[bug fix] convert github repository_owner to lowercase (#6673)

* disable mmap to fix memcpy crash, add missed cmd in guide, fix softmax

* refactor to disable mmap for SYCL backend

* fix compile error in other os

* refactor the solution, use host buf to fix it, instead of disable mmap

* keep to support mmap()

* use host buff to reduce malloc times

* revert to malloc/free solution, for threaad safe

.clang-tidy

@@ -12,6 +12,7 @@ Checks: >
     -readability-implicit-bool-conversion,
     -readability-magic-numbers,
     -readability-uppercase-literal-suffix,
+    -readability-simplify-boolean-expr,
     clang-analyzer-*,
     -clang-analyzer-security.insecureAPI.DeprecatedOrUnsafeBufferHandling,
     performance-*,

.devops/full-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -26,8 +26,10 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
 
 RUN make
 

.devops/full-rocm.Dockerfile

@@ -40,6 +40,11 @@ ENV LLAMA_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
 
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 RUN make
 
 ENTRYPOINT ["/app/.devops/tools.sh"]

.devops/full.Dockerfile

@@ -3,7 +3,7 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential python3 python3-pip git
+    apt-get install -y build-essential python3 python3-pip git libcurl4-openssl-dev
 
 COPY requirements.txt   requirements.txt
 COPY requirements       requirements
@@ -15,6 +15,9 @@ WORKDIR /app
 
 COPY . .
 
+ENV LLAMA_CURL=1
+
+
 RUN make
 
 ENV LC_ALL=C.utf8

.devops/llama-cpp-clblast.srpm.spec

@@ -1,5 +1,5 @@
 # SRPM for building from source and packaging an RPM for RPM-based distros.
-# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
 # Built and maintained by John Boero - boeroboy@gmail.com
 # In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
 

.devops/llama-cpp-cuda.srpm.spec

@@ -1,5 +1,5 @@
 # SRPM for building from source and packaging an RPM for RPM-based distros.
-# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
 # Built and maintained by John Boero - boeroboy@gmail.com
 # In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
 
@@ -12,7 +12,7 @@
 # 4. OpenCL/CLBLAST support simply requires the ICD loader and basic opencl libraries.
 #    It is up to the user to install the correct vendor-specific support.
 
-Name:           llama.cpp-cublas
+Name:           llama.cpp-cuda
 Version:        %( date "+%%Y%%m%%d" )
 Release:        1%{?dist}
 Summary:        CPU Inference of LLaMA model in pure C/C++ (no CUDA/OpenCL)
@@ -32,16 +32,16 @@ CPU inference for Meta's Lllama2 models using default options.
 %setup -n llama.cpp-master
 
 %build
-make -j LLAMA_CUBLAS=1
+make -j LLAMA_CUDA=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacppcublas
-cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
-cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
+cp -p main %{buildroot}%{_bindir}/llamacppcuda
+cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
-%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacublas.service
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
 [Unit]
 Description=Llama.cpp server, CPU only (no GPU support in this build).
 After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.target
@@ -49,7 +49,7 @@ After=syslog.target network.target local-fs.target remote-fs.target nss-lookup.t
 [Service]
 Type=simple
 EnvironmentFile=/etc/sysconfig/llama
-ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS
+ExecStart=/usr/bin/llamacppcudaserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,10 +67,10 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacppcublas
-%{_bindir}/llamacppcublasserver
-%{_bindir}/llamacppcublassimple
-/usr/lib/systemd/system/llamacublas.service
+%{_bindir}/llamacppcuda
+%{_bindir}/llamacppcudaserver
+%{_bindir}/llamacppcudasimple
+/usr/lib/systemd/system/llamacuda.service
 %config /etc/sysconfig/llama
 
 %pre

.devops/llama-cpp.srpm.spec

@@ -1,5 +1,5 @@
 # SRPM for building from source and packaging an RPM for RPM-based distros.
-# https://fedoraproject.org/wiki/How_to_create_an_RPM_package
+# https://docs.fedoraproject.org/en-US/quick-docs/creating-rpm-packages
 # Built and maintained by John Boero - boeroboy@gmail.com
 # In honor of Seth Vidal https://www.redhat.com/it/blog/thank-you-seth-vidal
 

.devops/main-cuda.Dockerfile

@@ -20,8 +20,8 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
 
 RUN make
 

.devops/nix/package.nix

@@ -4,13 +4,14 @@
   config,
   stdenv,
   mkShell,
+  runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
   python3,
   mpi,
-  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
+  blas,
   cudaPackages,
   darwin,
   rocmPackages,
@@ -23,7 +24,7 @@
     useOpenCL
     useRocm
     useVulkan
-  ],
+  ] && blas.meta.available,
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
   useMpi ? false, # Increases the runtime closure size by ~700M
@@ -35,7 +36,8 @@
   # It's necessary to consistently use backendStdenv when building with CUDA support,
   # otherwise we get libstdc++ errors downstream.
   effectiveStdenv ? if useCuda then cudaPackages.backendStdenv else stdenv,
-  enableStatic ? effectiveStdenv.hostPlatform.isStatic
+  enableStatic ? effectiveStdenv.hostPlatform.isStatic,
+  precompileMetalShaders ? false
 }@inputs:
 
 let
@@ -65,10 +67,15 @@ let
     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
@@ -87,6 +94,11 @@ let
     ]
   );
 
+  xcrunHost = runCommand "xcrunHost" {} ''
+    mkdir -p $out/bin
+    ln -s /usr/bin/xcrun $out/bin
+  '';
+
   # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
   # separately
   darwinBuildInputs =
@@ -150,13 +162,18 @@ effectiveStdenv.mkDerivation (
     postPatch = ''
       substituteInPlace ./ggml-metal.m \
         --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-
-      # TODO: Package up each Python script or service appropriately.
-      # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
-      # we could make those *.py into setuptools' entrypoints
-      substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
+      substituteInPlace ./ggml-metal.m \
+        --replace '[bundle pathForResource:@"default" ofType:@"metallib"];' "@\"$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
@@ -173,6 +190,8 @@ effectiveStdenv.mkDerivation (
       ]
       ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
         glibc.static
+      ] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
+        xcrunHost
       ];
 
     buildInputs =
@@ -181,6 +200,7 @@ effectiveStdenv.mkDerivation (
       ++ optionals useMpi [ mpi ]
       ++ optionals useOpenCL [ clblast ]
       ++ optionals useRocm rocmBuildInputs
+      ++ optionals useBlas [ blas ]
       ++ optionals useVulkan vulkanBuildInputs;
 
     cmakeFlags =
@@ -191,7 +211,7 @@ effectiveStdenv.mkDerivation (
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
-        (cmakeBool "LLAMA_CUBLAS" useCuda)
+        (cmakeBool "LLAMA_CUDA" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
@@ -216,14 +236,16 @@ effectiveStdenv.mkDerivation (
         # Should likely use `rocmPackages.clr.gpuTargets`.
         "-DAMDGPU_TARGETS=gfx803;gfx900;gfx906:xnack-;gfx908:xnack-;gfx90a:xnack+;gfx90a:xnack-;gfx940;gfx941;gfx942;gfx1010;gfx1012;gfx1030;gfx1100;gfx1101;gfx1102"
       ]
-      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
-      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
+      ++ optionals useMetalKit [
+        (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1")
+        (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
+      ];
 
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''
-      mv $out/bin/main $out/bin/llama
-      mv $out/bin/server $out/bin/llama-server
+      mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
+      mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
       mkdir -p $out/include
       cp $src/llama.h $out/include/
     '';

.devops/server-cuda.Dockerfile

@@ -12,7 +12,7 @@ FROM ${BASE_CUDA_DEV_CONTAINER} as build
 ARG CUDA_DOCKER_ARCH=all
 
 RUN apt-get update && \
-    apt-get install -y build-essential git
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -20,13 +20,18 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable cuBLAS
-ENV LLAMA_CUBLAS=1
+# Enable CUDA
+ENV LLAMA_CUDA=1
+# Enable cURL
+ENV LLAMA_CURL=1
 
 RUN make
 
 FROM ${BASE_CUDA_RUN_CONTAINER} as runtime
 
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 COPY --from=build /app/server /server
 
 ENTRYPOINT [ "/server" ]

.devops/server-intel.Dockerfile

@@ -4,7 +4,7 @@ FROM intel/oneapi-basekit:$ONEAPI_VERSION as build
 
 ARG LLAMA_SYCL_F16=OFF
 RUN apt-get update && \
-    apt-get install -y git
+    apt-get install -y git libcurl4-openssl-dev
 
 WORKDIR /app
 
@@ -16,11 +16,14 @@ RUN mkdir build && \
         echo "LLAMA_SYCL_F16 is set" && \
         export OPT_SYCL_F16="-DLLAMA_SYCL_F16=ON"; \
     fi && \
-    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx ${OPT_SYCL_F16} && \
+    cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_CURL=ON ${OPT_SYCL_F16} && \
     cmake --build . --config Release --target server
 
 FROM intel/oneapi-basekit:$ONEAPI_VERSION as runtime
 
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 COPY --from=build /app/build/bin/server /server
 
 ENV LC_ALL=C.utf8

.devops/server-rocm.Dockerfile

@@ -40,6 +40,11 @@ ENV LLAMA_HIPBLAS=1
 ENV CC=/opt/rocm/llvm/bin/clang
 ENV CXX=/opt/rocm/llvm/bin/clang++
 
+# Enable cURL
+ENV LLAMA_CURL=1
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 RUN make
 
 ENTRYPOINT [ "/app/server" ]

.devops/server-vulkan.Dockerfile

@@ -11,12 +11,16 @@ RUN wget -qO - https://packages.lunarg.com/lunarg-signing-key-pub.asc | apt-key
     apt update -y && \
     apt-get install -y vulkan-sdk
 
+# Install cURL
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 # Build it
 WORKDIR /app
 COPY . .
 RUN mkdir build && \
     cd build && \
-    cmake .. -DLLAMA_VULKAN=1 && \
+    cmake .. -DLLAMA_VULKAN=1 -DLLAMA_CURL=1 && \
     cmake --build . --config Release --target server
 
 # Clean up

.devops/server.Dockerfile

@@ -3,16 +3,21 @@ ARG UBUNTU_VERSION=22.04
 FROM ubuntu:$UBUNTU_VERSION as build
 
 RUN apt-get update && \
-    apt-get install -y build-essential git
+    apt-get install -y build-essential git libcurl4-openssl-dev
 
 WORKDIR /app
 
 COPY . .
 
+ENV LLAMA_CURL=1
+
 RUN make
 
 FROM ubuntu:$UBUNTU_VERSION as runtime
 
+RUN apt-get update && \
+    apt-get install -y libcurl4-openssl-dev
+
 COPY --from=build /app/server /server
 
 ENV LC_ALL=C.utf8

.github/workflows/bench.yml

@@ -0,0 +1,300 @@
+# Benchmark
+name: Benchmark
+
+on:
+  workflow_dispatch:
+    inputs:
+      gpu-series:
+        description: 'Azure GPU series to run with'
+        required: true
+        type: choice
+        options:
+          - Standard_NC4as_T4_v3
+          - Standard_NC24ads_A100_v4
+          - Standard_NC80adis_H100_v5
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
+      duration:
+        description: 'Duration of the bench'
+        type: string
+        default: 10m
+
+  push:
+    branches:
+      - master
+    paths: ['llama.cpp', 'ggml.c', 'ggml-backend.c', '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']
+  schedule:
+    -  cron: '04 2 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}-${{ github.event.inputs.sha }}
+  cancel-in-progress: true
+
+jobs:
+  bench-server-baseline:
+    runs-on: Standard_NC4as_T4_v3
+    env:
+      RUNNER_LABEL: Standard_NC4as_T4_v3 # FIXME Do not find a way to not duplicate it
+      N_USERS: 8
+      DURATION: 10m
+
+    strategy:
+      matrix:
+        model: [phi-2]
+        ftype: [q4_0, q8_0, f16]
+        include:
+          - model: phi-2
+            ftype: q4_0
+            pr_comment_enabled: "true"
+
+    if: ${{ github.event.inputs.gpu-series == 'Standard_NC4as_T4_v3' || github.event.schedule || github.event.pull_request || github.head_ref == 'master' || github.ref_name == 'master' || github.event.push.ref == 'refs/heads/master' }}
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Install python env
+        id: pipenv
+        run: |
+          cd examples/server/bench
+          python3 -m venv venv
+          source venv/bin/activate
+          pip install -r requirements.txt
+
+      - name: Prometheus
+        id: install_prometheus
+        run: |
+          wget --quiet https://github.com/prometheus/prometheus/releases/download/v2.51.0/prometheus-2.51.0.linux-amd64.tar.gz
+          tar xzf prometheus*.tar.gz --strip-components=1
+          ./prometheus --config.file=examples/server/bench/prometheus.yml &
+          while ! nc -z localhost 9090; do
+            sleep 0.1
+          done
+
+      - name: Set up Go
+        uses: actions/setup-go@v5
+        with:
+          go-version: '1.21'
+
+      - name: Install k6 and xk6-sse
+        id: k6_installation
+        run: |
+          cd examples/server/bench
+          go install go.k6.io/xk6/cmd/xk6@latest
+          xk6 build master \
+              --with github.com/phymbert/xk6-sse
+
+      - name: Build
+        id: cmake_build
+        run: |
+          set -eux
+          mkdir build
+          cd build
+          cmake .. \
+              -DLLAMA_NATIVE=OFF \
+              -DLLAMA_BUILD_SERVER=ON \
+              -DLLAMA_CURL=ON \
+              -DLLAMA_CUBLAS=ON \
+              -DCUDAToolkit_ROOT=/usr/local/cuda \
+              -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc \
+              -DCMAKE_CUDA_ARCHITECTURES=75 \
+              -DLLAMA_FATAL_WARNINGS=OFF \
+              -DLLAMA_ALL_WARNINGS=OFF \
+              -DCMAKE_BUILD_TYPE=Release;
+          cmake --build . --config Release -j $(nproc) --target server
+
+      - name: Download the dataset
+        id: download_dataset
+        run: |
+          cd examples/server/bench
+          wget --quiet https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json
+
+      - name: Server bench
+        id: server_bench
+        run: |
+          set -eux
+
+          cd examples/server/bench
+          source venv/bin/activate
+          python bench.py \
+              --runner-label ${{ env.RUNNER_LABEL }} \
+              --name ${{ github.job }} \
+              --branch ${{ github.head_ref || github.ref_name }} \
+              --commit ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha }} \
+              --scenario script.js \
+              --duration ${{ github.event.inputs.duration || env.DURATION }} \
+              --hf-repo ggml-org/models	 \
+              --hf-file ${{ matrix.model }}/ggml-model-${{ matrix.ftype }}.gguf \
+              --model-path-prefix /models \
+              --parallel ${{ env.N_USERS }} \
+              -ngl 33 \
+              --batch-size 2048 \
+              --ubatch-size	256 \
+              --ctx-size 16384 \
+              --n-prompts 1000 \
+              --max-prompt-tokens 1024 \
+              --max-tokens 2048
+
+          cat results.github.env >> $GITHUB_ENV
+
+          # Remove dataset as we do not want it in the artefact
+          rm ShareGPT_V3_unfiltered_cleaned_split.json
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          compression-level: 9
+          path: |
+            examples/server/bench/*.jpg
+            examples/server/bench/*.json
+            examples/server/bench/*.log
+
+      - name: Commit status
+        uses: Sibz/github-status-action@v1
+        with:
+          authToken: ${{secrets.GITHUB_TOKEN}}
+          sha: ${{ inputs.sha || github.event.pull_request.head.sha || github.sha }}
+          context: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          description: |
+            ${{ env.BENCH_RESULTS }}
+          state: 'success'
+
+      - name: Upload benchmark images
+        uses: devicons/public-upload-to-imgur@v2.2.2
+        continue-on-error: true # Important as it looks unstable: 503
+        id: imgur_step
+        with:
+          client_id: ${{secrets.IMGUR_CLIENT_ID}}
+          path: |
+            examples/server/bench/prompt_tokens_seconds.jpg
+            examples/server/bench/predicted_tokens_seconds.jpg
+            examples/server/bench/kv_cache_usage_ratio.jpg
+            examples/server/bench/requests_processing.jpg
+
+      - name: Extract mermaid
+        id: set_mermaid
+        run: |
+          set -eux
+
+          cd examples/server/bench
+          PROMPT_TOKENS_SECONDS=$(cat prompt_tokens_seconds.mermaid)
+          echo "PROMPT_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PROMPT_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          PREDICTED_TOKENS_SECONDS=$(cat predicted_tokens_seconds.mermaid)
+          echo "PREDICTED_TOKENS_SECONDS<<EOF" >> $GITHUB_ENV
+          echo "$PREDICTED_TOKENS_SECONDS" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          KV_CACHE_USAGE_RATIO=$(cat kv_cache_usage_ratio.mermaid)
+          echo "KV_CACHE_USAGE_RATIO<<EOF" >> $GITHUB_ENV
+          echo "$KV_CACHE_USAGE_RATIO" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+          REQUESTS_PROCESSING=$(cat requests_processing.mermaid)
+          echo "REQUESTS_PROCESSING<<EOF" >> $GITHUB_ENV
+          echo "$REQUESTS_PROCESSING" >> $GITHUB_ENV
+          echo "EOF" >> $GITHUB_ENV
+
+      - name: Extract image url
+        id: extract_image_url
+        continue-on-error: true
+        run: |
+          set -eux
+
+          echo "IMAGE_O=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[0] }}" >> $GITHUB_ENV
+          echo "IMAGE_1=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[1] }}" >> $GITHUB_ENV
+          echo "IMAGE_2=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[2] }}" >> $GITHUB_ENV
+          echo "IMAGE_3=${{ fromJSON(steps.imgur_step.outputs.imgur_urls)[3] }}" >> $GITHUB_ENV
+
+      - name: Comment PR
+        uses: mshick/add-pr-comment@v2
+        id: comment_pr
+        if: ${{ github.event.pull_request != '' && matrix.pr_comment_enabled == 'true' }}
+        with:
+          message-id: bench-server-${{ github.job }}-${{ env.RUNNER_LABEL }}-${{ matrix.model }}-${{ matrix.ftype }}
+          message: |
+            <p align="center">
+
+            📈 **llama.cpp server** for _${{ github.job }}_ on _${{ env.RUNNER_LABEL }}_ for `${{ matrix.model }}`-`${{ matrix.ftype }}`: **${{ env.BENCH_ITERATIONS}} iterations** 🚀
+
+            </p>
+
+            <details>
+
+            <summary>Expand details for performance related PR only</summary>
+
+            - Concurrent users: ${{ env.N_USERS }}, duration: ${{ github.event.inputs.duration || env.DURATION }}
+            - HTTP request          : avg=${{ env.HTTP_REQ_DURATION_AVG }}ms        p(95)=${{ env.HTTP_REQ_DURATION_P_95_ }}ms fails=${{ env.HTTP_REQ_FAILED_PASSES }}, finish reason: stop=${{ env.LLAMACPP_COMPLETIONS_STOP_RATE_PASSES }} truncated=${{ env.LLAMACPP_COMPLETIONS_TRUNCATED_RATE_PASSES }}
+            - Prompt processing (pp): avg=${{ env.LLAMACPP_PROMPT_PROCESSING_SECOND_AVG }}tk/s p(95)=${{ env.LLAMACPP_PROMPT_PROCESSING_SECOND_P_95_ }}tk/s
+            - Token generation  (tg): avg=${{ env.LLAMACPP_TOKENS_SECOND_AVG }}tk/s p(95)=${{ env.LLAMACPP_TOKENS_SECOND_P_95_ }}tk/s
+            - ${{ env.BENCH_GRAPH_XLABEL }}
+
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_O }}" alt="prompt_tokens_seconds" />
+
+            <details>
+
+            <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PROMPT_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_1 }}" alt="predicted_tokens_seconds"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.PREDICTED_TOKENS_SECONDS }}
+            ```
+
+            </details>
+
+            </p>
+
+            <details>
+
+            <summary>Details</summary>
+
+            <p align="center">
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_2 }}" alt="kv_cache_usage_ratio" />
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.KV_CACHE_USAGE_RATIO }}
+            ```
+
+            </details>
+
+            <img width="100%" height="100%" src="${{ env.IMAGE_3 }}" alt="requests_processing"/>
+
+            <details>
+                <summary>More</summary>
+
+            ```mermaid
+            ${{ env.REQUESTS_PROCESSING }}
+            ```
+
+            </details>
+
+            </p>
+            </details>
+            </details>

.github/workflows/build.yml

@@ -15,19 +15,140 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
 jobs:
-  ubuntu-focal-make:
-    runs-on: ubuntu-20.04
+  macOS-latest-cmake-arm64:
+    runs-on: macos-14
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L 'main|curl' --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(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
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
+
+  macOS-latest-cmake-x64:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+
+      - name: Dependencies
+        id: depends
+        continue-on-error: true
+        run: |
+          brew update
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          mkdir build
+          cd build
+          # 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 -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF -DLLAMA_CURL=ON ..
+          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
+
+      - name: Test
+        id: cmake_test
+        run: |
+          cd build
+          ctest -L main --verbose --timeout 900
+
+      - name: Determine tag name
+        id: tag
+        shell: bash
+        run: |
+          BUILD_NUMBER="$(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
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp LICENSE ./build/bin/
+          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
+
+  ubuntu-focal-make:
+    runs-on: ubuntu-20.04
+    env:
+      LLAMA_NODE_AVAILABLE: true
+      LLAMA_PYTHON_AVAILABLE: true
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -35,6 +156,14 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential gcc-8
 
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "20"
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+
       - name: Build
         id: make_build
         env:
@@ -54,7 +183,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -76,27 +205,38 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
         run: |
           sudo apt-get update
-          sudo apt-get install build-essential
+          sudo apt-get install build-essential libcurl4-openssl-dev
 
       - name: Build
         id: cmake_build
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON
+          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_CURL=ON
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L 'main|curl' --verbose --timeout 900
+
+      - name: Test llama2c conversion
+        id: llama2c_test
+        run: |
+          cd build
+          echo "Fetch tokenizer"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
+          echo "Fetch llama2c model"
+          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
+          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
+          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
 
 #  ubuntu-latest-cmake-sanitizer:
 #    runs-on: ubuntu-latest
@@ -111,7 +251,7 @@ jobs:
 #    steps:
 #      - name: Clone
 #        id: checkout
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        id: depends
@@ -145,7 +285,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -173,7 +313,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -219,7 +359,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -260,7 +400,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -280,7 +420,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -311,7 +451,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -455,7 +595,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -585,23 +725,23 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
+          name: llama-bin-win-${{ matrix.build }}-x64.zip
 
-  windows-latest-cmake-cublas:
+  windows-latest-cmake-cuda:
     runs-on: windows-latest
 
     strategy:
       matrix:
         cuda: ['12.2.0', '11.7.1']
-        build: ['cublas']
+        build: ['cuda']
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -617,7 +757,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUDA=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
@@ -641,10 +781,10 @@ jobs:
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-cu${{ matrix.cuda }}-x64.zip
+          name: llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
       - name: Copy and pack Cuda runtime
         run: |
@@ -655,13 +795,14 @@ jobs:
 
       - name: Upload Cuda runtime
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          path: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
+          name: cudart-llama-bin-win-cu${{ matrix.cuda }}-x64.zip
 
   windows-latest-cmake-sycl:
     runs-on: windows-latest
+
     defaults:
       run:
         shell: bash
@@ -670,11 +811,10 @@ jobs:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
 
-
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -685,12 +825,38 @@ jobs:
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
 
+      - 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
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          path: llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          name: llama-bin-win-sycl-x64.zip
+
   ios-xcode-build:
     runs-on: macos-latest
 
     steps:
       - name: Checkout code
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build Xcode project
         run: xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' build
@@ -700,7 +866,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -723,7 +889,7 @@ jobs:
 #    runs-on: macos-12
 #    steps:
 #    - name: Clone
-#      uses: actions/checkout@v3
+#      uses: actions/checkout@v4
 #
 #    - name: Build
 #      uses: cross-platform-actions/action@v0.19.0
@@ -747,12 +913,14 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
-      - windows-latest-cmake-cublas
+      - windows-latest-cmake-cuda
+      - macOS-latest-cmake-arm64
+      - macOS-latest-cmake-x64
 
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -771,7 +939,13 @@ jobs:
 
       - name: Download artifacts
         id: download-artifact
-        uses: actions/download-artifact@v3
+        uses: actions/download-artifact@v4
+        with:
+          path: ./artifact
+
+      - name: Move artifacts
+        id: move_artifacts
+        run: mkdir -p ./artifact/release && mv ./artifact/*/*.zip ./artifact/release
 
       - name: Create release
         id: create_release
@@ -790,7 +964,7 @@ jobs:
             const path = require('path');
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact')) {
+            for (let file of await fs.readdirSync('./artifact/release')) {
               if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
@@ -798,7 +972,7 @@ jobs:
                   repo: context.repo.repo,
                   release_id: release_id,
                   name: file,
-                  data: await fs.readFileSync(`./artifact/${file}`)
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
                 });
               }
             }
@@ -812,7 +986,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -836,7 +1010,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -860,7 +1034,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -890,7 +1064,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -906,7 +1080,7 @@ jobs:
 #          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ matrix.build }} -p:platform=${{ matrix.arch }}
 #
 #      - name: Upload binaries
-#        uses: actions/upload-artifact@v1
+#        uses: actions/upload-artifact@v4
 #        with:
 #          name: llama-bin-${{ matrix.arch }}
 #          path: build/bin/${{ matrix.build }}
@@ -929,7 +1103,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -961,7 +1135,7 @@ jobs:
 #
 #      - name: Upload binaries
 #        if: matrix.blas == 'ON'
-#        uses: actions/upload-artifact@v1
+#        uses: actions/upload-artifact@v4
 #        with:
 #          name: llama-blas-bin-${{ matrix.arch }}
 #          path: build/bin/${{ matrix.build }}
@@ -975,7 +1149,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |

.github/workflows/close-issue.yml

@@ -19,5 +19,5 @@ jobs:
           close-issue-message: "This issue was closed because it has been inactive for 14 days since being marked as stale."
           days-before-pr-stale: -1
           days-before-pr-close: -1
-          operations-per-run: 1000
+          operations-per-run: 10000
           repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/code-coverage.yml

@@ -5,12 +5,16 @@ env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   run:
     runs-on: ubuntu-20.04
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         run: |

.github/workflows/docker.yml

@@ -15,6 +15,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub
@@ -42,7 +46,7 @@ jobs:
           - { tag: "server-intel", dockerfile: ".devops/server-intel.Dockerfile", platforms: "linux/amd64" }
     steps:
       - name: Check out the repo
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up QEMU
         uses: docker/setup-qemu-action@v2
@@ -87,6 +91,12 @@ jobs:
             echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
           fi
 
+      - name: Downcase github.repository_owner
+        run: |
+          echo "repository_owner_lowercase=${GITHUB_REPOSITORY_OWNER@L}" >> $GITHUB_ENV
+        env:
+          GITHUB_REPOSITORY_OWNER: '${{ github.repository_owner }}'
+
       - name: Build and push Docker image (versioned)
         if: github.event_name == 'push'
         uses: docker/build-push-action@v4
@@ -94,7 +104,7 @@ jobs:
           context: .
           push: true
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ env.COMMIT_SHA }}"
+          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)
@@ -103,5 +113,5 @@ jobs:
           context: .
           push: ${{ github.event_name == 'push' }}
           platforms: ${{ matrix.config.platforms }}
-          tags: "ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ github.repository_owner }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
+          tags: "ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }},ghcr.io/${{ env.repository_owner_lowercase }}/llama.cpp:${{ matrix.config.tag }}-${{ steps.tag.outputs.name }}"
           file: ${{ matrix.config.dockerfile }}

.github/workflows/editorconfig.yml

@@ -14,10 +14,14 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   editorconfig:
     runs-on: ubuntu-latest
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
       - uses: editorconfig-checker/action-editorconfig-checker@main
       - run: editorconfig-checker

.github/workflows/gguf-publish.yml

@@ -24,9 +24,9 @@ jobs:
     runs-on: ubuntu-latest
 
     steps:
-    - uses: actions/checkout@v3
+    - uses: actions/checkout@v4
     - name: Set up Python
-      uses: actions/setup-python@v2
+      uses: actions/setup-python@v5
       with:
         python-version: '3.9.x'
     - name: Install dependencies

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

@@ -17,6 +17,10 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/*.nix', 'flake.lock']
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   nix-build-aarch64:
     runs-on: ubuntu-latest

.github/workflows/nix-ci.yml

@@ -8,6 +8,10 @@ on:
   pull_request:
     types: [opened, synchronize, reopened]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   nix-eval:
     strategy:

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

@@ -16,15 +16,19 @@ on:
       - 'requirements.txt'
       - 'requirements/*.txt'
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   python-check-requirements:
     runs-on: ubuntu-latest
     name: check-requirements
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script

.github/workflows/python-lint.yml

@@ -2,15 +2,19 @@ name: flake8 Lint
 
 on: [push, pull_request]
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   flake8-lint:
     runs-on: ubuntu-latest
     name: Lint
     steps:
       - name: Check out source repository
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v4
+        uses: actions/setup-python@v5
         with:
           python-version: "3.11"
       - name: flake8 Lint

.github/workflows/server.yml

@@ -4,6 +4,10 @@ name: Server
 on:
   workflow_dispatch: # allows manual triggering
     inputs:
+      sha:
+        description: 'Commit SHA1 to build'
+        required: false
+        type: string
       slow_tests:
         description: 'Run slow tests'
         required: true
@@ -11,12 +15,16 @@ on:
   push:
     branches:
       - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  pull_request:
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
+  pull_request_target:
     types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
+    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/**.*']
   schedule:
-    -  cron: '0 0 * * *'
+    -  cron: '2 4 * * *'
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
 
 jobs:
   server:
@@ -31,7 +39,6 @@ jobs:
         include:
           - build_type: Release
             sanitizer: ""
-            disabled_on_pr: true
       fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
     container:
@@ -41,25 +48,45 @@ jobs:
       options: --cpus 4
 
     steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
       - name: Dependencies
         id: depends
         run: |
           apt-get update
           apt-get -y install \
             build-essential \
+            xxd \
             git \
             cmake \
             python3-pip \
+            curl \
             wget \
             language-pack-en \
             libcurl4-openssl-dev
 
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}
+
+      - name: Verify server deps
+        id: verify_server_deps
+        run: |
+          git config --global --add safe.directory $(realpath .)
+          cd examples/server
+          git ls-files --others --modified
+          git status
+          ./deps.sh
+          git status
+          not_ignored_files="$(git ls-files --others --modified)"
+          echo "Modified files: ${not_ignored_files}"
+          if [ -n "${not_ignored_files}" ]; then
+            echo "Repository is dirty or server deps are not built as expected"
+            echo "${not_ignored_files}"
+            exit 1
+          fi
+
       - name: Build
         id: cmake_build
         run: |
@@ -99,7 +126,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 

.github/workflows/zig-build.yml

@@ -6,6 +6,10 @@ on:
     branches:
       - master
 
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
 jobs:
   build:
     strategy:
@@ -14,7 +18,7 @@ jobs:
         runs-on: [ubuntu-latest, macos-latest, windows-latest]
     runs-on: ${{ matrix.runs-on }}
     steps:
-      - uses: actions/checkout@v3
+      - uses: actions/checkout@v4
         with:
           submodules: recursive
           fetch-depth: 0

.gitignore

@@ -11,7 +11,10 @@
 *.gcda
 *.dot
 *.bat
+*.tmp
 *.metallib
+*.etag
+*.lastModified
 .DS_Store
 .build/
 .cache/
@@ -45,8 +48,10 @@ models-mnt
 /convert-llama2c-to-ggml
 /embd-input-test
 /embedding
+/eval-callback
 /gguf
 /gguf-llama-simple
+/gguf-split
 /gritlm
 /imatrix
 /infill
@@ -55,6 +60,9 @@ models-mnt
 /llava-cli
 /lookahead
 /lookup
+/lookup-create
+/lookup-merge
+/lookup-stats
 /main
 /metal
 /passkey
@@ -70,6 +78,7 @@ models-mnt
 /batched-bench
 /export-lora
 /finetune
+/retrieval
 /speculative
 /parallel
 /train-text-from-scratch

