_tokens;
 static bool is_interacting = false;
 
 static void write_logfile(
-    const llama_context * ctx, const gpt_params & params, const llama_model * model,
+    const llama_context * ctx, const common_params & params, const llama_model * model,
     const std::vector<llama_token> & input_tokens, const std::string & output,
     const std::vector<llama_token> & output_tokens
 ) {
@@ -54,7 +56,7 @@ static void write_logfile(
 
     const bool success = fs_create_directory_with_parents(params.logdir);
     if (!success) {
-        fprintf(stderr, "%s: warning: failed to create logdir %s, cannot write logfile\n",
+        LOG_ERR("%s: warning: failed to create logdir %s, cannot write logfile\n",
                 __func__, params.logdir.c_str());
         return;
     }
@@ -63,7 +65,7 @@ static void write_logfile(
     FILE * logfile = fopen(logfile_path.c_str(), "w");
 
     if (logfile == NULL) {
-        fprintf(stderr, "%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
+        LOG_ERR("%s: failed to open logfile %s\n", __func__, logfile_path.c_str());
         return;
     }
 
@@ -81,7 +83,7 @@ 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);
 }
 
@@ -92,9 +94,14 @@ static void sigint_handler(int signo) {
             is_interacting = true;
         } else {
             console::cleanup();
-            printf("\n");
-            llama_print_timings(*g_ctx);
+            LOG("\n");
+            common_perf_print(*g_ctx, *g_smpl);
             write_logfile(*g_ctx, *g_params, *g_model, *g_input_tokens, g_output_ss->str(), *g_output_tokens);
+
+            // make sure all logs are flushed
+            LOG("Interrupted by user\n");
+            common_log_pause(common_log_main());
+
             _exit(130);
         }
     }
@@ -102,121 +109,107 @@ static void sigint_handler(int signo) {
 #endif
 
 int main(int argc, char ** argv) {
-    gpt_params params;
-    llama_sampling_params & sparams = params.sparams;
+    common_params params;
     g_params = &params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        gpt_params_print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_INFILL)) {
         return 1;
     }
 
-#ifndef LOG_DISABLE_LOGS
-    log_set_target(log_filename_generator("infill", "log"));
-    LOG_TEE("Log start\n");
-    log_dump_cmdline(argc, argv);
-#endif // LOG_DISABLE_LOGS
+    common_init();
+
+    auto & sparams = params.sparams;
 
     console::init(params.simple_io, params.use_color);
     atexit([]() { console::cleanup(); });
 
     if (params.logits_all) {
-        printf("\n************\n");
-        printf("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
-        printf("************\n\n");
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use the 'perplexity' tool for perplexity calculations\n", __func__);
+        LOG_ERR("************\n\n");
 
         return 0;
     }
 
     if (params.embedding) {
-        printf("\n************\n");
-        printf("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
-        printf("************\n\n");
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use the 'embedding' tool for embedding calculations\n", __func__);
+        LOG_ERR("************\n\n");
 
         return 0;
     }
 
     if (params.n_ctx != 0 && params.n_ctx < 8) {
-        LOG_TEE("%s: warning: minimum context size is 8, using minimum size.\n", __func__);
+        LOG_WRN("%s: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;
     }
+
     if (!params.interactive_first && (params.input_prefix.empty() && params.input_suffix.empty())) {
-        printf("\n************\n");
-        printf("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
-        printf("************\n\n");
+        LOG_ERR("\n************\n");
+        LOG_ERR("%s: please use '--interactive_first' or specify '--in_prefix' and/or '--in_suffix'\n", __func__);
+        LOG_ERR("************\n\n");
 
         return 0;
     }
 
     if (params.rope_freq_base != 0.0) {
-        LOG_TEE("%s: warning: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
+        LOG_WRN("%s: changing RoPE frequency base to %g.\n", __func__, params.rope_freq_base);
     }
 
     if (params.rope_freq_scale != 0.0) {
-        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
+        LOG_WRN("%s: 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);
-
-    LOG("%s: llama backend init\n", __func__);
+    LOG_INF("%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;
+    common_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__);
-    llama_init_result llama_init = llama_init_from_gpt_params(params);
+    LOG_INF("%s: load the model and apply lora adapter, if any\n", __func__);
+    common_init_result llama_init = common_init_from_params(params);
 
     model = llama_init.model;
     ctx = llama_init.context;
 
     if (model == NULL) {
-        LOG_TEE("%s: error: unable to load model\n", __func__);
+        LOG_ERR("%s: unable to load model\n", __func__);
         return 1;
     }
 
     const int n_ctx_train = llama_n_ctx_train(model);
     const int n_ctx = llama_n_ctx(ctx);
-    LOG("n_ctx: %d\n", n_ctx);
+    LOG_DBG("n_ctx: %d\n", n_ctx);
 
     if (n_ctx > n_ctx_train) {
-        LOG_TEE("%s: warning: model was trained on only %d context tokens (%d specified)\n",
-                __func__, n_ctx_train, n_ctx);
+        LOG_WRN("%s: model was trained on only %d context tokens (%d specified)\n", __func__, n_ctx_train, n_ctx);
     }
 
     // print system information
     {
-        LOG_TEE("\n");
-        LOG_TEE("%s\n", gpt_params_get_system_info(params).c_str());
+        LOG_INF("\n");
+        LOG_INF("%s\n", common_params_get_system_info(params).c_str());
     }
-    const bool add_bos = llama_should_add_bos_token(model);
-    GGML_ASSERT(llama_add_eos_token(model) != 1);
-    LOG("add_bos: %d\n", add_bos);
+    const bool add_bos = llama_add_bos_token(model);
+    GGML_ASSERT(!llama_add_eos_token(model));
 
     std::vector<llama_token> embd_inp;
     std::vector<llama_token> embd_end;
-    std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-    std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+    std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+    std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-    GGML_ASSERT(llama_token_prefix(model) >= 0);
-    GGML_ASSERT(llama_token_suffix(model) >= 0);
+    GGML_ASSERT(llama_token_fim_pre(model) >= 0);
+    GGML_ASSERT(llama_token_fim_suf(model) >= 0);
 
-    inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-    inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+    inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+    inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
     embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
     embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -225,23 +218,24 @@ int main(int argc, char ** argv) {
     }
     embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
 
-    const llama_token middle_token = llama_token_middle(model);
+    const llama_token middle_token = llama_token_fim_mid(model);
     if (middle_token >= 0) {
         embd_inp.push_back(middle_token);
     }
 
-    LOG("prefix: \"%s\"\n", log_tostr(params.input_prefix));
-    LOG("suffix: \"%s\"\n", log_tostr(params.input_suffix));
-    LOG("tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+    LOG_DBG("add_bos: %d\n", add_bos);
+    LOG_DBG("prefix: \"%s\"\n", params.input_prefix.c_str());
+    LOG_DBG("suffix: \"%s\"\n", params.input_suffix.c_str());
+    LOG_DBG("tokens: %s\n", string_from(ctx, embd_inp).c_str());
 
     // Should not run without any tokens
     if (embd_inp.empty()) {
         embd_inp.push_back(llama_token_bos(model));
-        LOG("embd_inp was considered empty and bos was added: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd_inp).c_str());
+        LOG_WRN("embd_inp was considered empty and bos was added: %s\n", string_from(ctx, embd_inp).c_str());
     }
 
     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);
+        LOG_ERR("%s: prompt is too long (%d tokens, max %d)\n", __func__, (int) embd_inp.size(), n_ctx - 4);
         return 1;
     }
 
@@ -250,9 +244,8 @@ int main(int argc, char ** argv) {
         params.n_keep = (int)embd_inp.size();
     }
 
-    LOG("inp_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_pfx).c_str());
-    LOG("inp_sfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, inp_sfx).c_str());
-
+    LOG_INF("inp_pfx: %s\n", string_from(ctx, inp_pfx).c_str());
+    LOG_INF("inp_sfx: %s\n", string_from(ctx, inp_sfx).c_str());
 
     // enable interactive mode if interactive start is specified
     if (params.interactive_first) {
@@ -260,21 +253,21 @@ int main(int argc, char ** argv) {
     }
 
     if (params.verbose_prompt) {
-        LOG_TEE("\n");
-        LOG_TEE("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
-        LOG_TEE("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        LOG_INF("\n");
+        LOG_INF("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        LOG_INF("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
         for (int i = 0; i < (int) embd_inp.size(); i++) {
-            LOG_TEE("%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+            LOG_INF("%6d -> '%s'\n", embd_inp[i], common_token_to_piece(ctx, embd_inp[i]).c_str());
         }
 
         if (params.n_keep > 0) {
-        LOG_TEE("%s: static prompt based on n_keep: '", __func__);
+        LOG_INF("%s: static prompt based on n_keep: '", __func__);
             for (int i = 0; i < params.n_keep; i++) {
-                LOG_TEE("%s", llama_token_to_piece(ctx, embd_inp[i]).c_str());
+                LOG_CNT("%s", common_token_to_piece(ctx, embd_inp[i]).c_str());
             }
-            LOG_TEE("'\n");
+            LOG_CNT("'\n");
         }
-        LOG_TEE("\n");
+        LOG_INF("\n");
     }
 
     if (params.interactive) {
@@ -291,30 +284,30 @@ int main(int argc, char ** argv) {
         SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
 #endif
 
-        LOG_TEE("%s: interactive mode on.\n", __func__);
+        LOG_INF("%s: interactive mode on.\n", __func__);
 
         if (params.input_prefix_bos) {
-            LOG_TEE("Input prefix with BOS\n");
+            LOG_INF("Input prefix with BOS\n");
         }
 
         if (!params.input_prefix.empty()) {
-            LOG_TEE("Input prefix: '%s'\n", params.input_prefix.c_str());
+            LOG_INF("Input prefix: '%s'\n", params.input_prefix.c_str());
         }
 
         if (!params.input_suffix.empty()) {
-            LOG_TEE("Input suffix: '%s'\n", params.input_suffix.c_str());
+            LOG_INF("Input suffix: '%s'\n", params.input_suffix.c_str());
         }
     }
-    LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
-    LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
-    LOG_TEE("\n\n");
+    smpl = common_sampler_init(model, sparams);
 
-    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");
-    }
+    LOG_INF("sampler seed: %u\n",     common_sampler_get_seed(smpl));
+    LOG_INF("sampler params: \n%s\n", sparams.print().c_str());
+    LOG_INF("sampler chain: %s\n",    common_sampler_print(smpl).c_str());
+
+    LOG_INF("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    LOG_INF("\n");
+    LOG_INF("\n#####  Infill mode  #####\n\n");
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
@@ -325,11 +318,11 @@ int main(int argc, char ** argv) {
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
-        LOG_TEE("== Running in interactive mode. ==\n");
+        LOG_INF("== Running in interactive mode. ==\n");
 #if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
-        LOG_TEE(       " - Press Ctrl+C to interject at any time.\n");
+        LOG_INF(       " - Press Ctrl+C to interject at any time.\n");
 #endif
-        LOG_TEE(       "%s\n", control_message);
+        LOG_INF(       "%s\n", control_message);
 
         is_interacting = params.interactive_first;
     }
@@ -349,8 +342,6 @@ 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()) {
@@ -364,9 +355,8 @@ int main(int argc, char ** argv) {
                 embd.resize(max_embd_size);
 
                 console::set_display(console::error);
-                printf("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
+                LOG_WRN("<<input too long: skipped %d token%s>>", skipped_tokens, skipped_tokens != 1 ? "s" : "");
                 console::set_display(console::reset);
-                fflush(stdout);
             }
 
             // infinite text generation via context swapping
@@ -375,14 +365,14 @@ int main(int argc, char ** argv) {
             // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
             if (n_past + (int) embd.size() > n_ctx) {
                 if (params.n_predict == -2) {
-                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                    LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
                     break;
                 }
 
                 const int n_left    = n_past - params.n_keep - 1;
                 const int n_discard = n_left/2;
 
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
                 llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
@@ -390,9 +380,9 @@ int main(int argc, char ** argv) {
 
                 n_past -= n_discard;
 
-                LOG("after swap: n_past = %d\n", n_past);
+                LOG_DBG("after swap: n_past = %d\n", n_past);
 
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
 
             }
 
@@ -404,16 +394,16 @@ int main(int argc, char ** argv) {
                     n_eval = params.n_batch;
                 }
 
-                LOG("eval: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+                LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
                 if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
-                    LOG_TEE("%s : failed to eval\n", __func__);
+                    LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }
 
                 n_past += n_eval;
 
-                LOG("n_past = %d\n", n_past);
+                LOG_DBG("n_past = %d\n", n_past);
             }
 
         }
@@ -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 = common_sampler_sample(smpl, ctx, -1);
 
-            llama_sampling_accept(ctx_sampling, ctx, id, true);
+            common_sampler_accept(smpl, id, true);
 
-            LOG("last: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, ctx_sampling->prev).c_str());
+            // LOG_DBG("last: %s\n", string_from(ctx, smpl->prev.to_vector()).c_str());
 
             embd.push_back(id);
 
@@ -435,16 +425,16 @@ int main(int argc, char ** argv) {
             // decrement remaining sampling budget
             --n_remain;
 
-            LOG("n_remain: %d\n", n_remain);
+            LOG_DBG("n_remain: %d\n", n_remain);
         } else {
             // some user input remains from prompt or interaction, forward it to processing
-            LOG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
+            LOG_DBG("embd_inp.size(): %d, n_consumed: %d\n", (int) embd_inp.size(), n_consumed);
             while ((int) embd_inp.size() > n_consumed) {
                 embd.push_back(embd_inp[n_consumed]);
 
                 // push the prompt in the sampling context in order to apply repetition penalties later
                 // for the prompt, we don't apply grammar rules
-                llama_sampling_accept(ctx_sampling, ctx, embd_inp[n_consumed], false);
+                common_sampler_accept(smpl, embd_inp[n_consumed], false);
 
                 ++n_consumed;
                 if ((int) embd.size() >= params.n_batch) {
@@ -456,8 +446,8 @@ int main(int argc, char ** argv) {
         // display text
         if (input_echo) {
             for (auto id : embd) {
-                const std::string token_str = llama_token_to_piece(ctx, id);
-                printf("%s", token_str.c_str());
+                const std::string token_str = common_token_to_piece(ctx, id);
+                LOG("%s", token_str.c_str());
 
                 if (embd.size() > 1) {
                     input_tokens.push_back(id);
@@ -466,7 +456,6 @@ int main(int argc, char ** argv) {
                     output_ss << token_str;
                 }
             }
-            fflush(stdout);
         }
         // reset color to default if we there is no pending user input
         if (input_echo && (int) embd_inp.size() == n_consumed) {
@@ -476,13 +465,12 @@ 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 ((common_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());
+                    LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
-                fflush(stdout);
-                printf("\n");
+                LOG("\n");
                 console::set_display(console::user_input);
                 std::string buffer;
                 std::string line;
@@ -517,11 +505,11 @@ int main(int argc, char ** argv) {
                 }
 
                 // tokenize new prefix and suffix
-                std::vector<llama_token> inp_pfx = ::llama_tokenize(ctx, params.input_prefix, false);
-                std::vector<llama_token> inp_sfx = ::llama_tokenize(ctx, params.input_suffix, false);
+                std::vector<llama_token> inp_pfx = common_tokenize(ctx, params.input_prefix, false);
+                std::vector<llama_token> inp_sfx = common_tokenize(ctx, params.input_suffix, false);
 