AUTHORS

@@ -0,0 +1,655 @@
+# date: Tue Apr  9 09:17:14 EEST 2024
+# this file is auto-generated by scripts/gen-authors.sh
+
+0cc4m <picard12@live.de>
+0xspringtime <110655352+0xspringtime@users.noreply.github.com>
+2f38b454 <dxf@protonmail.com>
+3ooabkhxtn <31479382+3ooabkhxtn@users.noreply.github.com>
+44670 <44670@users.noreply.github.com>
+AN Long <aisk@users.noreply.github.com>
+AT <manyoso@users.noreply.github.com>
+Aarni Koskela <akx@iki.fi>
+Aaron Miller <apage43@ninjawhale.com>
+Aaryaman Vasishta <aaryaman.vasishta@amd.com>
+Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
+Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com>
+Adithya Balaji <adithya.b94@gmail.com>
+AdithyanI <adithyan.i4internet@gmail.com>
+Adrian <smith.adriane@gmail.com>
+Adrian Hesketh <a-h@users.noreply.github.com>
+AidanBeltonS <87009434+AidanBeltonS@users.noreply.github.com>
+Aisuko <urakiny@gmail.com>
+Alberto <57916483+albbus-stack@users.noreply.github.com>
+Alex <awhill19@icloud.com>
+Alex Azarov <alex@azarov.by>
+Alex Azarov <alexander.azarov@mapbox.com>
+Alex Klinkhamer <from.github.com.917@grencez.dev>
+Alex Klinkhamer <git@grencez.dev>
+Alex Nguyen <tiendung@users.noreply.github.com>
+Alex Petenchea <alex.petenchea@gmail.com>
+Alex Renda <alexrenda@users.noreply.github.com>
+Alex von Gluck IV <kallisti5@unixzen.com>
+Alexey Parfenov <zxed@alkatrazstudio.net>
+Ali Chraghi <63465728+alichraghi@users.noreply.github.com>
+Ali Nehzat <ali.nehzat@thanks.dev>
+Ali Tariq <ali.tariq@10xengineers.ai>
+Alon <alonfaraj@gmail.com>
+AlpinDale <52078762+AlpinDale@users.noreply.github.com>
+AmirAli Mirian <37371367+amiralimi@users.noreply.github.com>
+Ananta Bastola <anantarajbastola@gmail.com>
+Anas Ahouzi <112881240+aahouzi@users.noreply.github.com>
+András Salamon <ott2@users.noreply.github.com>
+Andrei <abetlen@gmail.com>
+Andrew Canis <andrew.canis@gmail.com>
+Andrew Duffy <a10y@users.noreply.github.com>
+Andrew Godfrey <AndrewGodfrey@users.noreply.github.com>
+Arik Poznanski <arikpoz@users.noreply.github.com>
+Artem <guinmoon@gmail.com>
+Artyom Lebedev <vagran.ast@gmail.com>
+Asbjørn Olling <asbjornolling@gmail.com>
+Ásgeir Bjarni Ingvarsson <asgeir@fundinn.org>
+Ashok Gelal <401055+ashokgelal@users.noreply.github.com>
+Ashraful Islam <ashraful.meche@gmail.com>
+Atsushi Tatsuma <yoshoku@outlook.com>
+Austin <77757836+teleprint-me@users.noreply.github.com>
+AustinMroz <austinmroz@utexas.edu>
+BADR <contact@pythops.com>
+Bach Le <bach@bullno1.com>
+Bailey Chittle <39804642+bachittle@users.noreply.github.com>
+BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
+Behnam M <58621210+ibehnam@users.noreply.github.com>
+Ben Garney <bengarney@users.noreply.github.com>
+Ben Siraphob <bensiraphob@gmail.com>
+Ben Williams <ben@719ben.com>
+Benjamin Lecaillon <84293038+blecaillon@users.noreply.github.com>
+Bernat Vadell <hounter.caza@gmail.com>
+Bodo Graumann <mail@bodograumann.de>
+Bono Lv <lvscar@users.noreply.github.com>
+Borislav Stanimirov <b.stanimirov@abv.bg>
+Branden Butler <bwtbutler@hotmail.com>
+Brian <mofosyne@gmail.com>
+Bruce MacDonald <brucewmacdonald@gmail.com>
+CJ Pais <cj@cjpais.com>
+CRD716 <crd716@gmail.com>
+Cameron <csteele@steelecameron.com>
+Cameron Kaiser <classilla@users.noreply.github.com>
+Casey Primozic <casey@cprimozic.net>
+Casey Primozic <me@ameo.link>
+CausalLM <148736309+CausalLM@users.noreply.github.com>
+Cebtenzzre <cebtenzzre@gmail.com>
+Chad Brewbaker <crb002@gmail.com>
+Cheng Shao <terrorjack@type.dance>
+Chris Kuehl <ckuehl@ckuehl.me>
+Christian Demsar <christian@github.email.demsar.us>
+Christian Demsar <crasm@git.vczf.us>
+Christian Falch <875252+chrfalch@users.noreply.github.com>
+Christian Kögler <ck3d@gmx.de>
+Clark Saben <76020733+csaben@users.noreply.github.com>
+Clint Herron <hanclinto@gmail.com>
+Cuong Trinh Manh <nguoithichkhampha@gmail.com>
+DAN™ <dranger003@gmail.com>
+Damian Stewart <d@damianstewart.com>
+Dane Madsen <dane_madsen@hotmail.com>
+DaniAndTheWeb <57776841+DaniAndTheWeb@users.noreply.github.com>
+Daniel Bevenius <daniel.bevenius@gmail.com>
+Daniel Drake <drake@endlessos.org>
+Daniel Hiltgen <dhiltgen@users.noreply.github.com>
+Daniel Illescas Romero <illescas.daniel@protonmail.com>
+DannyDaemonic <DannyDaemonic@gmail.com>
+Dat Quoc Nguyen <2412555+datquocnguyen@users.noreply.github.com>
+Dave Della Costa <ddellacosta+github@gmail.com>
+David Friehs <david@friehs.info>
+David Kennedy <dakennedyd@gmail.com>
+David Pflug <david@pflug.email>
+David Renshaw <dwrenshaw@gmail.com>
+David Sommers <12738+databyte@users.noreply.github.com>
+David Yang <davidyang6us@gmail.com>
+Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
+Dean <Dean.Sinaean@gmail.com>
+Deins <deinsegle@gmail.com>
+Didzis Gosko <didzis@users.noreply.github.com>
+Don Mahurin <dmahurin@users.noreply.github.com>
+DooWoong Lee (David) <manics99@naver.com>
+Doomsdayrs <38189170+Doomsdayrs@users.noreply.github.com>
+Douglas Hanley <thesecretaryofwar@gmail.com>
+Dr. Tom Murphy VII Ph.D <499244+tom7@users.noreply.github.com>
+Ebey Abraham <ebey97@gmail.com>
+Ed Lee <edilee@mozilla.com>
+Ed Lepedus <ed.lepedus@googlemail.com>
+Edward Taylor <edeetee@gmail.com>
+Elbios <141279586+Elbios@users.noreply.github.com>
+Engininja2 <139037756+Engininja2@users.noreply.github.com>
+Equim <sayaka@ekyu.moe>
+Eric Sommerlade <es0m@users.noreply.github.com>
+Eric Zhang <34133756+EZForever@users.noreply.github.com>
+Erik Garrison <erik.garrison@gmail.com>
+Erik Scholz <Green-Sky@users.noreply.github.com>
+Ettore Di Giacinto <mudler@users.noreply.github.com>
+Evan Jones <evan.q.jones@gmail.com>
+Evan Miller <emmiller@gmail.com>
+Eve <139727413+netrunnereve@users.noreply.github.com>
+Evgeny Kurnevsky <kurnevsky@gmail.com>
+Ewout ter Hoeven <E.M.terHoeven@student.tudelft.nl>
+ExtReMLapin <3909752+ExtReMLapin@users.noreply.github.com>
+FK <sozforex@gmail.com>
+Fabian <cmdrf@users.noreply.github.com>
+Fabio R. Sluzala <Fabio3rs@users.noreply.github.com>
+Faez Shakil <faez.shakil@gmail.com>
+FantasyGmm <16450052+FantasyGmm@users.noreply.github.com>
+Fattire <528174+fat-tire@users.noreply.github.com>
+Felix <stenbackfelix@gmail.com>
+Finn Voorhees <finnvoorhees@gmail.com>
+Firat <firatkiral@gmail.com>
+Folko-Ven <71110216+Folko-Ven@users.noreply.github.com>
+Foul-Tarnished <107711110+Foul-Tarnished@users.noreply.github.com>
+Francisco Melo <43780565+francis2tm@users.noreply.github.com>
+FrankHB <frankhb1989@gmail.com>
+Frederik Vogel <Schaltfehler@users.noreply.github.com>
+Gabe Goodhart <gabe.l.hart@gmail.com>
+GainLee <perfecter.gen@gmail.com>
+Galunid <karolek1231456@gmail.com>
+Gary Linscott <glinscott@gmail.com>
+Gary Mulder <gjmulder@gmail.com>
+Genkagaku.GPT <hlhr202@163.com>
+Georgi Gerganov <ggerganov@gmail.com>
+Gilad S <giladgd@users.noreply.github.com>
+GiviMAD <GiviMAD@users.noreply.github.com>
+Govlzkoy <gotope@users.noreply.github.com>
+Guillaume "Vermeille" Sanchez <Guillaume.V.Sanchez@gmail.com>
+Guillaume Wenzek <gwenzek@users.noreply.github.com>
+Guoteng <32697156+SolenoidWGT@users.noreply.github.com>
+Gustavo Rocha Dias <91472747+gustrd@users.noreply.github.com>
+Halalaluyafail3 <55773281+Halalaluyafail3@users.noreply.github.com>
+Haohui Mai <ricetons@gmail.com>
+Haoxiang Fei <tonyfettes@tonyfettes.com>
+Harald Fernengel <harald.fernengel@here.com>
+Hatsune Miku <129688334+at8u@users.noreply.github.com>
+Henk Poley <HenkPoley@gmail.com>
+Henri Vasserman <henv@hot.ee>
+Henrik Forstén <henrik.forsten@gmail.com>
+Herman Semenov <GermanAizek@yandex.ru>
+Hesen Peng <hesen.peng@gmail.com>
+Hoang Nguyen <hugo53@users.noreply.github.com>
+Hongyu Ouyang <96765450+casavaca@users.noreply.github.com>
+Howard Su <howard0su@gmail.com>
+Hua Jiang <allenhjiang@outlook.com>
+Huawei Lin <huaweilin.cs@gmail.com>
+Ian Bull <irbull@eclipsesource.com>
+Ian Bull <irbull@gmail.com>
+Ian Scrivener <github@zilogy.asia>
+Ido S <ido.pluto@gmail.com>
+IgnacioFDM <ignaciofdm@gmail.com>
+Igor Okulist <okigan@gmail.com>
+Ikko Eltociear Ashimine <eltociear@gmail.com>
+Ilya Kurdyukov <59548320+ilyakurdyukov@users.noreply.github.com>
+Ionoclast Laboratories <brigham@ionoclast.com>
+Isaac McFadyen <isaac@imcf.me>
+IsaacDynamo <61521674+IsaacDynamo@users.noreply.github.com>
+Ivan Komarov <Ivan.Komarov@dfyz.info>
+Ivan Stepanov <ivanstepanovftw@gmail.com>
+JH23X <165871467+JH23X@users.noreply.github.com>
+Jack Mousseau <jmousseau@users.noreply.github.com>
+JackJollimore <130917767+JackJollimore@users.noreply.github.com>
+Jag Chadha <jagtesh@gmail.com>
+Jakub N <jakubniemczyk97@gmail.com>
+James Reynolds <magnusviri@users.noreply.github.com>
+Jan Boon <jan.boon@kaetemi.be>
+Jan Boon <kaetemi@gmail.com>
+Jan Ploski <jpl@plosquare.com>
+Jannis Schönleber <joennlae@gmail.com>
+Jared Van Bortel <cebtenzzre@gmail.com>
+Jared Van Bortel <jared@nomic.ai>
+Jason McCartney <jmac@theroot.org>
+Jean-Christophe Hoelt <hoelt@fovea.cc>
+Jean-Michaël Celerier <jeanmichael.celerier+github@gmail.com>
+Jed Fox <git@jedfox.com>
+Jeffrey Quesnelle <emozilla@nousresearch.com>
+Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
+Jhen-Jie Hong <iainst0409@gmail.com>
+Jiahao Li <liplus17@163.com>
+Jian Liao <jianliao@users.noreply.github.com>
+JidongZhang-THU <1119708529@qq.com>
+Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
+Jiří Podivín <66251151+jpodivin@users.noreply.github.com>
+Johannes Gäßler <johannesg@5d6.de>
+Johannes Rudolph <johannes.rudolph@gmail.com>
+John <78893154+cmp-nct@users.noreply.github.com>
+John Balis <phobossystems@gmail.com>
+John Smith <67539080+kingsidelee@users.noreply.github.com>
+JohnnyB <jboero@users.noreply.github.com>
+Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com>
+Jorge A <161275481+jorgealias@users.noreply.github.com>
+Jose Maldonado <63384398+yukiteruamano@users.noreply.github.com>
+Joseph Stahl <1269177+josephst@users.noreply.github.com>
+Joyce <joycebrum@google.com>
+Juan Calderon-Perez <835733+gaby@users.noreply.github.com>
+Judd <foldl@users.noreply.github.com>
+Julius Arkenberg <arki05@users.noreply.github.com>
+Jun Jie <71215065+junnjiee16@users.noreply.github.com>
+Juraj Bednar <juraj@bednar.io>
+Justin Parker <jparkerweb@gmail.com>
+Justin Suess <justin.suess@westpoint.edu>
+Justine Tunney <jtunney@gmail.com>
+Juuso Alasuutari <juuso.alasuutari@gmail.com>
+KASR <karim.asrih@gmail.com>
+Kamil Tomšík <info@tomsik.cz>
+Karsten Weiss <knweiss@gmail.com>
+Karthick <j.karthic2004@gmail.com>
+Karthik Kumar Viswanathan <195178+guilt@users.noreply.github.com>
+Karthik Sethuraman <k.seth1993@gmail.com>
+Kasumi <90275229+kasumi-1@users.noreply.github.com>
+Kawrakow <48489457+ikawrakow@users.noreply.github.com>
+Keiichi Tabata <keiichi.tabata@outlook.com>
+Kenvix ⭐ <kenvixzure@live.com>
+Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
+Kevin Ji <1146876+kevinji@users.noreply.github.com>
+Kevin Kwok <antimatter15@gmail.com>
+Kevin Lo <kevlo@kevlo.org>
+Kolen Cheung <ickc@users.noreply.github.com>
+Konstantin Herud <konstantin.herud@denkbares.com>
+Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
+Kunshang Ji <kunshang.ji@intel.com>
+Kyle Liang <liangmanlai@gmail.com>
+Kyle Mistele <kyle@mistele.com>
+Kylin <56434533+KyL0N@users.noreply.github.com>
+Lars Grammel <lars.grammel@gmail.com>
+Laura <Tijntje_7@msn.com>
+Lee <44310445+lx200916@users.noreply.github.com>
+Lee Drake <b.lee.drake@gmail.com>
+Leng Yue <lengyue@lengyue.me>
+LeonEricsson <70749762+LeonEricsson@users.noreply.github.com>
+Leonardo Neumann <leonardo@neumann.dev.br>
+Li Tan <tanliboy@gmail.com>
+Linwei Wang <wanix1988@gmail.com>
+LoganDark <github@logandark.mozmail.com>
+LostRuins <39025047+LostRuins@users.noreply.github.com>
+Luciano <lucianostrika44@gmail.com>
+Luo Tian <lt@basecity.com>
+M. Yusuf Sarıgöz <yusufsarigoz@gmail.com>
+Maarten ter Huurne <maarten@treewalker.org>
+Mack Straight <eiz@users.noreply.github.com>
+Maël Kerbiriou <m431.kerbiriou@gmail.com>
+MaggotHATE <clay1326@gmail.com>
+Marc Köhlbrugge <subscriptions@marckohlbrugge.com>
+Marco Matthies <71844+marcom@users.noreply.github.com>
+Marcus Dunn <51931484+MarcusDunn@users.noreply.github.com>
+Marian Cepok <marian.cepok@gmail.com>
+Mark Fairbairn <thebaron88@gmail.com>
+Marko Tasic <mtasic85@gmail.com>
+Martin Krasser <krasserm@googlemail.com>
+Martin Schwaighofer <mschwaig@users.noreply.github.com>
+Marvin Gießing <marvin.giessing@gmail.com>
+Mateusz Charytoniuk <mateusz.charytoniuk@protonmail.com>
+Matheus C. França <matheus-catarino@hotmail.com>
+Matheus Gabriel Alves Silva <matheusgasource@gmail.com>
+Mathieu Nayrolles <MathieuNls@users.noreply.github.com>
+Mathijs de Bruin <mathijs@mathijsfietst.nl>
+Matt Clayton <156335168+mattjcly@users.noreply.github.com>
+Matt Pulver <matt.pulver@heavy.ai>
+Matteo Boschini <12133566+mbosc@users.noreply.github.com>
+Matthew Tejo <matthew.tejo@gmail.com>
+Matvey Soloviev <blackhole89@gmail.com>
+Maxime <672982+maximegmd@users.noreply.github.com>
+Maximilian Winter <maximilian.winter.91@gmail.com>
+Meng Zhang <meng@tabbyml.com>
+Meng, Hengyu <hengyu.meng@intel.com>
+Merrick Christensen <merrick.christensen@gmail.com>
+Michael Coppola <m18coppola@gmail.com>
+Michael Hueschen <m@mhueschen.dev>
+Michael Kesper <mkesper@schokokeks.org>
+Michael Klimenko <mklimenko29@gmail.com>
+Michael Podvitskiy <podvitskiymichael@gmail.com>
+Michael Potter <NanoTekGuy@Gmail.com>
+Michaël de Vries <vriesdemichael@gmail.com>
+Mihai <mihai.chirculescu@yahoo.com>
+Mike <ytianhui2004@gmail.com>
+Minsoo Cheong <54794500+mscheong01@users.noreply.github.com>
+Mirko185 <mirkosig@gmail.com>
+Mirror Azure <54669636+MirrorAzure@users.noreply.github.com>
+Miwa / Ensan <63481257+ensan-hcl@users.noreply.github.com>
+Mohammadreza Hendiani <hendiani.mohammadreza@gmail.com>
+Murilo Santana <mvrilo@gmail.com>
+Musab Gultekin <musabgultekin@users.noreply.github.com>
+Nam D. Tran <42194884+namtranase@users.noreply.github.com>
+NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
+Nebula <infinitewormhole@gmail.com>
+Neo Zhang Jianyu <jianyu.zhang@intel.com>
+Neuman Vong <neuman.vong@gmail.com>
+Nexesenex <124105151+Nexesenex@users.noreply.github.com>
+Niall Coates <1349685+Niall-@users.noreply.github.com>
+Nicolai Weitkemper <kontakt@nicolaiweitkemper.de>
+Nigel Bosch <pnigelb@gmail.com>
+Niklas Korz <niklas@niklaskorz.de>
+Nindaleth <Nindaleth@users.noreply.github.com>
+Oleksandr Nikitin <oleksandr@tvori.info>
+Oleksii Maryshchenko <oleksii.maryshchenko@gmail.com>
+Olivier Chafik <ochafik@users.noreply.github.com>
+Ondřej Čertík <ondrej@certik.us>
+Ouadie EL FAROUKI <ouadie.elfarouki@codeplay.com>
+Paul Tsochantaris <ptsochantaris@icloud.com>
+Pavol Rusnak <pavol@rusnak.io>
+Pedro Cuenca <pedro@huggingface.co>
+Peter Sugihara <peter@campsh.com>
+Phil H <5756783+phiharri@users.noreply.github.com>
+Philip Taron <philip.taron@gmail.com>
+Phillip Kravtsov <phillip@kravtsov.net>
+Pierre Alexandre SCHEMBRI <pa.schembri@gmail.com>
+Pierrick Hymbert <pierrick.hymbert@gmail.com>
+Przemysław Pawełczyk <przemoc@gmail.com>
+Qin Yue Chen <71813199+chenqiny@users.noreply.github.com>
+Qingyou Meng <meng.qingyou@gmail.com>
+Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
+RJ Adriaansen <adriaansen@eshcc.eur.nl>
+Radoslav Gerganov <rgerganov@gmail.com>
+Radosław Gryta <radek.gryta@gmail.com>
+Rahul Vivek Nair <68507071+RahulVivekNair@users.noreply.github.com>
+Rand Xie <randxiexyy29@gmail.com>
+Randall Fitzgerald <randall@dasaku.net>
+Reinforce-II <fate@eastal.com>
+Riceball LEE <snowyu.lee@gmail.com>
+Richard Kiss <him@richardkiss.com>
+Richard Roberson <richardr1126@gmail.com>
+Rick G <26732651+TheFlipbook@users.noreply.github.com>
+Rickard Edén <rickardeden@gmail.com>
+Rickard Hallerbäck <rickard.hallerback@gmail.com>
+Rickey Bowers Jr <bitRAKE@gmail.com>
+Riley Stewart <ristew@users.noreply.github.com>
+Rinne <AsakusaRinne@gmail.com>
+Rinne <liu_yaohui1998@126.com>
+Robert Brisita <986796+rbrisita@users.noreply.github.com>
+Robert Sung-wook Shin <edp1096@users.noreply.github.com>
+Robey Holderith <robey@flaminglunchbox.net>
+Robyn <robyngraf@users.noreply.github.com>
+Roger Meier <r.meier@siemens.com>
+Roland <14355895+rbur0425@users.noreply.github.com>
+Romain D <90720+Artefact2@users.noreply.github.com>
+Romain Neutron <romain@neutron.io>
+Roman Parykin <donderom@gmail.com>
+Ron Evans <ron@hybridgroup.com>
+Ron Jailall <rojailal@gmail.com>
+Ronny Brendel <ronnybrendel@gmail.com>
+Ronsor <ronsor@ronsor.pw>
+Rowan Hart <rowanbhart@gmail.com>
+Rune <43761327+Rune-AI@users.noreply.github.com>
+Ryan Landay <rlanday@gmail.com>
+Ryder Wishart <ryderwishart@gmail.com>
+Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
+SakuraUmi <yukinon244@gmail.com>
+Salvador E. Tropea <stropea@inti.gob.ar>
+Sam Spilsbury <smspillaz@gmail.com>
+Sami Farin <3876865+Safari77@users.noreply.github.com>
+Samuel Maynard <samwmaynard@gmail.com>
+Sang-Kil Park <sang.park@42dot.ai>
+Seb C <47074056+Sebby37@users.noreply.github.com>
+Sebastián A <sebastian.aedo29@gmail.com>
+SebastianApel <13675545+SebastianApel@users.noreply.github.com>
+Senemu <10880819+Senemu@users.noreply.github.com>
+Sergey Alirzaev <zl29ah@gmail.com>
+Sergio López <slp@sinrega.org>
+SeungWon Jeong <65549245+redlion0929@users.noreply.github.com>
+ShadovvBeast <ShadovvBeast@gmail.com>
+Shakhar Dasgupta <shakhardasgupta@gmail.com>
+Shangning Xu <32517059+xushangning@users.noreply.github.com>
+Shijie <821898965@qq.com>
+Shintarou Okada <kokuzen@gmail.com>
+Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
+Shouzheng Liu <lshzh.hi@gmail.com>
+Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
+Simon Willison <swillison@gmail.com>
+Siwen Yu <yusiwen@gmail.com>
+Sky Yan <skyan83@gmail.com>
+Slaren <2141330+slaren@users.noreply.github.com>
+Slava Primenko <primenko.s@gmail.com>
+SoftwareRenderer <138734813+SoftwareRenderer@users.noreply.github.com>
+Someone <sergei.kozlukov@aalto.fi>
+Someone Serge <sergei.kozlukov@aalto.fi>
+Sourab Mangrulkar <13534540+pacman100@users.noreply.github.com>
+Spencer Sutton <spencersutton@users.noreply.github.com>
+Srinivas Billa <nivibilla@gmail.com>
+Stefan Sydow <stefan@sydow.email>
+Stephan Walter <stephan@walter.name>
+Stephen Nichols <snichols@users.noreply.github.com>
+Steve Grubb <ausearch.1@gmail.com>
+Steven Roussey <sroussey@gmail.com>
+Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
+Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
+SuperUserNameMan <yoann@terminajones.com>
+Tai Duc Nguyen <taiducnguyen.drexel@gmail.com>
+Taikono-Himazin <kazu@po.harenet.ne.jp>
+Tameem <113388789+AhmadTameem@users.noreply.github.com>
+Tamotsu Takahashi <ttakah+github@gmail.com>
+Thái Hoàng Tâm <75922889+RoyalHeart@users.noreply.github.com>
+Thatcher Chamberlin <j.thatcher.c@gmail.com>
+Theia Vogel <theia@vgel.me>
+Thérence <13496987+Royalphax@users.noreply.github.com>
+Thibault Terrasson <thibault.terrasson@gmail.com>
+Thomas Klausner <wiz@gatalith.at>
+Tim Miller <drasticactions@users.noreply.github.com>
+Timmy Knight <r2d2fish@gmail.com>
+Timothy Cronin <40186632+4imothy@users.noreply.github.com>
+Ting Lou <ting.lou@gmail.com>
+Ting Sun <suntcrick@gmail.com>
+Tobias Lütke <tobi@shopify.com>
+Tom C <tom.corelis@gmail.com>
+Tom Jobbins <784313+TheBloke@users.noreply.github.com>
+Tomas <tom.tomas.36478119@gmail.com>
+Tomáš Pazdiora <tomas.pazdiora@gmail.com>
+Tristan Ross <rosscomputerguy@protonmail.com>
+Tungsten842 <886724vf@anonaddy.me>
+Tungsten842 <quantmint@protonmail.com>
+Tushar <ditsuke@protonmail.com>
+UEXTM.com <84163508+uextm@users.noreply.github.com>
+Uzo Nweke <uzoechi@gmail.com>
+Vaibhav Srivastav <vaibhavs10@gmail.com>
+Val Kharitonov <mail@kharvd.com>
+Valentin Konovalov <valle.ketsujin@gmail.com>
+Valentyn Bezshapkin <61702053+valentynbez@users.noreply.github.com>
+Victor Z. Peng <ziliangdotme@gmail.com>
+Vlad <spitfireage@gmail.com>
+Vladimir <bogdad@gmail.com>
+Vladimir Malyutin <first-leon@yandex.ru>
+Vladimir Zorin <vladimir@deviant.guru>
+Volodymyr Vitvitskyi <72226+signalpillar@users.noreply.github.com>
+WangHaoranRobin <56047610+WangHaoranRobin@users.noreply.github.com>
+Weird Constructor <weirdconstructor@gmail.com>
+Welby Seely <welbyseely@gmail.com>
+Wentai Zhang <rchardx@gmail.com>
+WillCorticesAI <150854901+WillCorticesAI@users.noreply.github.com>
+Willy Tarreau <w@1wt.eu>
+Wu Jian Ping <wujjpp@hotmail.com>
+Wu Jian Ping <wujp@greatld.com>
+Xiake Sun <xiake.sun@intel.com>
+Xiang (Kevin) Li <kevinli020508@gmail.com>
+Xiao-Yong Jin <jinxiaoyong@gmail.com>
+XiaotaoChen <chenxiaotao1234@gmail.com>
+Xiaoyi Chen <cxychina@gmail.com>
+Xingchen Song(宋星辰) <xingchensong1996@163.com>
+Xuan Son Nguyen <thichthat@gmail.com>
+Yann Follet <131855179+YannFollet@users.noreply.github.com>
+Yiming Cui <conandiy@vip.qq.com>
+Yishuo Wang <MeouSker77@outlook.com>
+Yueh-Po Peng <94939112+y10ab1@users.noreply.github.com>
+Yui <dev@sleepyyui.com>
+Yusuf Kağan Hanoğlu <hanoglu@yahoo.com>
+Yuval Peled <31162840+Yuval-Peled@users.noreply.github.com>
+ZHAOKAI WANG <sanxianwei@163.com>
+Zane Shannon <z@zcs.me>
+Zay <95888118+isaiahbjork@users.noreply.github.com>
+Zenix <zenixls2@gmail.com>
+Zhang Peiyuan <a1286225768@gmail.com>
+ZhouYuChen <zhouyuchen@naver.com>
+Ziad Ben Hadj-Alouane <zied.benhadjalouane@gmail.com>
+Ziang Wu <97337387+ZiangWu-77@users.noreply.github.com>
+Zsapi <martin1.zsapka@gmail.com>
+a-n-n-a-l-e-e <150648636+a-n-n-a-l-e-e@users.noreply.github.com>
+adel boussaken <netdur@gmail.com>
+afrideva <95653597+afrideva@users.noreply.github.com>
+akawrykow <142945436+akawrykow@users.noreply.github.com>
+alexpinel <93524949+alexpinel@users.noreply.github.com>
+alonfaraj <alonfaraj@gmail.com>
+andrijdavid <david@geek.mg>
+anon998 <131767832+anon998@users.noreply.github.com>
+anzz1 <anzz1@live.com>
+apaz <aarpazdera@gmail.com>
+apcameron <37645737+apcameron@users.noreply.github.com>
+arcrank <arcrank@gmail.com>
+arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
+at8u <129688334+at8u@users.noreply.github.com>
+automaticcat <daogiatuank54@gmail.com>
+bandoti <141645996+bandoti@users.noreply.github.com>
+beiller <beiller@gmail.com>
+bhubbb <79117352+bhubbb@users.noreply.github.com>
+bmwl <brian.marshall@tolko.com>
+bobqianic <129547291+bobqianic@users.noreply.github.com>
+bryanSwk <93190252+bryanSwk@users.noreply.github.com>
+bsilvereagle <bsilvereagle@users.noreply.github.com>
+bssrdf <merlintiger@hotmail.com>
+byte-6174 <88070277+byte-6174@users.noreply.github.com>
+cebtenzzre <cebtenzzre@gmail.com>
+chaihahaha <chai836275709@gmail.com>
+chiranko <96988916+chiranko@users.noreply.github.com>
+clibdev <52199778+clibdev@users.noreply.github.com>
+clyang <clyang@clyang.net>
+cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
+coezbek <c.oezbek@gmail.com>
+comex <comexk@gmail.com>
+compilade <113953597+compilade@users.noreply.github.com>
+crasm <crasm@git.vczf.net>
+crasm <crasm@git.vczf.us>
+daboe01 <daboe01@googlemail.com>
+david raistrick <keen99@users.noreply.github.com>
+ddpasa <112642920+ddpasa@users.noreply.github.com>
+deepdiffuser <112834445+deepdiffuser@users.noreply.github.com>
+divinity76 <divinity76@gmail.com>
+dotpy314 <33351922+dotpy314@users.noreply.github.com>
+drbh <david.richard.holtz@gmail.com>
+ds5t5 <145942675+ds5t5@users.noreply.github.com>
+dylan <canardleteer@users.noreply.github.com>
+eastriver <lee@eastriver.dev>
+ebraminio <ebraminio@gmail.com>
+eiery <19350831+eiery@users.noreply.github.com>
+eric8607242 <e0928021388@gmail.com>
+fraxy-v <65565042+fraxy-v@users.noreply.github.com>
+github-actions[bot] <github-actions[bot]@users.noreply.github.com>
+gliptic <gliptic@users.noreply.github.com>
+goerch <jhr.walter@t-online.de>
+grahameth <96447521+grahameth@users.noreply.github.com>
+gwjr <502526+gwjr@users.noreply.github.com>
+h-h-h-h <13482553+h-h-h-h@users.noreply.github.com>
+hankcs <cnhankmc@gmail.com>
+hoangmit <hoangmit@users.noreply.github.com>
+hongbo.mo <352280764@qq.com>
+howlger <eclipse@voormann.de>
+howlger <github@voormann.de>
+hutli <6594598+hutli@users.noreply.github.com>
+hutli <hutli@hutli.hu>
+hutli <jensstaermose@hotmail.com>
+hxer7963 <hxer7963@gmail.com>
+hydai <z54981220@gmail.com>
+iSma <ismail.senhaji@gmail.com>
+iacore <74560659+iacore@users.noreply.github.com>
+igarnier <igarnier@protonmail.com>
+iohub <rickyang.pro@gmail.com>
+jacobi petrucciani <8117202+jpetrucciani@users.noreply.github.com>
+jameswu2014 <545426914@qq.com>
+jneem <joeneeman@gmail.com>
+johnson442 <56517414+johnson442@users.noreply.github.com>
+jon-chuang <9093549+jon-chuang@users.noreply.github.com>
+jp-x-g <jpxg-dev@protonmail.com>
+jwj7140 <32943891+jwj7140@users.noreply.github.com>
+kaizau <kaizau@users.noreply.github.com>
+kalomaze <66376113+kalomaze@users.noreply.github.com>
+kang <tpdns9032100@gmail.com>
+katsu560 <118887472+katsu560@users.noreply.github.com>
+kchro3 <62481661+kchro3@users.noreply.github.com>
+khimaros <me@khimaros.com>
+kiltyj <kiltyj@gmail.com>
+klosax <131523366+klosax@users.noreply.github.com>
+kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
+kunnis <kunnis@users.noreply.github.com>
+kuronekosaiko <EvanChanJ@163.com>
+kuvaus <22169537+kuvaus@users.noreply.github.com>
+kwin1412 <42286931+kwin1412@users.noreply.github.com>
+l3utterfly <gc.pthzfoldr@gmail.com>
+ldwang <ftgreat@163.com>
+le.chang <cljs118@126.com>
+leejet <leejet714@gmail.com>
+limitedAtonement <limitedAtonement@users.noreply.github.com>
+lon <114724657+longregen@users.noreply.github.com>
+m3ndax <adrian.goessl@outlook.com>
+maddes8cht <55592906+maddes8cht@users.noreply.github.com>
+makomk <makosoft@googlemail.com>
+manikbhandari <mbbhandarimanik2@gmail.com>
+mdrokz <mohammadmunshi@gmail.com>
+mgroeber9110 <45620825+mgroeber9110@users.noreply.github.com>
+minarchist <minarchist@users.noreply.github.com>
+mj-shifu <77107165+mj-shifu@users.noreply.github.com>
+mmyjona <jonathan.gonse@gmail.com>
+momonga <115213907+mmnga@users.noreply.github.com>
+moritzbrantner <31051084+moritzbrantner@users.noreply.github.com>
+mzcu <milos.cubrilo@gmail.com>
+nanahi <130121847+na-na-hi@users.noreply.github.com>
+ngc92 <7938269+ngc92@users.noreply.github.com>
+nhamanasu <45545786+nhamanasu@users.noreply.github.com>
+niansa/tuxifan <anton-sa@web.de>
+niansa/tuxifan <tuxifan@posteo.de>
+ningshanwutuobang <ningshanwutuobang@gmail.com>
+nold <Nold360@users.noreply.github.com>
+nopperl <54780682+nopperl@users.noreply.github.com>
+nusu-github <29514220+nusu-github@users.noreply.github.com>
+olexiyb <olexiyb@gmail.com>
+oobabooga <112222186+oobabooga@users.noreply.github.com>
+opparco <parco.opaai@gmail.com>
+ostix360 <55257054+ostix360@users.noreply.github.com>
+perserk <perserk@gmail.com>
+postmasters <namnguyen@google.com>
+pudepiedj <pudepiedj@gmail.com>
+qingfengfenga <41416092+qingfengfenga@users.noreply.github.com>
+qouoq <qouoq@fastmail.com>
+qunash <anzoria@gmail.com>
+rabidcopy <rabidcopy@yahoo.com>
+rankaiyx <rankaiyx@rankaiyx.com>
+rhjdvsgsgks <26178113+rhjdvsgsgks@users.noreply.github.com>
+rhuddleston <ryan.huddleston@percona.com>
+rimoliga <53384203+rimoliga@users.noreply.github.com>
+runfuture <runfuture@users.noreply.github.com>
+sandyiscool <sandyiscool@gmail.com>
+semidark <me@semidark.net>
+sharpHL <132747147+sharpHL@users.noreply.github.com>
+shibe2 <shibe@tuta.io>
+singularity <12184989+singularity-s0@users.noreply.github.com>
+sjinzh <sjinzh@gmail.com>
+slaren <2141330+slaren@users.noreply.github.com>
+slaren <slarengh@gmail.com>
+snadampal <87143774+snadampal@users.noreply.github.com>
+staviq <staviq@gmail.com>
+stduhpf <stephduh@live.fr>
+swittk <switt1995@gmail.com>
+takov751 <40316768+takov751@users.noreply.github.com>
+tarcey <cey.tarik@gmail.com>
+texmex76 <40733439+texmex76@users.noreply.github.com>
+thement <40525767+thement@users.noreply.github.com>
+tjohnman <tjohnman@users.noreply.github.com>
+tslmy <tslmy@users.noreply.github.com>
+ubik2 <ubik2@users.noreply.github.com>
+uint256_t <konndennsa@gmail.com>
+uint256_t <maekawatoshiki1017@gmail.com>
+unbounded <haakon@likedan.net>
+valiray <133289098+valiray@users.noreply.github.com>
+vodkaslime <646329483@qq.com>
+vvhg1 <94630311+vvhg1@users.noreply.github.com>
+vxiiduu <73044267+vxiiduu@users.noreply.github.com>
+wbpxre150 <100937007+wbpxre150@users.noreply.github.com>
+whoreson <139810751+whoreson@users.noreply.github.com>
+wonjun Jang <strutive07@gmail.com>
+wzy <32936898+Freed-Wu@users.noreply.github.com>
+xaedes <xaedes@gmail.com>
+xaedes <xaedes@googlemail.com>
+xloem <0xloem@gmail.com>
+yangli2 <yangli2@gmail.com>
+yuiseki <yuiseki@gmail.com>
+zakkor <edward.partenie@gmail.com>
+zhouwg <6889919+zhouwg@users.noreply.github.com>
+zrm <trustiosity.zrm@gmail.com>
+源文雨 <41315874+fumiama@users.noreply.github.com>
+Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>

CMakeLists.txt

@@ -89,8 +89,8 @@ endif()
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
-#option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
+option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
+option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
@@ -99,6 +99,7 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
+option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
 option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
@@ -112,6 +113,9 @@ option(LLAMA_METAL                           "llama: use Metal"
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
 option(LLAMA_METAL_SHADER_DEBUG              "llama: compile Metal with -fno-fast-math"         OFF)
 option(LLAMA_METAL_EMBED_LIBRARY             "llama: embed Metal library"                       OFF)
+set(LLAMA_METAL_MACOSX_VERSION_MIN "" CACHE STRING
+                                             "llama: metal minimum macOS version")
+set(LLAMA_METAL_STD "" CACHE STRING          "llama: metal standard version (-std flag)")
 option(LLAMA_KOMPUTE                         "llama: use Kompute"                               OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
@@ -249,6 +253,16 @@ if (LLAMA_METAL)
             set(XC_FLAGS -O3)
         endif()
 
+        # Append macOS metal versioning flags
+        if (LLAMA_METAL_MACOSX_VERSION_MIN)
+            message(STATUS "Adding -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
+            list(APPEND XC_FLAGS -mmacosx-version-min=${LLAMA_METAL_MACOSX_VERSION_MIN})
+        endif()
+        if (LLAMA_METAL_STD)
+            message(STATUS "Adding -std=${LLAMA_METAL_STD} flag to metal compilation")
+            list(APPEND XC_FLAGS -std=${LLAMA_METAL_STD})
+        endif()
+
         add_custom_command(
             OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
             COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
@@ -359,18 +373,25 @@ if (LLAMA_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
+    message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
+    set(LLAMA_CUDA ON)
+endif()
+
+if (LLAMA_CUDA)
     cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
     if (CUDAToolkit_FOUND)
-        message(STATUS "cuBLAS found")
+        message(STATUS "CUDA found")
 
         enable_language(CUDA)
 
         set(GGML_HEADERS_CUDA ggml-cuda.h)
-        set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
-        add_compile_definitions(GGML_USE_CUBLAS)
+        file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
+        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
+
+        add_compile_definitions(GGML_USE_CUDA)
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -387,6 +408,9 @@ if (LLAMA_CUBLAS)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
+        if (LLAMA_CUDA_NO_PEER_COPY)
+            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
@@ -416,7 +440,7 @@ if (LLAMA_CUBLAS)
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     else()
-        message(WARNING "cuBLAS not found")
+        message(WARNING "CUDA not found")
     endif()
 endif()
 
@@ -515,9 +539,11 @@ if (LLAMA_HIPBLAS)
     message(STATUS "HIP and hipBLAS found")
 
     set(GGML_HEADERS_ROCM ggml-cuda.h)
-    set(GGML_SOURCES_ROCM ggml-cuda.cu)
 
-    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+    file(GLOB GGML_SOURCES_ROCM "ggml-cuda/*.cu")
+    list(APPEND GGML_SOURCES_ROCM "ggml-cuda.cu")
+
+    add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUDA)
 
     if (LLAMA_HIP_UMA)
         add_compile_definitions(GGML_HIP_UMA)
@@ -531,11 +557,15 @@ if (LLAMA_HIPBLAS)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
 
+    if (LLAMA_CUDA_NO_PEER_COPY)
+        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
+    endif()
+
     add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
     add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
     add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
 
-    set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
 
     if (LLAMA_STATIC)
         message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
@@ -818,7 +848,7 @@ endif()
 
 set(CUDA_CXX_FLAGS "")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     set(CUDA_FLAGS -use_fast_math)
 
     if (LLAMA_FATAL_WARNINGS)
@@ -1043,7 +1073,7 @@ endif()
 add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
 add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     list(APPEND CUDA_CXX_FLAGS ${ARCH_FLAGS})
     list(JOIN CUDA_CXX_FLAGS " " CUDA_CXX_FLAGS_JOINED)  # pass host compiler flags as a single argument
     if (NOT CUDA_CXX_FLAGS_JOINED STREQUAL "")
@@ -1153,6 +1183,7 @@ add_library(llama
             llama.h
             unicode.h
             unicode.cpp
+            unicode-data.cpp
             )
 
 target_include_directories(llama PUBLIC .)
@@ -1248,6 +1279,12 @@ if (LLAMA_METAL)
             GROUP_READ
             WORLD_READ
         DESTINATION ${CMAKE_INSTALL_BINDIR})
+    if (NOT LLAMA_METAL_EMBED_LIBRARY)
+        install(
+            FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+            DESTINATION ${CMAKE_INSTALL_BINDIR}
+        )
+    endif()
 endif()
 
 #

LICENSE

@@ -1,6 +1,6 @@
 MIT License
 
-Copyright (c) 2023 Georgi Gerganov
+Copyright (c) 2023-2024 The ggml authors
 
 Permission is hereby granted, free of charge, to any person obtaining a copy
 of this software and associated documentation files (the "Software"), to deal

Makefile

@@ -1,15 +1,16 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
-	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
+	simple batched batched-bench save-load-state server gguf gguf-split eval-callback llama-bench libllava.a llava-cli baby-llama beam-search  \
+	retrieval speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
 	tests/test-llama-grammar tests/test-grammar-parser tests/test-double-float tests/test-grad0 tests/test-opt \
 	tests/test-quantize-fns tests/test-quantize-perf tests/test-sampling tests/test-tokenizer-0-llama          \
 	tests/test-tokenizer-0-falcon tests/test-tokenizer-1-llama tests/test-tokenizer-1-bpe tests/test-rope      \
-	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease
+	tests/test-backend-ops tests/test-model-load-cancel tests/test-autorelease                                 \
+	tests/test-json-schema-to-grammar tests/test-grammar-integration
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -389,14 +390,20 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
+# LLAMA_CUBLAS is deprecated and will be removed in the future
+	LLAMA_CUDA := 1
+endif
+
+ifdef LLAMA_CUDA
 	ifneq ('', '$(wildcard /opt/cuda)')
 		CUDA_PATH ?= /opt/cuda
 	else
 		CUDA_PATH ?= /usr/local/cuda
 	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I$(CUDA_PATH)/include -I$(CUDA_PATH)/targets/$(UNAME_M)-linux/include
+	MK_CPPFLAGS  += -DGGML_USE_CUDA -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/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
+	OBJS         += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
 	MK_NVCCFLAGS += -use_fast_math
 ifdef LLAMA_FATAL_WARNINGS
 	MK_NVCCFLAGS += -Werror all-warnings
@@ -451,19 +458,30 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-#ifdef LLAMA_CUDA_CUBLAS
-#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
-#endif # LLAMA_CUDA_CUBLAS
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.h
+
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+define NVCC_COMPILE
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUDA -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
 else
+define NVCC_COMPILE
 	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+endef # NVCC_COMPILE
 endif # JETSON_EOL_MODULE_DETECT
-endif # LLAMA_CUBLAS
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(NVCC_COMPILE)
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
+	$(NVCC_COMPILE)
+
+endif # LLAMA_CUDA
 
 ifdef LLAMA_CLBLAST
 
@@ -509,7 +527,6 @@ ggml-vulkan.o: ggml-vulkan.cpp ggml-vulkan.h
 endif # LLAMA_VULKAN
 
 ifdef LLAMA_HIPBLAS
-
 	ifeq ($(wildcard /opt/rocm),)
 		ROCM_PATH	?= /usr
 		GPU_TARGETS ?= $(shell $(shell which amdgpu-arch))
@@ -521,7 +538,7 @@ ifdef LLAMA_HIPBLAS
 	LLAMA_CUDA_DMMV_X       ?= 32
 	LLAMA_CUDA_MMV_Y        ?= 1
 	LLAMA_CUDA_KQUANTS_ITER ?= 2
-	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUDA
 ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
@@ -534,9 +551,18 @@ endif # LLAMA_HIP_UMA
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
+ifdef LLAMA_CUDA_NO_PEER_COPY
+	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
+endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+	OBJS        += $(patsubst %.cu,%.o,$(wildcard ggml-cuda/*.cu))
+
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml.h ggml-backend.h ggml-backend-impl.h ggml-common.h $(wildcard ggml-cuda/*.cuh)
 	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
+ggml-cuda/%.o: ggml-cuda/%.cu ggml-cuda/%.cuh ggml.h ggml-common.h ggml-cuda/common.cuh
+	$(HIPCC) $(CXXFLAGS) $(HIPFLAGS) -x hip -c -o $@ $<
+
 endif # LLAMA_HIPBLAS
 
 ifdef LLAMA_METAL
@@ -589,7 +615,7 @@ override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
 # identify CUDA host compiler
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
 CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
@@ -614,19 +640,26 @@ $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
 $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef LLAMA_CUBLAS
+ifdef LLAMA_CUDA
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
 CUDA_VERSION := $(shell $(NVCC) --version | grep -oP 'release (\K[0-9]+\.[0-9])')
 ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
 ifndef CUDA_DOCKER_ARCH
 ifndef CUDA_POWER_ARCH
-$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via environment variable CUDA_DOCKER_ARCH, e.g. by running "export CUDA_DOCKER_ARCH=compute_XX" on Unix-like systems, where XX is the minimum compute capability that the code needs to run on. A list with compute capabilities can be found here: https://developer.nvidia.com/cuda-gpus )
 endif # CUDA_POWER_ARCH
 endif # CUDA_DOCKER_ARCH
 endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-endif # LLAMA_CUBLAS
+endif # LLAMA_CUDA
 $(info )
 
+ifdef LLAMA_CUBLAS
+$(info !!!!)
+$(info LLAMA_CUBLAS is deprecated and will be removed in the future. Use LLAMA_CUDA instead.)
+$(info !!!!)
+$(info )
+endif
+
 #
 # Build library
 #
@@ -646,7 +679,10 @@ ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 unicode.o: unicode.cpp unicode.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
+unicode-data.o: unicode-data.cpp unicode-data.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -666,9 +702,15 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 train.o: common/train.cpp common/train.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
+ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
@@ -676,7 +718,8 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf ggml-cuda/*.o
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -745,7 +788,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp json-schema-to-grammar.o common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
@@ -757,6 +800,10 @@ gguf-split: examples/gguf-split/gguf-split.cpp ggml.o llama.o $(COMMON_DEPS) $(O
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -794,6 +841,10 @@ export-lora: examples/export-lora/export-lora.cpp ggml.o common/common.h $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+retrieval: examples/retrieval/retrieval.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -806,11 +857,21 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
+	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
+
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+gbnf-validator: examples/gbnf-validator/gbnf-validator.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -861,10 +922,18 @@ tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o grammar-
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-grammar-integration: tests/test-grammar-integration.cpp ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
+	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

Package.swift

@@ -32,6 +32,7 @@ let package = Package(
                 "ggml.c",
                 "llama.cpp",
                 "unicode.cpp",
+                "unicode-data.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",

README-sycl.md

@@ -3,32 +3,42 @@
 - [Background](#background)
 - [News](#news)
 - [OS](#os)
-- [Intel GPU](#intel-gpu)
+- [Hardware](#hardware)
 - [Docker](#docker)
 - [Linux](#linux)
 - [Windows](#windows)
 - [Environment Variable](#environment-variable)
-- [Known Issue](#known-issue)
-- [Q&A](#q&a)
-- [Todo](#todo)
+- [Known Issue](#known-issues)
+- [Q&A](#qa)
+- [TODO](#todo)
 
 ## Background
 
-SYCL is a higher-level programming model to improve programming productivity on various hardware accelerators—such as CPUs, GPUs, and FPGAs. It is a single-source embedded domain-specific language based on pure C++17.
+**SYCL** is a high-level parallel programming model designed to improve developers productivity writing code across various hardware accelerators such as CPUs, GPUs, and FPGAs. It is a single-source language designed for heterogeneous computing and based on standard C++17.
 