-                inp_pfx.insert(inp_pfx.begin(), llama_token_prefix(model));
-                inp_sfx.insert(inp_sfx.begin(), llama_token_suffix(model));
+                inp_pfx.insert(inp_pfx.begin(), llama_token_fim_pre(model));
+                inp_sfx.insert(inp_sfx.begin(), llama_token_fim_suf(model));
 
                 embd_inp = params.spm_infill ? inp_sfx : inp_pfx;
                 embd_end = params.spm_infill ? inp_pfx : inp_sfx;
@@ -538,35 +526,33 @@ int main(int argc, char ** argv) {
                 n_remain = params.n_predict;
                 n_past = 0;
                 n_consumed = 0;
-                // LOG_TEE("took new input\n");
                 is_interacting = false;
             }
             // deal with end of generation tokens in interactive mode
-            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
-                LOG("found EOS token\n");
+            else if (llama_token_is_eog(model, common_sampler_last(smpl))) {
+                LOG_DBG("found EOS token\n");
 
                 if (params.interactive) {
 
                     is_interacting = true;
-                    printf("\n");
+                    LOG("\n");
                     console::set_display(console::user_input);
-                    fflush(stdout);
                }
             }
 
             if (n_past > 0 && is_interacting && !params.interactive) {
-                LOG("waiting for user input\n");
+                LOG_DBG("waiting for user input\n");
 
                 if (params.input_prefix_bos) {
-                    LOG("adding input prefix BOS token\n");
+                    LOG_DBG("adding input prefix BOS token\n");
                     embd_inp.push_back(llama_token_bos(model));
                 }
 
                 std::string buffer;
                 if (!params.input_prefix.empty()) {
-                    LOG("appending input prefix: '%s'\n", params.input_prefix.c_str());
+                    LOG_DBG("appending input prefix: '%s'\n", params.input_prefix.c_str());
                     buffer += params.input_prefix;
-                    printf("%s", buffer.c_str());
+                    LOG("%s", buffer.c_str());
                 }
 
                 std::string line;
@@ -584,30 +570,30 @@ int main(int argc, char ** argv) {
                 if (buffer.length() > 1) {
                     // append input suffix if any
                     if (!params.input_suffix.empty()) {
-                        LOG("appending input suffix: '%s'\n", params.input_suffix.c_str());
+                        LOG_DBG("appending input suffix: '%s'\n", params.input_suffix.c_str());
                         buffer += params.input_suffix;
-                        printf("%s", params.input_suffix.c_str());
+                        LOG("%s", params.input_suffix.c_str());
                     }
 
-                    LOG("buffer: '%s'\n", buffer.c_str());
+                    LOG_DBG("buffer: '%s'\n", buffer.c_str());
 
                     const size_t original_size = embd_inp.size();
 
-                    const auto line_inp = ::llama_tokenize(ctx, buffer, false);
-                    LOG("input tokens: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, line_inp).c_str());
+                    const auto line_inp = common_tokenize(ctx, buffer, false);
+                    LOG_DBG("input tokens: %s\n", string_from(ctx, line_inp).c_str());
 
                     embd_inp.insert(embd_inp.end(), line_inp.begin(), line_inp.end());
 
                     for (size_t i = original_size; i < embd_inp.size(); ++i) {
                         const llama_token token = embd_inp[i];
                         output_tokens.push_back(token);
-                        output_ss << llama_token_to_piece(ctx, token);
+                        output_ss << common_token_to_piece(ctx, token);
                     }
 
                     n_remain -= line_inp.size();
-                    LOG("n_remain: %d\n", n_remain);
+                    LOG_DBG("n_remain: %d\n", n_remain);
                 } else {
-                    LOG("empty line, passing control back\n");
+                    LOG_DBG("empty line, passing control back\n");
                 }
 
                 input_echo = false; // do not echo this again
@@ -615,7 +601,7 @@ int main(int argc, char ** argv) {
 
             if (n_past > 0) {
                 if (is_interacting) {
-                    llama_sampling_reset(ctx_sampling);
+                    common_sampler_reset(smpl);
                 }
                 is_interacting = false;
             }
@@ -634,22 +620,18 @@ int main(int argc, char ** argv) {
         }
     }
     if (!params.interactive && n_remain <= 0) {
-        printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
-        fflush(stdout);
+        LOG("%s", common_token_to_piece(ctx, llama_token_eot(model)).c_str());
     }
 
-    llama_print_timings(ctx);
+    LOG("\n");
+    common_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);
+    common_sampler_free(smpl);
     llama_backend_free();
 
-#ifndef LOG_DISABLE_LOGS
-    LOG_TEE("Log end\n");
-#endif // LOG_DISABLE_LOGS
-
     return 0;
 }

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());
+    if (llama_supports_rpc()) {
+        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("  --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];
@@ -409,6 +439,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_k.insert(params.type_k.end(), types.begin(), types.end());
         } else if (arg == "-ctv" || arg == "--cache-type-v") {
             if (++i >= argc) {
@@ -425,6 +458,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 types.push_back(gt);
             }
+            if (invalid_param) {
+                break;
+            }
             params.type_v.insert(params.type_v.end(), types.begin(), types.end());
         } else if (arg == "-t" || arg == "--threads") {
             if (++i >= argc) {
@@ -433,6 +469,27 @@ 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;
@@ -440,7 +497,7 @@ 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());
-        } else if (arg == "-rpc" || arg == "--rpc") {
+        } else if (llama_supports_rpc() && (arg == "-rpc" || arg == "--rpc")) {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
@@ -467,6 +524,9 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
                 }
                 modes.push_back(mode);
             }
+            if (invalid_param) {
+                break;
+            }
             params.split_mode.insert(params.split_mode.end(), modes.begin(), modes.end());
         } else if (arg == "-mg" || arg == "--main-gpu") {
             if (++i >= argc) {
@@ -541,6 +601,18 @@ 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 +627,8 @@ 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 +659,9 @@ 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 +675,9 @@ 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 +747,10 @@ 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 +764,9 @@ 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 +793,9 @@ 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 +822,9 @@ 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 +861,9 @@ 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 +890,9 @@ 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 +970,14 @@ 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 +986,7 @@ 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 +995,7 @@ 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 +1028,8 @@ 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 +1097,39 @@ 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 +1138,7 @@ 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 +1161,25 @@ 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 +1188,7 @@ struct markdown_printer : public printer {
             return -30;
         }
         if (field == "t/s") {
-            return 16;
+            return 20;
         }
         if (field == "size" || field == "params") {
             return 10;
@@ -1149,6 +1270,15 @@ 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 +1480,8 @@ 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 +1515,8 @@ 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 +1536,13 @@ 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 +1568,40 @@ 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 +1611,15 @@ 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 +1637,11 @@ 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/build.gradle.kts

@@ -18,6 +18,7 @@ android {
         }
         externalNativeBuild {
             cmake {
+                arguments += "-DLLAMA_BUILD_COMMON=ON"
                 arguments += "-DCMAKE_BUILD_TYPE=Release"
                 cppFlags += listOf()
                 arguments += listOf()

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;
 
@@ -186,11 +186,11 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
     for (nri = 0; nri < nr; nri++) {
         LOGi("Benchmark prompt processing (pp)");
 
-        llama_batch_clear(*batch);
+        common_batch_clear(*batch);
 
         const int n_tokens = pp;
         for (i = 0; i < n_tokens; i++) {
-            llama_batch_add(*batch, 0, i, { 0 }, false);
+            common_batch_add(*batch, 0, i, { 0 }, false);
         }
 
         batch->logits[batch->n_tokens - 1] = true;
@@ -210,9 +210,9 @@ Java_android_llama_cpp_LLamaAndroid_bench_1model(
         const auto t_tg_start = ggml_time_us();
         for (i = 0; i < tg; i++) {
 
-            llama_batch_clear(*batch);
+            common_batch_clear(*batch);
             for (j = 0; j < pl; j++) {
-                llama_batch_add(*batch, 0, i, { j }, true);
+                common_batch_add(*batch, 0, i, { j }, true);
             }
 
             LOGi("llama_decode() text generation: %d", i);
@@ -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) {
@@ -340,7 +357,7 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
 
-    const auto tokens_list = llama_tokenize(context, text, 1);
+    const auto tokens_list = common_tokenize(context, text, 1);
 
     auto n_ctx = llama_n_ctx(context);
     auto n_kv_req = tokens_list.size() + (n_len - tokens_list.size());
@@ -352,14 +369,14 @@ Java_android_llama_cpp_LLamaAndroid_completion_1init(
     }
 
     for (auto id : tokens_list) {
-        LOGi("%s", llama_token_to_piece(context, id).c_str());
+        LOGi("%s", common_token_to_piece(context, id).c_str());
     }
 
-    llama_batch_clear(*batch);
+    common_batch_clear(*batch);
 