-oneAPI is a specification that is open and standards-based, supporting multiple architecture types including but not limited to GPU, CPU, and FPGA. The spec has both direct programming and API-based programming paradigms.
+**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:
 
-Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
+- **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 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.
 
-To avoid to re-invent the wheel, this code refer other code paths in llama.cpp (like OpenBLAS, cuBLAS, CLBlast). We use a open-source tool [SYCLomatic](https://github.com/oneapi-src/SYCLomatic) (Commercial release [Intel® DPC++ Compatibility Tool](https://www.intel.com/content/www/us/en/developer/tools/oneapi/dpc-compatibility-tool.html)) migrate to SYCL.
+### Llama.cpp + SYCL
 
-The llama.cpp for SYCL is used to support Intel GPUs.
+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*).
 
-For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
+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, CLBlast 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.
 
 ## News
 
+- 2024.4
+  - Support data types: GGML_TYPE_IQ4_NL, GGML_TYPE_IQ4_XS, GGML_TYPE_IQ3_XXS, GGML_TYPE_IQ3_S, GGML_TYPE_IQ2_XXS, GGML_TYPE_IQ2_XS, GGML_TYPE_IQ2_S, GGML_TYPE_IQ1_S, GGML_TYPE_IQ1_M.
+
 - 2024.3
+  - Release binary files of Windows.
+  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
   - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
   - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
   - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
@@ -44,216 +54,234 @@ For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 
 ## OS
 
-|OS|Status|Verified|
-|-|-|-|
-|Linux|Support|Ubuntu 22.04, Fedora Silverblue 39|
-|Windows|Support|Windows 11|
+| OS      | Status  | Verified                           |
+|---------|---------|------------------------------------|
+| Linux   | Support | Ubuntu 22.04, Fedora Silverblue 39 |
+| Windows | Support | Windows 11                         |
 
 
-## Intel GPU
+## Hardware
 
-### Verified
+### Intel GPU
 
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Intel Data Center Max Series| Support| Max 1550|
-|Intel Data Center Flex Series| Support| Flex 170|
-|Intel Arc Series| Support| Arc 770, 730M|
-|Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
-|Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
+**Verified devices**
 
-Note: If the EUs (Execution Unit) in iGPU is less than 80, the inference speed will be too slow to use.
+| Intel GPU                     | Status  | Verified Model                        |
+|-------------------------------|---------|---------------------------------------|
+| Intel Data Center Max Series  | Support | Max 1550, 1100                        |
+| Intel Data Center Flex Series | Support | Flex 170                              |
+| Intel Arc Series              | Support | Arc 770, 730M                         |
+| Intel built-in Arc GPU        | Support | built-in Arc GPU in Meteor Lake       |
+| Intel iGPU                    | Support | iGPU in i5-1250P, i7-1260P, i7-1165G7 |
 
-### Memory
+*Notes:*
 
-The memory is a limitation to run LLM on GPUs.
+- **Memory**
+  - The device memory is a limitation when running a large model. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/main`.
 
-When run llama.cpp, there is print log to show the applied memory on GPU. You could know how much memory to be used in your case. Like `llm_load_tensors:            buffer size =  3577.56 MiB`.
+  - Please make sure the GPU shared memory from the host is large enough to account for the model's size. For e.g. the *llama-2-7b.Q4_0* requires at least 8.0GB for integrated GPU and 4.0GB for discrete GPU.
 
-For iGPU, please make sure the shared memory from host memory is enough. For llama-2-7b.Q4_0, recommend the host memory is 8GB+.
+- **Execution Unit (EU)**
+  - If the iGPU has less than 80 EUs, the inference speed will likely be too slow for practical use.
 
-For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
+### Other Vendor GPU
 
-## Nvidia GPU
+**Verified devices**
 
-### Verified
-
-|Intel GPU| Status | Verified Model|
-|-|-|-|
-|Ampere Series| Support| A100|
-
-### oneMKL for CUDA
-
-The current oneMKL release does not contain the oneMKL cuBlas backend.
-As a result for Nvidia GPU's oneMKL must be built from source.
-
-```
-git clone https://github.com/oneapi-src/oneMKL
-cd oneMKL
-mkdir build
-cd build
-cmake -G Ninja .. -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON
-ninja
-// Add paths as necessary
-```
+| Nvidia GPU               | Status  | Verified Model |
+|--------------------------|---------|----------------|
+| Ampere Series            | Support | A100, A4000    |
+| Ampere Series *(Mobile)* | Support | RTX 40 Series  |
 
 ## Docker
+The docker build option is currently limited to *intel GPU* targets.
 
-Note:
-- Only docker on Linux is tested. Docker on WSL may not work.
-- You may need to install Intel GPU driver on the host machine (See the [Linux](#linux) section to know how to do that)
-
-### Build the image
-
-You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
-
-
+### Build image
 ```sh
-# For F16:
-#docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
-
-# Or, for F32:
-docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
-
-# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example
+# Using FP16
+docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
 ```
 
-### Run
+*Notes*:
+
+To build in default FP32 *(Slower than FP16 alternative)*, you can remove the `--build-arg="LLAMA_SYCL_F16=ON"` argument from the previous command.
+
+You can also use the `.devops/server-intel.Dockerfile`, which builds the *"server"* alternative.
+
+### Run container
 
 ```sh
-# Firstly, find all the DRI cards:
+# First, find all the DRI cards
 ls -la /dev/dri
-# Then, pick the card that you want to use.
-
-# For example with "/dev/dri/card1"
+# Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
 docker run -it --rm -v "$(pwd):/app:Z" --device /dev/dri/renderD128:/dev/dri/renderD128 --device /dev/dri/card1:/dev/dri/card1 llama-cpp-sycl -m "/app/models/YOUR_MODEL_FILE" -p "Building a website can be done in 10 simple steps:" -n 400 -e -ngl 33
 ```
 
+*Notes:*
+- Docker has been tested successfully on native Linux. WSL support has not been verified yet.
+- You may need to install Intel GPU driver on the **host** machine *(Please refer to the [Linux configuration](#linux) for details)*.
+
 ## Linux
 
-### Setup Environment
+### I. Setup Environment
 
-1. Install Intel GPU driver.
+1. **Install GPU drivers**
 
-a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
+  - **Intel GPU**
 
-Note: for iGPU, please install the client GPU driver.
+Intel data center GPUs drivers installation guide and download page can be found here: [Get intel dGPU Drivers](https://dgpu-docs.intel.com/driver/installation.html#ubuntu-install-steps).
 
-b. Add user to group: video, render.
+*Note*: for client GPUs *(iGPU & Arc A-Series)*, please refer to the [client iGPU driver installation](https://dgpu-docs.intel.com/driver/client/overview.html).
+
+Once installed, add the user(s) to the `video` and `render` groups.
 
 ```sh
-sudo usermod -aG render username
-sudo usermod -aG video username
+sudo usermod -aG render $USER
+sudo usermod -aG video $USER
 ```
 
-Note: re-login to enable it.
+*Note*: logout/re-login for the changes to take effect.
 
-c. Check
+Verify installation through `clinfo`:
 
 ```sh
 sudo apt install clinfo
 sudo clinfo -l
 ```
 
-Output (example):
+Sample output:
 
-```
+```sh
 Platform #0: Intel(R) OpenCL Graphics
  `-- Device #0: Intel(R) Arc(TM) A770 Graphics
 
-
 Platform #0: Intel(R) OpenCL HD Graphics
  `-- Device #0: Intel(R) Iris(R) Xe Graphics [0x9a49]
 ```
 
-2. Install Intel® oneAPI Base toolkit.
+- **Nvidia GPU**
 
-a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cuda)-* are installed.
 
-Recommend to install to default folder: **/opt/intel/oneapi**.
+2. **Install Intel® oneAPI Base toolkit**
 
-Following guide use the default folder as example. If you use other folder, please modify the following guide info with your folder.
+- **For Intel GPU**
 
-b. Check
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
+
+Please follow the instructions for downloading and installing the Toolkit for Linux, and preferably keep the default installation values unchanged, notably the installation path *(`/opt/intel/oneapi` by default)*.
+
+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.
+
+- **Adding support to Nvidia GPUs**
+
+**oneAPI Plugin**: In order to enable SYCL support on Nvidia GPUs, please install the [Codeplay oneAPI Plugin for Nvidia GPUs](https://developer.codeplay.com/products/oneapi/nvidia/download). User should also make sure the plugin version matches the installed base toolkit one *(previous step)* for a seamless "oneAPI on Nvidia GPU" setup.
+
+
+**oneMKL for cuBlas**: The current oneMKL releases *(shipped with the oneAPI base-toolkit)* do not contain the cuBLAS backend. A build from source of the upstream [oneMKL](https://github.com/oneapi-src/oneMKL) with the *cuBLAS* backend enabled is thus required to run it on Nvidia GPUs.
 
 ```sh
-source /opt/intel/oneapi/setvars.sh
+git clone https://github.com/oneapi-src/oneMKL
+cd oneMKL
+mkdir -p buildWithCublas && cd buildWithCublas
+cmake ../ -DCMAKE_CXX_COMPILER=icpx -DCMAKE_C_COMPILER=icx -DENABLE_MKLGPU_BACKEND=OFF -DENABLE_MKLCPU_BACKEND=OFF -DENABLE_CUBLAS_BACKEND=ON -DTARGET_DOMAINS=blas
+make
+```
 
+
+3. **Verify installation and environment**
+
+In order to check the available SYCL devices on the machine, please use the `sycl-ls` command.
+```sh
+source /opt/intel/oneapi/setvars.sh
 sycl-ls
 ```
 
-There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+- **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:
 
-Output (example):
 ```
 [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]
-
 ```
 
-2. Build locally:
+- **Nvidia GPU**
 
-Note:
-- You can choose between **F16** and **F32** build. F16 is faster for long-prompt inference.
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+Similarly, user targeting Nvidia GPUs should expect at least one SYCL-CUDA device [`ext_oneapi_cuda:gpu`] as bellow:
+```
+[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]
+```
 
+### II. Build llama.cpp
+
+#### Intel GPU
 ```sh
-mkdir -p build
-cd build
+# Export relevant ENV variables
 source /opt/intel/oneapi/setvars.sh
 
-# For FP16:
-#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+# Build LLAMA with MKL BLAS acceleration for intel GPU
+mkdir -p build && cd build
 
-# Or, for FP32:
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Option 1: Use FP16 for better performance in long-prompt  inference
+cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+# Or without "--build", run "make" next
 
-# For Nvidia GPUs
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+# Option 2: Use FP32 by default
+cmake --build .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
-# Build example/main only
-#cmake --build . --config Release --target main
-
-# Or, build all binary
-cmake --build . --config Release -v
-
-cd ..
+#build all binary
+cmake --build . --config Release -j -v
 ```
 
-or
-
+#### Nvidia GPU
 ```sh
-./examples/sycl/build.sh
+# Export relevant ENV variables
+export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
+export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
+export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
+
+# Build LLAMA with Nvidia BLAS acceleration through SYCL
+mkdir -p build && cd build
+
+# Option 1: Use FP16 for better performance in long-prompt  inference
+cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
+# Option 2: Use FP32 by default
+cmake --build .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+#build all binary
+cmake --build . --config Release -j -v
+
 ```
 
-### Run
+### III. Run the inference
 
-1. Put model file to folder **models**
+1. Retrieve and prepare model
 
-You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) 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
 
-```
+```sh
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List device ID
+3. List devices information
 
-Run without parameter:
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```sh
 ./build/bin/ls-sycl-device
-
-# or running the "main" executable and look at the output log:
-
-./build/bin/main
 ```
-
-Check the ID in startup log, like:
-
+A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
 ```
 found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -267,22 +295,22 @@ found 6 SYCL devices:
 | 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+| Attribute              | Note                                                        |
+|------------------------|-------------------------------------------------------------|
+| compute capability 1.3 | Level-zero driver/runtime, recommended                      |
+| compute capability 3.0 | OpenCL driver/runtime, slower than level-zero in most cases |
 
-4. Device selection and execution of llama.cpp
+4. Launch inference
 
 There are two device selection modes:
 
-- Single device: Use one device assigned by user.
-- Multiple devices: Automatically choose the devices with the same biggest Max compute units.
+- Single device: Use one device target specified by the user.
+- Multiple devices: Automatically select the devices with the same largest Max compute-units.
 
-|Device selection|Parameter|
-|-|-|
-|Single device|--split-mode none --main-gpu DEVICE_ID |
-|Multiple devices|--split-mode layer (default)|
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
 
 Examples:
 
@@ -302,74 +330,63 @@ or run by script:
 ```sh
 ZES_ENABLE_SYSMAN=1 ./build/bin/main -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
 ```
-or run by script:
+
+Otherwise, you can run the script:
 
 ```sh
 ./examples/sycl/run_llama2.sh
 ```
 
-Note:
+*Notes:*
 
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
-
-5. Verify the device ID in output
-
-Verify to see if the selected GPU is shown in the output, like:
-
-```
+```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
 Or
-```
+```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
-
 ## Windows
 
-### Setup Environment
+### I. Setup Environment
 
-1. Install Intel GPU driver.
+1. Install GPU driver
 
-Please install Intel GPU driver by official guide: [Install GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
+Intel GPU drivers instructions guide and download page can be found here: [Get intel GPU Drivers](https://www.intel.com/content/www/us/en/products/docs/discrete-gpus/arc/software/drivers.html).
 
-Note: **The driver is mandatory for compute function**.
+2. Install Visual Studio
 
-2. Install Visual Studio.
+If you already have a recent version of Microsoft Visual Studio, you can skip this step. Otherwise, please refer to the official download page for [Microsoft Visual Studio](https://visualstudio.microsoft.com/).
 
-Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact oneAPI environment enabling in Windows.
+3. Install Intel® oneAPI Base toolkit
 
-3. Install Intel® oneAPI Base toolkit.
+The base toolkit can be obtained from the official [Intel® oneAPI Base Toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) page.
 
-a. Please follow the procedure in [Get the Intel® oneAPI Base Toolkit ](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html).
+Please follow the instructions for downloading and installing the Toolkit for Windows, and preferably keep the default installation values unchanged, notably the installation path *(`C:\Program Files (x86)\Intel\oneAPI` by default)*.
 
-Recommend to install to default folder: **C:\Program Files (x86)\Intel\oneAPI**.
-
-Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
+Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
 
 b. Enable oneAPI running environment:
 
-- In Search, input 'oneAPI'.
+- Type "oneAPI" in the search bar, then open the `Intel oneAPI command prompt for Intel 64 for Visual Studio 2022` App.
 
-Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
-
-- In Run:
-
-In CMD:
+- On the command prompt, enable the runtime environment with the following:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-c. Check GPU
+c. Verify installation
 
-In oneAPI command line:
+In the oneAPI command line, run the following to print the available SYCL devices:
 
 ```
 sycl-ls
 ```
 
-There should be one or more level-zero devices. Please confirm that at least one GPU is present, like **[ext_oneapi_level_zero:gpu:0]**.
+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:
 
 Output (example):
 ```
@@ -379,7 +396,7 @@ Output (example):
 [ext_oneapi_level_zero:gpu:0] Intel(R) Level-Zero, Intel(R) Iris(R) Xe Graphics 1.3 [1.3.28044]
 ```
 
-4. Install cmake & make
+4. Install build tools
 
 a. Download & install cmake for Windows: https://cmake.org/download/
 
@@ -389,76 +406,53 @@ b. Download & install mingw-w64 make for Windows provided by w64devkit
 
 - Extract `w64devkit` on your pc.
 
-- Add the **bin** folder path in the Windows system PATH environment, like `C:\xxx\w64devkit\bin\`.
+- Add the **bin** folder path in the Windows system PATH environment (for e.g. `C:\xxx\w64devkit\bin\`).
 
-### Build locally:
+### II. Build llama.cpp
 
-In oneAPI command line window:
+On the oneAPI command line window, step into the llama.cpp main directory and run the following:
 
 ```
 mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-::  for FP16
-::  faster for long-prompt inference
-::  cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
-::  for FP32
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
-
-
-::  build example/main only
-::  make main
-
-::  build all binary
 make -j
-cd ..
 ```
 
-or
-
-```
+Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
+```sh
 .\examples\sycl\win-build-sycl.bat
 ```
 
-Note:
+*Notes:*
 
-- By default, it will build for all binary files. It will take more time. To reduce the time, we recommend to build for **example/main** only.
+- By default, calling `make` will build all target binary files. In case of a minimal experimental setup, the user can build the inference executable only through `make main`.
 
-### Run
+### III. Run the inference
 
-1. Put model file to folder **models**
+1. Retrieve and prepare model
 
-You could download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) as example.
+You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
 
 2. Enable oneAPI running environment
 
-- In Search, input 'oneAPI'.
-
-Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
-
-- In Run:
-
-In CMD:
+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 device ID
+3. List devices information
 
-Run without parameter:
+Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
 
 ```
 build\bin\ls-sycl-device.exe
-
-or
-
-build\bin\main.exe
 ```
 
-Check the ID in startup log, like:
-
+The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
 ```
 found 6 SYCL devices:
 |  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
@@ -473,23 +467,23 @@ found 6 SYCL devices:
 
 ```
 
-|Attribute|Note|
-|-|-|
-|compute capability 1.3|Level-zero running time, recommended |
-|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
+| Attribute              | Note                                                      |
+|------------------------|-----------------------------------------------------------|
+| compute capability 1.3 | Level-zero running time, recommended                      |
+| compute capability 3.0 | OpenCL running time, slower than level-zero in most cases |
 
 
-4. Device selection and execution of llama.cpp
+4. Launch inference
 
 There are two device selection modes:
 
 - Single device: Use one device assigned by user.
 - Multiple devices: Automatically choose the devices with the same biggest Max compute units.
 
-|Device selection|Parameter|
-|-|-|
-|Single device|--split-mode none --main-gpu DEVICE_ID |
-|Multiple devices|--split-mode layer (default)|
+| Device selection | Parameter                              |
+|------------------|----------------------------------------|
+| Single device    | --split-mode none --main-gpu DEVICE_ID |
+| Multiple devices | --split-mode layer (default)           |
 
 Examples:
 
@@ -504,7 +498,7 @@ build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p "Building a website can be
 ```
 build\bin\main.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
 ```
-or run by script:
+Otherwise, run the following wrapper script:
 
 ```
 .\examples\sycl\win-run-llama2.bat
@@ -512,19 +506,13 @@ or run by script:
 
 Note:
 
-- By default, mmap is used to read model file. In some cases, it leads to the hang issue. Recommend to use parameter **--no-mmap** to disable mmap() to skip this issue.
+- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
 
-
-
-5. Verify the device ID in output
-
-Verify to see if the selected GPU is shown in the output, like:
-
-```
+```sh
 detect 1 SYCL GPUs: [0] with top Max compute units:512
 ```
 Or
-```
+```sh
 use 1 SYCL GPUs: [0] with Max compute units:512
 ```
 
@@ -532,67 +520,51 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 #### Build
 
-|Name|Value|Function|
-|-|-|-|
-|LLAMA_SYCL|ON (mandatory)|Enable build with SYCL code path. <br>For FP32/FP16, LLAMA_SYCL=ON is mandatory.|
-|LLAMA_SYCL_F16|ON (optional)|Enable FP16 build with SYCL code path. Faster for long-prompt inference. <br>For FP32, not set it.|
-|CMAKE_C_COMPILER|icx|Use icx compiler for SYCL code path|
-|CMAKE_CXX_COMPILER|icpx (Linux), icx (Windows)|use icpx/icx for SYCL code path|
+| Name               | Value                             | Function                                    |
+|--------------------|-----------------------------------|---------------------------------------------|
+| LLAMA_SYCL         | ON (mandatory)                    | Enable build with SYCL code path.           |
+| LLAMA_SYCL_TARGET  | INTEL *(default)* \| NVIDIA       | Set the SYCL target device type.            |
+| LLAMA_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. |
 
-#### Running
+#### Runtime
 
+| Name              | Value            | Function                                                                                                                  |
+|-------------------|------------------|---------------------------------------------------------------------------------------------------------------------------|
+| GGML_SYCL_DEBUG   | 0 (default) or 1 | Enable log function by macro: GGML_SYCL_DEBUG                                                                             |
+| ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 
-|Name|Value|Function|
-|-|-|-|
-|GGML_SYCL_DEBUG|0 (default) or 1|Enable log function by macro: GGML_SYCL_DEBUG|
-|ZES_ENABLE_SYSMAN| 0 (default) or 1|Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer|
+## Known Issues
 
-## Known Issue
-
-- Hang during startup
-
-  llama.cpp use mmap as default way to read model file and copy to GPU. In some system, memcpy will be abnormal and block.
-
-  Solution: add **--no-mmap** or **--mmap 0**.
-
-- Split-mode: [row] is not supported
-
-  It's on developing.
+- `Split-mode:[row]` is not supported.
 
 ## Q&A
 
-Note: please add prefix **[SYCL]** in issue title, so that we will check it as soon as possible.
-
-
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
 
-  Miss to enable oneAPI running environment.
+  - Potential cause: Unavailable oneAPI installation or not set ENV variables.
+  - Solution: Install *oneAPI base toolkit* and enable its ENV through: `source /opt/intel/oneapi/setvars.sh`.
 
-  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
+- General compiler error:
 
-- In Windows, no result, not error.
+  - Remove **build** folder or try a clean-build.
 
-  Miss to enable oneAPI running environment.
+- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
 
-- Meet compile error.
+  Please double-check with `sudo sycl-ls`.
 
-  Remove folder **build** and try again.
-
-- I can **not** see **[ext_oneapi_level_zero:gpu:0]** afer install GPU driver in Linux.
-
-  Please run **sudo sycl-ls**.
-
-  If you see it in result, please add video/render group to your ID:
+  If it's present in the list, please add video/render group to your user then **logout/login** or restart your system:
 
   ```
-  sudo usermod -aG render username
-  sudo usermod -aG video username
+  sudo usermod -aG render $USER
+  sudo usermod -aG video $USER
   ```
+  Otherwise, please double-check the GPU driver installation steps.
 
-  Then **relogin**.
+### **GitHub contribution**:
+Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
 
-  If you do not see it, please check the installation GPU steps again.
-
-## Todo
+## TODO
 
 - Support row layer split for multiple card runs.

README.md

@@ -10,6 +10,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
+- [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
@@ -17,7 +19,10 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
-- Multi-GPU pipeline parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
+- **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
@@ -89,6 +94,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] LLaMA 2 🦙🦙
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
+- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
 - [X] Falcon
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca) and [Chinese LLaMA-2 / Alpaca-2](https://github.com/ymcui/Chinese-LLaMA-Alpaca-2)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
@@ -112,7 +118,12 @@ Typically finetunes of the base models below are supported as well.
 - [x] [CodeShell](https://github.com/WisdomShell/codeshell)
 - [x] [Gemma](https://ai.google.dev/gemma)
 - [x] [Mamba](https://github.com/state-spaces/mamba)
+- [x] [Xverse](https://huggingface.co/models?search=xverse)
 - [x] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
+- [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)
+
+(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
 **Multimodal models:**
 
@@ -136,6 +147,7 @@ Typically finetunes of the base models below are supported as well.
 - 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)
+- Rust (more features): [edgenai/llama_cpp-rs](https://github.com/edgenai/llama_cpp-rs)
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
@@ -145,6 +157,7 @@ Typically finetunes of the base models below are supported as well.
 - Java: [kherud/java-llama.cpp](https://github.com/kherud/java-llama.cpp)
 - Zig: [deins/llama.cpp.zig](https://github.com/Deins/llama.cpp.zig)
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
+- PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
 
 **UI:**
 
@@ -165,11 +178,18 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
+- [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
 - [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
 - [Msty](https://msty.app) (proprietary)
 - [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
+- [KanTV](https://github.com/zhouwg/kantv?tab=readme-ov-file)(Apachev2.0 or later)
+- [Dot](https://github.com/alexpinel/Dot) (GPL)
+- [MindMac](https://mindmac.app) (proprietary)
+- [KodiBot](https://github.com/firatkiral/kodibot) (GPL)
+- [eva](https://github.com/ylsdamxssjxxdd/eva) (MIT)
+*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---
 
@@ -445,44 +465,41 @@ Building the program with BLAS support may lead to some performance improvements
 
   Check [Optimizing and Running LLaMA2 on Intel® CPU](https://www.intel.com/content/www/us/en/content-details/791610/optimizing-and-running-llama2-on-intel-cpu.html) for more information.
 
-- #### cuBLAS
+- #### CUDA
 
-  This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
+  This provides GPU acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads).
 
   For Jetson user, if you have Jetson Orin, you can try this: [Offical Support](https://www.jetson-ai-lab.com/tutorial_text-generation.html). If you are using an old model(nano/TX2), need some additional operations before compiling.
 
   - Using `make`:
     ```bash
-    make LLAMA_CUBLAS=1
+    make LLAMA_CUDA=1
     ```
   - Using `CMake`:
 
     ```bash
     mkdir build
     cd build
-    cmake .. -DLLAMA_CUBLAS=ON
+    cmake .. -DLLAMA_CUDA=ON
     cmake --build . --config Release
     ```
 
   The environment variable [`CUDA_VISIBLE_DEVICES`](https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#env-vars) can be used to specify which GPU(s) will be used. The following compilation options are also available to tweak performance:
 
-<!---
-  | LLAMA_CUDA_CUBLAS       | Boolean                |   false | Use cuBLAS instead of custom CUDA kernels for prompt processing. Faster for all quantization formats except for q4_0 and q8_0, especially for k-quants. Increases VRAM usage (700 MiB for 7b, 970 MiB for 13b, 1430 MiB for 33b). |
---->
-  | Option                         | Legal values           | Default | Description |
-  |--------------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_FORCE_DMMV          | Boolean                |   false | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
-  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 |      32 | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y               | Positive integer       |       1 | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. |
-  | LLAMA_CUDA_F16                 | Boolean                |   false | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs. |
-  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 |       2 | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
-  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       |     128 | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial. |
+  | Option                         | Legal values           | Default | Description                                                                                                                                                                                                                                                                             |
+  |--------------------------------|------------------------|---------|-----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | LLAMA_CUDA_FORCE_DMMV          | Boolean                | false   | Force the use of dequantization + matrix vector multiplication kernels instead of using kernels that do matrix vector multiplication on quantized data. By default the decision is made based on compute capability (MMVQ for 6.1/Pascal/GTX 1000 or higher). Does not affect k-quants. |
+  | LLAMA_CUDA_DMMV_X              | Positive integer >= 32 | 32      | Number of values in x direction processed by the CUDA dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants.                                         |
+  | LLAMA_CUDA_MMV_Y               | Positive integer       | 1       | Block size in y direction for the CUDA mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended.                                                                                                                                         |
+  | LLAMA_CUDA_F16                 | Boolean                | false   | If enabled, use half-precision floating point arithmetic for the CUDA dequantization + mul mat vec kernels and for the q4_1 and q5_1 matrix matrix multiplication kernels. Can improve performance on relatively recent GPUs.                                                           |
+  | LLAMA_CUDA_KQUANTS_ITER        | 1 or 2                 | 2       | Number of values processed per iteration and per CUDA thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                                                                     |
+  | LLAMA_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                         |
 
 - #### hipBLAS
 
   This provides BLAS acceleration on HIP-supported AMD GPUs.
   Make sure to have ROCm installed.
-  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/en/latest/deploy/linux/quick_start.html).
+  You can download it from your Linux distro's package manager or from here: [ROCm Quick Start (Linux)](https://rocm.docs.amd.com/projects/install-on-linux/en/latest/tutorial/quick-start.html#rocm-install-quick).
 
   - Using `make`:
     ```bash
@@ -499,7 +516,7 @@ Building the program with BLAS support may lead to some performance improvements
 
   - Using `make` (example for target gfx1030, build with 16 CPU threads):
     ```bash
-    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gxf1030
+    make -j16 LLAMA_HIPBLAS=1 LLAMA_HIP_UMA=1 AMDGPU_TARGETS=gfx1030
     ```
 
   - Using `CMake` for Windows (using x64 Native Tools Command Prompt for VS, and assuming a gfx1100-compatible AMD GPU):
@@ -507,7 +524,7 @@ Building the program with BLAS support may lead to some performance improvements
     set PATH=%HIP_PATH%\bin;%PATH%
     mkdir build
     cd build
-    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ ..
+    cmake -G Ninja -DAMDGPU_TARGETS=gfx1100 -DLLAMA_HIPBLAS=ON -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DCMAKE_BUILD_TYPE=Release ..
     cmake --build .
     ```
     Make sure that `AMDGPU_TARGETS` is set to the GPU arch you want to compile for. The above example uses `gfx1100` that corresponds to Radeon RX 7900XTX/XT/GRE. You can find a list of targets [here](https://llvm.org/docs/AMDGPUUsage.html#processors)
@@ -518,11 +535,11 @@ Building the program with BLAS support may lead to some performance improvements
   If your GPU is not officially supported you can use the environment variable [`HSA_OVERRIDE_GFX_VERSION`] set to a similar GPU, for example 10.3.0 on RDNA2 (e.g. gfx1030, gfx1031, or gfx1035) or 11.0.0 on RDNA3.
   The following compilation options are also available to tweak performance (yes, they refer to CUDA, not HIP, because it uses the same code as the cuBLAS version above):
 
-  | Option                  | Legal values           | Default | Description |
-  |-------------------------|------------------------|---------|-------------|
-  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 |      32 | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_MMV_Y        | Positive integer       |       1 | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants. |
-  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 |       2 | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs. |
+  | Option                  | Legal values           | Default | Description                                                                                                                                                                                                                                    |
+  |-------------------------|------------------------|---------|------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+  | LLAMA_CUDA_DMMV_X       | Positive integer >= 32 | 32      | Number of values in x direction processed by the HIP dequantization + matrix vector multiplication kernel per iteration. Increasing this value can improve performance on fast GPUs. Power of 2 heavily recommended. Does not affect k-quants. |
+  | LLAMA_CUDA_MMV_Y        | Positive integer       | 1       | Block size in y direction for the HIP mul mat vec kernels. Increasing this value can improve performance on fast GPUs. Power of 2 recommended. Does not affect k-quants.                                                                       |
+  | LLAMA_CUDA_KQUANTS_ITER | 1 or 2                 | 2       | Number of values processed per iteration and per HIP thread for Q2_K and Q6_K quantization formats. Setting this value to 1 can improve performance for slow GPUs.                                                                             |
 
 - #### CLBlast
 
@@ -730,11 +747,11 @@ From the unzipped folder, open a terminal/cmd window here and place a pre-conver
 As the models are currently fully loaded into memory, you will need adequate disk space to save them and sufficient RAM to load them. At the moment, memory and disk requirements are the same.
 
 | Model | Original size | Quantized size (Q4_0) |
-|------:|--------------:|-----------------------:|
-|    7B |         13 GB |                 3.9 GB |
-|   13B |         24 GB |                 7.8 GB |
-|   30B |         60 GB |                19.5 GB |
-|   65B |        120 GB |                38.5 GB |
+|------:|--------------:|----------------------:|
+|    7B |         13 GB |                3.9 GB |
+|   13B |         24 GB |                7.8 GB |
+|   30B |         60 GB |               19.5 GB |
+|   65B |        120 GB |               38.5 GB |
 
 ### Quantization
 
@@ -742,7 +759,7 @@ Several quantization methods are supported. They differ in the resulting model d
 
 *(outdated)*
 
-| Model | Measure      | F16    | Q4_0   | Q4_1   | Q5_0   | Q5_1   | Q8_0   |
+| Model | Measure      |    F16 |   Q4_0 |   Q4_1 |   Q5_0 |   Q5_1 |   Q8_0 |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
 |    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
 |    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |

SECURITY.md

@@ -0,0 +1,67 @@
+# Security Policy
+
+ - [**Using llama.cpp securely**](#using-llamacpp-securely)
+   - [Untrusted models](#untrusted-models)
+   - [Untrusted inputs](#untrusted-inputs)
+   - [Data privacy](#data-privacy)
+   - [Untrusted environments or networks](#untrusted-environments-or-networks)
+   - [Multi-Tenant environments](#multi-tenant-environments)
+ - [**Reporting a vulnerability**](#reporting-a-vulnerability)
+
+## Using llama.cpp securely
+
+### Untrusted models
+Be careful when running untrusted models. This classification includes models created by unknown developers or utilizing data obtained from unknown sources.
+
+*Always execute untrusted models within a secure, isolated environment such as a sandbox* (e.g., containers, virtual machines). This helps protect your system from potentially malicious code.
+
+> [!NOTE]
+> The trustworthiness of a model is not binary. You must always determine the proper level of caution depending on the specific model and how it matches your use case and risk tolerance.
+
+### Untrusted inputs
+
+Some models accept various input formats (text, images, audio, etc.). The libraries converting these inputs have varying security levels, so it's crucial to isolate the model and carefully pre-process inputs to mitigate script injection risks.
+
+For maximum security when handling untrusted inputs, you may need to employ the following:
+
+* Sandboxing: Isolate the environment where the inference happens.
+* Pre-analysis: Check how the model performs by default when exposed to prompt injection (e.g. using [fuzzing for prompt injection](https://github.com/FonduAI/awesome-prompt-injection?tab=readme-ov-file#tools)). This will give you leads on how hard you will have to work on the next topics.
+* Updates: Keep both LLaMA C++ and your libraries updated with the latest security patches.
+* Input Sanitation: Before feeding data to the model, sanitize inputs rigorously. This involves techniques such as:
+    * Validation: Enforce strict rules on allowed characters and data types.
+    * Filtering: Remove potentially malicious scripts or code fragments.
+    * Encoding: Convert special characters into safe representations.
+    * Verification: Run tooling that identifies potential script injections (e.g. [models that detect prompt injection attempts](https://python.langchain.com/docs/guides/safety/hugging_face_prompt_injection)).
+
+### Data privacy
+
+To protect sensitive data from potential leaks or unauthorized access, it is crucial to sandbox the model execution. This means running the model in a secure, isolated environment, which helps mitigate many attack vectors.
+
+### Untrusted environments or networks
+
+If you can't run your models in a secure and isolated environment or if it must be exposed to an untrusted network, make sure to take the following security precautions:
+* Confirm the hash of any downloaded artifact (e.g. pre-trained model weights) matches a known-good value
+* Encrypt your data if sending it over the network.
+
+### Multi-Tenant environments
+
+If you intend to run multiple models in parallel with shared memory, it is your responsibility to ensure the models do not interact or access each other's data. The primary areas of concern are tenant isolation, resource allocation, model sharing and hardware attacks.
+
+1. Tenant Isolation: Models should run separately with strong isolation methods to prevent unwanted data access. Separating networks is crucial for isolation, as it prevents unauthorized access to data or models and malicious users from sending graphs to execute under another tenant's identity.
+
+2. Resource Allocation: A denial of service caused by one model can impact the overall system health. Implement safeguards like rate limits, access controls, and health monitoring.
+
+3. Model Sharing: In a multitenant model sharing design, tenants and users must understand the security risks of running code provided by others. Since there are no reliable methods to detect malicious models, sandboxing the model execution is the recommended approach to mitigate the risk.
+
+4. Hardware Attacks: GPUs or TPUs can also be attacked. [Researches](https://scholar.google.com/scholar?q=gpu+side+channel) has shown that side channel attacks on GPUs are possible, which can make data leak from other models or processes running on the same system at the same time.
+
+## Reporting a vulnerability
+
+Beware that none of the topics under [Using llama.cpp securely](#using-llamacpp-securely) are considered vulnerabilities of LLaMA C++.
+
+<!-- normal version -->
+However, If you have discovered a security vulnerability in this project, please report it privately. **Do not disclose it as a public issue.** This gives us time to work with you to fix the issue before public exposure, reducing the chance that the exploit will be used before a patch is released.
+
+Please disclose it as a private [security advisory](https://github.com/ggerganov/llama.cpp/security/advisories/new).
+
+A team of volunteers on a reasonable-effort basis maintains this project. As such, please give us at least 90 days to work on a fix before public exposure.

build.zig

@@ -116,24 +116,26 @@ pub fn build(b: *std.build.Builder) !void {
     const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
     const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
     const unicode = make.obj("unicode", "unicode.cpp");
+    const unicode_data = make.obj("unicode-data", "unicode-data.cpp");
     const llama = make.obj("llama", "llama.cpp");
     const buildinfo = make.obj("common", "common/build-info.cpp");
     const common = make.obj("common", "common/common.cpp");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
+    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
     const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, common, buildinfo, sampling, grammar_parser, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -40,7 +40,7 @@ if [ ! -z ${GG_BUILD_METAL} ]; then
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUDA=1"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
@@ -412,8 +412,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} -DLLAMA_CUBLAS=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} -DLLAMA_CUDA=1 .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                                           ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     python3 ../convert.py ${path_models}
 
@@ -575,7 +575,7 @@ function gg_run_embd_bge_small {
     cd ${SRC}
 
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/config.json
-    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/tokenizer.model
+    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/tokenizer_config.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/special_tokens_map.json
     gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/resolve/main/pytorch_model.bin

common/CMakeLists.txt

@@ -47,6 +47,8 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
+set(TARGET json-schema-to-grammar)
+add_library(${TARGET} OBJECT json-schema-to-grammar.cpp json-schema-to-grammar.h)
 
 set(TARGET common)
 
@@ -60,8 +62,11 @@ add_library(${TARGET} STATIC
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
+    json.hpp
     train.h
     train.cpp
+    ngram-cache.h
+    ngram-cache.cpp
     )
 
 if (BUILD_SHARED_LIBS)

common/common.cpp

@@ -16,6 +16,7 @@
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
+#include <codecvt>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -27,7 +28,6 @@
 #ifndef NOMINMAX
 #   define NOMINMAX
 #endif
-#include <codecvt>
 #include <locale>
 #include <windows.h>
 #include <fcntl.h>
@@ -39,18 +39,21 @@
 #endif
 #if defined(LLAMA_USE_CURL)
 #include <curl/curl.h>
+#include <curl/easy.h>
+#include <thread>
+#include <future>
 #endif
 
 #if defined(_MSC_VER)
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL))
-#define GGML_USE_CUBLAS_SYCL
+#if (defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL))
+#define GGML_USE_CUDA_SYCL
 #endif
 
-#if (defined(GGML_USE_CUBLAS) || defined(GGML_USE_SYCL)) || defined(GGML_USE_VULKAN)
-#define GGML_USE_CUBLAS_SYCL_VULKAN
+#if (defined(GGML_USE_CUDA) || defined(GGML_USE_SYCL)) || defined(GGML_USE_VULKAN)
+#define GGML_USE_CUDA_SYCL_VULKAN
 #endif
 
 #if defined(LLAMA_USE_CURL)
@@ -61,7 +64,7 @@
 #else
 #include <sys/syslimits.h>
 #endif
-#define LLAMA_CURL_MAX_PATH_LENGTH PATH_MAX
+#define LLAMA_CURL_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 #define LLAMA_CURL_MAX_HEADER_LENGTH 256
 #endif // LLAMA_USE_CURL
 
@@ -101,7 +104,7 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-void process_escapes(std::string& input) {
+void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
 
@@ -154,8 +157,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     return result;
 }
 
-static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int & i, bool & invalid_param) {
-    std::string arg = argv[i];
+bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param) {
     llama_sampling_params& sparams = params.sparams;
 
     if (arg == "-s" || arg == "--seed") {
@@ -648,14 +650,6 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
         params.model = argv[i];
         return true;
     }
-    if (arg == "-mu" || arg == "--model-url") {
-        if (++i >= argc) {
-            invalid_param = true;
-            return true;
-        }
-        params.model_url = argv[i];
-        return true;
-    }
     if (arg == "-md" || arg == "--model-draft") {
         if (++i >= argc) {
             invalid_param = true;
@@ -672,6 +666,30 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
         params.model_alias = argv[i];
         return true;
     }
+    if (arg == "-mu" || arg == "--model-url") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.model_url = argv[i];
+        return true;
+    }
+    if (arg == "-hfr" || arg == "--hf-repo") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.hf_repo = argv[i];
+        return true;
+    }
+    if (arg == "-hff" || arg == "--hf-file") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.hf_file = argv[i];
+        return true;
+    }
     if (arg == "--lora") {
         if (++i >= argc) {
             invalid_param = true;
@@ -843,9 +861,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
             return true;
         }
         params.main_gpu = std::stoi(argv[i]);
-#ifndef GGML_USE_CUBLAS_SYCL
-        fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the main GPU has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
+#ifndef GGML_USE_CUDA_SYCL
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL. Setting the main GPU has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL
         return true;
     }
     if (arg == "--split-mode" || arg == "-sm") {
@@ -871,9 +889,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
             invalid_param = true;
             return true;
         }
-#ifndef GGML_USE_CUBLAS_SYCL
-        fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL. Setting the split mode has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
+#ifndef GGML_USE_CUDA_SYCL
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL. Setting the split mode has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL
         return true;
     }
     if (arg == "--tensor-split" || arg == "-ts") {
@@ -899,9 +917,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
                 params.tensor_split[i] = 0.0f;
             }
         }
-#ifndef GGML_USE_CUBLAS_SYCL_VULKAN
-        fprintf(stderr, "warning: llama.cpp was compiled without cuBLAS/SYCL/Vulkan. Setting a tensor split has no effect.\n");
-#endif // GGML_USE_CUBLAS_SYCL
+#ifndef GGML_USE_CUDA_SYCL_VULKAN
+        fprintf(stderr, "warning: llama.cpp was compiled without CUDA/SYCL/Vulkan. Setting a tensor split has no effect.\n");
+#endif // GGML_USE_CUDA_SYCL_VULKAN
         return true;
     }
     if (arg == "--no-mmap") {
@@ -948,6 +966,22 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
         }
         return true;
     }
+    if (arg == "-lcs" || arg == "--lookup-cache-static") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.lookup_cache_static = argv[i];
+        return true;
+    }
+    if (arg == "-lcd" || arg == "--lookup-cache-dynamic") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        params.lookup_cache_dynamic = argv[i];
+        return true;
+    }
     if (arg == "--save-all-logits" || arg == "--kl-divergence-base") {
         if (++i >= argc) {
             invalid_param = true;
@@ -1028,8 +1062,8 @@ static bool gpt_params_find_arg(int argc, char ** argv, gpt_params & params, int
         params.ignore_eos = true;
         return true;
     }
-    if (arg == "--no-penalize-nl") {
-        sparams.penalize_nl = false;
+    if (arg == "--penalize-nl") {
+        sparams.penalize_nl = true;
         return true;
     }
     if (arg == "-l" || arg == "--logit-bias") {
@@ -1201,13 +1235,15 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             std::replace(arg.begin(), arg.end(), '_', '-');
         }
 
-        if (!gpt_params_find_arg(argc, argv, params, i, invalid_param)) {
+        if (!gpt_params_find_arg(argc, argv, arg, params, i, invalid_param)) {
             throw std::invalid_argument("error: unknown argument: " + arg);
         }
     }
+
     if (invalid_param) {
         throw std::invalid_argument("error: invalid parameter for argument: " + arg);
     }
+
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
              params.instruct)) {
@@ -1215,6 +1251,11 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
         throw std::invalid_argument("error: --prompt-cache-all not supported in interactive mode yet\n");
     }
 
+    // short-hand to avoid specifying --hf-file -> default it to --model
+    if (!params.hf_repo.empty() && params.hf_file.empty()) {
+        params.hf_file = params.model;
+    }
+
     if (params.escape) {
         process_escapes(params.prompt);
         process_escapes(params.input_prefix);
@@ -1332,7 +1373,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -dt N, --defrag-thold N\n");
     printf("                        KV cache defragmentation threshold (default: %.1f, < 0 - disabled)\n", params.defrag_thold);
     printf("  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
-    printf("  --no-penalize-nl      do not penalize newline token\n");
+    printf("  --penalize-nl         penalize newline tokens\n");
     printf("  --temp N              temperature (default: %.1f)\n", (double)sparams.temp);
     printf("  --all-logits          return logits for all tokens in the batch (default: disabled)\n");
     printf("  --hellaswag           compute HellaSwag score over random tasks from datafile supplied with -f\n");
@@ -1404,12 +1445,20 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        layer range to apply the control vector(s) to, start and end inclusive\n");
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
-    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
-    printf("                        model download url (default: %s)\n", params.model_url.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding\n");
+    printf("                        draft model for speculative decoding (default: unused)\n");
+    printf("  -mu MODEL_URL, --model-url MODEL_URL\n");
+    printf("                        model download url (default: unused)\n");
+    printf("  -hfr REPO, --hf-repo REPO\n");
+    printf("                        Hugging Face model repository (default: unused)\n");
+    printf("  -hff FILE, --hf-file FILE\n");
+    printf("                        Hugging Face model file (default: unused)\n");
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
+    printf("  -lcs FNAME, --lookup-cache-static FNAME\n");
+    printf("                        path to static lookup cache to use for lookup decoding (not updated by generation)\n");
+    printf("  -lcd FNAME, --lookup-cache-dynamic FNAME\n");
+    printf("                        path to dynamic lookup cache to use for lookup decoding (updated by generation)\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
@@ -1451,6 +1500,77 @@ std::string gpt_random_prompt(std::mt19937 & rng) {
     GGML_UNREACHABLE();
 }
 
+// Validate if a filename is safe to use
+// To validate a full path, split the path by the OS-specific path separator, and validate each part with this function
+bool validate_file_name(const std::string & filename) {
+    if (!filename.length()) {
+        // Empty filename invalid
+        return false;
+    }
+    if (filename.length() > 255) {
+        // Limit at common largest possible filename on Linux filesystems
+        // to avoid unnecessary further validation
+        // (On systems with smaller limits it will be caught by the OS)
+        return false;
+    }
+
+    std::u32string filename_utf32;
+    try {
+        std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
+        filename_utf32 = converter.from_bytes(filename);
+
+        // If the reverse conversion mismatches, it means overlong UTF-8 sequences were used,
+        // or invalid encodings were encountered. Reject such attempts
+        std::string filename_reencoded = converter.to_bytes(filename_utf32);
+        if (filename_reencoded != filename) {
+            return false;
+        }
+    } catch (const std::exception &) {
+        return false;
+    }
+
+    // Check for forbidden codepoints:
+    // - Control characters
+    // - Unicode equivalents of illegal characters
+    // - UTF-16 surrogate pairs
+    // - UTF-8 replacement character
+    // - Byte order mark (BOM)
+    // - Illegal characters: / \ : * ? " < > |
+    for (char32_t c : filename_utf32) {
+        if (c <= 0x1F // Control characters (C0)
+            || c == 0x7F // Control characters (DEL)
+            || (c >= 0x80 && c <= 0x9F) // Control characters (C1)
+            || c == 0xFF0E // Fullwidth Full Stop (period equivalent)
+            || c == 0x2215 // Division Slash (forward slash equivalent)
+            || c == 0x2216 // Set Minus (backslash equivalent)
+            || (c >= 0xD800 && c <= 0xDFFF) // UTF-16 surrogate pairs
+            || c == 0xFFFD // Replacement Character (UTF-8)
+            || c == 0xFEFF // Byte Order Mark (BOM)
+            || c == '/' || c == '\\' || c == ':' || c == '*' // Illegal characters
+            || c == '?' || c == '"' || c == '<' || c == '>' || c == '|') {
+            return false;
+        }
+    }
+
+    // Reject any leading or trailing ' ', or any trailing '.', these are stripped on Windows and will cause a different filename
+    // Unicode and other whitespace is not affected, only 0x20 space
+    if (filename.front() == ' ' || filename.back() == ' ' || filename.back() == '.') {
+        return false;
+    }
+
+    // Reject any ".." (currently stricter than necessary, it should be fine to just check for == ".." instead)
+    if (filename.find("..") != std::string::npos) {
+        return false;
+    }
+
+    // Reject "."
+    if (filename == ".") {
+        return false;
+    }
+
+    return true;
+}
+
 //
 // String utils
 //
@@ -1590,6 +1710,9 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
     if (s == "q4_1") {
         return GGML_TYPE_Q4_1;
     }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
     if (s == "q5_0") {
         return GGML_TYPE_Q5_0;
     }
@@ -1622,6 +1745,8 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     cparams.yarn_orig_ctx     = params.yarn_orig_ctx;
     cparams.pooling_type      = params.pooling_type;
     cparams.defrag_thold      = params.defrag_thold;
+    cparams.cb_eval           = params.cb_eval;
+    cparams.cb_eval_user_data = params.cb_eval_user_data;
     cparams.offload_kqv       = !params.no_kv_offload;
 
     cparams.type_k = kv_cache_type_from_str(params.cache_type_k);
@@ -1653,25 +1778,13 @@ void llama_batch_add(
 
 #ifdef LLAMA_USE_CURL
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model,
-                                              struct llama_model_params params) {
-    // Basic validation of the model_url
-    if (!model_url || strlen(model_url) == 0) {
-        fprintf(stderr, "%s: invalid model_url\n", __func__);
-        return NULL;
-    }
-
-    // Initialize libcurl globally
-    auto curl = curl_easy_init();
-
-    if (!curl) {
-        fprintf(stderr, "%s: error initializing libcurl\n", __func__);
-        return NULL;
-    }
+static bool llama_download_file(CURL * curl, const char * url, const char * path) {
+    bool force_download = false;
 
     // Set the URL, allow to follow http redirection
-    curl_easy_setopt(curl, CURLOPT_URL, model_url);
+    curl_easy_setopt(curl, CURLOPT_URL, url);
     curl_easy_setopt(curl, CURLOPT_FOLLOWLOCATION, 1L);
+
 #if defined(_WIN32)
     // CURLSSLOPT_NATIVE_CA tells libcurl to use standard certificate store of
     //   operating system. Currently implemented under MS-Windows.
@@ -1680,16 +1793,16 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
 
     // Check if the file already exists locally
     struct stat model_file_info;
-    auto file_exists = (stat(path_model, &model_file_info) == 0);
+    auto file_exists = (stat(path, &model_file_info) == 0);
 
     // If the file exists, check for ${path_model}.etag or ${path_model}.lastModified files
     char etag[LLAMA_CURL_MAX_HEADER_LENGTH] = {0};
-    char etag_path[LLAMA_CURL_MAX_PATH_LENGTH] = {0};
-    snprintf(etag_path, sizeof(etag_path), "%s.etag", path_model);
+    char etag_path[PATH_MAX] = {0};
+    snprintf(etag_path, sizeof(etag_path), "%s.etag", path);
 
     char last_modified[LLAMA_CURL_MAX_HEADER_LENGTH] = {0};
-    char last_modified_path[LLAMA_CURL_MAX_PATH_LENGTH] = {0};
-    snprintf(last_modified_path, sizeof(last_modified_path), "%s.lastModified", path_model);
+    char last_modified_path[PATH_MAX] = {0};
+    snprintf(last_modified_path, sizeof(last_modified_path), "%s.lastModified", path);
 
     if (file_exists) {
         auto * f_etag = fopen(etag_path, "r");
@@ -1697,7 +1810,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
             if (!fgets(etag, sizeof(etag), f_etag)) {
                 fprintf(stderr, "%s: unable to read file %s\n", __func__, etag_path);
             } else {
-                fprintf(stderr, "%s: previous model file found %s: %s\n", __func__, etag_path, etag);
+                fprintf(stderr, "%s: previous file found %s: %s\n", __func__, etag_path, etag);
             }
             fclose(f_etag);
         }
@@ -1707,7 +1820,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
             if (!fgets(last_modified, sizeof(last_modified), f_last_modified)) {
                 fprintf(stderr, "%s: unable to read file %s\n", __func__, last_modified_path);
             } else {
-                fprintf(stderr, "%s: previous model file found %s: %s\n", __func__, last_modified_path,
+                fprintf(stderr, "%s: previous file found %s: %s\n", __func__, last_modified_path,
                         last_modified);
             }
             fclose(f_last_modified);
@@ -1725,6 +1838,11 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
         auto header_callback = [](char * buffer, size_t /*size*/, size_t n_items, void * userdata) -> size_t {
             llama_load_model_from_url_headers *headers = (llama_load_model_from_url_headers *) userdata;
 
+            // Convert header field name to lowercase
+            for (size_t i = 0; i < n_items && buffer[i] != ':'; ++i) {
+                buffer[i] = tolower(buffer[i]);
+            }
+
             const char * etag_prefix = "etag: ";
             if (strncmp(buffer, etag_prefix, strlen(etag_prefix)) == 0) {
                 strncpy(headers->etag, buffer + strlen(etag_prefix), n_items - strlen(etag_prefix) - 2); // Remove CRLF
@@ -1747,7 +1865,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
         if (res != CURLE_OK) {
             curl_easy_cleanup(curl);
             fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
-            return NULL;
+            return false;
         }
 
         long http_code = 0;
@@ -1755,30 +1873,34 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
         if (http_code != 200) {
             // HEAD not supported, we don't know if the file has changed
             // force trigger downloading
-            file_exists = false;
+            force_download = true;
             fprintf(stderr, "%s: HEAD invalid http status code received: %ld\n", __func__, http_code);
         }
     }
 
     // If the ETag or the Last-Modified headers are different: trigger a new download
-    if (!file_exists || strcmp(etag, headers.etag) != 0 || strcmp(last_modified, headers.last_modified) != 0) {
-        char path_model_temporary[LLAMA_CURL_MAX_PATH_LENGTH] = {0};
-        snprintf(path_model_temporary, sizeof(path_model_temporary), "%s.downloadInProgress", path_model);
+    bool should_download = !file_exists
+        || force_download
+        || (strlen(headers.etag) > 0 && strcmp(etag, headers.etag) != 0)
+        || (strlen(headers.last_modified) > 0 && strcmp(last_modified, headers.last_modified) != 0);
+    if (should_download) {
+        char path_temporary[PATH_MAX] = {0};
+        snprintf(path_temporary, sizeof(path_temporary), "%s.downloadInProgress", path);
         if (file_exists) {
-            fprintf(stderr, "%s: deleting previous downloaded model file: %s\n", __func__, path_model);
-            if (remove(path_model) != 0) {
+            fprintf(stderr, "%s: deleting previous downloaded file: %s\n", __func__, path);
+            if (remove(path) != 0) {
                 curl_easy_cleanup(curl);
-                fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path_model);
-                return NULL;
+                fprintf(stderr, "%s: unable to delete file: %s\n", __func__, path);
+                return false;
             }
         }
 
         // Set the output file
-        auto * outfile = fopen(path_model_temporary, "wb");
+        auto * outfile = fopen(path_temporary, "wb");
         if (!outfile) {
             curl_easy_cleanup(curl);
-            fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path_model);
-            return NULL;
+            fprintf(stderr, "%s: error opening local file for writing: %s\n", __func__, path);
+            return false;
         }
 
         typedef size_t(*CURLOPT_WRITEFUNCTION_PTR)(void * data, size_t size, size_t nmemb, void * fd);
@@ -1792,15 +1914,30 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
         //  display download progress
         curl_easy_setopt(curl, CURLOPT_NOPROGRESS, 0L);
 
+        // helper function to hide password in URL
+        auto llama_download_hide_password_in_url = [](const std::string & url) -> std::string {
+            std::size_t protocol_pos = url.find("://");
+            if (protocol_pos == std::string::npos) {
+                return url;  // Malformed URL
+            }
+
+            std::size_t at_pos = url.find('@', protocol_pos + 3);
+            if (at_pos == std::string::npos) {
+                return url;  // No password in URL
+            }
+
+            return url.substr(0, protocol_pos + 3) + "********" + url.substr(at_pos);
+        };
+
         // start the download
-        fprintf(stderr, "%s: downloading model from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
-                model_url, path_model, headers.etag, headers.last_modified);
+        fprintf(stderr, "%s: downloading from %s to %s (server_etag:%s, server_last_modified:%s)...\n", __func__,
+                llama_download_hide_password_in_url(url).c_str(), path, headers.etag, headers.last_modified);
         auto res = curl_easy_perform(curl);
         if (res != CURLE_OK) {
             fclose(outfile);
             curl_easy_cleanup(curl);
             fprintf(stderr, "%s: curl_easy_perform() failed: %s\n", __func__, curl_easy_strerror(res));
-            return NULL;
+            return false;
         }
 
         long http_code = 0;
@@ -1809,7 +1946,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
             fclose(outfile);
             curl_easy_cleanup(curl);
             fprintf(stderr, "%s: invalid http status code received: %ld\n", __func__, http_code);
-            return NULL;
+            return false;
         }
 
         // Clean up
@@ -1821,7 +1958,7 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
             if (etag_file) {
                 fputs(headers.etag, etag_file);
                 fclose(etag_file);
-                fprintf(stderr, "%s: model etag saved %s: %s\n", __func__, etag_path, headers.etag);
+                fprintf(stderr, "%s: file etag saved %s: %s\n", __func__, etag_path, headers.etag);
             }
         }
 
@@ -1831,42 +1968,172 @@ struct llama_model * llama_load_model_from_url(const char * model_url, const cha
             if (last_modified_file) {
                 fputs(headers.last_modified, last_modified_file);
                 fclose(last_modified_file);
-                fprintf(stderr, "%s: model last modified saved %s: %s\n", __func__, last_modified_path,
+                fprintf(stderr, "%s: file last modified saved %s: %s\n", __func__, last_modified_path,
                         headers.last_modified);
             }
         }
 
-        if (rename(path_model_temporary, path_model) != 0) {
+        if (rename(path_temporary, path) != 0) {
+            curl_easy_cleanup(curl);
+            fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_temporary, path);
+            return false;
+        }
+    }
+
+    return true;
+}
+
+struct llama_model * llama_load_model_from_url(
+        const char * model_url,
+        const char * path_model,
+        const struct llama_model_params & params) {
+    // Basic validation of the model_url
+    if (!model_url || strlen(model_url) == 0) {
+        fprintf(stderr, "%s: invalid model_url\n", __func__);
+        return NULL;
+    }
+
+    // Initialize libcurl
+    auto * curl = curl_easy_init();
+
+    if (!curl) {
+        fprintf(stderr, "%s: error initializing libcurl\n", __func__);
+        return NULL;
+    }
+
+    if (!llama_download_file(curl, model_url, path_model)) {
+        return NULL;
+    }
+
+    // check for additional GGUFs split to download
+    int n_split = 0;
+    {
+        struct gguf_init_params gguf_params = {
+            /*.no_alloc = */ true,
+            /*.ctx      = */ NULL,
+        };
+        auto * ctx_gguf = gguf_init_from_file(path_model, gguf_params);
+        if (!ctx_gguf) {
+            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, path_model);
             curl_easy_cleanup(curl);
-            fprintf(stderr, "%s: unable to rename file: %s to %s\n", __func__, path_model_temporary, path_model);
             return NULL;
         }
+
+        auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
+        if (key_n_split >= 0) {
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
+        }
+
+        gguf_free(ctx_gguf);
     }
 
     curl_easy_cleanup(curl);
 
+    if (n_split > 1) {
+        char split_prefix[PATH_MAX] = {0};
+        char split_url_prefix[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+
+        // Verify the first split file format
+        // and extract split URL and PATH prefixes
+        {
+            if (!llama_split_prefix(split_prefix, sizeof(split_prefix), path_model, 0, n_split)) {
+                fprintf(stderr, "\n%s: unexpected model file name: %s"
+                                " n_split=%d\n", __func__, path_model, n_split);
+                return NULL;
+            }
+
+            if (!llama_split_prefix(split_url_prefix, sizeof(split_url_prefix), model_url, 0, n_split)) {
+                fprintf(stderr, "\n%s: unexpected model url: %s"
+                                " n_split=%d\n", __func__, model_url, n_split);
+                return NULL;
+            }
+        }
+
+        // Prepare download in parallel
+        std::vector<std::future<bool>> futures_download;
+        for (int idx = 1; idx < n_split; idx++) {
+            futures_download.push_back(std::async(std::launch::async, [&split_prefix, &split_url_prefix, &n_split](int download_idx) -> bool {
+                char split_path[PATH_MAX] = {0};
+                llama_split_path(split_path, sizeof(split_path), split_prefix, download_idx, n_split);
+
+                char split_url[LLAMA_CURL_MAX_URL_LENGTH] = {0};
+                llama_split_path(split_url, sizeof(split_url), split_url_prefix, download_idx, n_split);
+
+                auto * curl = curl_easy_init();
+                bool res = llama_download_file(curl, split_url, split_path);
+                curl_easy_cleanup(curl);
+
+                return res;
+            }, idx));
+        }
+
+        // Wait for all downloads to complete
+        for (auto & f : futures_download) {
+            if (!f.get()) {
+                return NULL;
+            }
+        }
+    }
+
     return llama_load_model_from_file(path_model, params);
 }
 
+struct llama_model * llama_load_model_from_hf(
+        const char * repo,
+        const char * model,
+        const char * path_model,
+        const struct llama_model_params & params) {
+    // construct hugging face model url:
+    //
+    //  --repo ggml-org/models --file tinyllama-1.1b/ggml-model-f16.gguf
+    //    https://huggingface.co/ggml-org/models/resolve/main/tinyllama-1.1b/ggml-model-f16.gguf
+    //
+    //  --repo TheBloke/Mixtral-8x7B-v0.1-GGUF --file mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //    https://huggingface.co/TheBloke/Mixtral-8x7B-v0.1-GGUF/resolve/main/mixtral-8x7b-v0.1.Q4_K_M.gguf
+    //
+
+    std::string model_url = "https://huggingface.co/";
+    model_url += repo;
+    model_url += "/resolve/main/";
+    model_url += model;
+
+    return llama_load_model_from_url(model_url.c_str(), path_model, params);
+}
+
 #else
 
-struct llama_model * llama_load_model_from_url(const char * /*model_url*/, const char * /*path_model*/,
-                                              struct llama_model_params /*params*/) {
+struct llama_model * llama_load_model_from_url(
+        const char * /*model_url*/,
+        const char * /*path_model*/,
+        const struct llama_model_params & /*params*/) {
     fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from an url not supported.\n", __func__);
     return nullptr;
 }
 
+struct llama_model * llama_load_model_from_hf(
+        const char * /*repo*/,
+        const char * /*model*/,
+        const char * /*path_model*/,
+        const struct llama_model_params & /*params*/) {
+    fprintf(stderr, "%s: llama.cpp built without libcurl, downloading from Hugging Face not supported.\n", __func__);
+    return nullptr;
+}
+
 #endif // LLAMA_USE_CURL
 
 std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_params(gpt_params & params) {
     auto mparams = llama_model_params_from_gpt_params(params);
 
     llama_model * model = nullptr;
-    if (!params.model_url.empty()) {
+
+    if (!params.hf_repo.empty() && !params.hf_file.empty()) {
+        model = llama_load_model_from_hf(params.hf_repo.c_str(), params.hf_file.c_str(), params.model.c_str(), mparams);
+    } else if (!params.model_url.empty()) {
         model = llama_load_model_from_url(params.model_url.c_str(), params.model.c_str(), mparams);
     } else {
         model = llama_load_model_from_file(params.model.c_str(), mparams);
     }
+
     if (model == NULL) {
         fprintf(stderr, "%s: error: failed to load model '%s'\n", __func__, params.model.c_str());
         return std::make_tuple(nullptr, nullptr);
@@ -1906,7 +2173,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
     }
 
     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
-        const std::string& lora_adapter = std::get<0>(params.lora_adapter[i]);
+        const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
         float lora_scale = std::get<1>(params.lora_adapter[i]);
         int err = llama_model_apply_lora_from_file(model,
                                              lora_adapter.c_str(),
@@ -1927,7 +2194,7 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
         params.sparams.logit_bias[llama_token_eos(model)] = -INFINITY;
     }
 
-    {
+    if (params.warmup) {
         LOG("warming up the model with an empty run\n");
 
         std::vector<llama_token> tmp = { llama_token_bos(model), llama_token_eos(model), };
@@ -1947,23 +2214,23 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special) {
-    return llama_tokenize(llama_get_model(ctx), text, add_bos, special);
+                        bool   add_special,
+                        bool   parse_special) {
+    return llama_tokenize(llama_get_model(ctx), text, add_special, parse_special);
 }
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special) {
+                        bool   add_special,
+                        bool   parse_special) {
     // upper limit for the number of tokens
-    int n_tokens = text.length() + add_bos;
+    int n_tokens = text.length() + 2 * add_special;
     std::vector<llama_token> result(n_tokens);
-    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
+    n_tokens = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_bos, special);
+        int check = llama_tokenize(model, text.data(), text.length(), result.data(), result.size(), add_special, parse_special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -2188,7 +2455,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "cpu_has_avx512: %s\n",      ggml_cpu_has_avx512()      ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vbmi: %s\n", ggml_cpu_has_avx512_vbmi() ? "true" : "false");
     fprintf(stream, "cpu_has_avx512_vnni: %s\n", ggml_cpu_has_avx512_vnni() ? "true" : "false");
-    fprintf(stream, "cpu_has_cublas: %s\n",      ggml_cpu_has_cublas()      ? "true" : "false");
+    fprintf(stream, "cpu_has_cuda: %s\n",        ggml_cpu_has_cuda()        ? "true" : "false");
     fprintf(stream, "cpu_has_vulkan: %s\n",      ggml_cpu_has_vulkan()      ? "true" : "false");
     fprintf(stream, "cpu_has_clblast: %s\n",     ggml_cpu_has_clblast()     ? "true" : "false");
     fprintf(stream, "cpu_has_kompute: %s\n",     ggml_cpu_has_kompute()     ? "true" : "false");
@@ -2290,7 +2557,7 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "n_predict: %d # default: -1 (unlimited)\n", params.n_predict);
     fprintf(stream, "n_probs: %d # only used by server binary, default: 0\n", sparams.n_probs);
     fprintf(stream, "no_mmap: %s # default: false\n", !params.use_mmap ? "true" : "false");
-    fprintf(stream, "no_penalize_nl: %s # default: false\n", !sparams.penalize_nl ? "true" : "false");
+    fprintf(stream, "penalize_nl: %s # default: false\n", sparams.penalize_nl ? "true" : "false");
     fprintf(stream, "ppl_output_type: %d # default: 0\n", params.ppl_output_type);
     fprintf(stream, "ppl_stride: %d # default: 0\n", params.ppl_stride);
     fprintf(stream, "presence_penalty: %f # default: 0.0\n", sparams.penalty_present);

common/common.h

@@ -80,6 +80,9 @@ struct gpt_params {
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
     float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
 
+    ggml_backend_sched_eval_callback cb_eval = nullptr;
+    void * cb_eval_user_data                 = nullptr;
+
     ggml_numa_strategy numa = GGML_NUMA_STRATEGY_DISABLED;
 
     llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
@@ -88,18 +91,22 @@ struct gpt_params {
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
-    std::string model_url         = ""; // model url to download
-    std::string model_draft       = "";                              // draft model for speculative decoding
-    std::string model_alias       = "unknown"; // model alias
-    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 model                = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft          = "";  // draft model for speculative decoding
+    std::string model_alias          = "unknown"; // model alias
+    std::string model_url            = "";  // model url to download
+    std::string hf_repo              = "";  // HF repo
+    std::string hf_file              = "";  // HF file
+    std::string prompt               = "";
+    std::string prompt_file          = "";  // store the external prompt file name
+    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix         = "";  // string to prefix user inputs with
+    std::string input_suffix         = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
-    std::string logdir            = "";  // directory in which to save YAML log files
-    std::string logits_file       = "";  // file for saving *all* logits
+    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::vector<llama_model_kv_override> kv_overrides;
 
@@ -139,7 +146,7 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