     // evaluate the initial prompt
     for (auto i = 0; i < tokens_list.size(); i++) {
-        llama_batch_add(*batch, tokens_list[i], i, { 0 }, false);
+        common_batch_add(*batch, tokens_list[i], i, { 0 }, false);
     }
 
     // llama_decode will output logits only for the last token of the prompt
@@ -381,38 +398,28 @@ 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) {
         return nullptr;
     }
 
-    auto new_token_chars = llama_token_to_piece(context, new_token_id);
+    auto new_token_chars = common_token_to_piece(context, new_token_id);
     cached_token_chars += new_token_chars;
 
     jstring new_token = nullptr;
@@ -424,8 +431,8 @@ Java_android_llama_cpp_LLamaAndroid_completion_1loop(
         new_token = env->NewStringUTF("");
     }
 
-    llama_batch_clear(*batch);
-    llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
+    common_batch_clear(*batch);
+    common_batch_add(*batch, new_token_id, n_cur, { 0 }, true);
 
     env->CallVoidMethod(intvar_ncur, la_int_var_inc);
 

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/llama.vim

@@ -1,135 +0,0 @@
-" Requires an already running llama.cpp server
-" To install either copy or symlink to ~/.vim/autoload/llama.vim
-" Then start with either :call llama#doLlamaGen(),
-" or add a keybind to your vimrc such as
-" nnoremap Z :call llama#doLlamaGen()<CR>
-" Similarly, you could add an insert mode keybind with
-" inoremap <C-B> <Cmd>call llama#doLlamaGen()<CR>
-"
-" g:llama_api_url, g:llama_api_key and g:llama_overrides can be configured in your .vimrc
-" let g:llama_api_url = "192.168.1.10:8080"
-" llama_overrides can also be set through buffer/window scopes. For instance
-" autocmd filetype python let b:llama_overrides = {"temp": 0.2}
-" Could be added to your .vimrc to automatically set a lower temperature when
-" editing a python script
-" Additionally, an override dict can be stored at the top of a file
-" !*{"stop": ["User:"]}
-" Could be added to the start of your chatlog.txt to set the stopping token
-" These parameter dicts are merged together from lowest to highest priority:
-" server default -> g:llama_overrides -> w:llama_overrides ->
-" b:llama_overrides -> in file (!*) overrides
-"
-" Sublists (like logit_bias and stop) are overridden, not merged
-" Example override:
-" !*{"logit_bias": [[13, -5], [2, false]], "temperature": 1, "top_k": 5, "top_p": 0.5, "n_predict": 256, "repeat_last_n": 256, "repeat_penalty": 1.17647}
-if !exists("g:llama_api_url")
-    let g:llama_api_url= "127.0.0.1:8080"
-endif
-if !exists("g:llama_overrides")
-   let g:llama_overrides = {}
-endif
-const s:querydata = {"n_predict": 256, "stop": [ "\n" ], "stream": v:true }
-const s:curlcommand = ['curl','--data-raw', "{\"prompt\":\"### System:\"}", '--silent', '--no-buffer', '--request', 'POST', '--url', g:llama_api_url .. '/completion', '--header', "Content-Type: application/json"]
-let s:linedict = {}
-
-func s:callbackHandler(bufn, channel, msg)
-   if len(a:msg) < 3
-      return
-   elseif a:msg[0] == "d"
-      let l:msg = a:msg[6:-1]
-   else
-      let l:msg = a:msg
-   endif
-   let l:decoded_msg = json_decode(l:msg)
-   let l:newtext = split(l:decoded_msg['content'], "\n", 1)
-   if len(l:newtext) > 0
-      call setbufline(a:bufn, s:linedict[a:bufn], getbufline(a:bufn, s:linedict[a:bufn])[0] .. newtext[0])
-   else
-      echo "nothing genned"
-   endif
-   if len(newtext) > 1
-      let l:failed = appendbufline(a:bufn, s:linedict[a:bufn], newtext[1:-1])
-      let s:linedict[a:bufn] = s:linedict[a:bufn] + len(newtext)-1
-   endif
-   if has_key(l:decoded_msg, "stop") && l:decoded_msg.stop
-       echo "Finished generation"
-   endif
-endfunction
-
-func llama#doLlamaGen()
-   if exists("b:job")
-      if job_status(b:job) == "run"
-         call job_stop(b:job)
-         return
-      endif
-   endif
-
-   let l:cbuffer = bufnr("%")
-   let s:linedict[l:cbuffer] = line('$')
-   let l:buflines = getbufline(l:cbuffer, 1, 1000)
-   let l:querydata = copy(s:querydata)
-   call extend(l:querydata, g:llama_overrides)
-   if exists("w:llama_overrides")
-      call extend(l:querydata, w:llama_overrides)
-   endif
-   if exists("b:llama_overrides")
-      call extend(l:querydata, b:llama_overrides)
-   endif
-   if l:buflines[0][0:1] == '!*'
-      let l:userdata = json_decode(l:buflines[0][2:-1])
-      call extend(l:querydata, l:userdata)
-      let l:buflines = l:buflines[1:-1]
-   endif
-   let l:querydata.prompt = join(l:buflines, "\n")
-   let l:curlcommand = copy(s:curlcommand)
-   if exists("g:llama_api_key")
-       call extend(l:curlcommand, ['--header', 'Authorization: Bearer ' .. g:llama_api_key])
-   endif
-   let l:curlcommand[2] = json_encode(l:querydata)
-   let b:job = job_start(l:curlcommand, {"callback": function("s:callbackHandler", [l:cbuffer])})
-endfunction
-
-" Echos the tokkenization of the provided string , or cursor to end of word
-" Onus is placed on the user to include the preceding space
-func llama#tokenizeWord(...)
-    if (a:0 > 0)
-        let l:input = a:1
-    else
-        exe "normal \"*ye"
-        let l:input = @*
-    endif
-    let l:querydata = {"content": l:input}
-    let l:curlcommand = copy(s:curlcommand)
-    let l:curlcommand[2] = json_encode(l:querydata)
-    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
-   let s:token_job = job_start(l:curlcommand, {"callback": function("s:tokenizeWordCallback", [l:input])})
-endfunction
-
-func s:tokenizeWordCallback(plaintext, channel, msg)
-    echo '"' .. a:plaintext ..'" - ' .. string(json_decode(a:msg).tokens)
-endfunction
-
-
-" Echos the token count of the entire buffer (or provided string)
-" Example usage :echo llama#tokenCount()
-func llama#tokenCount(...)
-    if (a:0 > 0)
-        let l:buflines = a:1
-    else
-        let l:buflines = getline(1,1000)
-        if l:buflines[0][0:1] == '!*'
-            let l:buflines = l:buflines[1:-1]
-        endif
-        let l:buflines = join(l:buflines, "\n")
-    endif
-    let l:querydata = {"content": l:buflines}
-    let l:curlcommand = copy(s:curlcommand)
-    let l:curlcommand[2] = json_encode(l:querydata)
-    let l:curlcommand[8] = g:llama_api_url .. "/tokenize"
-   let s:token_job = job_start(l:curlcommand, {"callback": "s:tokenCountCallback"})
-endfunction
-
-func s:tokenCountCallback(channel, msg)
-    let resp = json_decode(a:msg)
-    echo len(resp.tokens)
-endfunction

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

@@ -15,9 +15,9 @@ 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
-python ./convert-hf-to-gguf.py ../MiniCPM-Llama3-V-2_5/model
+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
 ./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M

examples/llava/README-minicpmv2.6.md

@@ -0,0 +1,107 @@
+## MiniCPM-V 2.6
+
+### Prepare models and code
+
+Download [MiniCPM-V-2_6](https://huggingface.co/openbmb/MiniCPM-V-2_6) PyTorch model from huggingface to "MiniCPM-V-2_6" folder.
+
+Clone llama.cpp:
+```bash
+git clone git@github.com:OpenBMB/llama.cpp.git
+cd llama.cpp
+git checkout minicpmv-main
+```
+
+### Usage of MiniCPM-V 2.6
+
+Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-V-2_6-gguf) by us)
+
+```bash
+python ./examples/llava/minicpmv-surgery.py -m ../MiniCPM-V-2_6
+python ./examples/llava/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-V-2_6 --minicpmv-projector ../MiniCPM-V-2_6/minicpmv.projector --output-dir ../MiniCPM-V-2_6/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5 --minicpmv_version 3
+python ./convert_hf_to_gguf.py ../MiniCPM-V-2_6/model
+
+# quantize int4 version
+./llama-quantize ../MiniCPM-V-2_6/model/ggml-model-f16.gguf ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf Q4_K_M
+```
+
+Build for Linux or Mac
+
+```bash
+make
+make llama-minicpmv-cli
+```
+
+Inference on Linux or Mac
+```
+# run f16 version
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-f16.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
+
+# run quantized int4 version
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+
+# or run in interactive mode
+./llama-minicpmv-cli -m ../MiniCPM-V-2_6/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-V-2_6/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
+```
+
+### Video
+Install FFmpeg
+```
+brew install ffmpeg
+brew install pkg-config
+```
+
+### Android
+
+#### Build on Android device using Termux
+We found that build on Android device would bring better runtime performance, so we recommend to build on device.
+
+[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
+
+Install tools in Termux:
+```
+apt update && apt upgrade -y
+apt install git make cmake
+```
+
+It's recommended to move your model inside the `~/` directory for best performance:
+```
+cd storage/downloads
+mv model.gguf ~/
+```
+
+#### Building the Project using Android NDK
+Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
+
+Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
+
+```bash
+mkdir build-android
+cd build-android
+export NDK=/your_ndk_path
+cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
+make
+```
+
+Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
+
+Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
+
+(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
+```
+$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
+$cd /data/data/com.termux/files/home/bin
+$chmod +x ./*
+```
+
+Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
+
+```
+$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
+$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
+```
+
+Now, you can start chatting:
+```
+$cd /data/data/com.termux/files/home/bin
+$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg  -p "What is in the image?"
+```

examples/llava/clip.cpp

@@ -3,7 +3,6 @@
 // I'll gradually clean and extend it
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
-#include "log.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -20,6 +19,10 @@
 #include "ggml-cann.h"
 #endif
 
+#ifdef GGML_USE_VULKAN
+#include "ggml-vulkan.h"
+#endif
+
 #define STB_IMAGE_IMPLEMENTATION
 #include "stb_image.h"
 
@@ -36,6 +39,11 @@
 #include <cinttypes>
 #include <limits>
 
+#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_DBG(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+
 //#define CLIP_DEBUG_FUNCTIONS
 
 // RGB uint8 image
@@ -81,6 +89,7 @@ static std::string format(const char * fmt, ...) {
 #define KEY_HAS_VIS_ENC         "clip.has_vision_encoder"
 #define KEY_HAS_LLAVA_PROJ      "clip.has_llava_projector"
 #define KEY_HAS_MINICPMV_PROJ   "clip.has_minicpmv_projector"
+#define KEY_MINICPMV_VERSION    "clip.minicpmv_version"
 #define KEY_USE_GELU            "clip.use_gelu"
 #define KEY_N_EMBD              "clip.%s.embedding_length"
 #define KEY_N_FF                "clip.%s.feed_forward_length"
@@ -160,7 +169,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
 static int get_key_idx(const gguf_context * ctx, const char * key) {
     int i = gguf_find_key(ctx, key);
     if (i == -1) {
-        LOG_TEE("key %s not found in file\n", key);
+        LOG_ERR("key %s not found in file\n", key);
         throw std::runtime_error(format("Missing required key: %s", key));
     }
 
@@ -211,13 +220,19 @@ 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) {
@@ -259,7 +274,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
 
 static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
     size_t tensor_size = ggml_nbytes(tensor);
-    LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+    LOG_INF("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
             prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
             tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
 }
@@ -277,7 +292,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
 static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
         return;
     }
 
@@ -296,7 +311,7 @@ static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::s
 static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
+        LOG_ERR("Failed to open file for writing: %s\n", filename.c_str());
         return;
     }
 
@@ -526,6 +541,7 @@ struct clip_ctx {
     bool has_vision_encoder  = false;
     bool has_llava_projector = false;
     bool has_minicpmv_projector = false;
+    int minicpmv_version = 2;
 
     struct clip_vision_model vision_model;
     projector_type proj_type = PROJECTOR_TYPE_MLP;
@@ -556,7 +572,7 @@ struct clip_ctx {
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        LOG_ERR("This gguf file seems to have no vision encoder\n");
         return nullptr;
     }
 
@@ -570,7 +586,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         if (load_image_size == nullptr) {
             load_image_size = clip_image_size_init();
         }
-        LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        LOG_DBG("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
         image_size_width  = load_image_size->width;
         image_size_height = load_image_size->height;
         if (is_inf) {
@@ -641,7 +657,12 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     if (ctx->has_minicpmv_projector) {
         int pos_w = image_size_width/patch_size;
         int pos_h = image_size_height/patch_size;
-        pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        if (ctx->minicpmv_version == 2) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
+        }
+        else if (ctx->minicpmv_version == 3) {
+            pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 3584, pos_w * pos_h, 1);
+        }
         ggml_set_name(pos_embed, "pos_embed");
         ggml_set_input(pos_embed);
     }
@@ -768,8 +789,8 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             embeddings = ggml_gelu(ctx0, embeddings);
             embeddings = ggml_mul_mat(ctx0, model.mm_2_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_2_b);
-
-        } else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
+        }
+        else if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
             embeddings = ggml_mul_mat(ctx0, model.mm_0_w, embeddings);
             embeddings = ggml_add(ctx0, embeddings, model.mm_0_b);
             // ggml_tensor_printf(embeddings, "mm_0_w",0,true,false);
@@ -949,10 +970,20 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             }
 
             { // attention
-                const int hidden_size = 4096;
+                int hidden_size = 4096;
                 const int d_head = 128;
-                const int n_head = hidden_size/d_head;
-                const int num_query = 96;
+                int n_head = hidden_size/d_head;
+                int num_query = 96;
+                if (ctx->minicpmv_version == 2) {
+                    hidden_size = 4096;
+                    n_head = hidden_size/d_head;
+                    num_query = 96;
+                }
+                else if (ctx->minicpmv_version == 3) {
+                    hidden_size = 3584;
+                    n_head = hidden_size/d_head;
+                    num_query = 64;
+                }
 
                 struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
                 Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
@@ -1020,21 +1051,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         const int idx_name = gguf_find_key(ctx, KEY_NAME);
         if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
             const std::string name = gguf_get_val_str(ctx, idx_name);
-            LOG_TEE("%s: model name:   %s\n", __func__, name.c_str());
+            LOG_INF("%s: model name:   %s\n", __func__, name.c_str());
         }
-        LOG_TEE("%s: description:  %s\n", __func__, description.c_str());
-        LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
-        LOG_TEE("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
-        LOG_TEE("%s: n_tensors:    %d\n", __func__, n_tensors);
-        LOG_TEE("%s: n_kv:         %d\n", __func__, n_kv);
-        LOG_TEE("%s: ftype:        %s\n", __func__, ftype_str.c_str());
-        LOG_TEE("\n");
+        LOG_INF("%s: description:  %s\n", __func__, description.c_str());
+        LOG_INF("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+        LOG_INF("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
+        LOG_INF("%s: n_tensors:    %d\n", __func__, n_tensors);
+        LOG_INF("%s: n_kv:         %d\n", __func__, n_kv);
+        LOG_INF("%s: ftype:        %s\n", __func__, ftype_str.c_str());
+        LOG_INF("\n");
     }
     const int n_tensors = gguf_get_n_tensors(ctx);
 
     // kv
     const int n_kv = gguf_get_n_kv(ctx);
-    LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+    LOG_INF("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
         __func__, n_kv, n_tensors, fname);
     {
         std::map<enum ggml_type, uint32_t> n_type;
@@ -1045,7 +1076,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             n_type[type]++;
         }
 
-        LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        LOG_INF("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
         for (int i = 0; i < n_kv; i++) {
             const char * name           = gguf_get_key(ctx, i);
             const enum gguf_type type   = gguf_get_kv_type(ctx, i);
@@ -1061,7 +1092,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             }
             replace_all(value, "\n", "\\n");
 
-            LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+            LOG_INF("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
         }
 