-    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
+    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
@@ -152,6 +159,7 @@ struct gpt_params {
     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
 
     std::string cache_type_k = "f16"; // KV cache data type for the K
     std::string cache_type_v = "f16"; // KV cache data type for the V
@@ -167,12 +175,16 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
 
 void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
 
+bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_params & params, int & i, bool & invalid_param);
+
 std::string get_system_info(const gpt_params & params);
 
 std::string gpt_random_prompt(std::mt19937 & rng);
 
 void process_escapes(std::string& input);
 
+bool validate_file_name(const std::string & filename);
+
 //
 // String utils
 //
@@ -192,8 +204,8 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model,
-                                                         struct llama_model_params     params);
+struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
+struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
 
 // Batch utils
 
@@ -215,14 +227,14 @@ void llama_batch_add(
 std::vector<llama_token> llama_tokenize(
   const struct llama_context * ctx,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special = false);
+                        bool   add_special,
+                        bool   parse_special = false);
 
 std::vector<llama_token> llama_tokenize(
     const struct llama_model * model,
            const std::string & text,
-                        bool   add_bos,
-                        bool   special = false);
+                        bool   add_special,
+                        bool   parse_special = false);
 
 // tokenizes a token into a piece
 // should work similar to Python's `tokenizer.id_to_piece`
@@ -302,3 +314,10 @@ 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);
+
+//
+// Split utils
+//
+static const char * const LLM_KV_SPLIT_NO            = "split.no";
+static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
+static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

common/json-schema-to-grammar.cpp

@@ -0,0 +1,764 @@
+#include "json-schema-to-grammar.h"
+#include <algorithm>
+#include <fstream>
+#include <map>
+#include <regex>
+#include <sstream>
+#include <string>
+#include <unordered_map>
+#include <unordered_set>
+#include <vector>
+
+using json = nlohmann::ordered_json;
+
+template <typename Iterator>
+static std::string join(Iterator begin, Iterator end, const std::string & separator);
+
+static std::string repeat(const std::string & str, size_t n);
+
+static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "", bool item_rule_is_literal = false) {
+    if (separator_rule.empty()) {
+        if (min_items == 0 && max_items == 1) {
+            return item_rule + "?";
+        } else if (min_items == 1 && max_items == std::numeric_limits<int>::max()) {
+            return item_rule + "+";
+        }
+    }
+
+    std::string result;
+    if (min_items > 0) {
+        if (item_rule_is_literal && separator_rule.empty()) {
+            result = "\"" + repeat(std::string(item_rule.begin() + 1, item_rule.end() - 1), min_items) + "\"";
+        } else {
+            std::vector<std::string> items(min_items, item_rule);
+            result = join(items.begin(), items.end(), separator_rule.empty() ? " " : " " + separator_rule + " ");
+        }
+    }
+
+    std::function<std::string(int, bool)> opt_repetitions = [&](int up_to_n, bool prefix_with_sep) -> std::string {
+        auto content = prefix_with_sep && !separator_rule.empty() ? separator_rule + " " + item_rule : item_rule;
+
+        if (up_to_n == 0) {
+            return "";
+        } else if (up_to_n == 1) {
+            return "(" + content + ")?";
+        } else if (!separator_rule.empty() && !prefix_with_sep) {
+            return "(" + content + " " + opt_repetitions(up_to_n - 1, true) + ")?";
+        } else {
+            std::string res = repeat("(" + content + " ", up_to_n);
+            // strip trailing space
+            res = res.substr(0, res.length() - 1);
+            res += repeat(")?", up_to_n);
+            return res;
+        }
+    };
+
+    if (min_items > 0 && max_items != min_items) {
+        result += " ";
+    }
+
+    if (max_items != std::numeric_limits<int>::max()) {
+        result += opt_repetitions(max_items - min_items, min_items > 0);
+    } else {
+        std::string item_operator = "(" + (separator_rule.empty() ? "" : separator_rule + " ") + item_rule + ")";
+        if (min_items == 0 && !separator_rule.empty()) {
+            result = "(" + item_rule + " " + item_operator + "*)?";
+        } else {
+            result += item_operator + "*";
+        }
+    }
+
+    return result;
+}
+
+const std::string SPACE_RULE = "\" \"?";
+
+struct BuiltinRule {
+    std::string content;
+    std::vector<std::string> deps;
+};
+
+const std::string _up_to_15_digits = build_repetition("[0-9]", 0, 15);
+
+std::unordered_map<std::string, BuiltinRule> PRIMITIVE_RULES = {
+    {"boolean", {"(\"true\" | \"false\") space", {}}},
+    {"decimal-part", {"[0-9] " + _up_to_15_digits, {}}},
+    {"integral-part", {"[0-9] | [1-9] " + _up_to_15_digits, {}}},
+    {"number", {"(\"-\"? integral-part) (\".\" decimal-part)? ([eE] [-+]? integral-part)? space", {"integral-part", "decimal-part"}}},
+    {"integer", {"(\"-\"? integral-part) space", {"integral-part"}}},
+    {"value", {"object | array | string | number | boolean | null", {"object", "array", "string", "number", "boolean", "null"}}},
+    {"object", {"\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space", {"string", "value"}}},
+    {"array", {"\"[\" space ( value (\",\" space value)* )? \"]\" space", {"value"}}},
+    {"uuid", {"\"\\\"\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] "
+                "\"-\" [0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F][0-9a-fA-F] \"\\\"\" space", {}}},
+    {"char",   {"[^\"\\\\] | \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])", {}}},
+    {"string", {"\"\\\"\" char* \"\\\"\" space", {"char"}}},
+    {"null", {"\"null\" space", {}}},
+};
+
+std::unordered_map<std::string, BuiltinRule> STRING_FORMAT_RULES = {
+    {"date", {"[0-9] [0-9] [0-9] [0-9] \"-\" ( \"0\" [1-9] | \"1\" [0-2] ) \"-\" ( \"0\" [1-9] | [1-2] [0-9] | \"3\" [0-1] )", {}}},
+    {"time", {"([01] [0-9] | \"2\" [0-3]) \":\" [0-5] [0-9] \":\" [0-5] [0-9] ( \".\" [0-9] [0-9] [0-9] )? ( \"Z\" | ( \"+\" | \"-\" ) ( [01] [0-9] | \"2\" [0-3] ) \":\" [0-5] [0-9] )", {}}},
+    {"date-time", {"date \"T\" time", {"date", "time"}}},
+    {"date-string", {"\"\\\"\" date \"\\\"\" space", {"date"}}},
+    {"time-string", {"\"\\\"\" time \"\\\"\" space", {"time"}}},
+    {"date-time-string", {"\"\\\"\" date-time \"\\\"\" space", {"date-time"}}}
+};
+
+static bool is_reserved_name(const std::string & name) {
+    static std::unordered_set<std::string> RESERVED_NAMES;
+    if (RESERVED_NAMES.empty()) {
+        RESERVED_NAMES.insert("root");
+        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
+        for (const auto &p : STRING_FORMAT_RULES) RESERVED_NAMES.insert(p.first);
+    }
+    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
+}
+
+std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
+std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
+std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
+std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
+    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
+};
+
+std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
+std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
+
+template <typename Iterator>
+std::string join(Iterator begin, Iterator end, const std::string & separator) {
+    std::ostringstream result;
+    if (begin != end) {
+        result << *begin;
+        for (Iterator it = begin + 1; it != end; ++it) {
+            result << separator << *it;
+        }
+    }
+    return result.str();
+}
+
+static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
+    std::vector<std::string> tokens;
+    size_t start = 0;
+    size_t end = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        tokens.push_back(str.substr(start, end - start));
+        start = end + delimiter.length();
+        end = str.find(delimiter, start);
+    }
+
+    tokens.push_back(str.substr(start));
+
+    return tokens;
+}
+
+static std::string repeat(const std::string & str, size_t n) {
+    if (n == 0) {
+        return "";
+    }
+
+    std::string result;
+    result.reserve(str.length() * n);
+
+    for (size_t i = 0; i < n; ++i) {
+        result += str;
+    }
+
+    return result;
+}
+
+static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
+    std::smatch match;
+    std::string result;
+
+    std::string::const_iterator searchStart(input.cbegin());
+    std::string::const_iterator searchEnd(input.cend());
+
+    while (std::regex_search(searchStart, searchEnd, match, regex)) {
+        result.append(searchStart, searchStart + match.position());
+        result.append(replacement(match));
+        searchStart = match.suffix().first;
+    }
+
+    result.append(searchStart, searchEnd);
+
+    return result;
+}
+
+static std::string format_literal(const std::string & literal) {
+    std::string escaped = replacePattern(literal, GRAMMAR_LITERAL_ESCAPE_RE, [&](const std::smatch & match) {
+        char c = match.str()[0];
+        return GRAMMAR_LITERAL_ESCAPES.at(c);
+    });
+    return "\"" + escaped + "\"";
+}
+
+
+class SchemaConverter {
+private:
+    std::function<json(const std::string &)> _fetch_json;
+    bool _dotall;
+    std::map<std::string, std::string> _rules;
+    std::unordered_map<std::string, json> _refs;
+    std::unordered_set<std::string> _refs_being_resolved;
+    std::vector<std::string> _errors;
+    std::vector<std::string> _warnings;
+
+    std::string _add_rule(const std::string & name, const std::string & rule) {
+        std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
+        if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
+            _rules[esc_name] = rule;
+            return esc_name;
+        } else {
+            int i = 0;
+            while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
+                i++;
+            }
+            std::string key = esc_name + std::to_string(i);
+            _rules[key] = rule;
+            return key;
+        }
+    }
+
+    std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
+        std::vector<std::string> rules;
+        for (size_t i = 0; i < alt_schemas.size(); i++) {
+            rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
+        }
+        return join(rules.begin(), rules.end(), " | ");
+    }
+
+    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
+        if (!(pattern.front() == '^' && pattern.back() == '$')) {
+            _errors.push_back("Pattern must start with '^' and end with '$'");
+            return "";
+        }
+        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
+        std::unordered_map<std::string, std::string> sub_rule_ids;
+
+        size_t i = 0;
+        size_t length = sub_pattern.length();
+
+        using literal_or_rule = std::pair<std::string, bool>;
+        auto to_rule = [&](const literal_or_rule & ls) {
+            auto is_literal = ls.second;
+            auto s = ls.first;
+            return is_literal ? "\"" + s + "\"" : s;
+        };
+        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
+            size_t start = i;
+            std::vector<literal_or_rule> seq;
+
+            auto get_dot = [&]() {
+                std::string rule;
+                if (_dotall) {
+                    rule = "[\\U00000000-\\U0010FFFF]";
+                } else {
+                    rule = "[^\\x0A\\x0D]";
+                }
+                return _add_rule("dot", rule);
+            };
+
+            // Joins the sequence, merging consecutive literals together.
+            auto join_seq = [&]() {
+                std::vector<literal_or_rule> ret;
+
+                std::string literal;
+                auto flush_literal = [&]() {
+                    if (literal.empty()) {
+                        return false;
+                    }
+                    ret.push_back(std::make_pair(literal, true));
+                    literal.clear();
+                    return true;
+                };
+
+                for (const auto & item : seq) {
+                    auto is_literal = item.second;
+                    if (is_literal) {
+                        literal += item.first;
+                    } else {
+                        flush_literal();
+                        ret.push_back(item);
+                    }
+                }
+                flush_literal();
+
+                std::vector<std::string> results;
+                for (const auto & item : ret) {
+                    results.push_back(to_rule(item));
+                }
+                return std::make_pair(join(results.begin(), results.end(), " "), false);
+            };
+
+            while (i < length) {
+                char c = sub_pattern[i];
+                if (c == '.') {
+                    seq.push_back(std::make_pair(get_dot(), false));
+                    i++;
+                } else if (c == '(') {
+                    i++;
+                    if (i < length) {
+                        if (sub_pattern[i] == '?') {
+                            _warnings.push_back("Unsupported pattern syntax");
+                        }
+                    }
+                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
+                } else if (c == ')') {
+                    i++;
+                    if (start > 0 && sub_pattern[start - 1] != '(') {
+                        _errors.push_back("Unbalanced parentheses");
+                    }
+                    return join_seq();
+                } else if (c == '[') {
+                    std::string square_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != ']') {
+                        if (sub_pattern[i] == '\\') {
+                            square_brackets += sub_pattern.substr(i, 2);
+                            i += 2;
+                        } else {
+                            square_brackets += sub_pattern[i];
+                            i++;
+                        }
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced square brackets");
+                    }
+                    square_brackets += ']';
+                    i++;
+                    seq.push_back(std::make_pair(square_brackets, false));
+                } else if (c == '|') {
+                    seq.push_back(std::make_pair("|", false));
+                    i++;
+                } else if (c == '*' || c == '+' || c == '?') {
+                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
+                    i++;
+                } else if (c == '{') {
+                    std::string curly_brackets = std::string(1, c);
+                    i++;
+                    while (i < length && sub_pattern[i] != '}') {
+                        curly_brackets += sub_pattern[i];
+                        i++;
+                    }
+                    if (i >= length) {
+                        _errors.push_back("Unbalanced curly brackets");
+                    }
+                    curly_brackets += '}';
+                    i++;
+                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
+                    int min_times = 0;
+                    int max_times = std::numeric_limits<int>::max();
+                    try {
+                        if (nums.size() == 1) {
+                            min_times = max_times = std::stoi(nums[0]);
+                        } else if (nums.size() != 2) {
+                            _errors.push_back("Wrong number of values in curly brackets");
+                        } else {
+                            if (!nums[0].empty()) {
+                                min_times = std::stoi(nums[0]);
+                            }
+                            if (!nums[1].empty()) {
+                                max_times = std::stoi(nums[1]);
+                            }
+                        }
+                    } catch (const std::invalid_argument & e) {
+                        _errors.push_back("Invalid number in curly brackets");
+                        return std::make_pair("", false);
+                    }
+                    auto &last = seq.back();
+                    auto &sub = last.first;
+                    auto sub_is_literal = last.second;
+
+                    if (!sub_is_literal) {
+                        std::string & sub_id = sub_rule_ids[sub];
+                        if (sub_id.empty()) {
+                            sub_id = _add_rule(name + "-" + std::to_string(sub_rule_ids.size()), sub);
+                        }
+                        sub = sub_id;
+                    }
+                    seq.back().first = build_repetition(
+                        sub_is_literal ? "\"" + sub + "\"" : sub,
+                        min_times,
+                        max_times,
+                        "",
+                        sub_is_literal
+                    );
+                    seq.back().second = false;
+                } else {
+                    std::string literal;
+                    auto is_non_literal = [&](char c) {
+                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
+                    };
+                    while (i < length) {
+                        if (sub_pattern[i] == '\\' && i < length - 1) {
+                            char next = sub_pattern[i + 1];
+                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
+                                i++;
+                                literal += sub_pattern[i];
+                                i++;
+                            } else {
+                                literal += sub_pattern.substr(i, 2);
+                                i += 2;
+                            }
+                        } else if (sub_pattern[i] == '"') {
+                            literal += "\\\"";
+                            i++;
+                        } else if (!is_non_literal(sub_pattern[i]) &&
+                                (i == length - 1 || literal.empty() || sub_pattern[i + 1] == '.' || !is_non_literal(sub_pattern[i + 1]))) {
+                            literal += sub_pattern[i];
+                            i++;
+                        } else {
+                            break;
+                        }
+                    }
+                    if (!literal.empty()) {
+                        seq.push_back(std::make_pair(literal, true));
+                    }
+                }
+            }
+            return join_seq();
+        };
+        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
+    }
+
+    std::string _resolve_ref(const std::string & ref) {
+        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
+        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
+            _refs_being_resolved.insert(ref);
+            json resolved = _refs[ref];
+            ref_name = visit(resolved, ref_name);
+            _refs_being_resolved.erase(ref);
+        }
+        return ref_name;
+    }
+
+    std::string _build_object_rule(
+        const std::vector<std::pair<std::string, json>> & properties,
+        const std::unordered_set<std::string> & required,
+        const std::string & name,
+        const json & additional_properties)
+    {
+        std::vector<std::string> required_props;
+        std::vector<std::string> optional_props;
+        std::unordered_map<std::string, std::string> prop_kv_rule_names;
+        for (const auto & kv : properties) {
+            const auto &prop_name = kv.first;
+            const auto &prop_schema = kv.second;
+
+            std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name);
+            prop_kv_rule_names[prop_name] = _add_rule(
+                name + (name.empty() ? "" : "-") + prop_name + "-kv",
+                format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name
+            );
+            if (required.find(prop_name) != required.end()) {
+                required_props.push_back(prop_name);
+            } else {
+                optional_props.push_back(prop_name);
+            }
+        }
+        if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
+            std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
+            std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
+            std::string kv_rule = _add_rule(sub_name + "-kv", _add_primitive("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            prop_kv_rule_names["*"] = kv_rule;
+            optional_props.push_back("*");
+        }
+
+        std::string rule = "\"{\" space ";
+        for (size_t i = 0; i < required_props.size(); i++) {
+            if (i > 0) {
+                rule += " \",\" space ";
+            }
+            rule += prop_kv_rule_names[required_props[i]];
+        }
+
+        if (!optional_props.empty()) {
+            rule += " (";
+            if (!required_props.empty()) {
+                rule += " \",\" space ( ";
+            }
+
+            std::function<std::string(const std::vector<std::string> &, bool)> get_recursive_refs = [&](const std::vector<std::string> & ks, bool first_is_optional) {
+                std::string res;
+                if (ks.empty()) {
+                    return res;
+                }
+                std::string k = ks[0];
+                std::string kv_rule_name = prop_kv_rule_names[k];
+                if (k == "*") {
+                    res = _add_rule(
+                        name + (name.empty() ? "" : "-") + "additional-kvs",
+                        kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
+                    );
+                } else if (first_is_optional) {
+                    res = "( \",\" space " + kv_rule_name + " )?";
+                } else {
+                    res = kv_rule_name;
+                }
+                if (ks.size() > 1) {
+                    res += " " + _add_rule(
+                        name + (name.empty() ? "" : "-") + k + "-rest",
+                        get_recursive_refs(std::vector<std::string>(ks.begin() + 1, ks.end()), true)
+                    );
+                }
+                return res;
+            };
+
+            for (size_t i = 0; i < optional_props.size(); i++) {
+                if (i > 0) {
+                    rule += " | ";
+                }
+                rule += get_recursive_refs(std::vector<std::string>(optional_props.begin() + i, optional_props.end()), false);
+            }
+            if (!required_props.empty()) {
+                rule += " )";
+            }
+            rule += " )?";
+        }
+
+        rule += " \"}\" space";
+
+        return rule;
+    }
+
+    std::string _add_primitive(const std::string & name, const BuiltinRule & rule) {
+        auto n = _add_rule(name, rule.content);
+        for (const auto & dep : rule.deps) {
+            BuiltinRule dep_rule;
+            auto it = PRIMITIVE_RULES.find(dep);
+            if (it == PRIMITIVE_RULES.end()) {
+                it = STRING_FORMAT_RULES.find(dep);
+                if (it == STRING_FORMAT_RULES.end()) {
+                    _errors.push_back("Rule " + dep + " not known");
+                    continue;
+                }
+            }
+            if (_rules.find(dep) == _rules.end()) {
+                _add_primitive(dep, it->second);
+            }
+        }
+        return n;
+    }
+
+public:
+    SchemaConverter(
+        const std::function<json(const std::string &)> & fetch_json,
+        bool dotall)
+          : _fetch_json(fetch_json), _dotall(dotall)
+    {
+        _rules["space"] = SPACE_RULE;
+    }
+
+    void resolve_refs(json & schema, const std::string & url) {
+        /*
+        * Resolves all $ref fields in the given schema, fetching any remote schemas,
+        * replacing each $ref with absolute reference URL and populates _refs with the
+        * respective referenced (sub)schema dictionaries.
+        */
+        std::function<void(json &)> visit_refs = [&](json & n) {
+            if (n.is_array()) {
+                for (auto & x : n) {
+                    visit_refs(x);
+                }
+            } else if (n.is_object()) {
+                if (n.contains("$ref")) {
+                    std::string ref = n["$ref"];
+                    if (_refs.find(ref) == _refs.end()) {
+                        json target;
+                        if (ref.find("https://") == 0) {
+                            std::string base_url = ref.substr(0, ref.find('#'));
+                            auto it = _refs.find(base_url);
+                            if (it != _refs.end()) {
+                                target = it->second;
+                            } else {
+                                // Fetch the referenced schema and resolve its refs
+                                auto referenced = _fetch_json(ref);
+                                resolve_refs(referenced, base_url);
+                                _refs[base_url] = referenced;
+                            }
+                            if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) {
+                                return;
+                            }
+                        } else if (ref.find("#/") == 0) {
+                            target = schema;
+                            n["$ref"] = url + ref;
+                            ref = url + ref;
+                        } else {
+                            _errors.push_back("Unsupported ref: " + ref);
+                            return;
+                        }
+                        std::string pointer = ref.substr(ref.find('#') + 1);
+                        std::vector<std::string> tokens = split(pointer, "/");
+                        for (size_t i = 1; i < tokens.size(); ++i) {
+                            std::string sel = tokens[i];
+                            if (target.is_null() || !target.contains(sel)) {
+                                _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
+                                return;
+                            }
+                            target = target[sel];
+                        }
+                        _refs[ref] = target;
+                    }
+                } else {
+                    for (auto & kv : n.items()) {
+                        visit_refs(kv.value());
+                    }
+                }
+            }
+        };
+
+        visit_refs(schema);
+    }
+
+    std::string _generate_constant_rule(const json & value) {
+        return format_literal(value.dump());
+    }
+
+    std::string visit(const json & schema, const std::string & name) {
+        json schema_type = schema.contains("type") ? schema["type"] : json();
+        std::string schema_format = schema.contains("format") ? schema["format"].get<std::string>() : "";
+        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
+
+        if (schema.contains("$ref")) {
+            return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
+        } else if (schema.contains("oneOf") || schema.contains("anyOf")) {
+            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
+        } else if (schema_type.is_array()) {
+            std::vector<json> schema_types;
+            for (const auto & t : schema_type) {
+                schema_types.push_back({{"type", t}});
+            }
+            return _add_rule(rule_name, _generate_union_rule(name, schema_types));
+        } else if (schema.contains("const")) {
+            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
+        } else if (schema.contains("enum")) {
+            std::vector<std::string> enum_values;
+            for (const auto & v : schema["enum"]) {
+                enum_values.push_back(_generate_constant_rule(v));
+            }
+            return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
+        } else if ((schema_type.is_null() || schema_type == "object")
+                && (schema.contains("properties") ||
+                    (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
+            std::unordered_set<std::string> required;
+            if (schema.contains("required") && schema["required"].is_array()) {
+                for (const auto & item : schema["required"]) {
+                    if (item.is_string()) {
+                        required.insert(item.get<std::string>());
+                    }
+                }
+            }
+            std::vector<std::pair<std::string, json>> properties;
+            if (schema.contains("properties")) {
+                for (const auto & prop : schema["properties"].items()) {
+                    properties.emplace_back(prop.key(), prop.value());
+                }
+            }
+            return _add_rule(rule_name,
+                _build_object_rule(
+                    properties, required, name,
+                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
+        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
+            std::unordered_set<std::string> required;
+            std::vector<std::pair<std::string, json>> properties;
+            std::string hybrid_name = name;
+            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
+                if (comp_schema.contains("$ref")) {
+                    add_component(_refs[comp_schema["$ref"]], is_required);
+                } else if (comp_schema.contains("properties")) {
+                    for (const auto & prop : comp_schema["properties"].items()) {
+                        properties.emplace_back(prop.key(), prop.value());
+                        if (is_required) {
+                            required.insert(prop.key());
+                        }
+                    }
+                } else {
+                  // todo warning
+                }
+            };
+            for (auto & t : schema["allOf"]) {
+                if (t.contains("anyOf")) {
+                    for (auto & tt : t["anyOf"]) {
+                        add_component(tt, false);
+                    }
+                } else {
+                    add_component(t, true);
+                }
+            }
+            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
+        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
+            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
+            if (items.is_array()) {
+                std::string rule = "\"[\" space ";
+                for (size_t i = 0; i < items.size(); i++) {
+                    if (i > 0) {
+                        rule += " \",\" space ";
+                    }
+                    rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i));
+                }
+                rule += " \"]\" space";
+                return _add_rule(rule_name, rule);
+            } else {
+                std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
+                int min_items = schema.contains("minItems") ? schema["minItems"].get<int>() : 0;
+                json max_items_json = schema.contains("maxItems") ? schema["maxItems"] : json();
+                int max_items = max_items_json.is_number_integer() ? max_items_json.get<int>() : std::numeric_limits<int>::max();
+
+                return _add_rule(rule_name, "\"[\" space " + build_repetition(item_rule_name, min_items, max_items, "\",\" space") + " \"]\" space");
+            }
+        } else if ((schema_type.is_null() || schema_type == "string") && schema.contains("pattern")) {
+            return _visit_pattern(schema["pattern"], rule_name);
+        } else if ((schema_type.is_null() || schema_type == "string") && std::regex_match(schema_format, std::regex("^uuid[1-5]?$"))) {
+            return _add_primitive(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
+        } else if ((schema_type.is_null() || schema_type == "string") && STRING_FORMAT_RULES.find(schema_format + "-string") != STRING_FORMAT_RULES.end()) {
+            auto prim_name = schema_format + "-string";
+            return _add_rule(rule_name, _add_primitive(prim_name, STRING_FORMAT_RULES.at(prim_name)));
+        } else if (schema_type == "string" && (schema.contains("minLength") || schema.contains("maxLength"))) {
+            std::string char_rule = _add_primitive("char", PRIMITIVE_RULES.at("char"));
+            int min_len = schema.contains("minLength") ? schema["minLength"].get<int>() : 0;
+            int max_len = schema.contains("maxLength") ? schema["maxLength"].get<int>() : std::numeric_limits<int>::max();
+            return _add_rule(rule_name, "\"\\\"\" " + build_repetition(char_rule, min_len, max_len) + " \"\\\"\" space");
+        } else if (schema.empty() || schema_type == "object") {
+            return _add_rule(rule_name, _add_primitive("object", PRIMITIVE_RULES.at("object")));
+        } else {
+            if (!schema_type.is_string() || PRIMITIVE_RULES.find(schema_type.get<std::string>()) == PRIMITIVE_RULES.end()) {
+                _errors.push_back("Unrecognized schema: " + schema.dump());
+                return "";
+            }
+            // TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+            return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+        }
+    }
+
+    void check_errors() {
+        if (!_errors.empty()) {
+            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
+        }
+        if (!_warnings.empty()) {
+            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
+        }
+    }
+
+    std::string format_grammar() {
+        std::stringstream ss;
+        for (const auto & kv : _rules) {
+            ss << kv.first << " ::= " << kv.second << std::endl;
+        }
+        return ss.str();
+    }
+};
+
+std::string json_schema_to_grammar(const json & schema) {
+    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
+    auto copy = schema;
+    converter.resolve_refs(copy, "input");
+    converter.visit(copy, "");
+    converter.check_errors();
+    return converter.format_grammar();
+}

common/json-schema-to-grammar.h

@@ -0,0 +1,4 @@
+#pragma once
+#include "json.hpp"
+
+std::string json_schema_to_grammar(const nlohmann::ordered_json& schema);

common/log.h

@@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
     #define LOG_IMPL(str, ...)                                                                                      \
     do {                                                                                                            \
         if (LOG_TARGET != nullptr)                                                                                  \
@@ -257,7 +257,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG_TEE() INSTEAD
 //
-#ifndef _MSC_VER
+#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
     #define LOG_TEE_IMPL(str, ...)                                                                                                      \
     do {                                                                                                                                \
         if (LOG_TARGET != nullptr)                                                                                                      \
@@ -566,6 +566,7 @@ inline void log_print_usage()
     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)

common/ngram-cache.cpp

@@ -0,0 +1,282 @@
+#include "ngram-cache.h"
+#include "common.h"
+#include "log.h"
+
+#include <cstdint>
+#include <fstream>
+
+void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
+                              std::vector<llama_token> & inp, int nnew, bool print_progress) {
+    const int64_t t_start_ms = ggml_time_ms();
+    const int64_t inp_size = inp.size();
+
+    const int64_t n_todo = inp_size * (ngram_max - ngram_min + 1);
+    int64_t n_done = 0;
+
+    for (int64_t ngram_size = ngram_min; ngram_size <= ngram_max; ++ngram_size) {
+        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);
+            const llama_token token = inp[i];
+
+            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
+            if (part_it == ngram_cache.end()) {
+                llama_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);
+                if (token_count_it == part_it->second.end()) {
+                    part_it->second.emplace(token, 1);
+                } else {
+                    token_count_it->second++;
+                }
+            }
+            ++n_done;
+
+            if (print_progress && n_done % 10000000 == 0) {
+                const int64_t t_now_ms = ggml_time_ms();
+                const int64_t eta_ms   = (inp_size*(ngram_max-ngram_min+1) - n_done) * (t_now_ms - t_start_ms) / n_done;
+                const int64_t eta_min  = eta_ms / (60*1000);
+                const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+                fprintf(stderr, "%s: %" PRId64 "/%" PRId64 " done, ETA: %02" PRId64 ":%02" PRId64 "\n", __func__, n_done, n_todo, eta_min, eta_s);
+            }
+        }
+    }
+}
+
+// Helper function to get a token from the combined, speculative sequence of inp and draft.
+static llama_token get_token(const std::vector<llama_token> & inp, const std::vector<llama_token> & draft, const size_t i) {
+    return i < inp.size() ? inp[i] : draft[1 + i - inp.size()];
+}
+
+// If sample size or percentage are below these thresholds the draft is aborted early:
+constexpr int    draft_min_sample_size_lax[LLAMA_NGRAM_MAX] = { 2,  2,  1,  1};
+constexpr int        draft_min_percent_lax[LLAMA_NGRAM_MAX] = {66, 50, 50, 50};
+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);
+    if (part_static_it == nc_static.end()) {
+        return -1;
+    }
+    const llama_ngram_cache_part part_static = part_static_it->second;
+
+    int max_count_static  = 0;
+    int sum_count_static  = 0;
+    llama_token max_token = -1;
+
+    for (std::pair<llama_token, int> token_count_static : part_static) {
+        const llama_token token = token_count_static.first;
+        const int32_t count_static  = token_count_static.second;
+
+        if (count_static > max_count_static) {
+            max_token        = token;
+            max_count_static = count_static;
+        }
+        sum_count_static += count_static;
+    }
+
+    if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
+        return -1;
+    }
+    if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
+        return -1;
+    }
+    return max_token;
+}
+
+// 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,
+    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];
+
+        llama_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;
+
+        int max_count_primary = 0;
+        int max_count_static  = 0;
+        int sum_count_primary = 0;
+        llama_token max_token = -1;
+
+        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);
+
+            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;
+
+            if (count_primary*count_static > max_count_primary*max_count_static) {
+                max_token         = token;
+                max_count_primary = count_primary;
+                max_count_static  = count_static;
+            }
+            sum_count_primary += count_primary;
+        }
+
+        if (sum_count_primary < min_sample_size[i]) {
+            continue;
+        }
+        if (100*max_count_primary < min_percent[i]*sum_count_primary) {
+            continue;;
+        }
+        drafted_token = max_token;
+    }
+
+    return drafted_token;
+}
+
+void llama_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
+) {
+    GGML_ASSERT(draft.size() == 1);
+    const int inp_size = inp.size();
+
+    if (inp_size < LLAMA_NGRAM_STATIC) {
+        return;
+    }
+
+    while ((int) draft.size()-1 < n_draft) {
+        llama_token drafted_token = -1;
+
+        const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
+        llama_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;
+        if (part_static_it != nc_static.end()) {
+            part_static = part_static_it->second;
+        }
+
+        // cd = context + dynamic
+        std::vector<llama_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;
+            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);
+            }
+            ngrams_cd.push_back(ngram_cd);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
+        }
+        if (drafted_token == -1) {
+            drafted_token = try_draft(nc_static, ngram_static);
+        }
+
+        if (drafted_token == -1) {
+            break;
+        }
+
+        LOG(" - draft candidate: token=%d\n", drafted_token);
+        draft.push_back(drafted_token);
+    }
+}
+
+void llama_ngram_cache_save(llama_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;
+        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 *>(&ntokens), sizeof(int32_t));
+        for (std::pair<llama_token, int32_t> item2 : token_counts) {
+            const llama_token token = item2.first;
+            const int32_t     count = item2.second;
+            GGML_ASSERT(count > 0);
+
+            file_out.write(reinterpret_cast<const char *>(&token), sizeof(llama_token));
+            file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
+        }
+    }
+
+}
+
+llama_ngram_cache llama_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;
+
+    llama_ngram ngram;
+    int32_t     ntokens;
+    llama_token token;
+    int32_t     count;
+
+    char * ngramc   = reinterpret_cast<char*>(&ngram);
+    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))) {
+        GGML_ASSERT(!hashmap_file.eof());
+        GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
+        GGML_ASSERT(ntokens > 0);
+        llama_ngram_cache_part token_counts;
+
+        for (int i = 0; i < ntokens; ++i) {
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(tokenc, sizeof(llama_token)));
+            GGML_ASSERT(!hashmap_file.eof());
+            GGML_ASSERT(hashmap_file.read(countc, sizeof(int32_t)));
+            GGML_ASSERT(count > 0);
+            token_counts.emplace(token, count);
+        }
+
+        ngram_cache.emplace(ngram, token_counts);
+    }
+    GGML_ASSERT(hashmap_file.eof());
+
+    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;
+
+        llama_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;
+        }
+
+        for (std::pair<llama_token, int32_t> token_count : part) {
+            const llama_token token = token_count.first;
+            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);
+            if (token_count_merged_it == part_merged_it->second.end()) {
+                part_merged_it->second.emplace(token, count);
+                continue;
+            }
+
+            token_count_merged_it->second += count;
+        }
+    }
+}

common/ngram-cache.h

@@ -0,0 +1,94 @@
+#pragma once
+
+#include "llama.h"
+
+#include <unordered_map>
+#include <string>
+#include <vector>
+
+#define LLAMA_NGRAM_MIN    1
+#define LLAMA_NGRAM_MAX    4
+#define LLAMA_NGRAM_STATIC 2
+
+// Data structures to map n-grams to empirical token probabilities:
+
+struct llama_ngram {
+    llama_token tokens[LLAMA_NGRAM_MAX];
+
+    llama_ngram() {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            tokens[i] = -1;
+        }
+    }
+
+    llama_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 {
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            if (tokens[i] != other.tokens[i]) {
+                return false;
+            }
+        }
+        return true;
+    }
+};
+
+struct llama_ngram_hash_function {
+    size_t operator()(const llama_ngram & ngram) const {
+        size_t hash = 0;
+        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
+            hash ^= std::hash<llama_token>{}(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;
+
+// n-gram -> empirical distribution of following tokens
+typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
+
+
+// Update an ngram cache with tokens.
+// ngram_cache:         the cache to modify.
+// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data.
+// inp_data:            the token sequence with which to update ngram_cache.
+// nnew:                how many new tokens have been appended to inp_data since the last call to this function.
+// print_progress:      whether to print progress to stderr.
+//
+// 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);
+
+// Try to draft tokens from ngram caches.
+// inp:                the tokens generated so far.
+// draft:              the token sequence to draft. Expected to initially contain the previously sampled token.
+// n_draft:            maximum number of tokens to add to draft.
+// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic.
+// 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(
+    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);
+
+// 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);
+
+// Load an ngram cache saved with llama_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);
+
+// 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);

common/sampling.cpp

@@ -168,77 +168,20 @@ static llama_token llama_sampling_sample_impl(
                   bool is_resampling) {  // Add a parameter to indicate if we are resampling
     const llama_sampling_params & params = ctx_sampling->params;
 
-    const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
-
     const float   temp            = params.temp;
-    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 int     mirostat        = params.mirostat;
     const float   mirostat_tau    = params.mirostat_tau;
     const float   mirostat_eta    = params.mirostat_eta;
-    const bool    penalize_nl     = params.penalize_nl;
-
-    auto & prev = ctx_sampling->prev;
-    auto & cur  = ctx_sampling->cur;
 
+    std::vector<float> original_logits;
+    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !is_resampling, &original_logits);
+    if (!is_resampling) {
+        GGML_ASSERT(!original_logits.empty());
+    }
     llama_token id = 0;
-
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
-
-    if (!is_resampling) {
-        // Only make a copy of the original logits if we are not in the resampling phase, not sure if I actually have to do this.
-        original_logits = std::vector<float>(logits, logits + llama_n_vocab(llama_get_model(ctx_main)));
-    }
-
-    // apply params.logit_bias map
-    for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
-        logits[it->first] += it->second;
-    }
-
-    if (ctx_cfg) {
-        float * logits_guidance = llama_get_logits_ith(ctx_cfg, idx);
-        llama_sample_apply_guidance(ctx_main, logits, logits_guidance, params.cfg_scale);
-    }
-
-    cur.clear();
-
-    for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
-        cur.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
-    }
-
-    llama_token_data_array cur_p = { cur.data(), cur.size(), false };
-
-    // apply penalties
-    const auto& penalty_tokens = params.use_penalty_prompt_tokens ? params.penalty_prompt_tokens : prev;
-    const int penalty_tokens_used_size = std::min((int)penalty_tokens.size(), penalty_last_n);
-    if (penalty_tokens_used_size) {
-        const float nl_logit = logits[llama_token_nl(llama_get_model(ctx_main))];
-
-        llama_sample_repetition_penalties(ctx_main, &cur_p,
-                penalty_tokens.data() + penalty_tokens.size() - penalty_tokens_used_size,
-                penalty_tokens_used_size, penalty_repeat, penalty_freq, penalty_present);
-
-        if (!penalize_nl) {
-            for (size_t idx = 0; idx < cur_p.size; idx++) {
-                if (cur_p.data[idx].id == llama_token_nl(llama_get_model(ctx_main))) {
-                    cur_p.data[idx].logit = nl_logit;
-                    break;
-                }
-            }
-        }
-    }
-
-    // If we are in the resampling phase, apply grammar checks before sampling logic
-    if (is_resampling && ctx_sampling->grammar != NULL) {
-        llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
-    }
-
     if (temp < 0.0) {
         // greedy sampling, with probs
         llama_sample_softmax(ctx_main, &cur_p);
@@ -302,11 +245,13 @@ static llama_token llama_sampling_sample_impl(
     return id;
 }
 
-static llama_token_data_array llama_sample_probability_distribution_impl(
+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) {
+                  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));
@@ -315,6 +260,7 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
     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;
@@ -323,8 +269,10 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    // Declare original_logits at the beginning of the function scope
-    std::vector<float> original_logits;
+    if (apply_grammar && original_logits != NULL) {
+        // Only make a copy of the original logits if we are not applying grammar checks, not sure if I actually have to do this.
+        *original_logits = {logits, logits + llama_n_vocab(llama_get_model(ctx_main))};
+    }
 
     // apply params.logit_bias map
     for (auto it = params.logit_bias.begin(); it != params.logit_bias.end(); it++) {
@@ -364,12 +312,11 @@ static llama_token_data_array llama_sample_probability_distribution_impl(
         }
     }
 
-    // apply grammar checks
-    if (ctx_sampling->grammar != NULL) {
+    // apply grammar checks before sampling logic
+    if (apply_grammar && ctx_sampling->grammar != NULL) {
         llama_sample_grammar(ctx_main, &cur_p, ctx_sampling->grammar);
     }
 
-    llama_sample_softmax(ctx_main, &cur_p);
     return cur_p;
 }
 
@@ -382,12 +329,14 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
-llama_token_data_array llama_sampling_probability_distribution(
+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) {
-    return llama_sample_probability_distribution_impl(ctx_sampling,ctx_main, ctx_cfg, idx);
+                  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(

common/sampling.h

@@ -129,14 +129,16 @@ llama_token llama_sampling_sample(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,
         struct llama_context * ctx_cfg,
-        int idx = 0);
+        int idx = -1);
 
-// returns the probability that token of given id will be sampled
-llama_token_data_array llama_sampling_probability_distribution(
+// 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);
+        int idx = 0,
+        bool apply_grammar = true,
+        std::vector<float> * original_logits = nullptr);
 
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,

convert-hf-to-gguf.py

@@ -23,7 +23,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-from convert import HfVocab
+from convert import LlamaHfVocab, permute
 
 
 ###### MODEL DEFINITIONS ######
@@ -43,17 +43,18 @@ AnyModel = TypeVar("AnyModel", bound="type[Model]")
 class Model(ABC):
     _model_classes: dict[str, type[Model]] = {}
 
-    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
+    def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool, use_temp_file: bool):
         self.dir_model = dir_model
         self.ftype = ftype
         self.fname_out = fname_out
         self.is_big_endian = is_big_endian
         self.endianess = gguf.GGUFEndian.BIG if is_big_endian else gguf.GGUFEndian.LITTLE
+        self.use_temp_file = use_temp_file
         self.is_safetensors = self._is_model_safetensors()
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
-        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
+        self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=self.use_temp_file)
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
 
     @property
@@ -93,31 +94,42 @@ class Model(ABC):
 
         if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
             self.gguf_writer.add_context_length(n_ctx)
+            print(f"gguf: context length = {n_ctx}")
 
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         self.gguf_writer.add_embedding_length(n_embd)
+        print(f"gguf: embedding length = {n_embd}")
 
         if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
+            print(f"gguf: feed forward length = {n_ff}")
 
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
         self.gguf_writer.add_head_count(n_head)
+        print(f"gguf: head count = {n_head}")
 
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
+            print(f"gguf: key-value head count = {n_head_kv}")
 
         if (rope_theta := self.hparams.get("rope_theta")) is not None:
             self.gguf_writer.add_rope_freq_base(rope_theta)
+            print(f"gguf: rope theta = {rope_theta}")
         if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
             self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
+            print(f"gguf: rms norm epsilon = {f_rms_eps}")
         if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
             self.gguf_writer.add_layer_norm_eps(f_norm_eps)
+            print(f"gguf: layer norm epsilon = {f_norm_eps}")
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
+            print(f"gguf: expert count = {n_experts}")
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
+            print(f"gguf: experts used count = {n_experts_used}")
 
         self.gguf_writer.add_file_type(self.ftype)
+        print(f"gguf: file type = {self.ftype}")
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -149,7 +161,7 @@ class Model(ABC):
                 data = data.astype(np.float32)
 
             # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+            if self.ftype == 1 and data_dtype == np.float16 and (n_dims == 1 or new_name.endswith("_norm.weight")):
                 data = data.astype(np.float32)
 
             # if f16 desired, convert any float32 2-dim weight tensors to float16
@@ -216,15 +228,14 @@ class Model(ABC):
             return ("pytorch_model.bin",)
         return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
 
-    def _set_vocab_gpt2(self):
-        dir_model = self.dir_model
-        hparams = self.hparams
-        tokens: list[bytearray] = []
+    # used for GPT-2 BPE and WordPiece vocabs
+    def get_basic_vocab(self) -> tuple[list[str], list[int]]:
+        tokens: list[str] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
-        tokenizer = AutoTokenizer.from_pretrained(dir_model)
-        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        tokenizer = AutoTokenizer.from_pretrained(self.dir_model)
+        vocab_size = self.hparams.get("vocab_size", len(tokenizer.vocab))
         assert max(tokenizer.vocab.values()) < vocab_size
 
         reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
@@ -232,8 +243,7 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                pad_token = f"[PAD{i}]".encode('utf-8')
-                tokens.append(bytearray(pad_token))
+                tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -245,17 +255,21 @@ class Model(ABC):
                 tokens.append(reverse_vocab[i])
                 toktypes.append(gguf.TokenType.NORMAL)
 
+        return tokens, toktypes
+
+    def _set_vocab_gpt2(self) -> None:
+        tokens, toktypes = self.get_basic_vocab()
         self.gguf_writer.add_tokenizer_model("gpt2")
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
 
-        special_vocab = gguf.SpecialVocab(dir_model, load_merges=True)
+        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=True)
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def _set_vocab_qwen(self):
         dir_model = self.dir_model
         hparams = self.hparams
-        tokens: list[bytearray] = []
+        tokens: list[str] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
@@ -280,8 +294,7 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                pad_token = f"[PAD{i}]".encode("utf-8")
-                tokens.append(bytearray(pad_token))
+                tokens.append(f"[PAD{i}]")
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -314,13 +327,12 @@ class Model(ABC):
         toktypes: list[int] = []
 
         if not tokenizer_path.is_file():
-            print(f'Error: Missing {tokenizer_path}', file=sys.stderr)
-            sys.exit(1)
+            raise FileNotFoundError(f"File not found: {tokenizer_path}")
 
         tokenizer = SentencePieceProcessor(str(tokenizer_path))
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
-        for token_id in range(vocab_size):
+        for token_id in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.get_score(token_id)
@@ -345,9 +357,13 @@ class Model(ABC):
                 added_tokens_json = json.load(f)
 
                 for key in added_tokens_json:
-                    tokens.append(key.encode("utf-8"))
-                    scores.append(-1000.0)
-                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+                    key = key.encode("utf-8")
+                    if key not in tokens:
+                        tokens.append(key)
+                        scores.append(-1000.0)
+                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
+
+        assert len(tokens) == vocab_size
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_token_list(tokens)
@@ -357,12 +373,8 @@ class Model(ABC):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _set_vocab_hf(self):
-        path = self.dir_model
-        added_tokens_path = self.dir_model
-        vocab = HfVocab(