         // print type counts
@@ -1070,7 +1101,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 continue;
             }
 
-            LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+            LOG_INF("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
         }
     }
 
@@ -1085,13 +1116,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             size_t tensor_size = ggml_nbytes(cur);
             model_size += tensor_size;
             if (verbosity >= 3) {
-                LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+                LOG_INF("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
                        __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
             }
         }
     }
 
-    clip_ctx * new_clip = new clip_ctx;
+    clip_ctx * new_clip = new clip_ctx{};
 
     // update projector type
     {
@@ -1112,23 +1143,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
 #ifdef GGML_USE_CUDA
     new_clip->backend = ggml_backend_cuda_init(0);
-    LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
+    LOG_INF("%s: CLIP using CUDA backend\n", __func__);
 #endif
 
 #ifdef GGML_USE_METAL
     new_clip->backend = ggml_backend_metal_init();
-    LOG_TEE("%s: CLIP using Metal backend\n", __func__);
+    LOG_INF("%s: CLIP using Metal backend\n", __func__);
 #endif
 
 #ifdef GGML_USE_CANN
     new_clip->backend = ggml_backend_cann_init(0);
-    LOG_TEE("%s: CLIP using CANN backend\n", __func__);
+    LOG_INF("%s: CLIP using CANN backend\n", __func__);
 #endif
 
+#ifdef GGML_USE_VULKAN
+    new_clip->backend = ggml_backend_vk_init(0);
+    LOG_INF("%s: CLIP using Vulkan backend\n", __func__);
+#endif
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
-        LOG_TEE("%s: CLIP using CPU backend\n", __func__);
+        LOG_INF("%s: CLIP using CPU backend\n", __func__);
     }
 
     // model size and capabilities
@@ -1149,6 +1184,11 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
         }
 
+        idx = gguf_find_key(ctx, KEY_MINICPMV_VERSION);
+        if (idx != -1) {
+            new_clip->minicpmv_version = gguf_get_val_i32(ctx, idx);
+        }
+
         // GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
 
         GGML_ASSERT(new_clip->has_vision_encoder);
@@ -1158,16 +1198,16 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
 
         if (verbosity >= 1) {
-            LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
-            LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
-            LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            LOG_TEE("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
-            LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
-            LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+            LOG_INF("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
+            LOG_INF("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+            LOG_INF("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_INF("%s: minicpmv_projector:  %d\n", __func__, new_clip->has_minicpmv_projector);
+            LOG_INF("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+            LOG_INF("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
-    LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+    LOG_INF("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
 
     // load tensors
     {
@@ -1180,7 +1220,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         new_clip->ctx_data = ggml_init(params);
         if (!new_clip->ctx_data) {
-            LOG_TEE("%s: ggml_init() failed\n", __func__);
+            LOG_ERR("%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1188,7 +1228,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         auto fin = std::ifstream(fname, std::ios::binary);
         if (!fin) {
-            LOG_TEE("cannot open model file for loading tensors\n");
+            LOG_ERR("cannot open model file for loading tensors\n");
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1210,7 +1250,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
-                LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
+                LOG_ERR("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
                 gguf_free(ctx);
                 return nullptr;
@@ -1281,23 +1321,23 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         if (verbosity >= 2) {
-            LOG_TEE("\n%s: vision model hparams\n", __func__);
-            LOG_TEE("image_size         %d\n", hparams.image_size);
-            LOG_TEE("patch_size         %d\n", hparams.patch_size);
-            LOG_TEE("v_hidden_size      %d\n", hparams.hidden_size);
-            LOG_TEE("v_n_intermediate   %d\n", hparams.n_intermediate);
-            LOG_TEE("v_projection_dim   %d\n", hparams.projection_dim);
-            LOG_TEE("v_n_head           %d\n", hparams.n_head);
-            LOG_TEE("v_n_layer          %d\n", hparams.n_layer);
-            LOG_TEE("v_eps              %f\n", hparams.eps);
-            LOG_TEE("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
-            LOG_TEE("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
-            LOG_TEE("v_image_grid_pinpoints: ");
+            LOG_INF("\n%s: vision model hparams\n", __func__);
+            LOG_INF("image_size         %d\n", hparams.image_size);
+            LOG_INF("patch_size         %d\n", hparams.patch_size);
+            LOG_INF("v_hidden_size      %d\n", hparams.hidden_size);
+            LOG_INF("v_n_intermediate   %d\n", hparams.n_intermediate);
+            LOG_INF("v_projection_dim   %d\n", hparams.projection_dim);
+            LOG_INF("v_n_head           %d\n", hparams.n_head);
+            LOG_INF("v_n_layer          %d\n", hparams.n_layer);
+            LOG_INF("v_eps              %f\n", hparams.eps);
+            LOG_INF("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+            LOG_INF("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+            LOG_INF("v_image_grid_pinpoints: ");
             for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
-                LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
+                LOG_INF("%d ", hparams.image_grid_pinpoints[i]);
             }
-            LOG_TEE("\n");
-            LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+            LOG_INF("\n");
+            LOG_INF("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
 
         }
 
@@ -1335,7 +1375,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
         } catch(const std::exception& /*e*/) {
-            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
+            LOG_ERR("%s: failed to load vision model tensors\n", __func__);
         }
 
         // LLaVA projection
@@ -1364,7 +1404,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             } catch (std::runtime_error & /*e*/) { }
             try {
                 vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
-                // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
+                // LOG_INF("%s: image_newline tensor (llava-1.6) found\n", __func__);
             } catch (std::runtime_error & /*e*/) { }
         } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
@@ -1465,7 +1505,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
-        LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+        LOG_INF("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
     return new_clip;
@@ -1516,7 +1556,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
     int nx, ny, nc;
     auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
     if (!data) {
-        LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
+        LOG_ERR("%s: failed to load image '%s'\n", __func__, fname);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1528,7 +1568,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
     int nx, ny, nc;
     auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
     if (!data) {
-        LOG_TEE("%s: failed to decode image bytes\n", __func__);
+        LOG_ERR("%s: failed to decode image bytes\n", __func__);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1587,7 +1627,7 @@ 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));
 }
 
@@ -1718,7 +1758,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int> & or
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // LOG_INF("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -1791,10 +1831,6 @@ 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}) {
@@ -1840,7 +1876,7 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
     const int multiple = fmin(ceil(ratio), max_slice_nums);
 
     std::vector<std::vector<clip_image_u8 *>> images;
-    LOG_TEE("%s: multiple %d\n", __func__, multiple);
+    LOG_INF("%s: multiple %d\n", __func__, multiple);
     images.push_back(std::vector<clip_image_u8 *>());
 
     if (multiple <= 1) {
@@ -1855,17 +1891,17 @@ static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_imag
         clip_image_u8 * source_image = clip_image_u8_init();
         bicubic_resize(*img, *source_image, best_size.first, best_size.second);
         // source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
-        LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
+        LOG_INF("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
         images[images.size()-1].push_back(source_image);
 
         std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
-        LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
+        LOG_INF("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
 
         auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
         clip_image_u8 * refine_image = clip_image_u8_init();
         bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
 
-        LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
+        LOG_INF("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
 
         // split_to_patches
         int width = refine_image->nx;
@@ -1910,17 +1946,19 @@ int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
 // returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
 // res_imgs memory is being allocated here, previous allocations will be freed if found
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
-    if (clip_is_minicpmv(ctx)) {
-        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img);
+
+    if(clip_is_minicpmv(ctx)){
+        int max_slice_nums = 9;
+        std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img, max_slice_nums);
         res_imgs->size = 0;
-        for (size_t i = 0; i < imgs.size(); ++i) {
+        for (size_t i = 0; i < imgs.size(); ++i){
             res_imgs->size += imgs[i].size();
         }
         res_imgs->data = new clip_image_f32[res_imgs->size];
         int idx = 0;
         for (size_t i = 0; i < imgs.size(); ++i) {
             for (size_t j = 0; j < imgs[i].size(); ++j) {
-                LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
+                LOG_DBG("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
                 clip_image_f32 * res = clip_image_f32_init();
                 normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
                 res_imgs->data[idx++] = *res;
@@ -1932,7 +1970,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
 
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        LOG_ERR("This gguf file seems to have no vision encoder\n");
         return false;
     }
     auto & params = ctx->vision_model.hparams;
@@ -2009,7 +2047,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             }
 
             for (size_t i = 0; i < patches.size(); i++) {
-                // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+                // LOG_DBG("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
                 clip_image_u8_free(patches[i]);
             }
 
@@ -2146,7 +2184,12 @@ int clip_n_patches(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
     } else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
-        n_patches = 96;
+        if (ctx->minicpmv_version == 2) {
+            n_patches = 96;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            n_patches = 64;
+        }
     }
 
     return n_patches;
@@ -2240,7 +2283,7 @@ static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, co
 
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        LOG_ERR("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -2252,7 +2295,7 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
 
 bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
     if (!ctx->has_vision_encoder) {
-        LOG_TEE("This gguf file seems to have no vision encoder\n");
+        LOG_ERR("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -2282,6 +2325,11 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
     const int patch_size    = hparams.patch_size;
     const int num_patches   = ((image_size_width / patch_size) * (image_size_height / patch_size));
     const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
+    if(ctx->load_image_size==nullptr){
+        ctx->load_image_size= clip_image_size_init();
+    }
+    const int pos_w = ctx->load_image_size->width/patch_size;
+    const int pos_h = ctx->load_image_size->height/patch_size;
 
     {
         struct ggml_tensor * inp_raw = ggml_graph_get_tensor(gf, "inp_raw");
@@ -2316,8 +2364,18 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
             struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
             int* positions_data = (int*)malloc(ggml_nbytes(positions));
-            for (int i = 0; i < num_positions; i++) {
-                positions_data[i] = std::floor(70.0*i/num_positions);
+            int bucket_coords_h[70];
+            int bucket_coords_w[70];
+            for (int i = 0; i < pos_h; i++){
+                bucket_coords_h[i] = std::floor(70.0*i/pos_h);
+            }
+            for (int i = 0; i < pos_w; i++){
+                bucket_coords_w[i] = std::floor(70.0*i/pos_w);
+            }
+            for (int i = 0, id = 0; i < pos_h; i++){
+                for (int j = 0; j < pos_w; j++){
+                    positions_data[id++] = bucket_coords_h[i]*70 + bucket_coords_w[j];
+                }
             }
             ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
             free(positions_data);
@@ -2328,12 +2386,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             //    -> https://huggingface.co/Qwen/Qwen-VL/tree/main
             //    -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
             struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
-            if(ctx->load_image_size==nullptr){
-                ctx->load_image_size= clip_image_size_init();
-            }
-            int pos_w = ctx->load_image_size->width/patch_size;
-            int pos_h = ctx->load_image_size->height/patch_size;
             int embed_dim = 4096;
+            if (ctx->minicpmv_version == 2) {
+                embed_dim = 4096;
+            }
+            else if (ctx->minicpmv_version == 3) {
+                embed_dim = 3584;
+            }
             auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
 
             float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
@@ -2346,7 +2405,8 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
             ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
             free(pos_embed_data);
         }
-    } else {
+    }
+    else{
         {
             if (ctx->has_class_embedding) {
                 struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
@@ -2384,16 +2444,10 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
         ggml_backend_cpu_set_n_threads(ctx->backend, n_threads);
     }
 
-#ifdef GGML_USE_METAL
-    if (ggml_backend_is_metal(ctx->backend)) {
-        ggml_backend_metal_set_n_cb(ctx->backend, n_threads);
-    }
-#endif
-
     ggml_backend_graph_compute(ctx->backend, gf);
 
     // the last node is the embedding tensor
-    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = 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));
@@ -2465,7 +2519,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             new_type = type;
             if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
                 new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
-                // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+                // LOG_ERR("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
             }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
@@ -2484,7 +2538,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 f32_data = (float *)conv_buf.data();
                 break;
             default:
-                LOG_TEE("Please use an input file in f32 or f16\n");
+                LOG_ERR("Please use an input file in f32 or f16\n");
                 gguf_free(ctx_out);
                 return false;
             }
@@ -2511,7 +2565,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+        LOG_INF("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 
@@ -2527,8 +2581,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     gguf_free(ctx_out);
 
     {
-        LOG_TEE("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
-        LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+        LOG_INF("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        LOG_INF("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
     }
 
     return true;
@@ -2548,13 +2602,21 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         return ctx->vision_model.mm_3_b->ne[0];
     }
     if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
-        return 4096;
+        if (ctx->minicpmv_version == 2) {
+            return 4096;
+        }
+        else if (ctx->minicpmv_version == 3) {
+            return 3584;
+        }
     }
 
     std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
     throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
 }
 
-bool clip_is_minicpmv(const struct clip_ctx * ctx) {
-    return ctx->has_minicpmv_projector;
+int clip_is_minicpmv(const struct clip_ctx * ctx) {
+    if (ctx->has_minicpmv_projector) {
+        return ctx->minicpmv_version;
+    }
+    return 0;
 }

examples/llava/clip.h

@@ -85,7 +85,7 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
 
 CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
 
-CLIP_API bool clip_is_minicpmv(const struct clip_ctx * ctx);
+CLIP_API int clip_is_minicpmv(const struct clip_ctx * ctx);
 
 #ifdef __cplusplus
 }

examples/llava/convert_image_encoder_to_gguf.py

@@ -274,7 +274,7 @@ fout.add_bool("clip.use_gelu", use_gelu)
 
 
 if has_llava_projector:
-    model.vision_model.encoder.layers.pop(-1)  # pyright: ignore[reportAttributeAccessIssue]
+    model.vision_model.encoder.layers.pop(-1)
     projector = torch.load(args.llava_projector)
     for name, data in projector.items():
         name = get_tensor_name(name)
@@ -288,7 +288,7 @@ if has_llava_projector:
 
     print("Projector tensors added\n")
 
-state_dict = model.state_dict()  # pyright: ignore[reportAttributeAccessIssue]
+state_dict = model.state_dict()
 for name, data in state_dict.items():
     if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_projector):
         # we don't need this

examples/llava/llava-cli.cpp

@@ -1,14 +1,16 @@
-#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>
+#include <cstring>
 #include <vector>
 
 static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
@@ -19,7 +21,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
             n_eval = n_batch;
         }
         if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
-            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
         *n_past += n_eval;
@@ -35,21 +37,21 @@ static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
 
 static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
     std::string              str2     = str;
-    std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
+    std::vector<llama_token> embd_inp = common_tokenize(ctx_llama, str2, add_bos, true);
     eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
     return true;
 }
 
-static const char * sample(struct llama_sampling_context * ctx_sampling,
+static const char * sample(struct common_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 = common_sampler_sample(smpl, ctx_llama, -1);
+    common_sampler_accept(smpl, id, true);
     static std::string ret;
     if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
-        ret = llama_token_to_piece(ctx_llama, id);
+        ret = common_token_to_piece(ctx_llama, id);
     }
     eval_id(ctx_llama, id, n_past);
     return ret.c_str();
@@ -74,7 +76,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
     size_t img_base64_str_start, img_base64_str_end;
     find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
     if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
         return NULL;
     }
 
@@ -88,7 +90,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
 
     auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
     if (!embed) {
-        LOG_TEE("%s: could not load image from base64 string.\n", __func__);
+        LOG_ERR("%s: could not load image from base64 string.\n", __func__);
         return NULL;
     }
 
@@ -112,31 +114,29 @@ 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);
-
-    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");
+static void print_usage(int, char ** argv) {
+    LOG("\n example usage:\n");
+    LOG("\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("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
-static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params, const std::string & fname) {
+static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
 
     // load and preprocess the image
     llava_image_embed * embed = NULL;
     auto prompt = params->prompt;
     if (prompt_contains_image(prompt)) {
         if (!params->image.empty()) {
-            LOG_TEE("using base64 encoded image instead of command line image path\n");
+            LOG_INF("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__);
+            LOG_ERR("%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;
@@ -146,7 +146,7 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     return embed;
 }
 
-static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, gpt_params * params, const std::string & prompt) {
+static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
     int n_past = 0;
 