-            path, added_tokens_path if added_tokens_path.exists() else None
-        )
+    def _set_vocab_llama_hf(self):
+        vocab = LlamaHfVocab(self.dir_model)
         tokens = []
         scores = []
         toktypes = []
@@ -502,6 +514,17 @@ class BloomModel(Model):
 class MPTModel(Model):
     model_arch = gguf.MODEL_ARCH.MPT
 
+    def set_vocab(self):
+        try:
+            self._set_vocab_gpt2()
+        except Exception:
+            # Fallback for SEA-LION model
+            self._set_vocab_sentencepiece()
+            self.gguf_writer.add_add_bos_token(False)
+            self.gguf_writer.add_pad_token_id(3)
+            self.gguf_writer.add_eos_token_id(1)
+            self.gguf_writer.add_unk_token_id(0)
+
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -515,7 +538,10 @@ class MPTModel(Model):
         self.gguf_writer.add_layer_norm_eps(1e-5)
         if self.hparams["attn_config"]["clip_qkv"] is not None:
             self.gguf_writer.add_clamp_kqv(self.hparams["attn_config"]["clip_qkv"])
-        self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+        if self.hparams["attn_config"]["alibi"]:
+            self.gguf_writer.add_max_alibi_bias(self.hparams["attn_config"]["alibi_bias_max"])
+        else:
+            self.gguf_writer.add_max_alibi_bias(0.0)
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers"))
@@ -764,6 +790,148 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
+@Model.register("XverseForCausalLM")
+class XverseModel(Model):
+    model_arch = gguf.MODEL_ARCH.XVERSE
+
+    def set_vocab(self):
+        assert (self.dir_model / "tokenizer.json").is_file()
+        dir_model = self.dir_model
+        hparams = self.hparams
+
+        tokens: list[bytearray] = []
+        toktypes: list[int] = []
+
+        from transformers import AutoTokenizer
+        tokenizer = AutoTokenizer.from_pretrained(dir_model)
+        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))
+        assert max(tokenizer.vocab.values()) < vocab_size
+
+        reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}
+        added_vocab = tokenizer.get_added_vocab()
+
+        for token_id in range(vocab_size):
+            token_text = reverse_vocab[token_id].encode('utf-8')
+            # replace "\x00" to string with length > 0
+            if token_text == b"\x00":
+                toktype = gguf.TokenType.BYTE  # special
+                token_text = f"<{token_text}>".encode('utf-8')
+            elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):
+                toktype = gguf.TokenType.BYTE  # special
+            elif reverse_vocab[token_id] in added_vocab:
+                if tokenizer.added_tokens_decoder[token_id].special:
+                    toktype = gguf.TokenType.CONTROL
+                else:
+                    toktype = gguf.TokenType.USER_DEFINED
+            else:
+                toktype = gguf.TokenType.NORMAL
+
+            tokens.append(token_text)
+            toktypes.append(toktype)
+
+        self.gguf_writer.add_tokenizer_model("llama")
+        self.gguf_writer.add_token_list(tokens)
+        self.gguf_writer.add_token_types(toktypes)
+
+        special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))
+        special_vocab.add_to_gguf(self.gguf_writer)
+
+    def set_gguf_parameters(self):
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+        hf_repo = self.hparams.get("_name_or_path", "")
+
+        ctx_length = 0
+        if "max_sequence_length" in self.hparams:
+            ctx_length = self.hparams["max_sequence_length"]
+        elif "max_position_embeddings" in self.hparams:
+            ctx_length = self.hparams["max_position_embeddings"]
+        elif "model_max_length" in self.hparams:
+            ctx_length = self.hparams["model_max_length"]
+        else:
+            print("gguf: can not find ctx length parameter.")
+            sys.exit()
+
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_source_hf_repo(hf_repo)
+        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
+        self.gguf_writer.add_context_length(ctx_length)
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])
+        self.gguf_writer.add_head_count(head_count)
+        self.gguf_writer.add_head_count_kv(head_count_kv)
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
+
+        if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]:
+            if self.hparams["rope_scaling"].get("type") == "linear":
+                self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
+                self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
+
+    def write_tensors(self):
+        # Collect tensors from generator object
+        model_kv = dict(self.get_tensors())
+        block_count = self.hparams["num_hidden_layers"]
+        head_count = self.hparams["num_attention_heads"]
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        head_count_kv = self.hparams.get("num_key_value_heads", head_count)
+
+        for name, data_torch in model_kv.items():
+            # we don't need these
+            if name.endswith(".rotary_emb.inv_freq"):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            # HF models permute some of the tensors, so we need to undo that
+            if name.endswith(("q_proj.weight")):
+                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)
+            if name.endswith(("k_proj.weight")):
+                data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+            self.gguf_writer.add_tensor(new_name, data)
+
+    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("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
     model_arch = gguf.MODEL_ARCH.FALCON
@@ -1043,13 +1211,318 @@ class StableLMModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
 
 
-@Model.register("MixtralForCausalLM")
-class MixtralModel(Model):
+@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
+class LlamaModel(Model):
     model_arch = gguf.MODEL_ARCH.LLAMA
 
+    def set_vocab(self):
+        try:
+            self. _set_vocab_sentencepiece()
+        except FileNotFoundError:
+            self._set_vocab_llama_hf()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hparams = self.hparams
+        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
+        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
+
+    # Same as super class, but permuting q_proj, k_proj
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        n_head = self.hparams.get("num_attention_heads")
+        n_kv_head = self.hparams.get("num_key_value_heads")
+        n_experts = self.hparams.get("num_local_experts")
+        experts = dict()
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.numpy()
+
+            if name.endswith("q_proj.weight"):
+                data = permute(data, n_head, n_head)
+            if name.endswith("k_proj.weight"):
+                data = permute(data, n_head, n_kv_head)
+
+            data = data.squeeze()
+
+            # process the experts separately
+            if name.find("block_sparse_moe.experts") != -1:
+                experts[name] = data
+                if len(experts) >= n_experts:
+                    # merge the experts into a single 3d tensor
+                    for bid in range(block_count):
+                        for wid in range(1, 4):
+                            full = True
+                            for xid in range(n_experts):
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.w{wid}.weight"
+                                if ename not in experts:
+                                    full = False
+                                    break
+                            if not full:
+                                continue
+
+                            datas = []
+                            for xid in range(n_experts):
+                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.w{wid}.weight"
+                                datas.append(experts[ename])
+                                del experts[ename]
+
+                            data = np.stack(datas, axis=0)
+                            data_dtype = data.dtype
+
+                            if self.ftype == 0 and data_dtype == np.float16:
+                                data = data.astype(np.float32)
+
+                            if self.ftype == 1 and data_dtype == np.float32:
+                                data = data.astype(np.float16)
+
+                            merged_name = f"layers.{bid}.feed_forward.experts.w{wid}.weight"
+
+                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
+                            if new_name is None:
+                                print(f"Can not map tensor {name!r}")
+                                sys.exit()
+
+                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
+
+                            self.gguf_writer.add_tensor(new_name, data)
+                continue
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # 1d tensors need to be converted to float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+        if len(experts) > 0:
+            raise ValueError(f"Unprocessed experts: {experts.keys()}")
+
+
+@Model.register("GrokForCausalLM")
+class GrokModel(Model):
+    model_arch = gguf.MODEL_ARCH.GROK
+
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        self.gguf_writer.add_name("Grok")
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        n_experts = self.hparams.get("num_local_experts")
+        experts = dict()
+        for name, data_torch in self.get_tensors():
+            # we don't need these
+            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
+                continue
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # process the experts separately
+            if name.find(".moe.") != -1:
+                experts[name] = data
+                if len(experts) >= n_experts:
+                    # merge the experts into a single 3d tensor
+                    for bid in range(block_count):
+                        for wid in ["linear", "linear_1", "linear_v"]:
+                            full = True
+                            for xid in range(n_experts):
+                                ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
+                                if ename not in experts:
+                                    full = False
+                                    break
+                            if not full:
+                                continue
+
+                            datas = []
+                            for xid in range(n_experts):
+                                ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid}.weight"
+                                datas.append(experts[ename])
+                                del experts[ename]
+
+                            data = np.stack(datas, axis=0)
+                            data_dtype = data.dtype
+
+                            if self.ftype == 0 and data_dtype == np.float16:
+                                data = data.astype(np.float32)
+
+                            if self.ftype == 1 and data_dtype == np.float32:
+                                data = data.astype(np.float16)
+
+                            merged_name = f"transformer.decoder_layer.{bid}.moe.{wid}.weight"
+
+                            new_name = tensor_map.get_name(merged_name, try_suffixes=(".weight", ".bias"))
+                            if new_name is None:
+                                print(f"Can not map tensor {name!r}")
+                                sys.exit()
+
+                            print(f"{new_name}, n_dims = {len(data.shape)}, shape = {data.shape} --> {data.dtype}")
+
+                            self.gguf_writer.add_tensor(new_name, data)
+                continue
+
+            # map tensor names
+            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
+            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
+
+@Model.register("DbrxForCausalLM")
+class DbrxModel(Model):
+    model_arch = gguf.MODEL_ARCH.DBRX
+
+    def set_gguf_parameters(self):
+        ffn_config = self.hparams["ffn_config"]
+        attn_config = self.hparams["attn_config"]
+        self.gguf_writer.add_name(self.hparams["model_type"])
+        self.gguf_writer.add_block_count(self.hparams["n_layers"])
+
+        self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
+        self.gguf_writer.add_embedding_length(self.hparams["d_model"])
+        self.gguf_writer.add_feed_forward_length(ffn_config["ffn_hidden_size"])
+
+        self.gguf_writer.add_head_count(self.hparams["n_heads"])
+        self.gguf_writer.add_head_count_kv(attn_config["kv_n_heads"])
+
+        self.gguf_writer.add_rope_freq_base(attn_config["rope_theta"])
+
+        self.gguf_writer.add_clamp_kqv(attn_config["clip_qkv"])
+        self.gguf_writer.add_file_type(self.ftype)
+
+        self.gguf_writer.add_expert_count(ffn_config["moe_num_experts"])
+        self.gguf_writer.add_expert_used_count(ffn_config["moe_top_k"])
+
+        self.gguf_writer.add_layer_norm_eps(1e-5)
+
+        self.gguf_writer.add_file_type(self.ftype)
+        print(f"gguf: file type = {self.ftype}")
+
+    def write_tensors(self):
+        block_count = self.hparams.get("n_layers")
+        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
+        for name, data_torch in self.get_tensors():
+            n_expert = self.hparams["ffn_config"]["moe_num_experts"]
+            n_ff = self.hparams["ffn_config"]["ffn_hidden_size"]
+            n_embd = self.hparams["d_model"]
+
+            # Specific behavior for experts tensors: suffix .weight, view as 3D and transpose
+            # original implementation expects (n_expert, n_ff, n_embd) for all experts weights
+            # But llama.cpp moe graph works differently
+            # AND the dimensions in ggml are typically in the reverse order of the pytorch dimensions
+            # so (n_expert, n_ff, n_embd) in pytorch is {n_embd, n_ff, n_expert} in ggml_tensor
+            exp_tensor_names = {"ffn.experts.mlp.w1": None,       # LLM_TENSOR_FFN_GATE_EXPS ggml_tensor->ne{n_embd, n_ff,   n_expert}
+                                "ffn.experts.mlp.w2": (0, 2, 1),  # LLM_TENSOR_FFN_DOWN_EXPS ggml_tensor->ne{n_ff,   n_embd, n_expert}
+                                "ffn.experts.mlp.v1": None}       # LLM_TENSOR_FFN_UP_EXPS   ggml_tensor->ne{n_embd, n_ff,   n_expert}
+            experts = False
+            for exp_tensor_name in exp_tensor_names.keys():
+                if name.find(exp_tensor_name) != -1 and name.find(".weight") == -1:
+                    experts = True
+                    data_torch = data_torch.view(n_expert, n_ff, n_embd)
+                    if (permute_tensor := exp_tensor_names[exp_tensor_name]) is not None:
+                        data_torch = data_torch.permute(*permute_tensor)
+                    break
+
+            old_dtype = data_torch.dtype
+
+            # convert any unsupported data types to float32
+            if data_torch.dtype not in (torch.float16, torch.float32):
+                data_torch = data_torch.to(torch.float32)
+
+            data = data_torch.squeeze().numpy()
+
+            # map tensor names
+            # In MoE models the ffn tensors are typically most of the model weights,
+            # and need to be quantizable. Quantize expects tensor names to be suffixed by .weight.
+            # Every other model has the weight names ending in .weight,
+            # let's assume that is the convention which is not the case for dbrx:
+            # https://huggingface.co/databricks/dbrx-instruct/blob/main/model.safetensors.index.json#L15
+            new_name = tensor_map.get_name(name if not experts else name + ".weight", try_suffixes=(".weight",))
+            if new_name is None:
+                print(f"Can not map tensor {name!r}")
+                sys.exit()
+
+            n_dims = len(data.shape)
+            data_dtype = data.dtype
+
+            # Most of the codebase that takes in 1D tensors only handles F32 tensors
+            # and most of the outputs tensors are F32.
+            if data_dtype != np.float32 and n_dims == 1:
+                print(f"Can not map tensor {name!r}: all 1D tensors must be F32")
+                sys.exit()
+
+            # if f32 desired, convert any float16 to float32
+            if self.ftype == 0 and data_dtype == np.float16:
+                data = data.astype(np.float32)
+
+            # if f16 desired, convert any float32 2-dim weight tensors to float16
+            if self.ftype == 1 and data_dtype == np.float32 and n_dims > 1:
+                data = data.astype(np.float16)
+
+            print(f"{new_name}, n_dims = {n_dims}, shape = {data.shape}, {old_dtype} --> {data.dtype}")
+
+            self.gguf_writer.add_tensor(new_name, data)
+
 
 @Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
@@ -1069,7 +1542,7 @@ class MiniCPMModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
     def set_vocab(self):
-        self._set_vocab_hf()
+        self._set_vocab_llama_hf()
 
     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:
@@ -1670,37 +2143,25 @@ class BertModel(Model):
             self.gguf_writer.add_pooling_type(pooling_type)
 
     def set_vocab(self):
-        path = self.dir_model
-        added_tokens_path = self.dir_model if self.dir_model.exists() else None
-
-        # use huggingface vocab to get all tokens
-        vocab = HfVocab(path, added_tokens_path)
-        tokens, scores, toktypes = zip(*vocab.all_tokens())
-        assert len(tokens) == vocab.vocab_size
-        self.vocab_size = vocab.vocab_size
+        tokens, toktypes = self.get_basic_vocab()
+        self.vocab_size = len(tokens)
 
         # we need this to validate the size of the token_type embeddings
         # though currently we are passing all zeros to the token_type embeddings
-        n_token_types = len(set(toktypes))
-        self.gguf_writer.add_token_type_count(n_token_types)
+        self.gguf_writer.add_token_type_count(2)  # "Sequence A" or "Sequence B"
 
         # convert to phantom space vocab
-        def phantom(tok, typ):
-            if tok.startswith(b"[") and tok.endswith(b"]"):
+        def phantom(tok):
+            if tok.startswith("[") and tok.endswith("]"):
                 return tok
-            if tok.startswith(b"##"):
+            if tok.startswith("##"):
                 return tok[2:]
-            return b"\xe2\x96\x81" + tok
-        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
-
-        # set up bos and eos tokens (cls and sep)
-        self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
-        self.gguf_writer.add_eos_token_id(vocab.tokenizer.sep_token_id)
+            return "\u2581" + tok
+        tokens = list(map(phantom, tokens))
 
         # add vocab to gguf
         self.gguf_writer.add_tokenizer_model("bert")
         self.gguf_writer.add_token_list(tokens)
-        self.gguf_writer.add_token_scores(scores)
         self.gguf_writer.add_token_types(toktypes)
 
         # handle special tokens
@@ -1772,16 +2233,6 @@ class NomicBertModel(BertModel):
         super().set_gguf_parameters()
         self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
 
-    def get_tensors(self):
-        assert self.vocab_size is not None
-        for name, data in super().get_tensors():
-            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
-            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
-                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
-                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
-                data = data[:self.vocab_size, :]
-            yield name, data
-
 
 @Model.register("GemmaForCausalLM")
 class GemmaModel(Model):
@@ -1957,7 +2408,8 @@ class MambaModel(Model):
                 data = data.astype(np.float32)
 
             # if f16 desired, convert big float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
+            new_weight_name = new_name[:-len(".weight")] if new_name.endswith(".weight") else ""
+            if self.ftype == 1 and data_dtype == np.float32 and new_weight_name.endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
                 data = data.astype(np.float16)
 
             print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
@@ -2008,6 +2460,7 @@ def parse_args() -> argparse.Namespace:
         "model", type=Path,
         help="directory containing model file",
     )
+    parser.add_argument("--use-temp-file", action="store_true", help="use the tempfile library while processing (helpful when running out of memory, process killed)")
 
     return parser.parse_args()
 
@@ -2051,7 +2504,7 @@ def main() -> None:
 
     with torch.inference_mode():
         model_class = Model.from_model_architecture(hparams["architectures"][0])
-        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+        model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian, args.use_temp_file)
 
         print("Set model parameters")
         model_instance.set_gguf_parameters()

convert-persimmon-to-gguf.py

@@ -1,4 +1,6 @@
 #!/usr/bin/env python3
+from __future__ import annotations
+
 import argparse
 import os
 import sys
@@ -106,12 +108,12 @@ def main():
     tensor_map = gguf.get_tensor_name_map(arch, block_count)
     print(tensor_map)
     for name in tensors.keys():
-        data = tensors[name]
+        data_torch = tensors[name]
         if name.endswith(".self_attention.rotary_emb.inv_freq"):
             continue
-        old_dtype = data.dtype
+        old_dtype = data_torch.dtype
         # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data.to(torch.float32).squeeze().numpy()
+        data = data_torch.to(torch.float32).squeeze().numpy()
         new_name = tensor_map.get_name(name, try_suffixes = (".weight", ".bias"))
         if new_name is None:
             print("Can not map tensor '" + name + "'")

convert.py

@@ -16,13 +16,14 @@ import re
 import signal
 import struct
 import sys
+import textwrap
 import time
 import zipfile
-from abc import ABCMeta, abstractmethod
+from abc import ABC, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
+from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -32,7 +33,7 @@ if 'NO_LOCAL_GGUF' not in os.environ:
 import gguf
 
 if TYPE_CHECKING:
-    from typing import TypeAlias
+    from typing_extensions import Self, TypeAlias
 
 if hasattr(faulthandler, 'register') and hasattr(signal, 'SIGUSR1'):
     faulthandler.register(signal.SIGUSR1)
@@ -43,6 +44,9 @@ ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
 
+ADDED_TOKENS_FILE = 'added_tokens.json'
+FAST_TOKENIZER_FILE = 'tokenizer.json'
+
 #
 # data types
 #
@@ -135,7 +139,8 @@ class GGMLFileType(enum.IntEnum):
         dt = GGML_FILE_TYPE_TO_DATA_TYPE.get(self)
         if dt is None:
             raise ValueError(self)
-        # 1D tensors are always F32.
+        # Convert all 1D tensors to F32.  Most of the codebase that takes in 1D tensors only handles F32 tensors, and most of the outputs tensors are F32.
+        #  Also The 1d tensors aren't much of a performance/size issue.  So instead of having to have separate F32 and F16 implementations of both, just convert everything to F32 for now.
         return dt if len(tensor.shape) > 1 else DT_F32
 
 
@@ -188,8 +193,10 @@ class Params:
             n_layer = next(i for i in itertools.count() if f"layers.{i}.attention.wq.weight" not in model)
 
         if n_layer < 1:
-            raise Exception("failed to guess 'n_layer'. This model is unknown or unsupported.\n"
-                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
+            msg = """\
+                failed to guess 'n_layer'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
 
         n_head = n_embd // 128 # guessed
         n_mult = 256           # guessed
@@ -211,7 +218,8 @@ class Params:
 
     @staticmethod
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
-        config = json.load(open(config_path))
+        with open(config_path) as f:
+            config = json.load(f)
 
         rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
@@ -233,8 +241,10 @@ class Params:
         elif "max_position_embeddings" in config:
             n_ctx = config["max_position_embeddings"]
         else:
-            raise Exception("failed to guess 'n_ctx'. This model is unknown or unsupported.\n"
-                            "Suggestion: provide 'config.json' of the model in the same directory containing model files.")
+            msg = """\
+                failed to guess 'n_ctx'. This model is unknown or unsupported.
+                Suggestion: provide 'config.json' of the model in the same directory containing model files."""
+            raise KeyError(textwrap.dedent(msg))
 
         n_experts      = None
         n_experts_used = None
@@ -265,7 +275,8 @@ class Params:
     # {"dim": 8192, "multiple_of": 4096, "ffn_dim_multiplier": 1.3, "n_heads": 64, "n_kv_heads": 8, "n_layers": 80, "norm_eps": 1e-05, "vocab_size": -1}
     @staticmethod
     def loadOriginalParamsJson(model: LazyModel, config_path: Path) -> Params:
-        config = json.load(open(config_path))
+        with open(config_path) as f:
+            config = json.load(f)
 
         n_experts      = None
         n_experts_used = None
@@ -331,47 +342,86 @@ class Params:
 # vocab
 #
 
-class BpeVocab:
+@runtime_checkable
+class BaseVocab(Protocol):
+    tokenizer_model: ClassVar[str]
+    name: ClassVar[str]
+
+
+class NoVocab(BaseVocab):
+    tokenizer_model = "no_vocab"
+    name = "no_vocab"
+
+    def __repr__(self) -> str:
+        return "<NoVocab for a model without integrated vocabulary>"
+
+
+@runtime_checkable
+class Vocab(BaseVocab, Protocol):
+    vocab_size: int
+    added_tokens_dict: dict[str, int]
+    added_tokens_list: list[str]
+    fname_tokenizer: Path
+
+    def __init__(self, base_path: Path): ...
+    def all_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]: ...
+
+
+class BpeVocab(Vocab):
     tokenizer_model = "gpt2"
     name = "bpe"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.bpe_tokenizer = json.loads(open(str(fname_tokenizer), encoding="utf-8").read())
-        if isinstance(self.bpe_tokenizer.get('model'), dict):
-            self.vocab = self.bpe_tokenizer["model"]["vocab"]
-        else:
-            self.vocab = self.bpe_tokenizer
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
-        else:
-            # Fall back to trying to find the added tokens in tokenizer.json
-            tokenizer_json_file = fname_tokenizer.parent / 'tokenizer.json'
-            if not tokenizer_json_file.is_file():
-                added_tokens = {}
-            else:
-                tokenizer_json = json.load(open(tokenizer_json_file, encoding="utf-8"))
-                added_tokens = dict(
-                    (item['content'], item['id'])
-                    for item in tokenizer_json.get('added_tokens', [])
-                    # Added tokens here can be duplicates of the main vocabulary.
-                    if item['content'] not in self.bpe_tokenizer)
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
 
-        vocab_size: int = len(self.vocab)
-        expected_ids    = list(range(vocab_size, vocab_size + len(added_tokens)))
-        actual_ids      = sorted(added_tokens.values())
+        if (fname_tokenizer := base_path / 'vocab.json').exists():
+            # "slow" tokenizer
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                self.vocab = json.load(f)
+
+            try:
+                # FIXME: Verify that added tokens here _cannot_ overlap with the main vocab.
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        else:
+            # "fast" tokenizer
+            fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+
+            # if this fails, FileNotFoundError propagates to caller
+            with open(fname_tokenizer, encoding="utf-8") as f:
+                tokenizer_json = json.load(f)
+
+            tokenizer_model: dict[str, Any] = tokenizer_json['model']
+            if (
+                tokenizer_model['type'] != 'BPE' or tokenizer_model.get('byte_fallback', False)
+                or tokenizer_json['decoder']['type'] != 'ByteLevel'
+            ):
+                raise FileNotFoundError('Cannot find GPT-2 BPE tokenizer')
+
+            self.vocab = tokenizer_model["vocab"]
+
+            if (added := tokenizer_json.get('added_tokens')) is not None:
+                # Added tokens here can be duplicates of the main vocabulary.
+                added_tokens = {item['content']: item['id']
+                                for item in added
+                                if item['content'] not in self.vocab}
+
+        vocab_size   = len(self.vocab)
+        expected_ids = list(range(vocab_size, vocab_size + len(added_tokens)))
+        actual_ids   = sorted(added_tokens.values())
         if expected_ids != actual_ids:
             expected_end_id = vocab_size + len(actual_ids) - 1
-            raise Exception(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range {vocab_size} - {expected_end_id}; got {actual_ids}")
+            raise ValueError(f"Expected the {len(actual_ids)} added token ID(s) to be sequential in the range "
+                             f"{vocab_size} - {expected_end_id}; got {actual_ids}")
 
         items = sorted(added_tokens.items(), key=lambda text_idx: text_idx[1])
         self.added_tokens_dict    = added_tokens
         self.added_tokens_list    = [text for (text, idx) in items]
-        self.vocab_size_base: int = vocab_size
-        self.vocab_size: int      = self.vocab_size_base + len(self.added_tokens_list)
+        self.vocab_size_base      = vocab_size
+        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
         self.fname_tokenizer      = fname_tokenizer
-        self.fname_added_tokens   = fname_added_tokens
 
     def bpe_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {id: encoded_tok for encoded_tok, id in self.vocab.items()}
@@ -392,19 +442,25 @@ class BpeVocab:
         return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class SentencePieceVocab:
+class SentencePieceVocab(Vocab):
     tokenizer_model = "llama"
     name = "spm"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
-        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        added_tokens: dict[str, int]
-        if fname_added_tokens is not None:
-            added_tokens = json.load(open(fname_added_tokens, encoding="utf-8"))
-        else:
-            added_tokens = {}
+    def __init__(self, base_path: Path):
+        added_tokens: dict[str, int] = {}
+        if (fname_tokenizer := base_path / 'tokenizer.model').exists():
+            # normal location
+            try:
+                with open(base_path / ADDED_TOKENS_FILE, encoding="utf-8") as f:
+                    added_tokens = json.load(f)
+            except FileNotFoundError:
+                pass
+        elif not (fname_tokenizer := base_path.parent / 'tokenizer.model').exists():
+            # not found in alternate location either
+            raise FileNotFoundError('Cannot find tokenizer.model')
 
-        vocab_size: int = self.sentencepiece_tokenizer.vocab_size()
+        self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
+        vocab_size = self.sentencepiece_tokenizer.vocab_size()
 
         new_tokens       = {id: piece for piece, id in added_tokens.items() if id >= vocab_size}
         expected_new_ids = list(range(vocab_size, vocab_size + len(new_tokens)))
@@ -414,18 +470,17 @@ class SentencePieceVocab:
             raise ValueError(f"Expected new token IDs {expected_new_ids} to be sequential; got {actual_new_ids}")
 
         # Token pieces that were added to the base vocabulary.
-        self.added_tokens_dict = added_tokens
+        self.added_tokens_dict  = added_tokens
         self.added_tokens_list  = [new_tokens[id] for id in actual_new_ids]
         self.vocab_size_base    = vocab_size
         self.vocab_size         = self.vocab_size_base + len(self.added_tokens_list)
         self.fname_tokenizer    = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
 
     def sentencepiece_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         tokenizer = self.sentencepiece_tokenizer
         for i in range(tokenizer.vocab_size()):
             piece = tokenizer.id_to_piece(i)
-            text: bytes = piece.encode("utf-8")
+            text         = piece.encode("utf-8")
             score: float = tokenizer.get_score(i)
 
             toktype = gguf.TokenType.NORMAL
@@ -458,27 +513,40 @@ class SentencePieceVocab:
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class HfVocab:
+class LlamaHfVocab(Vocab):
     tokenizer_model = "llama"
     name = "hfft"
 
-    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
+    def __init__(self, base_path: Path):
+        fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+        # if this fails, FileNotFoundError propagates to caller
+        with open(fname_tokenizer, encoding='utf-8') as f:
+            tokenizer_json = json.load(f)
+
+        # pre-check so we know if we need transformers
+        tokenizer_model: dict[str, Any] = tokenizer_json['model']
+        if (
+            tokenizer_model['type'] != 'BPE' or not tokenizer_model.get('byte_fallback', False)
+            or tokenizer_json['decoder']['type'] != 'Sequence'
+        ):
+            raise FileNotFoundError('Cannot find Llama BPE tokenizer')
+
         try:
             from transformers import AutoTokenizer
         except ImportError as e:
             raise ImportError(
-                "To use HfVocab, please install the `transformers` package. "
+                "To use LlamaHfVocab, please install the `transformers` package. "
                 "You can install it with `pip install transformers`."
             ) from e
 
-        print("fname_tokenizer:", fname_tokenizer)
         # Allow the tokenizer to default to slow or fast versions.
         # Explicitly set tokenizer to use local paths.
         self.tokenizer = AutoTokenizer.from_pretrained(
-            fname_tokenizer,
-            cache_dir=fname_tokenizer,
+            base_path,
+            cache_dir=base_path,
             local_files_only=True,
         )
+        assert self.tokenizer.is_fast  # assume tokenizer.json is used
 
         # Initialize lists and dictionaries for added tokens
         self.added_tokens_list = []
@@ -506,8 +574,7 @@ class HfVocab:
         self.vocab_size_base = self.tokenizer.vocab_size
         self.vocab_size      = self.vocab_size_base + len(self.added_tokens_list)
 
-        self.fname_tokenizer    = fname_tokenizer
-        self.fname_added_tokens = fname_added_tokens
+        self.fname_tokenizer = fname_tokenizer
 
     def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {
@@ -559,18 +626,7 @@ class HfVocab:
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
-
-
-class NoVocab:
-    tokenizer_model = "no_vocab"
-    name = "no_vocab"
-
-    def __repr__(self) -> str:
-        return "<NoVocab for a model without integrated vocabulary>"
-
-
-Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab | NoVocab"
+        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
 #
@@ -588,17 +644,18 @@ def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
             .reshape(weights.shape))
 
 
-class Tensor(metaclass=ABCMeta):
+class Tensor(ABC):
+    ndarray: NDArray
     data_type: DataType
 
     @abstractmethod
-    def astype(self, data_type: DataType) -> Tensor: ...
+    def astype(self, data_type: DataType) -> Self: ...
     @abstractmethod
-    def permute(self, n_head: int, n_head_kv: int) -> Tensor: ...
+    def permute(self, n_head: int, n_head_kv: int) -> Self: ...
     @abstractmethod
-    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor: ...
+    def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> Self: ...
     @abstractmethod
-    def part(self, n_part: int) -> UnquantizedTensor: ...
+    def part(self, n_part: int) -> Self: ...
     @abstractmethod
     def to_ggml(self) -> GGMLCompatibleTensor: ...
 
@@ -610,18 +667,18 @@ def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
 
 
 class UnquantizedTensor(Tensor):
-    def __init__(self, ndarray: NDArray) -> None:
+    def __init__(self, ndarray: NDArray):
         assert isinstance(ndarray, np.ndarray)
         self.ndarray = ndarray
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
 
-    def astype(self, data_type: DataType) -> Tensor:
+    def astype(self, data_type: DataType) -> UnquantizedTensor:
         dtype = data_type.dtype
         if self.data_type == DT_BF16:
             self.ndarray = bf16_to_fp32(self.ndarray)
         return UnquantizedTensor(self.ndarray.astype(dtype))
 
-    def to_ggml(self) -> UnquantizedTensor:
+    def to_ggml(self) -> Self:
         return self
 
     def permute_part(self, n_part: int, n_head: int, n_head_kv: int) -> UnquantizedTensor:
@@ -689,7 +746,7 @@ class ModelPlus:
     model: LazyModel
     paths: list[Path]  # Where this was read from.
     format: Literal['ggml', 'torch', 'safetensors', 'none']
-    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
+    vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
 def merge_sharded(models: list[LazyModel]) -> LazyModel:
@@ -698,7 +755,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
     names = {name: None for model in models for name in model}
 
     def convert(name: str) -> LazyTensor:
-        lazy_tensors: list[LazyTensor] = [model[name] for model in models]
+        lazy_tensors = [model[name] for model in models]
         if len(lazy_tensors) == 1:
             # only one file; don't go through this procedure since there might
             # be quantized tensors
@@ -719,7 +776,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
 
         def load() -> UnquantizedTensor:
             ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors]
-            concatenated: NDArray = np.concatenate(ndarrays, axis=axis)
+            concatenated = np.concatenate(ndarrays, axis=axis)
             return UnquantizedTensor(concatenated)
         description = 'concatenated[[' + '] | ['.join(lt.description for lt in lazy_tensors) + ']]'
         return LazyTensor(load, concatenated_shape, lazy_tensors[0].data_type, description)
@@ -771,6 +828,15 @@ def part_lazy(lazy_tensor: LazyTensor, n_part: int) -> LazyTensor:
     return LazyTensor(load, s, lazy_tensor.data_type, 'part ' + lazy_tensor.description)
 
 
+def pack_experts_lazy(lazy_tensors: list[LazyTensor]) -> LazyTensor:
+    def load() -> Tensor:
+        tensors = [lazy_tensor.load() for lazy_tensor in lazy_tensors]
+        return UnquantizedTensor(np.array([tensor.ndarray for tensor in tensors]))
+    s = lazy_tensors[0].shape.copy()
+    s.insert(0, len(lazy_tensors))
+    return LazyTensor(load, s, lazy_tensors[0].data_type, 'pack_experts ' + ' | '.join(lt.description for lt in lazy_tensors))
+
+
 # Functionality that simulates `torch.load` but where individual tensors are
 # only loaded into memory on demand, not all at once.
 # PyTorch can't do this natively as of time of writing:
@@ -807,10 +873,10 @@ class LazyUnpickler(pickle.Unpickler):
 
         def load(offset: int, elm_count: int) -> NDArray:
             dtype = data_type.dtype
-            fp = self.zip_file.open(info)
-            fp.seek(offset * dtype.itemsize)
-            size = elm_count * dtype.itemsize
-            data = fp.read(size)
+            with self.zip_file.open(info) as fp:
+                fp.seek(offset * dtype.itemsize)
+                size = elm_count * dtype.itemsize
+                data = fp.read(size)
             assert len(data) == size
             return np.frombuffer(data, dtype)
         description = f'storage data_type={data_type} path-in-zip={filename} path={self.zip_file.filename}'
@@ -831,7 +897,7 @@ class LazyUnpickler(pickle.Unpickler):
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES: dict[tuple[str, str], Any] = {
+    CLASSES = {
         # getattr used here as a workaround for mypy not being smart enough to determine
         # the staticmethods have a __func__ attribute.
         ('torch._tensor', '_rebuild_from_type_v2'): getattr(rebuild_from_type_v2, '__func__'),
@@ -890,7 +956,7 @@ def lazy_load_safetensors_file(fp: IO[bytes], path: Path) -> ModelPlus:
 def must_read(fp: IO[bytes], length: int) -> bytes:
     ret = fp.read(length)
     if len(ret) < length:
-        raise Exception("unexpectedly reached end of file")
+        raise EOFError("unexpectedly reached end of file")
     return ret
 
 
@@ -948,13 +1014,14 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
             yield result
 
 
-def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
+def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) -> None:
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            f"The model's vocab size is set to -1 in params.json. Please update it manually.{f' Maybe {vocab.vocab_size}?' if hasattr(vocab, 'vocab_size') else ''}"
+            "The model's vocab size is set to -1 in params.json. Please update it manually."
+            + (f" Maybe {vocab.vocab_size}?" if isinstance(vocab, Vocab) else ""),
         )
-    if isinstance(vocab, NoVocab):
+    if not isinstance(vocab, Vocab):
         return  # model has no vocab
 
     # Check for a vocab size mismatch
@@ -979,11 +1046,11 @@ def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> N
     if vocab.vocab_size < params.n_vocab:
         msg += " Add the --pad-vocab option and try again."
 
-    raise Exception(msg)
+    raise ValueError(msg)
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
@@ -1034,8 +1101,6 @@ class OutputFile:
             self.gguf.add_file_type(params.ftype)
 
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
-        assert not isinstance(vocab, NoVocab)
-
         tokens = []
         scores = []
         toktypes = []
@@ -1135,7 +1200,7 @@ class OutputFile:
 
     @staticmethod
     def write_all(
-        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
     ) -> None:
@@ -1145,11 +1210,11 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        if isinstance(vocab, NoVocab):
-            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
-        else:
+        if isinstance(vocab, Vocab):
             of.add_meta_vocab(vocab)
             of.add_meta_special_vocab(svocab)
+        else:  # NoVocab
+            of.gguf.add_tokenizer_model(vocab.tokenizer_model)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1176,7 +1241,7 @@ def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileT
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
-    raise Exception(f"Unexpected combination of types: {name_to_type}")
+    raise ValueError(f"Unexpected combination of types: {name_to_type}")
 
 
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1186,10 +1251,26 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
 
 def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
-    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
+    should_skip = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
+    # merge experts into one tensor
+    if params.n_experts and params.n_experts > 0:
+        for i_l in range(params.n_layer):
+            for w in range(1, 4):
+                experts = []
+                for e in range(params.n_experts):
+                    if f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"])
+                        del tmp[f"layers.{i_l}.feed_forward.experts.{e}.w{w}.weight"]
+                    elif f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight" in model:
+                        experts.append(model[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"])
+                        del tmp[f"model.layers.{i_l}.block_sparse_moe.experts.{e}.w{w}.weight"]
+                    else:
+                        raise ValueError(f"Expert tensor not found: layers.{i_l}.feed_forward.experts.{e}.w{w}.weight")
+                tmp[f"layers.{i_l}.feed_forward.experts.w{w}.weight"] = pack_experts_lazy(experts)
+
     # HF models permut or pack some of the tensors, so we need to undo that
     for i in itertools.count():
         if f"model.layers.{i}.self_attn.q_proj.weight" in model:
@@ -1213,8 +1294,7 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) ->
             if skip_unknown:
                 print(f"Unexpected tensor name: {name} - skipping")
                 continue
-            else:
-                raise Exception(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
+            raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1231,7 +1311,7 @@ def nth_multifile_path(path: Path, n: int) -> Path | None:
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns: list[tuple[str, str]] = [
+    patterns = [
         # - x.00.pth, x.01.pth, etc.
         (r'\.[0-9]{2}\.pth$', f'.{n:02}.pth'),
         # - x-00001-of-00002.bin, x-00002-of-00002.bin, etc.
@@ -1270,16 +1350,16 @@ def load_some_model(path: Path) -> ModelPlus:
     # Be extra-friendly and accept either a file or a directory:
     if path.is_dir():
         # Check if it's a set of safetensors files first
-        globs = ["model-00001-of-*.safetensors", "model.safetensors"]
+        globs = ["model-00001-of-*.safetensors", "model.safetensors", "consolidated.safetensors"]
         files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try the PyTorch patterns too, with lower priority
             globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin"]
             files = [file for glob in globs for file in path.glob(glob)]
         if not files:
-            raise Exception(f"Can't find model in directory {path}")
+            raise FileNotFoundError(f"Can't find model in directory {path}")
         if len(files) > 1:
-            raise Exception(f"Found multiple models in {path}, not sure which to pick: {files}")
+            raise ValueError(f"Found multiple models in {path}, not sure which to pick: {files}")
         path = files[0]
 
     paths = find_multifile_paths(path)
@@ -1293,36 +1373,14 @@ def load_some_model(path: Path) -> ModelPlus:
 
 
 class VocabFactory:
-    _FILES = {"spm": "tokenizer.model", "bpe": "vocab.json", "hfft": "tokenizer.json"}
+    _VOCAB_CLASSES: list[type[Vocab]] = [SentencePieceVocab, BpeVocab, LlamaHfVocab]
 
     def __init__(self, path: Path):
         self.path = path
-        self.file_paths = self._detect_files()
-        print(f"Found vocab files: {self.file_paths}")
 
-    def _detect_files(self) -> dict[str, Path | None]:
-        def locate(file: str) -> Path | None:
-            if (path := self.path / file).exists():
-                return path
-            if (path := self.path.parent / file).exists():
-                return path
-            return None
-
-        return {vt: locate(f) for vt, f in self._FILES.items()}
-
-    def _select_file(self, vocab_types: list[str]) -> tuple[str, Path]:
-        for vtype in vocab_types:
-            try:
-                path = self.file_paths[vtype]
-            except KeyError:
-                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
-            if path is not None:
-                return vtype, path
-        raise FileNotFoundError(f"Could not find any of {[self._FILES[vt] for vt in vocab_types]}")
-
-    def _create_special_vocab(self, vocab: Vocab, model_parent_path: Path) -> gguf.SpecialVocab:
+    def _create_special_vocab(self, vocab: BaseVocab, model_parent_path: Path) -> gguf.SpecialVocab:
         load_merges = vocab.name == "bpe"
-        n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
+        n_vocab = vocab.vocab_size if isinstance(vocab, Vocab) else None
         return gguf.SpecialVocab(
             model_parent_path,
             load_merges=load_merges,
@@ -1331,27 +1389,29 @@ class VocabFactory:
         )
 
     def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
-        vocab_type, path = self._select_file(vocab_types)
-        print(f"Loading vocab file {path!r}, type {vocab_type!r}")
+        vocab_classes: dict[str, type[Vocab]] = {cls.name: cls for cls in self._VOCAB_CLASSES}
+        selected_vocabs: dict[str, type[Vocab]] = {}
+        for vtype in vocab_types:
+            try:
+                selected_vocabs[vtype] = vocab_classes[vtype]
+            except KeyError:
+                raise ValueError(f"Unsupported vocabulary type {vtype}") from None
 
-        added_tokens_path = path.parent / "added_tokens.json"
-        if vocab_type == "bpe":
-            return BpeVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        if vocab_type == "spm":
-            return SentencePieceVocab(
-                path, added_tokens_path if added_tokens_path.exists() else None
-            )
-        if vocab_type == "hfft":
-            return HfVocab(
-                path.parent, added_tokens_path if added_tokens_path.exists() else None
-            )
-        raise ValueError(vocab_type)
+        for vtype, cls in selected_vocabs.items():
+            try:
+                vocab = cls(self.path)
+                break
+            except FileNotFoundError:
+                pass  # ignore unavailable tokenizers
+        else:
+            raise FileNotFoundError(f"Could not find a tokenizer matching any of {vocab_types}")
 
-    def load_vocab(self, vocab_types: list[str], model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
-        vocab: Vocab
-        if len(vocab_types) == 1 and "no_vocab" in vocab_types:
+        print(f"Loaded vocab file {vocab.fname_tokenizer!r}, type {vocab.name!r}")
+        return vocab
+
+    def load_vocab(self, vocab_types: list[str] | None, model_parent_path: Path) -> tuple[BaseVocab, gguf.SpecialVocab]:
+        vocab: BaseVocab
+        if vocab_types is None:
             vocab = NoVocab()
         else:
             vocab = self._create_vocab_by_path(vocab_types)
@@ -1408,10 +1468,8 @@ def main(args_in: list[str] | None = None) -> None:
     parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
 
     args = parser.parse_args(args_in)
-    if args.no_vocab:
-        if args.vocab_only:
-            raise ValueError("no need to specify --vocab-only if using --no-vocab")
-        args.vocab_type = "no_vocab"
+    if args.no_vocab and args.vocab_only:
+        raise ValueError("--vocab-only does not make sense with --no-vocab")
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1433,10 +1491,12 @@ def main(args_in: list[str] | None = None) -> None:
     params = Params.load(model_plus)
     if params.n_ctx == -1:
         if args.ctx is None:
-            raise Exception("The model doesn't have a context size, and you didn't specify one with --ctx\n"
-                            "Please specify one with --ctx:\n"
-                            " - LLaMA v1: --ctx 2048\n"
-                            " - LLaMA v2: --ctx 4096\n")
+            msg = """\
+                The model doesn't have a context size, and you didn't specify one with --ctx
+                Please specify one with --ctx:
+                 - LLaMA v1: --ctx 2048
+                 - LLaMA v2: --ctx 4096"""
+            parser.error(textwrap.dedent(msg))
         params.n_ctx = args.ctx
 
     if args.outtype:
@@ -1451,9 +1511,11 @@ def main(args_in: list[str] | None = None) -> None:
     model_parent_path = model_plus.paths[0].parent
     vocab_path = Path(args.vocab_dir or args.model or model_parent_path)
     vocab_factory = VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type.split(","), model_parent_path)
+    vocab_types = None if args.no_vocab else args.vocab_type.split(",")
+    vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
 
     if args.vocab_only:
+        assert isinstance(vocab, Vocab)
         if not args.outfile:
             raise ValueError("need --outfile if using --vocab-only")
         outfile = args.outfile

docs/HOWTO-add-model.md

@@ -0,0 +1,119 @@
+## Add a new model architecture to `llama.cpp`
+
+Adding a model requires few steps:
+
+1. Convert the model to GGUF
+2. Define the model architecture in `llama.cpp`
+3. Build the GGML graph implementation
+
+After following these steps, you can open PR.
+
+Also, it is important to check that the examples and main ggml backends (CUDA, METAL, CPU) are working with the new architecture, especially:
+- [main](../examples/main)
+- [imatrix](../examples/imatrix)
+- [quantize](../examples/quantize)
+- [server](../examples/server)
+
+### 1. Convert the model to GGUF
+
+This step is done in python with a `convert` script using the [gguf](https://pypi.org/project/gguf/) library.
+Depending on the model architecture, you can use either [convert.py](../convert.py) or [convert-hf-to-gguf.py](../convert-hf-to-gguf.py).
+
+The convert script reads the model configuration, tokenizer, tensor names+data and converts them to GGUF metadata and tensors.
+
+The required steps to implement for an HF model are:
+
+1. Define the model `Model.register` annotation in a new `Model` subclass, example:
+
+```python
+@Model.register("MyModelForCausalLM")
+class MyModel(Model):
+    model_arch = gguf.MODEL_ARCH.GROK
+```
+
+2. Define the layout of the GGUF tensors in [constants.py](../gguf-py/gguf/constants.py)
+
+Add an enum entry in `MODEL_ARCH`, the model human friendly name in `MODEL_ARCH_NAMES` and the GGUF tensor names in `MODEL_TENSORS`.
+
+Example for `falcon` model:
+```python
+    MODEL_ARCH.FALCON: [
+        MODEL_TENSOR.TOKEN_EMBD,
+        MODEL_TENSOR.OUTPUT_NORM,
+        MODEL_TENSOR.OUTPUT,
+        MODEL_TENSOR.ATTN_NORM,
+        MODEL_TENSOR.ATTN_NORM_2,
+        MODEL_TENSOR.ATTN_QKV,
+        MODEL_TENSOR.ATTN_OUT,
+        MODEL_TENSOR.FFN_DOWN,
+        MODEL_TENSOR.FFN_UP,
+    ]
+```
+
+3. Map the original tensor names to the standardize equivalent in GGUF
+
+As a general rule, before adding a new tensor name to GGUF, be sure the equivalent naming does not already exist.
+
+Once you have found the GGUF tensor name equivalent, add it to the [tensor_mapping.py](../gguf-py/gguf/tensor_mapping.py) file.
+
+If the tensor name is part of a repetitive layer/block, the key word `bid` substitutes it.
+
+Example for the normalization tensor in attention layers:
+
+```python
+block_mappings_cfg: dict[MODEL_TENSOR, tuple[str, ...]] = {
+        # Attention norm
+        MODEL_TENSOR.ATTN_NORM: (
+            "gpt_neox.layers.{bid}.input_layernorm",                # gptneox
+            "transformer.h.{bid}.ln_1",                             # gpt2 gpt-j refact qwen
+            "transformer.blocks.{bid}.norm_1",                      # mpt
+            ...
+        )
+}
+```
+
+`transformer.blocks.{bid}.norm_1` will be mapped to `blk.{bid}.attn_norm` in GGUF.
+
+Depending on the model configuration, tokenizer, code and tensors layout, you will have to override:
+- `Model#set_gguf_parameters`
+- `Model#set_vocab`
+- `Model#write_tensors`
+
+NOTE: Tensor names must end with `.weight` suffix, that is the convention and several tools like `quantize` expect this to proceed the weights.
+
+### 2. Define the model architecture in `llama.cpp`
+
+The model params and tensors layout must be defined in `llama.cpp`:
+1. Define a new `llm_arch`
+2. Define the tensors layout in `LLM_TENSOR_NAMES`
+3. Add any non standard metadata in `llm_load_hparams`
+4. Create the tensors for inference in `llm_load_tensors`
+5. If the model has a RoPE operation, add the rope type in `llama_rope_type`
+
+NOTE: The dimensions in `ggml` are typically in the reverse order of the `pytorch` dimensions.
+
+### 3. Build the GGML graph implementation
+
+This is the funniest part, you have to provide the inference graph implementation of the new model architecture in `llama_build_graph`.
+
+Have a look to existing implementation like `build_llama`, `build_dbrx` or `build_bert`.
+
+When implementing a new graph, please note that the underlying `ggml` backends might not support them all, support of missing backend operations can be added in another PR.
+
+Note: to debug the inference graph: you can use [eval-callback](../examples/eval-callback).
+
+## GGUF specification
+
+https://github.com/ggerganov/ggml/blob/master/docs/gguf.md
+
+## Resources
+
+- YaRN RoPE scaling https://github.com/ggerganov/llama.cpp/pull/2268
+- support Baichuan serial models https://github.com/ggerganov/llama.cpp/pull/3009
+- support attention bias https://github.com/ggerganov/llama.cpp/pull/4283
+- Mixtral support https://github.com/ggerganov/llama.cpp/pull/4406
+- BERT embeddings https://github.com/ggerganov/llama.cpp/pull/5423
+- Grok-1 support https://github.com/ggerganov/llama.cpp/pull/6204
+- Command R Plus support https://github.com/ggerganov/llama.cpp/pull/6491
+- support arch DBRX https://github.com/ggerganov/llama.cpp/pull/6515
+- How to convert HuggingFace model to GGUF format https://github.com/ggerganov/llama.cpp/discussions/2948

docs/token_generation_performance_tips.md

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

examples/CMakeLists.txt

@@ -19,6 +19,7 @@ else()
     add_subdirectory(benchmark)
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
+    add_subdirectory(eval-callback)
     add_subdirectory(finetune)
     add_subdirectory(gritlm)
     add_subdirectory(gguf-split)
@@ -34,6 +35,7 @@ else()
     add_subdirectory(perplexity)
     add_subdirectory(quantize)
     add_subdirectory(quantize-stats)
+    add_subdirectory(retrieval)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
     add_subdirectory(passkey)

examples/batched-bench/README.md

@@ -10,16 +10,16 @@ There are 2 modes of operation:
 - `prompt is shared` - there is a common prompt of size `PP` used by all batches (i.e. `N_KV = PP + B*TG`)
 
 ```bash
-./batched-bench MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
+./batched-bench MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] [MMQ] <PP> <TG> <PL>
 
 # LLaMA 7B, F16, N_KV_MAX = 16384 (8GB), prompt not shared
-./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 0 99
+./batched-bench ./models/llama-7b/ggml-model-f16.gguf 16384 2048 512 0 99
 
 # LLaMA 7B, Q8_0, N_KV_MAX = 16384 (8GB), prompt is shared
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 1 99
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 16384 2048 512 1 99
 
 # custom set of batches
-./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 0 999 0 128,256,512 128,256 1,2,4,8,16,32
+./batched-bench ./models/llama-7b/ggml-model-q8_0.gguf 2048 512 512 0 999 0 128,256,512 128,256 1,2,4,8,16,32
 ```
 
 ## Sample results

examples/batched-bench/batched-bench.cpp

@@ -32,13 +32,15 @@ int main(int argc, char ** argv) {
     gpt_params params;
 
     if (argc == 1 || argv[1][0] == '-') {
-        printf("usage: %s MODEL_PATH [N_KV_MAX] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
+        printf("usage: %s MODEL_PATH [N_KV_MAX] [N_BATCH] [N_UBATCH] [IS_PP_SHARED] [NGL] <PP> <TG> <PL>\n" , argv[0]);
         printf("  <PP>, <TG> and PL are comma-separated lists of numbers without spaces\n\n");
-        printf("  example: %s ggml-model-f16.gguf 2048 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
+        printf("  example: %s ggml-model-f16.gguf 2048 2048 512 0 999 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
     }
 
     int n_kv_max     = 2048;
+    int n_batch      = 2048;
+    int n_ubatch     = 512;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
 
@@ -56,23 +58,31 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 4) {
-        is_pp_shared = std::atoi(argv[3]);
+        n_batch = std::atoi(argv[3]);
     }
 
     if (argc >= 5) {
-        n_gpu_layers = std::atoi(argv[4]);
+        n_ubatch = std::atoi(argv[4]);
     }
 
     if (argc >= 6) {
-        n_pp = parse_list(argv[5]);
+        is_pp_shared = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_tg = parse_list(argv[6]);
+        n_gpu_layers = std::atoi(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pl = parse_list(argv[7]);
+        n_pp = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_tg = parse_list(argv[8]);
+    }
+
+    if (argc >= 10) {
+        n_pl = parse_list(argv[9]);
     }
 
     // init LLM
@@ -100,7 +110,8 @@ int main(int argc, char ** argv) {
 
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
-    ctx_params.n_batch   = 512;
+    ctx_params.n_batch   = n_batch;
+    ctx_params.n_ubatch  = n_ubatch;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
@@ -158,7 +169,7 @@ int main(int argc, char ** argv) {
     }
 
     LOG_TEE("\n");
-    LOG_TEE("%s: n_kv_max = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, ctx_params.n_threads, ctx_params.n_threads_batch);
+    LOG_TEE("%s: n_kv_max = %d, n_batch = %d, n_ubatch = %d, is_pp_shared = %d, n_gpu_layers = %d, n_threads = %u, n_threads_batch = %u\n", __func__, n_kv_max, n_batch, n_ubatch, is_pp_shared, 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");

examples/batched/batched.cpp

@@ -48,6 +48,8 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
+    process_escapes(params.prompt);
+
     // init LLM
 
     llama_backend_init();
@@ -78,7 +80,7 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx = n_kv_req;
+    ctx_params.n_ctx   = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
     ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;

examples/convert-llama2c-to-ggml/README.md

@@ -21,6 +21,8 @@ An example command using a model from [karpathy/tinyllamas](https://huggingface.
 
 `$ ./convert-llama2c-to-ggml --copy-vocab-from-model llama-2-7b-chat.gguf.q2_K.bin --llama2c-model stories42M.bin --llama2c-output-model stories42M.gguf.bin`
 
+Note: The vocabulary for `stories260K.bin` should be its own tokenizer `tok512.bin` found in [karpathy/tinyllamas/stories260K](https://huggingface.co/karpathy/tinyllamas/tree/main/stories260K).
+
 Now you can use the model with a command like:
 
 `$ ./main -m stories42M.gguf.bin -p "One day, Lily met a Shoggoth" -n 500 -c 256`

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

@@ -1,6 +1,7 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
+#include "log.h"
 
 #include <unordered_map>
 #include <vector>
@@ -78,111 +79,101 @@ typedef struct {
 
 struct TransformerWeights {
     // token embedding table
-    float* token_embedding_table;    // (vocab_size, dim)
+    std::vector<float> token_embedding_table;    // (vocab_size, dim)
     // weights for rmsnorms
-    float* rms_att_weight; // (layer, dim) rmsnorm weights
-    float* rms_ffn_weight; // (layer, dim)
+    std::vector<float> rms_att_weight; // (layer, dim) rmsnorm weights
+    std::vector<float> rms_ffn_weight; // (layer, dim)
     // weights for matmuls
-    float* wq; // (layer, dim, dim)
-    float* wk; // (layer, dim, dim)
-    float* wv; // (layer, dim, dim)
-    float* wo; // (layer, dim, dim)
+    std::vector<float> wq; // (layer, dim, dim)
+    std::vector<float> wk; // (layer, dim, dim)
+    std::vector<float> wv; // (layer, dim, dim)
+    std::vector<float> wo; // (layer, dim, dim)
     // weights for ffn
-    float* w1; // (layer, hidden_dim, dim)
-    float* w2; // (layer, dim, hidden_dim)
-    float* w3; // (layer, hidden_dim, dim)
+    std::vector<float> w1; // (layer, hidden_dim, dim)
+    std::vector<float> w2; // (layer, dim, hidden_dim)
+    std::vector<float> w3; // (layer, hidden_dim, dim)
     // final rmsnorm
-    float* rms_final_weight; // (dim,)
+    std::vector<float> rms_final_weight; // (dim,)
     // freq_cis for RoPE relatively positional embeddings
-    // float* freq_cis_real; // (seq_len, dim/2)
-    // float* freq_cis_imag; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_real; // (seq_len, dim/2)
+    // std::vector<float> freq_cis_imag; // (seq_len, dim/2)
     // (optional) classifier weights for the logits, on the last layer
-    float* wcls;
-
-    ~TransformerWeights() {
-        delete[] token_embedding_table;
-        delete[] rms_att_weight;
-        delete[] rms_ffn_weight;
-        delete[] wq;
-        delete[] wk;
-        delete[] wv;
-        delete[] wo;
-        delete[] w1;
-        delete[] w2;
-        delete[] w3;
-        delete[] rms_final_weight;
-        delete[] wcls;
-    }
+    std::vector<float> wcls;
 };
 
-static void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
-    // we calloc instead of malloc to keep valgrind happy
-    w->token_embedding_table = new float[p->vocab_size * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->token_embedding_table\n",__func__,p->vocab_size , p->dim, p->vocab_size * p->dim);
+static void alloc_weights(TransformerWeights * w, const Config * p, bool shared_weights) {
+    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);
 
-    w->rms_att_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] 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_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);
 
-    w->rms_ffn_weight = new float[p->n_layers * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->wq = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->wk = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wk\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);
 
-    w->wv = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] Allocating [%d] x [%d] x [%d] = [%d] float space for w->wv\n",__func__, p->n_layers, p->dim, p->dim, p->n_layers * p->dim * p->dim);
+        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);
 
-    w->wo = new float[p->n_layers * p->dim * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->w1 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->w2 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->w3 = new float[p->n_layers * p->hidden_dim * p->dim]();
-    printf("[%s:AK] 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->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);
 
-    w->rms_final_weight = new float[p->dim]();
-    printf("[%s:AK] Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
+        w->rms_final_weight.resize(p->dim);
+        LOG("%s: Allocating [%d] float space for w->rms_final_weight\n",__func__,p->dim);
 
-    if (shared_weights) {
-        w->wcls = NULL;
-    } else {
-        w->wcls = new float[p->vocab_size * p->dim]();
-        printf("[%s:AK] Allocating [%d] x [%d] = [%d] float space for w->wcls\n",__func__,p->vocab_size , p->dim, p->vocab_size * 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);
+        }
+    }
+    catch (std::length_error &) {
+        die("Invalid configuration. Failed to allocate memory for weights");
     }
 }
 
-static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shared_weights) {
-    if (fread(w->token_embedding_table, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
-    if (fread(w->rms_att_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->wq, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wk, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wv, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->wo, sizeof(float), p->n_layers * p->dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->dim)) return 1;
-    if (fread(w->rms_ffn_weight, sizeof(float), p->n_layers * p->dim, f) != static_cast<size_t>(p->n_layers * p->dim)) return 1;
-    if (fread(w->w1, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->w2, sizeof(float), p->n_layers * p->hidden_dim * p->dim, f) != static_cast<size_t>(p->n_layers * p->hidden_dim * p->dim)) return 1;
-    if (fread(w->w3, sizeof(float), p->n_layers * p->dim * p->hidden_dim, f) != static_cast<size_t>(p->n_layers * p->dim * p->hidden_dim)) return 1;
-    if (fread(w->rms_final_weight, sizeof(float), p->dim, f) != static_cast<size_t>(p->dim)) return 1;
+static int checkpoint_init_weights(TransformerWeights * w, const Config * p, FILE * f, bool shared_weights) {
+    if (fread(w->token_embedding_table.data(), sizeof(float), w->token_embedding_table.size(), f) != w->token_embedding_table.size()) return 1;
+    if (fread(w->rms_att_weight.data(), sizeof(float), w->rms_att_weight.size(), f) != w->rms_att_weight.size()) return 1;
+    if (fread(w->wq.data(), sizeof(float), w->wq.size(), f) != w->wq.size()) return 1;
+    if (fread(w->wk.data(), sizeof(float), w->wk.size(), f) != w->wk.size()) return 1;
+    if (fread(w->wv.data(), sizeof(float), w->wv.size(), f) != w->wv.size()) return 1;
+    if (fread(w->wo.data(), sizeof(float), w->wo.size(), f) != w->wo.size()) return 1;
+    if (fread(w->rms_ffn_weight.data(), sizeof(float), w->rms_ffn_weight.size(), f) != w->rms_ffn_weight.size()) return 1;
+    if (fread(w->w1.data(), sizeof(float), w->w1.size(), f) != w->w1.size()) return 1;
+    if (fread(w->w2.data(), sizeof(float), w->w2.size(), f) != w->w2.size()) return 1;
+    if (fread(w->w3.data(), sizeof(float), w->w3.size(), f) != w->w3.size()) return 1;
+    if (fread(w->rms_final_weight.data(), sizeof(float), w->rms_final_weight.size(), f) != w->rms_final_weight.size()) return 1;
 
     // Skip freq_cis_real & freq_cis_imag
     int head_size = p->dim / p->n_heads;
     fseek(f, p->seq_len * head_size * sizeof(float), SEEK_CUR);
 
-    if (!shared_weights && fread(w->wcls, sizeof(float), p->vocab_size * p->dim, f) != static_cast<size_t>(p->vocab_size * p->dim)) return 1;
+    if (!shared_weights && fread(w->wcls.data(), sizeof(float), w->wcls.size(), f) != w->wcls.size()) return 1;
 
     // Check we didn't forget to read anything
     auto curr = ftell(f);
     fseek(f, 0, SEEK_END);
     auto end = ftell(f);
     if (curr != end) {
-        printf("Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", curr, end);
+        LOG("%s: Error: failed to read the checkpoint file to the end (curr = %ld, end =  %ld)\n", __func__, curr, end);
         return 1;
     }
 
@@ -190,20 +181,20 @@ static int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bo
 }
 
 static void print_sample_weights(TransformerWeights *w){
-    printf("----- Quick print of first of the weight vales of all the variables\n");
-    printf("%f\n", w->token_embedding_table[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    printf("%f\n", w->rms_ffn_weight[0]);
+    LOG("----- Quick print of first of the weight vales of all the variables\n");
+    LOG("%f\n", w->token_embedding_table[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    LOG("%f\n", w->rms_ffn_weight[0]);
 
-    printf("%f\n", w->wq[0]);
-    printf("%f\n", w->wk[0]);
-    printf("%f\n", w->wv[0]);
-    printf("%f\n", w->wo[0]);
-    printf("%f\n", w->w1[0]);
-    printf("%f\n", w->w2[0]);
-    printf("%f\n", w->w3[0]);
-    printf("%f\n", w->rms_att_weight[0]);
-    if (w->wcls) printf("%f\n", w->wcls[0]);
+    LOG("%f\n", w->wq[0]);
+    LOG("%f\n", w->wk[0]);
+    LOG("%f\n", w->wv[0]);
+    LOG("%f\n", w->wo[0]);
+    LOG("%f\n", w->w1[0]);
+    LOG("%f\n", w->w2[0]);
+    LOG("%f\n", w->w3[0]);
+    LOG("%f\n", w->rms_att_weight[0]);
+    if (!w->wcls.empty()) LOG("%f\n", w->wcls[0]);
 }
 ////////////////////////////////////////////////////////////////////////////////////////////////////////////
 
@@ -225,14 +216,16 @@ struct llama_vocab {
 };
 
 struct my_llama_hparams {
-    uint32_t n_vocab = 32000;
-    uint32_t n_ctx   = 512;   // this is provided as user input?
-    uint32_t n_embd  = 4096;
-    uint32_t n_ff    = 11008;
-    uint32_t n_mult  = 4;
-    uint32_t n_head  = 32;
-    uint32_t n_layer = 32;
-    uint32_t n_rot   = 64;
+    uint32_t n_vocab   = 32000;
+    uint32_t n_ctx     = 512;   // this is provided as user input?
+    uint32_t n_embd    = 4096;
+    uint32_t n_ff      = 11008;
+    uint32_t n_mult    = 4;
+    uint32_t n_head    = 32;
+    uint32_t n_head_kv = 32;
+    uint32_t n_layer   = 32;
+    uint32_t n_rot     = 64;
+
     bool operator!=(const my_llama_hparams& other) const {
         return memcmp(this, &other, sizeof(my_llama_hparams));
     }
@@ -325,14 +318,30 @@ struct train_params {
 };
 
 static void print_params(struct my_llama_hparams * params) {
-    printf("%s: n_vocab: %u\n", __func__, params->n_vocab);
-    printf("%s: n_ctx:   %u\n", __func__, params->n_ctx);
-    printf("%s: n_embd:  %u\n", __func__, params->n_embd);
-    printf("%s: n_mult:  %u\n", __func__, params->n_mult);
-    printf("%s: n_head:  %u\n", __func__, params->n_head);
-    printf("%s: n_ff:    %u\n", __func__, params->n_ff);
-    printf("%s: n_layer: %u\n", __func__, params->n_layer);
-    printf("%s: n_rot:   %u\n", __func__, params->n_rot);
+    LOG("%s: n_vocab:   %u\n", __func__, params->n_vocab);
+    LOG("%s: n_ctx:     %u\n", __func__, params->n_ctx);
+    LOG("%s: n_embd:    %u\n", __func__, params->n_embd);
+    LOG("%s: n_mult:    %u\n", __func__, params->n_mult);
+    LOG("%s: n_head:    %u\n", __func__, params->n_head);
+    LOG("%s: n_head_kv: %u\n", __func__, params->n_head_kv);
+    LOG("%s: n_ff:      %u\n", __func__, params->n_ff);
+    LOG("%s: n_layer:   %u\n", __func__, params->n_layer);
+    LOG("%s: n_rot:     %u\n", __func__, params->n_rot);
+}
+
+static void print_tensor_info(const struct ggml_context * ctx) {
+    for (auto t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        LOG("%s: Allocating ", __func__);
+        int64_t total = 1;
+        int i = 0;
+        for (; i < ggml_n_dims(t); ++i) {
+            if (i > 0) LOG("x ");
+            LOG("[%" PRId64 "] ", t->ne[i]);
+            total *= t->ne[i];
+        }
+        if (i > 1) LOG("= [%" PRId64 "] ", total);
+        LOG("float space for %s\n", ggml_get_name(t));
+    }
 }
 
 static void init_model(struct my_llama_model * model) {
@@ -342,6 +351,8 @@ static void init_model(struct my_llama_model * model) {
     const uint32_t n_layer = hparams.n_layer;
     const uint32_t n_vocab = hparams.n_vocab;
 
+    const uint32_t n_multiqueries = hparams.n_head_kv <= 0 || hparams.n_head_kv >= hparams.n_head ? 1 : hparams.n_head / hparams.n_head_kv;
+
     const uint32_t n_ff = hparams.n_ff;
     struct ggml_context * ctx = model->ctx;
 
@@ -350,25 +361,8 @@ static void init_model(struct my_llama_model * model) {
     model->train_tokens = 0;
 
     model->tok_embeddings = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x [%u] = [%u] float space for model->tok_embeddings\n",__func__,n_embd , n_vocab, n_embd * n_vocab);
-
     model->norm           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
-    printf("[%s:GG] Allocating [%u] float space for model->norm\n",__func__,n_embd);
-
     model->output         = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_vocab);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for model->output\n",__func__,n_embd, n_vocab, n_embd * n_vocab);
-
-    // printing the per-layer allocations here so we dont print in the for loop.
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wq for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wk for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wv for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.wo for [%u] layers\n",__func__, n_embd, n_embd, n_embd * n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] float space for layer.ffn_norm for [%u] layers\n",__func__,n_embd, n_layer);
-
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w1 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w2 for [%u] layers\n",__func__, n_embd, n_ff, n_ff * n_embd, n_layer);
-    printf("[%s:GG] Allocating [%u] x[%u] = [%u] float space for layer.w3 for [%u] layers\n",__func__, n_ff, n_embd, n_embd * n_ff, n_layer);
 
     ggml_set_name(model->tok_embeddings, "tok_embeddings.weight");
     ggml_set_name(model->norm,           "norm.weight");
@@ -383,8 +377,8 @@ static void init_model(struct my_llama_model * model) {
         layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
 
         layer.wq = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
-        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
+        layer.wk = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
+        layer.wv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd / n_multiqueries);
         layer.wo = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd);
 
         layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
@@ -406,6 +400,8 @@ static void init_model(struct my_llama_model * model) {
         ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
         ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
     }
+
+    print_tensor_info(ctx);
 }
 
 static float get_f32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
@@ -421,9 +417,9 @@ static int32_t get_i32_2d(struct ggml_tensor * tensor, int64_t i0, int64_t i1) {
 static void print_row(struct ggml_tensor * probs, int i) {
     for (int k = 0; k < probs->ne[0]; ++k) {
         float p = get_f32_2d(probs, k, i);
-        printf(" %f", p);
+        LOG(" %f", p);
     }
-    printf("\n");
+    LOG("\n");
 }
 
 static void print_matrix(struct ggml_tensor * probs) {
@@ -431,33 +427,12 @@ static void print_matrix(struct ggml_tensor * probs) {
     for (int i = 0; i < probs->ne[1]; ++i) {
         for (int k = 0; k < probs->ne[0]; ++k) {
             float p = get_f32_2d(probs, k, i);
-            printf(" %.2f", p);
+            LOG(" %.2f", p);
         }
-        printf("\n");
+        LOG("\n");
     }
 }
 
-#ifdef __GNUC__
-#ifdef __MINGW32__
-__attribute__((format(gnu_printf, 1, 2)))
-#else
-__attribute__((format(printf, 1, 2)))
-#endif
-#endif
-static std::string format(const char * fmt, ...) {
-    va_list ap, ap2;
-    va_start(ap, fmt);
-    va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
-    GGML_ASSERT(size >= 0 && size < INT_MAX);
-    std::vector<char> buf(size + 1);
-    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
-    GGML_ASSERT(size2 == size);
-    va_end(ap2);
-    va_end(ap);
-    return std::string(buf.data(), size);
-}
-
 struct llama_file {
     // use FILE * so we don't have to re-open the file to mmap
     FILE * fp;
@@ -549,8 +524,9 @@ static std::string llama_escape_whitespaces(const std::string & text) {
     return out.str();
 }
 
-static void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab) {
+static void load_vocab(const char * filename, const Config * config, struct llama_vocab * vocab) {
     if (is_ggml_file(filename)) {
+        LOG("%s: Loading vocabulary from gguf file %s\n", __func__, filename);
         struct ggml_context * ctx_data = NULL;
 
         struct gguf_init_params params = {
@@ -578,6 +554,9 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         const int * toktypes = (const int * ) gguf_get_arr_data(ctx, toktype_idx);
 
         const uint32_t n_vocab = gguf_get_arr_n(ctx, token_idx);
+        if (n_vocab != static_cast<uint32_t>(config->vocab_size)) {
+            die_fmt("vocab size mismatch: (gguf) %u != (llama2c) %d", n_vocab, config->vocab_size);
+        }
 
         vocab->id_to_token.resize(n_vocab);
 
@@ -595,7 +574,7 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
         gguf_free(ctx);
     } else {
         // assume llama2.c vocabulary
-        printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
+        LOG("%s: Assuming llama2.c vocabulary since %s is not a gguf file\n", __func__, filename);
         llama_file file(filename, "rb");
         if (!file.fp) {
             die_fmt("%s: %s", strerror(errno), filename);
@@ -638,38 +617,15 @@ static void load_vocab(const char *filename, Config *config, struct llama_vocab
 }
 
 static void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
-    int ct;
-    switch (ggml_n_dims(gg_weights)) {
-        case 1:
-            ct = 0;
-            for (int i0 = 0; i0 < gg_weights->ne[0]; i0++){
-                float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0]);
-                *ptr = karpathy_weights[ct];
-                ct++;
-            }
-            break;
-        case 2:
-            ct = 0;
-            for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                    float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1]);
-                    *ptr = karpathy_weights[ct];
-                    ct++;
-                }
-            }
-            break;
-        case 3:
-            ct = 0;
-            for (int i2 = 0; i2 < gg_weights->ne[2]; i2++) {
-                for (int i1 = 0; i1 < gg_weights->ne[1]; i1++) {
-                    for (int i0 = 0; i0 < gg_weights->ne[0]; i0++) {
-                        float * ptr = (float *) ((char *) gg_weights->data + i0*gg_weights->nb[0] + i1*gg_weights->nb[1] + i2*gg_weights->nb[2]);
-                        *ptr = karpathy_weights[ct];
-                        ct++;
-                    }
-                }
-            }
-            break;
+    int size = 1;
+    for (int dim = 0; dim < ggml_n_dims(gg_weights); ++dim) {
+        size *= gg_weights->ne[dim];
+    }
+    for (int ct = 0; ct < size; ++ct) {
+        int64_t i0 = 0; int64_t i1 = 0;
+        int64_t i2 = 0; int64_t i3 = 0;
+        ggml_unravel_index(gg_weights, ct, &i0, &i1, &i2, &i3);
+        ggml_set_f32_nd(gg_weights, i0, i1, i2, i3, karpathy_weights[ct]);
     }
 }
 
@@ -679,16 +635,18 @@ static void save_as_llama_model(
     // convert AK weights into GG weights one by one.
     // w->token_embedding_table -> model->tok_embeddings
     // float*                   -> struct ggml_tensor
-    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
-    convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
+    convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table.data());
+    convert_weights_ak_to_gg(model->output, !w->wcls.empty() ? w->wcls.data() : w->token_embedding_table.data());
 
-    convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
+    convert_weights_ak_to_gg(model->norm, w->rms_final_weight.data());
     //print_row(model->norm, 0);
 
     // for rms-att-weight
     int row_length = model->hparams.n_embd;
     int n_ff = model->hparams.n_ff;
 
+    const uint32_t n_multiqueries = model->hparams.n_head_kv <= 0 || model->hparams.n_head_kv >= model->hparams.n_head ? 1 : model->hparams.n_head / model->hparams.n_head_kv;
+
     for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
         auto & layer = model->layers[i];
         // 1d
@@ -697,9 +655,10 @@ static void save_as_llama_model(
 
         // from 3d matrix layer x dim x dim to 2d matrix dim x dim
         convert_weights_ak_to_gg(layer.wq            , &w->wq[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length]);
-        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length]);
         convert_weights_ak_to_gg(layer.wo            , &w->wo[i*row_length*row_length]);
+        // from 3d matrix layer x dim x dim to 2d matrix dim x dim / n_multiqueries
+        convert_weights_ak_to_gg(layer.wk            , &w->wk[i*row_length*row_length/n_multiqueries]);
+        convert_weights_ak_to_gg(layer.wv            , &w->wv[i*row_length*row_length/n_multiqueries]);
 
         convert_weights_ak_to_gg(layer.w1            , &w->w1[i*row_length*n_ff]);
         convert_weights_ak_to_gg(layer.w2            , &w->w2[i*n_ff*row_length]);
@@ -736,8 +695,8 @@ static void save_as_llama_model(
     gguf_set_val_u32(ctx, KV_EMBEDDING_LENGTH, model->hparams.n_embd);
     gguf_set_val_u32(ctx, KV_FEED_FORWARD_LENGTH, model->hparams.n_ff);
     gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
-    // n_head_kv is optional, default to n_head
-    // gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, ...);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT, model->hparams.n_head);
+    gguf_set_val_u32(ctx, KV_ATTENTION_HEAD_COUNT_KV, model->hparams.n_head_kv);
     gguf_set_val_u32(ctx, KV_BLOCK_COUNT, model->hparams.n_layer);
     gguf_set_val_u32(ctx, KV_ROPE_DIMENSION_COUNT, model->hparams.n_rot);
     gguf_set_val_f32(ctx, KV_ATTENTION_LAYERNORM_RMS_EPS, 1e-5f);
@@ -789,12 +748,12 @@ static void save_as_llama_model(
 
 static struct train_params get_default_train_params() {
     struct train_params params;
-    params.fn_vocab_model    = "models/7B/ggml-model-f16.gguf";
+    params.fn_vocab_model          = "models/7B/ggml-model-f16.gguf";
     params.fn_llama2c_output_model = "ak_llama_model.bin";
-    params.fn_train_data     = "shakespeare.txt";
-    params.fn_checkpoint_in  = "checkpoint.bin";
-    params.fn_checkpoint_out = "checkpoint.bin";
-    params.fn_model_out      = "ggml-checkpoint-f32.bin";
+    params.fn_train_data           = "shakespeare.txt";
+    params.fn_checkpoint_in        = "checkpoint.bin";
+    params.fn_checkpoint_out       = "checkpoint.bin";
+    params.fn_model_out            = "ggml-checkpoint-f32.bin";
 
     params.seed       =   -1;
 
@@ -829,8 +788,8 @@ static struct train_params get_default_train_params() {
     params.adam_alpha        = 1e-3f;
     params.adam_decay        = 1e-3f;
 
-    params.mem_model_gb   = 2;
-    params.mem_compute_gb = 24;
+    params.mem_model_gb    = 2;
+    params.mem_compute_gb  = 24;
     params.mem_compute0_gb = 8;
     params.mem_compute1_gb = 2;
 
@@ -916,19 +875,30 @@ int main(int argc, char ** argv) {
     if (!params_parse(argc, argv, &params)) {
         return 1;
     }
+    log_set_target(stdout);
     Config config;
     TransformerWeights weights = {};
     {
-        FILE *file = fopen(params.fn_llama2c_model, "rb");
-        if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
+        LOG("%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);
+            return 1;
+        }
         // read in the config header
-        if(fread(&config, sizeof(Config), 1, file) != 1) { return 1; }
+        if (fread(&config, sizeof(Config), 1, file) != 1) {
+            LOG("%s: Unable to read llama2c config from %s!\n",__func__,params.fn_llama2c_model);
+            return 1;
+        }
         auto shared_weights = config.vocab_size > 0;
         config.vocab_size = abs(config.vocab_size);
 
         // read in the Transformer weights
-        malloc_weights(&weights, &config, shared_weights);
-        if(checkpoint_init_weights(&weights, &config, file, shared_weights)) { return 1; }
+        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);
+            return 1;
+        }
         fclose(file);
     }
 
@@ -936,15 +906,18 @@ int main(int argc, char ** argv) {
     load_vocab(params.fn_vocab_model, &config, &vocab);
 
     struct my_llama_model model;
-    model.hparams.n_vocab = config.vocab_size; //llama_n_vocab(lctx);
-    model.hparams.n_ctx   = params.n_ctx;
-    model.hparams.n_embd  = config.dim; //params.n_embd;
-    model.hparams.n_ff    = config.hidden_dim;
-    model.hparams.n_mult  = 32;//params.n_mult;
-    model.hparams.n_head  = config.n_heads; //params.n_head;
-    model.hparams.n_layer = config.n_layers; //params.n_layer;
-    model.hparams.n_rot   = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+    model.hparams.n_vocab   = config.vocab_size; //llama_n_vocab(lctx);
+    model.hparams.n_ctx     = params.n_ctx;
+    model.hparams.n_embd    = config.dim; //params.n_embd;
+    model.hparams.n_ff      = config.hidden_dim;
+    model.hparams.n_mult    = 32;//params.n_mult;
+    model.hparams.n_head    = config.n_heads; //params.n_head;
+    model.hparams.n_head_kv = config.n_kv_heads;
+    model.hparams.n_layer   = config.n_layers; //params.n_layer;
+    model.hparams.n_rot     = std::min((uint32_t)params.n_rotmax, model.hparams.n_embd / model.hparams.n_head);
+
     print_params(&model.hparams);
+
     struct ggml_init_params lcparams;
     lcparams.mem_size   = 1024ll*1024ll*1024ll*((size_t) params.mem_model_gb);
     lcparams.mem_buffer = NULL;
@@ -956,7 +929,7 @@ int main(int argc, char ** argv) {
     model.name = basename(params.fn_llama2c_model);
     save_as_llama_model(&vocab, &model, &weights, params.fn_llama2c_output_model);
 
-    printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
+    LOG("%s: Saving llama.c model file %s in ggml format at %s\n", __func__, params.fn_llama2c_model, params.fn_llama2c_output_model);
 
     ggml_free(model.ctx);
     return 0;

examples/embedding/embedding.cpp

@@ -61,6 +61,8 @@ int main(int argc, char ** argv) {
     }
 
     params.embedding = true;
+    // For non-causal models, batch size must be equal to ubatch size
+    params.n_ubatch = params.n_batch;
 
     print_build_info();
 
@@ -114,15 +116,17 @@ int main(int argc, char ** argv) {
     for (const auto & prompt : prompts) {
         auto inp = ::llama_tokenize(ctx, prompt, true, false);
         if (inp.size() > n_batch) {
-            inp.resize(n_batch);
+            fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
+                    __func__, (long long int) inp.size(), (long long int) n_batch);
+            return 1;
         }
         inputs.push_back(inp);
     }
 
-    // add eos if not present
+    // add SEP if not present
     for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_token_eos(model)) {
-            inp.push_back(llama_token_eos(model));
+        if (inp.empty() || inp.back() != llama_token_sep(model)) {
+            inp.push_back(llama_token_sep(model));
         }
     }
 
@@ -174,25 +178,27 @@ int main(int argc, char ** argv) {
     float * out = emb + p * n_embd;
     batch_decode(ctx, batch, out, s, n_embd);
 
-    // print the first part of the embeddings
+    // 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 < std::min(16, n_embd); i++) {
+        for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
             fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
         }
         fprintf(stdout, "\n");
     }
 
     // print cosine similarity matrix
-    fprintf(stdout, "\n");
-    printf("cosine similarity matrix:\n\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 (n_prompts > 1) {
         fprintf(stdout, "\n");
+        printf("cosine similarity matrix:\n\n");
+        for (int i = 0; i < n_prompts; i++) {
+            for (int j = 0; j < n_prompts; j++) {
+                float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
+                fprintf(stdout, "%6.2f ", sim);
+            }
+            fprintf(stdout, "\n");
+        }
     }
 
     // clean up

examples/eval-callback/CMakeLists.txt

@@ -0,0 +1,9 @@
+set(TARGET eval-callback)
+add_executable(${TARGET} eval-callback.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TEST_TARGET test-eval-callback)
+add_test(NAME ${TEST_TARGET} COMMAND eval-callback --hf-repo ggml-org/models --hf-file tinyllamas/stories260K.gguf --model stories260K.gguf --prompt hello --seed 42 -ngl 0)
+set_property(TEST ${TEST_TARGET} PROPERTY LABELS eval-callback curl)

examples/eval-callback/README.md

@@ -0,0 +1,95 @@
+# llama.cpp/examples/eval-callback
+
+A simple example which demonstrates how to use callback during the inference.
+It simply prints to the console all operations and tensor data.
+
+Usage:
+
+```shell
+eval-callback \
+  --hf-repo ggml-org/models \
+  --hf-file phi-2/ggml-model-q4_0.gguf \
+  --model phi-2-q4_0.gguf \
+  --prompt hello \
+  --seed 42 \
+  -ngl 33
+```
+
+Will print:
+
+```shell
+llm_load_tensors: offloaded 33/33 layers to GPU
+...
+llama_new_context_with_model: n_ctx      = 512
+...
+llama_new_context_with_model:      CUDA0 compute buffer size =   105.00 MiB
+llama_new_context_with_model:  CUDA_Host compute buffer size =     6.01 MiB
+llama_new_context_with_model: graph nodes  = 1225
+llama_new_context_with_model: graph splits = 2
+ggml_debug:                 inp_embd = (f32)   GET_ROWS(token_embd.weight{2560, 51200, 1, 1}, inp_tokens{1, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.0181,   0.0272,   0.0272, ...],
+                                      ],
+                                     ]
+ggml_debug:                   norm-0 = (f32)       NORM(CUDA0#inp_embd#0{2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.6989,   1.0636,   1.0636, ...],
+                                      ],
+                                     ]
+ggml_debug:                 norm_w-0 = (f32)        MUL(norm-0{2560, 1, 1, 1}, blk.0.attn_norm.weight{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.1800,   0.2817,   0.2632, ...],
+                                      ],
+                                     ]
+ggml_debug:              attn_norm-0 = (f32)        ADD(norm_w-0{2560, 1, 1, 1}, blk.0.attn_norm.bias{2560, 1, 1, 1}}) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -0.1863,   0.2970,   0.2604, ...],
+                                      ],
+                                     ]
+ggml_debug:                   wqkv-0 = (f32)    MUL_MAT(blk.0.attn_qkv.weight{2560, 7680, 1, 1}, attn_norm-0{2560, 1, 1, 1}}) = {7680, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1238,   1.2876,  -1.8086, ...],
+                                      ],
+                                     ]
+ggml_debug:                   bqkv-0 = (f32)        ADD(wqkv-0{7680, 1, 1, 1}, blk.0.attn_qkv.bias{7680, 1, 1, 1}}) = {7680, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:            bqkv-0 (view) = (f32)       VIEW(bqkv-0{7680, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:                   Qcur-0 = (f32)       CONT(bqkv-0 (view){2560, 1, 1, 1}, }) = {2560, 1, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                      ],
+                                     ]
+ggml_debug:        Qcur-0 (reshaped) = (f32)    RESHAPE(Qcur-0{2560, 1, 1, 1}, }) = {80, 32, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                       [ -0.3608,   0.5076,  -1.8866, ...],
+                                       [  1.7643,   0.0273,  -2.1065, ...],
+                                       ...
+                                      ],
+                                     ]
+ggml_debug:                   Qcur-0 = (f32)       ROPE(Qcur-0 (reshaped){80, 32, 1, 1}, CUDA0#inp_pos#0{1, 1, 1, 1}}) = {80, 32, 1, 1}
+                                     [
+                                      [
+                                       [ -1.1135,   1.4604,  -1.9226, ...],
+                                       [ -0.3608,   0.5076,  -1.8866, ...],
+                                       [  1.7643,   0.0273,  -2.1065, ...],
+                                       ...
+                                      ],
+                                     ]
+```

examples/eval-callback/eval-callback.cpp

@@ -0,0 +1,195 @@
+#include "common.h"
+#include "llama.h"
+#include "ggml.h"
+
+#include <cstdio>
+#include <random>
+#include <string>
+#include <tuple>
+#include <vector>
+
+/**
+ * This the arbitrary data which will be passed to each callback.
+ * Later on we can for example add operation or tensor name filter from the CLI arg, or a file descriptor to dump the tensor.
+ */
+struct callback_data {
+    std::vector<uint8_t> data;
+};
+
+static std::string ggml_ne_string(const ggml_tensor * t) {
+    std::string str;
+    for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+        str += std::to_string(t->ne[i]);
+        if (i + 1 < GGML_MAX_DIMS) {
+            str += ", ";
+        }
+    }
+    return str;
+}
+
+static void ggml_print_tensor(uint8_t * data, ggml_type type, const int64_t * ne, const size_t * nb, int64_t n) {
+    GGML_ASSERT(n > 0);
+    float sum = 0;
+    for (int64_t i3 = 0; i3 < ne[3]; i3++) {
+        printf("                                     [\n");
+        for (int64_t i2 = 0; i2 < ne[2]; i2++) {
+            if (i2 == n && ne[2] > 2*n) {
+                printf("                                      ..., \n");
+                i2 = ne[2] - n;
+            }
+            printf("                                      [\n");
+            for (int64_t i1 = 0; i1 < ne[1]; i1++) {
+                if (i1 == n && ne[1] > 2*n) {
+                    printf("                                       ..., \n");
+                    i1 = ne[1] - n;
+                }
+                printf("                                       [");
+                for (int64_t i0 = 0; i0 < ne[0]; i0++) {
+                    if (i0 == n && ne[0] > 2*n) {
+                        printf("..., ");
+                        i0 = ne[0] - n;
+                    }
+                    size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
+                    float v;
+                    if (type == GGML_TYPE_F16) {
+                        v = ggml_fp16_to_fp32(*(ggml_fp16_t *) data + i);
+                    } else if (type == GGML_TYPE_F32) {
+                        v = *(float *) data + i;
+                    } else if (type == GGML_TYPE_I32) {
+                        v = (float) *(int32_t *) data + i;
+                    } else if (type == GGML_TYPE_I16) {
+                        v = (float) *(int16_t *) data + i;
+                    } else if (type == GGML_TYPE_I8) {
+                        v = (float) *(int8_t *) data + i;
+                    } else {
+                        GGML_ASSERT(false);
+                    }
+                    printf("%12.4f", v);
+                    sum += v;
+                    if (i0 < ne[0] - 1) printf(", ");
+                }
+                printf("],\n");
+            }
+            printf("                                      ],\n");
+        }
+        printf("                                     ]\n");
+        printf("                                     sum = %f\n", sum);
+    }
+}
+
+/**
+ * GGML operations callback during the graph execution.
+ *
+ * @param t current tensor
+ * @param ask when ask is true, the scheduler wants to know if we are interested in data from this tensor
+ *            if we return true, a follow-up call will be made with ask=false in which we can do the actual collection.
+ *            see ggml_backend_sched_eval_callback
+ * @param user_data user data to pass at each call back
+ * @return true to receive data or continue the graph, false otherwise
+ */
+static bool ggml_debug(struct ggml_tensor * t, bool ask, void * user_data) {
+    auto * cb_data = (callback_data *) user_data;
+
+    const struct ggml_tensor * src0 = t->src[0];
+    const struct ggml_tensor * src1 = t->src[1];
+
+    if (ask) {
+        return true; // Always retrieve data
+    }
+
+    char src1_str[128] = {0};
+    if (src1) {
+        sprintf(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());
+
+
+    // copy the data from the GPU memory if needed
+    const bool is_host = ggml_backend_buffer_is_host(t->buffer);
+
+    if (!is_host) {
+        auto n_bytes = ggml_nbytes(t);
+        cb_data->data.resize(n_bytes);
+        ggml_backend_tensor_get(t, cb_data->data.data(), 0, n_bytes);
+    }
+
+    if (!ggml_is_quantized(t->type)) {
+        uint8_t * data = is_host ? (uint8_t *) t->data : cb_data->data.data();
+        ggml_print_tensor(data, t->type, t->ne, t->nb, 3);
+    }
+
+    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));
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+        fprintf(stderr, "%s : failed to eval\n", __func__);
+        return false;
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    callback_data cb_data;
+
+    gpt_params params;
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    print_build_info();
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    // pass the callback to the backend scheduler
+    // it will be executed for each node during the graph computation
+    params.cb_eval = ggml_debug;
+    params.cb_eval_user_data = &cb_data;
+    params.warmup = false;
+
+    // init
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == nullptr || ctx == nullptr) {
+        fprintf(stderr, "%s : failed to init\n", __func__);
+        return 1;
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    bool OK = run(ctx, params);
+    if (!OK) {
+        return 1;
+    }
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/gbnf-validator/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET gbnf-validator)
+add_executable(${TARGET} gbnf-validator.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common grammar-parser llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/gbnf-validator/gbnf-validator.cpp

@@ -0,0 +1,132 @@
+#define LLAMA_API_INTERNAL
+
+#include "grammar-parser.h"
+#include "ggml.h"
+#include "llama.h"
+#include "unicode.h"
+
+#include <cstdio>
+#include <cstdlib>
+#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;
+
+    size_t pos = 0;
+    for (auto it = code_points.begin(), end = code_points.end() - 1; it != end; ++it) {
+        auto prev_stacks = grammar->stacks;
+        llama_grammar_accept(grammar->rules, prev_stacks, *it, grammar->stacks);
+        if (grammar->stacks.empty()) {
+            error_pos = pos;
+            error_msg = "Unexpected character '" + unicode_cpt_to_utf8(*it) + "'";
+            grammar->stacks = prev_stacks;
+            return false;
+        }
+        ++pos;
+    }
+
+    for (const auto & stack : grammar->stacks) {
+        if (stack.empty()) {
+            return true;
+        }
+    }
+
+    error_pos = pos;
+    error_msg = "Unexpected end of input";
+    return false;
+}
+
+static void print_error_message(const std::string & input_str, size_t error_pos, const std::string & error_msg) {
+    fprintf(stdout, "Input string is invalid according to the grammar.\n");
+    fprintf(stdout, "Error: %s at position %zu\n", error_msg.c_str(), error_pos);
+    fprintf(stdout, "\n");
+    fprintf(stdout, "Input string:\n");
+    fprintf(stdout, "%s", input_str.substr(0, error_pos).c_str());
+    if (error_pos < input_str.size()) {
+        fprintf(stdout, "\033[1;31m%c", input_str[error_pos]);
+        if (error_pos+1 < input_str.size()) {
+            fprintf(stdout, "\033[0;31m%s", input_str.substr(error_pos+1).c_str());
+        }
+        fprintf(stdout, "\033[0m\n");
+    }
+}
+
+int main(int argc, char** argv) {
+    if (argc != 3) {
+        fprintf(stdout, "Usage: %s <grammar_filename> <input_filename>\n", argv[0]);
+        return 1;
+    }
+
+    const std::string grammar_filename = argv[1];
+    const std::string input_filename = argv[2];
+
+    // Read the GBNF grammar file
+    FILE* grammar_file = fopen(grammar_filename.c_str(), "r");
+    if (!grammar_file) {
+        fprintf(stdout, "Failed to open grammar file: %s\n", grammar_filename.c_str());
+        return 1;
+    }
+
+    fseek(grammar_file, 0, SEEK_END);
+    size_t grammar_size = ftell(grammar_file);
+    fseek(grammar_file, 0, SEEK_SET);
+
+    std::string grammar_str(grammar_size, ' ');
+    fread(&grammar_str[0], 1, grammar_size, grammar_file);
+    fclose(grammar_file);
+
+    // 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"));
+
+    // Read the input file
+    FILE* input_file = fopen(input_filename.c_str(), "r");
+    if (!input_file) {
+        fprintf(stdout, "Failed to open input file: %s\n", input_filename.c_str());
+        return 1;
+    }
+
+    fseek(input_file, 0, SEEK_END);
+    size_t input_size = ftell(input_file);
+    fseek(input_file, 0, SEEK_SET);
+
+    std::string input_str(input_size, ' ');
+    fread(&input_str[0], 1, input_size, input_file);
+    fclose(input_file);
+
+    // 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);
+
+    if (is_valid) {
+        fprintf(stdout, "Input string is valid according to the grammar.\n");
+    } else {
+        print_error_message(input_str, error_pos, error_msg);
+    }
+
+    // Clean up
+    llama_grammar_free(grammar);
+
+    return 0;
+}

examples/gguf-split/gguf-split.cpp

@@ -1,37 +1,38 @@
 #include "llama.h"
-#include "ggml.h"
 #include "common.h"
 
 #include <algorithm>
 #include <cmath>
-#include <cstdint>
 #include <cstdlib>
 #include <fstream>
-#include <ios>
 #include <string>
 #include <vector>
 
 #include <stdio.h>
-#include <fcntl.h>
 #include <string.h>
+#include <climits>
+#include <stdexcept>
+
+#if defined(_WIN32)
+    #include <windows.h>
+    #ifndef PATH_MAX
+        #define PATH_MAX MAX_PATH
+    #endif
+    #include <io.h>
+#endif
 
 enum split_operation : uint8_t {
     SPLIT_OP_SPLIT,
     SPLIT_OP_MERGE,
 };
 
-static const char * const LLM_KV_GENERAL_SPLIT_I_SPLIT = "general.split";
-static const char * const LLM_KV_GENERAL_SPLIT_N_SPLIT = "general.split_count";
-
-static const int SPLIT_FILENAME_MAX = 256;
-
-static const char * const SPLIT_FILENAME_FORMAT = "%s-%05d-of-%05d.gguf";
-
 struct split_params {
     split_operation operation = SPLIT_OP_SPLIT;
+    size_t n_bytes_split = 0;
     int n_split_tensors = 128;
     std::string input;
     std::string output;
+    bool dry_run = false;
 };
 
 static void split_print_usage(const char * executable) {
@@ -42,15 +43,36 @@ static void split_print_usage(const char * executable) {
     printf("Apply a GGUF operation on IN to OUT.");
     printf("\n");
     printf("options:\n");
-    printf("  -h, --help            show this help message and exit\n");
-    printf("  --version             show version and build info\n");
-    printf("  --split               split GGUF to multiple GGUF (default)\n");
-    printf("  --split-max-tensors   max tensors in each split: default(%d)\n", default_params.n_split_tensors);
-    printf("  --merge               merge multiple GGUF to a single GGUF\n");
+    printf("  -h, --help              show this help message and exit\n");
+    printf("  --version               show version and build info\n");
+    printf("  --split                 split GGUF to multiple GGUF (enabled by default)\n");
+    printf("  --merge                 merge multiple GGUF to a single GGUF\n");
+    printf("  --split-max-tensors     max tensors in each split (default: %d)\n", default_params.n_split_tensors);
+    printf("  --split-max-size N(M|G) max size per split\n");
+    printf("  --dry-run               only print out a split plan and exit, without writing any new files\n");
     printf("\n");
 }
 
-static bool split_params_parse_ex(int argc, const char ** argv, split_params & params) {
+// return convert string, for example "128M" or "4G" to number of bytes
+static size_t split_str_to_n_bytes(std::string str) {
+    size_t n_bytes = 0;
+    int n;
+    if (str.back() == 'M') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = n * 1024 * 1024; // megabytes
+    } else if (str.back() == 'G') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = n * 1024 * 1024 * 1024; // gigabytes
+    } else {
+        throw std::invalid_argument("error: supported units are M (megabytes) or G (gigabytes), but got: " + std::string(1, str.back()));
+    }
+    if (n <= 0) {
+        throw std::invalid_argument("error: size must be a positive value");
+    }
+    return n_bytes;
+}
+
+static void split_params_parse_ex(int argc, const char ** argv, split_params & params) {
     std::string arg;
     const std::string arg_prefix = "--";
     bool invalid_param = false;
@@ -63,6 +85,8 @@ static bool 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);
@@ -72,23 +96,46 @@ static bool split_params_parse_ex(int argc, const char ** argv, split_params & p
             fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
             exit(0);
         }
+        if (arg == "--dry-run") {
+            arg_found = true;
+            params.dry_run = true;
+        }
 
+        if (is_op_set) {
+            throw std::invalid_argument("error: either --split or --merge can be specified, but not both");
+        }
         if (arg == "--merge") {
             arg_found = true;
+            is_op_set = true;
             params.operation = SPLIT_OP_MERGE;
         }
         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 (++arg_idx >= argc) {
                 invalid_param = true;
                 break;
             }
             arg_found = true;
+            is_mode_set = true;
             params.n_split_tensors = atoi(argv[arg_idx]);
         }
+        if (arg == "--split-max-size") {
+            if (++arg_idx >= argc) {
+                invalid_param = true;
+                break;
+            }
+            arg_found = true;
+            is_mode_set = true;
+            params.n_bytes_split = split_str_to_n_bytes(argv[arg_idx]);
+        }
 
         if (!arg_found) {
             throw std::invalid_argument("error: unknown argument: " + arg);
@@ -100,29 +147,22 @@ static bool split_params_parse_ex(int argc, const char ** argv, split_params & p
     }
 
     if (argc - arg_idx < 2) {
-        printf("%s: bad arguments\n", argv[0]);
-        split_print_usage(argv[0]);
-        return false;
+        throw std::invalid_argument("error: bad arguments");
     }
 
     params.input = argv[arg_idx++];
     params.output = argv[arg_idx++];
-
-    return true;
 }
 
 static bool split_params_parse(int argc, const char ** argv, split_params & params) {
     bool result = true;
     try {
-        if (!split_params_parse_ex(argc, argv, params)) {
-            split_print_usage(argv[0]);
-            exit(1);
-        }
+        split_params_parse_ex(argc, argv, params);
     }
     catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
         split_print_usage(argv[0]);
-        exit(1);
+        exit(EXIT_FAILURE);
     }
     return result;
 }
@@ -134,12 +174,6 @@ static void zeros(std::ofstream & file, size_t n) {
     }
 }
 
-static std::string split_file_name(const std::string & path, int i_split, int n_split) {
-    char f_split[SPLIT_FILENAME_MAX] = {0};
-    snprintf(f_split, sizeof(f_split), SPLIT_FILENAME_FORMAT, path.c_str(), i_split + 1, n_split);
-    return std::string(f_split);
-}
-
 struct split_strategy {
     const split_params params;
     std::ifstream & f_input;
@@ -147,15 +181,11 @@ struct split_strategy {
     struct ggml_context * ctx_meta = NULL;
     const int n_tensors;
 
-    const int n_split;
-    int i_split = 0;
+    // one ctx_out per one output file
+    std::vector<struct gguf_context *> ctx_outs;
 
-    int i_tensor = 0;
-
-    std::vector<uint8_t> read_data;
-
-    struct gguf_context * ctx_out;
-    std::ofstream fout;
+    // temporary buffer for reading in tensor data
+    std::vector<uint8_t> read_buf;
 
     split_strategy(const split_params & params,
             std::ifstream & f_input,
@@ -165,77 +195,141 @@ struct split_strategy {
         f_input(f_input),
         ctx_gguf(ctx_gguf),
         ctx_meta(ctx_meta),
-        n_tensors(gguf_get_n_tensors(ctx_gguf)),
-        n_split(std::ceil(1. * n_tensors / params.n_split_tensors)) {
+        n_tensors(gguf_get_n_tensors(ctx_gguf)) {
+
+        // because we need to know list of tensors for each file in advance, we will build all the ctx_out for all output splits
+        int i_split = -1;
+        struct gguf_context * ctx_out = NULL;
+        auto new_ctx_out = [&]() {
+            i_split++;
+            if (ctx_out != NULL) {
+                if (gguf_get_n_tensors(ctx_out) == 0) {
+                    fprintf(stderr, "error: one of splits have 0 tensors. Maybe size or tensors limit is too small\n");
+                    exit(EXIT_FAILURE);
+                }
+                ctx_outs.push_back(ctx_out);
+            }
+            ctx_out = gguf_init_empty();
+            // Save all metadata in first split only
+            if (i_split == 0) {
+                gguf_set_kv(ctx_out, ctx_gguf);
+            }
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
+            gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, 0); // placeholder
+            gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
+        };
+
+        // initialize ctx_out for the first split
+        new_ctx_out();
+
+        // process tensors one by one
+        size_t curr_tensors_size = 0; // current size by counting only tensors size (without metadata)
+        for (int i = 0; i < n_tensors; ++i) {
+            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+            // calculate the "imaginary" size = the current size + next tensor size
+            size_t n_bytes = GGML_PAD(ggml_nbytes(t), GGUF_DEFAULT_ALIGNMENT);
+            size_t next_tensors_size = curr_tensors_size + n_bytes;
+            if (should_split(i, next_tensors_size)) {
+                new_ctx_out();
+                curr_tensors_size = n_bytes;
+            } else {
+                curr_tensors_size = next_tensors_size;
+            }
+            gguf_add_tensor(ctx_out, t);
         }
 
-    bool should_split() const {
-        return i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
+        // push the last ctx_out
+        ctx_outs.push_back(ctx_out);
+
+        // set the correct n_split for all ctx_out
+        for (auto & ctx : ctx_outs) {
+            gguf_set_val_u16(ctx, LLM_KV_SPLIT_COUNT, ctx_outs.size());
+        }
     }
 
-    void split_start() {
-        ctx_out = gguf_init_empty();
-
-        // Save all metadata in first split only
-        if (i_split == 0) {
-            gguf_set_kv(ctx_out, ctx_gguf);
+    ~split_strategy() {
+        for (auto & ctx_out : ctx_outs) {
+            gguf_free(ctx_out);
         }
-        gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_I_SPLIT, i_split);
-        gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, n_split);
-
-        // populate the original tensors, so we get an initial metadata
-        for (int i = i_split * params.n_split_tensors; i < n_tensors && i < (i_split + 1) * params.n_split_tensors; ++i) {
-            struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
-            gguf_add_tensor(ctx_out, meta);
-        }
-
-        auto split_name = split_file_name(params.output, i_split, n_split);
-
-        fprintf(stderr, "%s: %s ...", __func__, split_name.c_str());
-        fout = std::ofstream(split_name, std::ios::binary);
-        fout.exceptions(std::ofstream::failbit); // fail fast on write errors
-
-        auto meta_size = gguf_get_meta_size(ctx_out);
-
-        // placeholder for the meta data
-        ::zeros(fout, meta_size);
-
-        i_split++;
     }
 
-    void next_tensor() {
-        const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
-        struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
-        auto n_bytes = ggml_nbytes(t);
-
-        if (read_data.size() < n_bytes) {
-            read_data.resize(n_bytes);
+    bool should_split(int i_tensor, size_t next_size) {
+        if (params.n_bytes_split > 0) {
+            // split by max size per file
+            return next_size > params.n_bytes_split;
+        } else {
+            // split by number of tensors per file
+            return i_tensor > 0 && i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
         }
-
-        auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor);
-        f_input.seekg(offset);
-        f_input.read((char *)read_data.data(), n_bytes);
-
-        t->data = read_data.data();
-
-        // write tensor data + padding
-        fout.write((const char *)t->data, n_bytes);
-        zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
-
-        i_tensor++;
     }
 
-    void split_end() {
-        // go back to beginning of file and write the updated metadata
-        fout.seekp(0);
-        std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
-        gguf_get_meta_data(ctx_out, data.data());
-        fout.write((const char *)data.data(), data.size());
+    void print_info() {
+        printf("n_split: %ld\n", ctx_outs.size());
+        int i_split = 0;
+        for (auto & ctx_out : ctx_outs) {
+            // re-calculate the real gguf size for each split (= metadata size + total size of all tensors)
+            size_t total_size = gguf_get_meta_size(ctx_out);
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_out, i));
+                total_size += ggml_nbytes(t);
+            }
+            total_size = total_size / 1024 / 1024; // convert to megabytes
+            printf("split %05d: n_tensors = %d, total_size = %ldM\n", i_split + 1, gguf_get_n_tensors(ctx_out), total_size);
+            i_split++;
+        }
+    }
 
-        fout.close();
-        gguf_free(ctx_out);
+    void write() {
+        int i_split = 0;
+        int n_split = ctx_outs.size();
+        for (auto & ctx_out : ctx_outs) {
+            // construct file path
+            char split_path[PATH_MAX] = {0};
+            llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
 
-        fprintf(stderr, "\033[3Ddone\n");
+            // open the output file
+            printf("Writing file %s ... ", split_path);
+            fflush(stdout);
+            std::ofstream fout = std::ofstream(split_path, std::ios::binary);
+            fout.exceptions(std::ofstream::failbit); // fail fast on write errors
+
+            // write metadata
+            std::vector<uint8_t> data(gguf_get_meta_size(ctx_out));
+            gguf_get_meta_data(ctx_out, data.data());
+            fout.write((const char *)data.data(), data.size());
+
+            // write tensors
+            for (int i = 0; i < gguf_get_n_tensors(ctx_out); ++i) {
+                // read tensor meta and prepare buffer
+                const char * t_name = gguf_get_tensor_name(ctx_out, i);
+                struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
+                auto n_bytes = ggml_nbytes(t);
+                read_buf.resize(n_bytes);
+
+                // calculate offset
+                auto i_tensor_in = gguf_find_tensor(ctx_gguf, t_name); // idx of tensor in the input file
+                auto offset = gguf_get_data_offset(ctx_gguf) + gguf_get_tensor_offset(ctx_gguf, i_tensor_in);
+
+                // copy tensor from input to output file
+                copy_file_to_file(f_input, fout, offset, n_bytes);
+                zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
+            }
+
+            printf("done\n");
+            // close the file
+            fout.close();
+            i_split++;
+        }
+    }
+
+    void copy_file_to_file(std::ifstream & f_in, std::ofstream & f_out, const size_t in_offset, const size_t len) {
+        // TODO: detect OS and use copy_file_range() here for better performance
+        if (read_buf.size() < len) {
+            read_buf.resize(len);
+        }
+        f_in.seekg(in_offset);
+        f_in.read((char *)read_buf.data(), len);
+        f_out.write((const char *)read_buf.data(), len);
     }
 };
 
@@ -250,37 +344,31 @@ static void gguf_split(const split_params & split_params) {
     std::ifstream f_input(split_params.input.c_str(), std::ios::binary);
     if (!f_input.is_open()) {
         fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_params.input.c_str());
-        exit(1);
+        exit(EXIT_FAILURE);
     }
 
     auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params);
     if (!ctx_gguf) {
         fprintf(stderr, "%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
-        exit(1);
+        exit(EXIT_FAILURE);
     }
 
+    // prepare the strategy
     split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
-    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
-            __func__, split_params.input.c_str(),
-            split_file_name(split_params.output, strategy.i_split, strategy.n_split).c_str(),
-            split_params.n_split_tensors);
+    int n_split = strategy.ctx_outs.size();
+    strategy.print_info();
 
-    strategy.split_start();
-
-    while (strategy.i_tensor < strategy.n_tensors) {
-        strategy.next_tensor();
-        if (strategy.should_split()) {
-            strategy.split_end();
-            strategy.split_start();
-        }
+    if (!split_params.dry_run) {
+        // write all output splits
+        strategy.write();
     }
-    strategy.split_end();
 
+    // done, clean up
     gguf_free(ctx_gguf);
     f_input.close();
 
     fprintf(stderr, "%s: %d gguf split written with a total of %d tensors.\n",
-            __func__, strategy.n_split, strategy.n_tensors);
+            __func__, n_split, strategy.n_tensors);
 }
 
 static void gguf_merge(const split_params & split_params) {
@@ -298,7 +386,9 @@ static void gguf_merge(const split_params & split_params) {
     std::vector<ggml_context *> ctx_metas;
     std::vector<gguf_context *> ctx_ggufs;
 
-    std::string split_prefix;
+    char split_path[PATH_MAX] = {0};
+    strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1);
+    char split_prefix[PATH_MAX] = {0};
 
     // First pass to find KV and tensors metadata
     for (int i_split = 0; i_split < n_split; i_split++) {
@@ -309,89 +399,66 @@ static void gguf_merge(const split_params & split_params) {
             /*.ctx      = */ &ctx_meta,
         };
 
-        auto split_name = split_params.input;
         if (i_split > 0) {
-            split_name = split_file_name(split_prefix, i_split, n_split);
+            llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
         }
-        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_name.c_str());
+        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path);
 
-        auto * ctx_gguf = gguf_init_from_file(split_name.c_str(), params);
+        auto * ctx_gguf = gguf_init_from_file(split_path, params);
         if (!ctx_gguf) {
             fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
-            exit(1);
+            exit(EXIT_FAILURE);
         }
         ctx_ggufs.push_back(ctx_gguf);
         ctx_metas.push_back(ctx_meta);
 
         if (i_split == 0) {
-            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_GENERAL_SPLIT_N_SPLIT);
+            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
             if (key_n_split < 0) {
                 fprintf(stderr,
                         "\n%s: input file does not contain %s metadata\n",
                         __func__,
-                        LLM_KV_GENERAL_SPLIT_N_SPLIT);
+                        LLM_KV_SPLIT_COUNT);
                 gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
                 gguf_free(ctx_out);
                 fout.close();
-                exit(1);
+                exit(EXIT_FAILURE);
             }
 
-            n_split = gguf_get_val_u8(ctx_gguf, key_n_split);
+            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
             if (n_split < 1) {
                 fprintf(stderr,
                         "\n%s: input file does not contain a valid split count %d\n",
                         __func__,
                         n_split);
                 gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
                 gguf_free(ctx_out);
                 fout.close();
-                exit(1);
+                exit(EXIT_FAILURE);
+            }
+
+            // Verify the file naming and extract split_prefix
+            if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) {
+                fprintf(stderr, "\n%s: unexpected input file name: %s"
+                                " i_split=%d"
+                                " n_split=%d\n", __func__,
+                        split_path, i_split, n_split);
+                gguf_free(ctx_gguf);
+                ggml_free(ctx_meta);
+                gguf_free(ctx_out);
+                fout.close();
+                exit(EXIT_FAILURE);
             }
 
             // Do not trigger merge if we try to merge again the output
-            gguf_set_val_u8(ctx_out, LLM_KV_GENERAL_SPLIT_N_SPLIT, 0);
+            gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0);
 
             // Set metadata from the first split
             gguf_set_kv(ctx_out, ctx_gguf);
         }
 
-        // Verify the file naming
-        {
-            int i_split_file = 0;
-            int n_split_file = 0;
-            const char * i_split_format = "-00000-of-00000.gguf";
-
-            if (split_name.size() < strlen(i_split_format)) {
-                fprintf(stderr, "\n%s: unexpected input file name: %s\n", __func__, split_params.input.c_str());
-                for (auto * _ctx_gguf : ctx_ggufs) {
-                    gguf_free(_ctx_gguf);
-                }
-                gguf_free(ctx_out);
-                fout.close();
-                exit(1);
-            }
-
-            split_prefix = split_name.substr(0, split_name.size() - strlen(i_split_format));
-
-            const char * split_name_c_str = split_name.c_str();
-            int n_part = sscanf(&split_name_c_str[0] + split_prefix.size(), "-%d-of-%d", &i_split_file, &n_split_file);
-
-            if (n_part != 2 || i_split_file - 1 != i_split || n_split_file != n_split) {
-                fprintf(stderr, "\n%s: unexpected input file name: %s"
-                                " i_split=%d i_split_file=%d"
-                                " n_split=%d n_split_file=%d\n", __func__,
-                        split_params.input.c_str(),
-                        i_split, i_split_file,
-                        n_split, n_split_file);
-                for (auto * _ctx_gguf : ctx_ggufs) {
-                    gguf_free(_ctx_gguf);
-                }
-                gguf_free(ctx_out);
-                fout.close();
-                exit(1);
-            }
-        }
-
         auto n_tensors = gguf_get_n_tensors(ctx_gguf);
         for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
             const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
@@ -411,18 +478,19 @@ static void gguf_merge(const split_params & split_params) {
 
     // Write tensors data
     for (int i_split = 0; i_split < n_split; i_split++) {
-        auto split_name = split_file_name(split_prefix, i_split, n_split);
-        std::ifstream f_input(split_name.c_str(), std::ios::binary);
+        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
+        std::ifstream f_input(split_path, std::ios::binary);
         if (!f_input.is_open()) {
-            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_name.c_str());
-            for (auto * _ctx_gguf : ctx_ggufs) {
-                gguf_free(_ctx_gguf);
+            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_path);
+            for (uint32_t i = 0; i < ctx_ggufs.size(); i++) {
+                gguf_free(ctx_ggufs[i]);
+                ggml_free(ctx_metas[i]);
             }
             gguf_free(ctx_out);
             fout.close();
-            exit(1);
+            exit(EXIT_FAILURE);
         }
-        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_name.c_str());
+        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
 
         auto * ctx_gguf = ctx_ggufs[i_split];
         auto * ctx_meta = ctx_metas[i_split];
@@ -469,10 +537,6 @@ static void gguf_merge(const split_params & split_params) {
 }
 
 int main(int argc, const char ** argv) {
-    if (argc < 3) {
-        split_print_usage(argv[0]);
-    }
-
     split_params params;
     split_params_parse(argc, argv, params);
 
@@ -481,8 +545,8 @@ int main(int argc, const char ** argv) {
             break;
         case SPLIT_OP_MERGE: gguf_merge(params);
             break;
-        default:split_print_usage(argv[0]);
-            exit(1);
+        default: split_print_usage(argv[0]);
+            exit(EXIT_FAILURE);
     }
 
     return 0;

examples/gguf/gguf.cpp

@@ -142,7 +142,7 @@ static bool gguf_ex_read_0(const std::string & fname) {
 }
 
 // read and create ggml_context containing the tensors and their data
-static bool gguf_ex_read_1(const std::string & fname) {
+static bool gguf_ex_read_1(const std::string & fname, bool check_data) {
     struct ggml_context * ctx_data = NULL;
 
     struct gguf_init_params params = {
@@ -206,7 +206,7 @@ static bool gguf_ex_read_1(const std::string & fname) {
             printf("\n\n");
 
             // check data
-            {
+            if (check_data) {
                 const float * data = (const float *) cur->data;
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
@@ -229,9 +229,16 @@ static bool gguf_ex_read_1(const std::string & fname) {
 
 int main(int argc, char ** argv) {
     if (argc < 3) {
-        printf("usage: %s data.gguf r|w\n", argv[0]);
+        printf("usage: %s data.gguf r|w [n]\n", argv[0]);
+        printf("r: read data.gguf file\n");
+        printf("w: write data.gguf file\n");
+        printf("n: no check of tensor data\n");
         return -1;
     }
+    bool check_data = true;
+    if (argc == 4) {
+        check_data = false;
+    }
 
     const std::string fname(argv[1]);
     const std::string mode (argv[2]);
@@ -242,7 +249,7 @@ int main(int argc, char ** argv) {
         GGML_ASSERT(gguf_ex_write(fname) && "failed to write gguf file");
     } else if (mode == "r") {
         GGML_ASSERT(gguf_ex_read_0(fname) && "failed to read gguf file");
-        GGML_ASSERT(gguf_ex_read_1(fname) && "failed to read gguf file");
+        GGML_ASSERT(gguf_ex_read_1(fname, check_data) && "failed to read gguf file");
     }
 
     return 0;

examples/imatrix/README.md

@@ -22,7 +22,7 @@ For faster computation, make sure to use GPU offloading via the `-ngl` argument
 ## Example
 