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
@@ -157,18 +157,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        LOG_TEE("system_prompt: %s\n", system_prompt.c_str());
+        LOG_INF("system_prompt: %s\n", system_prompt.c_str());
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
-        LOG_TEE("user_prompt: %s\n", user_prompt.c_str());
+        LOG_INF("user_prompt: %s\n", user_prompt.c_str());
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     } else {
@@ -176,9 +176,9 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
         if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
+            auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     }
@@ -189,21 +189,21 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
 
     // generate the response
 
-    LOG_TEE("\n");
+    LOG("\n");
 
-    struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    if (!ctx_sampling) {
-        fprintf(stderr, "%s: failed to initialize sampling subsystem\n", __func__);
+    struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sparams);
+    if (!smpl) {
+        LOG_ERR("%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
-        printf("%s", tmp);
+        LOG("%s", tmp);
         if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
         if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
         if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
@@ -211,25 +211,25 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         fflush(stdout);
     }
 
-    llama_sampling_free(ctx_sampling);
-    printf("\n");
+    common_sampler_free(smpl);
+    LOG("\n");
 }
 
-static struct llama_model * llava_init(gpt_params * params) {
+static struct llama_model * llava_init(common_params * params) {
     llama_backend_init();
     llama_numa_init(params->numa);
 
-    llama_model_params model_params = llama_model_params_from_gpt_params(*params);
+    llama_model_params model_params = common_model_params_to_llama(*params);
 
     llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
-        LOG_TEE("%s: error: unable to load model\n" , __func__);
+        LOG_ERR("%s: unable to load model\n" , __func__);
         return NULL;
     }
     return model;
 }
 
-static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
+static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
     const char * clip_path = params->mmproj.c_str();
 
     auto prompt = params->prompt;
@@ -240,17 +240,17 @@ static struct llava_context * llava_init_context(gpt_params * params, llama_mode
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
 
-    llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
+    llama_context_params ctx_params = common_context_params_to_llama(*params);
     ctx_params.n_ctx           = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
 
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
-        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
+        LOG_ERR("%s: failed to create the llama_context\n" , __func__);
         return NULL;
     }
 
-    auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
+    auto * ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
 
     ctx_llava->ctx_llama = ctx_llama;
     ctx_llava->ctx_clip = ctx_clip;
@@ -269,65 +269,54 @@ static void llava_free(struct llava_context * ctx_llava) {
     llama_backend_free();
 }
 
-static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
-    (void) level;
-    (void) user_data;
-    LOG_TEE("%s", text);
-}
-
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    gpt_params params;
+    common_params params;
 
-    if (!gpt_params_parse(argc, argv, params)) {
-        print_usage(argc, argv, params);
+    if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
         return 1;
     }
 
-#ifndef LOG_DISABLE_LOGS
-    log_set_target(log_filename_generator("llava", "log"));
-    LOG_TEE("Log start\n");
-    log_dump_cmdline(argc, argv);
-    llama_log_set(llama_log_callback_logTee, nullptr);
-#endif // LOG_DISABLE_LOGS
+    common_init();
 
     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);
+
+    auto * model = llava_init(&params);
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
         return 1;
     }
 
     if (prompt_contains_image(params.prompt)) {
-        auto ctx_llava = llava_init_context(&params, model);
+        auto * ctx_llava = llava_init_context(&params, model);
 
-        auto image_embed = load_image(ctx_llava, &params, "");
+        auto * image_embed = load_image(ctx_llava, &params, "");
 
         // 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);
     } else {
         for (auto & image : params.image) {
-            auto ctx_llava = llava_init_context(&params, model);
+            auto * ctx_llava = llava_init_context(&params, model);
 
-            auto image_embed = load_image(ctx_llava, &params, image);
+            auto * image_embed = load_image(ctx_llava, &params, image);
             if (!image_embed) {
-                std::cerr << "error: failed to load image " << image << ". Terminating\n\n";
+                LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
                 return 1;
             }
 
             // 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

@@ -1,13 +1,23 @@
 #include "clip.h"
-#include "common.h"
-#include "llama.h"
 #include "llava.h"
-#include "base64.hpp"
 
+#include "llama.h"
+
+#include <algorithm>
+#include <cerrno>
 #include <cstdio>
 #include <cstdlib>
+#include <cstring>
+#include <limits>
 #include <vector>
-#include <numeric>
+
+#define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
+#define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)
+
+#define LOG_INF(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
+#define LOG_WRN(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_ERR(...) do { fprintf(stderr, __VA_ARGS__); } while (0)
+#define LOG_DBG(...) do { fprintf(stdout, __VA_ARGS__); } while (0)
 
 // RGB uint8 image
 struct clip_image_u8 {
@@ -54,7 +64,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int>& ori
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // LOG_DBG("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -184,7 +194,7 @@ 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 ):
@@ -236,7 +246,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     img_res_v.size = 0;
     img_res_v.data = nullptr;
     if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
-        LOG_TEE("%s: unable to preprocess image\n", __func__);
+        LOG_ERR("%s: unable to preprocess image\n", __func__);
         delete[] img_res_v.data;
         return false;
     }
@@ -256,16 +266,23 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             load_image_size->width = img_res_v.data[i].nx;
             load_image_size->height = img_res_v.data[i].ny;
             clip_add_load_image_size(ctx_clip, load_image_size);
-            const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            bool encoded = false;
+            int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
+            if (has_minicpmv_projector == 2) {
+                encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
+            }
+            else if (has_minicpmv_projector == 3) {
+                encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
+            }
             if (!encoded) {
-                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
             const int64_t t_img_enc_steop_batch_us = ggml_time_us();
-            LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
+            LOG_INF("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         int n_img_pos_out = 0;
         for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -280,7 +297,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         load_image_size->width = img->nx;
         load_image_size->height = img->ny;
         clip_add_load_image_size(ctx_clip, load_image_size);
-        LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
+        LOG_INF("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
     }
     else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
         // flat / default llava-1.5 type embedding
@@ -288,7 +305,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
         delete[] img_res_v.data;
         if (!encoded) {
-            LOG_TEE("Unable to encode image\n");
+            LOG_ERR("Unable to encode image\n");
 
             return false;
         }
@@ -302,12 +319,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
             const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        LOG_TEE("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_INF("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
 
@@ -340,12 +357,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
-    LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+    LOG_INF("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
 
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
-    LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+    LOG_INF("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
 
     return true;
 }
@@ -355,7 +372,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
-        LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        LOG_ERR("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
         return false;
     }
     return true;
@@ -368,13 +385,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     }
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
     if (!image_embd) {
-        LOG_TEE("Unable to allocate memory for image embeddings\n");
+        LOG_ERR("Unable to allocate memory for image embeddings\n");
         return false;
     }
 
     int n_img_pos;
     if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
-        LOG_TEE("%s: cannot encode image, aborting\n", __func__);
+        LOG_ERR("%s: cannot encode image, aborting\n", __func__);
         free(image_embd);
         return false;
     }
@@ -394,7 +411,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         }
         llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
         if (llama_decode(ctx_llama, batch)) {
-            LOG_TEE("%s : failed to eval\n", __func__);
+            LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
         *n_past += n_eval;
@@ -406,7 +423,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
-        LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
+        LOG_ERR("%s: can't load image from bytes, is it a valid image?", __func__);
         return NULL;
     }
 
@@ -415,7 +432,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        LOG_TEE("%s: coulnd't embed the image\n", __func__);
+        LOG_ERR("%s: coulnd't embed the image\n", __func__);
         return NULL;
     }
 
@@ -429,7 +446,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
 static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
     auto file = fopen(path, "rb");
     if (file == NULL) {
-        LOG_TEE("%s: can't read file %s\n", __func__, path);
+        LOG_ERR("%s: can't read file %s\n", __func__, path);
         return false;
     }
 
@@ -439,7 +456,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
 
     auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
     if (buffer == NULL) {
-        LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        LOG_ERR("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
         perror("Memory allocation error");
         fclose(file);
         return false;
@@ -464,7 +481,7 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
     if (!loaded) {
-        LOG_TEE("%s: failed to load %s\n", __func__, image_path);
+        LOG_ERR("%s: failed to load %s\n", __func__, image_path);
         return NULL;
     }