 ```bash
-LLAMA_CUBLAS=1 make -j
+LLAMA_CUDA=1 make -j
 
 # generate importance matrix (imatrix.dat)
 ./imatrix -m ggml-model-f16.gguf -f train-data.txt -ngl 99

examples/imatrix/imatrix.cpp

@@ -50,29 +50,31 @@ private:
     void keep_imatrix(int ncall) const;
 };
 
+// remove any prefix and suffixes from the name
+// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
+static std::string filter_tensor_name(const char * name) {
+    std::string wname;
+    const char * p = strchr(name, '#');
+    if (p != NULL) {
+        p = p + 1;
+        const char * q = strchr(p, '#');
+        if (q != NULL) {
+            wname = std::string(p, q - p);
+        } else {
+            wname = p;
+        }
+    } else {
+        wname = name;
+    }
+    return wname;
+}
+
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
     GGML_UNUSED(user_data);
 
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
-
-    std::string wname;
-    {
-        // remove any prefix and suffixes from the name
-        // CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
-        const char * p = strchr(src0->name, '#');
-        if (p != NULL) {
-            p = p + 1;
-            const char * q = strchr(p, '#');
-            if (q != NULL) {
-                wname = std::string(p, q - p);
-            } else {
-                wname = p;
-            }
-        } else {
-            wname = src0->name;
-        }
-    }
+    std::string wname = filter_tensor_name(src0->name);
 
     // when ask is true, the scheduler wants to know if we are interested in data from this tensor
     // if we return true, a follow-up call will be made with ask=false in which we can do the actual collection
@@ -96,34 +98,36 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 
     const float * data = is_host ? (const float *) src1->data : m_src1_data.data();
 
+    // this has been adapted to the new format of storing merged experts in a single 3d tensor
+    // ref: https://github.com/ggerganov/llama.cpp/pull/6387
     if (t->op == GGML_OP_MUL_MAT_ID) {
         const int idx  = ((int32_t *) t->op_params)[0];
-        const int n_as = ((int32_t *) t->op_params)[1];
+        const ggml_tensor * ids = t->src[2];
+        const int n_as = src0->ne[2];
 
-        // the top-k selected expert ids are stored in the src0 tensor
-        // for simplicity, always copy src0 to host, because it is small
-        // take into account that src0 is not contiguous!
-        GGML_ASSERT(src0->ne[1] == src1->ne[1]);
-        GGML_ASSERT(n_as*ggml_nrows(src0)*sizeof(int) == GGML_PAD(ggml_nbytes(src0), n_as*sizeof(int)));
-        m_ids.resize(ggml_nbytes(src0)/sizeof(int));
-        ggml_backend_tensor_get(src0, m_ids.data(), 0, ggml_nbytes(src0));
+        // the top-k selected expert ids are stored in the ids tensor
+        // for simplicity, always copy ids to host, because it is small
+        GGML_ASSERT(ids->ne[1] == src1->ne[1]);
+        m_ids.resize(ggml_nbytes(ids)/sizeof(int));
+        ggml_backend_tensor_get(ids, m_ids.data(), 0, ggml_nbytes(ids));
+
+        auto & e = m_stats[wname];
+
+        ++e.ncall;
+        // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
+        //       using the following line, we can correct for that if needed by replacing the line above with:
+        //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
 
         // loop over all possible experts, regardless if they are used or not in the batch
-        // this is necessary to guarantee equal number of "ncall" for each tensor
         for (int ex = 0; ex < n_as; ++ex) {
-            src0 = t->src[2 + ex];
-            auto& e = m_stats[wname];
+            size_t e_start = ex*src1->ne[0];
             if (e.values.empty()) {
-                e.values.resize(src1->ne[0], 0);
+                e.values.resize(src1->ne[0]*n_as, 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]);
+            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);
                 exit(1); //GGML_ASSERT(false);
             }
-            // NOTE: since we select top-k experts, the number of calls for the expert tensors will be k times larger
-            //       using the following line, we can correct for that if needed
-            //if (idx == t->src[0]->ne[0] - 1) ++e.ncall;
-            ++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);
             }
@@ -133,7 +137,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
                 if (excur != ex) continue;
                 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.values[e_start + j] += x[j]*x[j];
                 }
             }
             if (e.ncall > m_last_call) {
@@ -343,12 +347,13 @@ static void process_logits(
 static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool compute_ppl, int from_chunk) {
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
     const int n_ctx = llama_n_ctx(ctx);
 
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+    std::vector<llama_token> tokens = ::llama_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());
@@ -421,6 +426,7 @@ static bool compute_imatrix(llama_context * ctx, const gpt_params & params, bool
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // 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__);
                 return false;
@@ -589,24 +595,18 @@ int main(int argc, char ** argv) {
     llama_backend_init();
     llama_numa_init(params.numa);
 
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-
-    llama_model * model = llama_load_model_from_file(params.model.c_str(), mparams);
-    if (model == NULL) {
-        fprintf(stderr, "%s: error: unable to load model\n", __func__);
-        return 1;
-    }
-
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
     // pass the callback to the backend scheduler
     // it will be executed for each node during the graph computation
-    cparams.cb_eval = ik_collect_imatrix;
-    cparams.cb_eval_user_data = NULL;
+    params.cb_eval = ik_collect_imatrix;
+    params.cb_eval_user_data = NULL;
+    params.warmup = false;
 
-    llama_context * ctx = llama_new_context_with_model(model, cparams);
-    if (ctx == NULL) {
-        fprintf(stderr, "%s: error: unable to create context\n", __func__);
+    // init
+    llama_model * model;
+    llama_context * ctx;
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == nullptr || ctx == nullptr) {
+        fprintf(stderr, "%s : failed to init\n", __func__);
         return 1;
     }
 

examples/infill/README.md

@@ -36,6 +36,11 @@ The `infill` program offers a seamless way to interact with LLaMA models, allowi
 
 ### Example
 
+Download a model that supports infill, for example CodeLlama:
+```console
+scripts/hf.sh --repo TheBloke/CodeLlama-13B-GGUF --file codellama-13b.Q5_K_S.gguf --outdir models
+```
+
 ```bash
 ./infill -t 10 -ngl 0 -m models/codellama-13b.Q5_K_S.gguf -c 4096 --temp 0.7 --repeat_penalty 1.1 -n 20 --in-prefix "def helloworld():\n    print(\"hell" --in-suffix "\n   print(\"goodbye world\")\n    "
 ```

examples/infill/infill.cpp

@@ -239,6 +239,7 @@ int main(int argc, char ** argv) {
         LOG_TEE("%s\n", 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);
 
     bool suff_rm_leading_spc = params.escape;
@@ -279,10 +280,10 @@ int main(int argc, char ** argv) {
     if (ctx_guidance) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos);
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true);
         LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true);
         LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
         original_prompt_len = original_inp.size();

examples/json-schema-pydantic-example.py

@@ -0,0 +1,74 @@
+# Usage:
+#! ./server -m some-model.gguf &
+#! pip install pydantic
+#! python json-schema-pydantic-example.py
+
+from pydantic import BaseModel, TypeAdapter
+from annotated_types import MinLen
+from typing import Annotated, List, Optional
+import json, requests
+
+if True:
+
+    def create_completion(*, response_model=None, endpoint="http://localhost:8080/v1/chat/completions", messages, **kwargs):
+        '''
+        Creates a chat completion using an OpenAI-compatible endpoint w/ JSON schema support
+        (llama.cpp server, llama-cpp-python, Anyscale / Together...)
+
+        The response_model param takes a type (+ supports Pydantic) and behaves just as w/ Instructor (see below)
+        '''
+        if response_model:
+            type_adapter = TypeAdapter(response_model)
+            schema = type_adapter.json_schema()
+            messages = [{
+                "role": "system",
+                "content": f"You respond in JSON format with the following schema: {json.dumps(schema, indent=2)}"
+            }] + messages
+            response_format={"type": "json_object", "schema": schema}
+
+        data = requests.post(endpoint, headers={"Content-Type": "application/json"},
+                             json=dict(messages=messages, response_format=response_format, **kwargs)).json()
+        if 'error' in data:
+            raise Exception(data['error']['message'])
+
+        content = data["choices"][0]["message"]["content"]
+        return type_adapter.validate_json(content) if type_adapter else content
+
+else:
+
+    # This alternative branch uses Instructor + OpenAI client lib.
+    # Instructor support streamed iterable responses, retry & more.
+    # (see https://python.useinstructor.com/)
+    #! pip install instructor openai
+    import instructor, openai
+    client = instructor.patch(
+        openai.OpenAI(api_key="123", base_url="http://localhost:8080"),
+        mode=instructor.Mode.JSON_SCHEMA)
+    create_completion = client.chat.completions.create
+
+
+if __name__ == '__main__':
+
+    class QAPair(BaseModel):
+        question: str
+        concise_answer: str
+        justification: str
+
+    class PyramidalSummary(BaseModel):
+        title: str
+        summary: str
+        question_answers: Annotated[List[QAPair], MinLen(2)]
+        sub_sections: Optional[Annotated[List['PyramidalSummary'], MinLen(2)]]
+
+    print("# Summary\n", create_completion(
+        model="...",
+        response_model=PyramidalSummary,
+        messages=[{
+            "role": "user",
+            "content": f"""
+                You are a highly efficient corporate document summarizer.
+                Create a pyramidal summary of an imaginary internal document about our company processes
+                (starting high-level, going down to each sub sections).
+                Keep questions short, and answers even shorter (trivia / quizz style).
+            """
+        }]))

examples/json-schema-to-grammar.py

@@ -1,147 +0,0 @@
-#!/usr/bin/env python3
-import argparse
-import json
-import re
-import sys
-
-# whitespace is constrained to a single space char to prevent model "running away" in
-# whitespace. Also maybe improves generation quality?
-SPACE_RULE = '" "?'
-
-PRIMITIVE_RULES = {
-    'boolean': '("true" | "false") space',
-    'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
-    'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
-    'string': r''' "\"" (
-        [^"\\] |
-        "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])
-      )* "\"" space ''',
-    'null': '"null" space',
-}
-
-INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
-GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"'}
-
-
-class SchemaConverter:
-    def __init__(self, prop_order):
-        self._prop_order = prop_order
-        self._rules = {'space': SPACE_RULE}
-
-    def _format_literal(self, literal):
-        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
-        )
-        return f'"{escaped}"'
-
-    def _add_rule(self, name, rule):
-        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
-        if esc_name not in self._rules or self._rules[esc_name] == rule:
-            key = esc_name
-        else:
-            i = 0
-            while f'{esc_name}{i}' in self._rules:
-                i += 1
-            key = f'{esc_name}{i}'
-        self._rules[key] = rule
-        return key
-
-    def visit(self, schema, name):
-        schema_type = schema.get('type')
-        rule_name = name or 'root'
-
-        if 'oneOf' in schema or 'anyOf' in schema:
-            rule = ' | '.join((
-                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
-                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
-            ))
-            return self._add_rule(rule_name, rule)
-
-        elif 'const' in schema:
-            return self._add_rule(rule_name, self._format_literal(schema['const']))
-
-        elif 'enum' in schema:
-            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
-            return self._add_rule(rule_name, rule)
-
-        elif schema_type == 'object' and 'properties' in schema:
-            # TODO: `required` keyword
-            prop_order = self._prop_order
-            prop_pairs = sorted(
-                schema['properties'].items(),
-                # sort by position in prop_order (if specified) then by key
-                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
-            )
-
-            rule = '"{" space'
-            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
-                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
-                if i > 0:
-                    rule += ' "," space'
-                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
-            rule += ' "}" space'
-
-            return self._add_rule(rule_name, rule)
-
-        elif schema_type == 'array' and 'items' in schema:
-            # TODO `prefixItems` keyword
-            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
-            list_item_operator = f'("," space {item_rule_name})'
-            successive_items = ""
-            min_items = schema.get("minItems", 0)
-            if min_items > 0:
-               first_item = f"({item_rule_name})"
-               successive_items = list_item_operator * (min_items - 1)
-               min_items -= 1
-            else:
-               first_item = f"({item_rule_name})?"
-            max_items = schema.get("maxItems")
-            if max_items is not None and max_items > min_items:
-                successive_items += (list_item_operator + "?") * (max_items - min_items - 1)
-            else:
-                successive_items += list_item_operator + "*"
-            rule = f'"[" space {first_item} {successive_items} "]" space'
-            return self._add_rule(rule_name, rule)
-
-        else:
-            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
-            return self._add_rule(
-                'root' if rule_name == 'root' else schema_type,
-                PRIMITIVE_RULES[schema_type]
-            )
-
-    def format_grammar(self):
-        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
-
-
-def main(args_in = None):
-    parser = argparse.ArgumentParser(
-        description='''
-            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
-            given JSON schema. Only a subset of JSON schema features are supported; more may be
-            added in the future.
-        ''',
-    )
-    parser.add_argument(
-        '--prop-order',
-        default=[],
-        type=lambda s: s.split(','),
-        help='''
-            comma-separated property names defining the order of precedence for object properties;
-            properties not specified here are given lower precedence than those that are, and are
-            sorted alphabetically
-        '''
-    )
-    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
-    args = parser.parse_args(args_in)
-
-    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
-    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
-    converter = SchemaConverter(prop_order)
-    converter.visit(schema, '')
-    print(converter.format_grammar())
-
-
-if __name__ == '__main__':
-    main()

examples/json_schema_to_grammar.py

@@ -0,0 +1,616 @@
+#!/usr/bin/env python3
+import argparse
+import itertools
+import json
+import re
+import sys
+from typing import Any, Dict, List, Set, Tuple, Union
+
+def _build_repetition(item_rule, min_items, max_items, separator_rule=None, item_rule_is_literal=False):
+    if not separator_rule:
+        if min_items == 0 and max_items == 1:
+            return f'{item_rule}?'
+        elif min_items == 1 and max_items is None:
+            return f'{item_rule}+'
+
+    result = ''
+
+    if min_items > 0:
+        if item_rule_is_literal and separator_rule is None:
+            result = '"' + (item_rule[1:-1] * min_items) + '"'
+        else:
+            result = (f' {separator_rule} ' if separator_rule else ' ').join([item_rule] * min_items)
+
+    def opt_repetitions(up_to_n, prefix_with_sep=False):
+        '''
+            - n=4, no sep:             '(a (a (a (a)?)?)?)?'
+            - n=4, sep=',', prefix:    '("," a ("," a ("," a ("," a)?)?)?)?'
+            - n=4, sep=',', no prefix: '(a ("," a ("," a ("," a)?)?)?)?'
+        '''
+
+        content = f'{separator_rule} {item_rule}' if prefix_with_sep and separator_rule else item_rule
+        if up_to_n == 0:
+            return ''
+        elif up_to_n == 1:
+            return f'({content})?'
+        elif separator_rule and not prefix_with_sep:
+            return f'({content} {opt_repetitions(up_to_n - 1, prefix_with_sep=True)})?'
+        else:
+            return (f'({content} ' * up_to_n).rstrip() + (')?' * up_to_n)
+
+    if min_items > 0 and max_items != min_items:
+        result += ' '
+
+    if max_items is not None:
+        result += opt_repetitions(max_items - min_items, prefix_with_sep=min_items > 0)
+    else:
+        item_operator = f'({separator_rule + " " if separator_rule else ""}{item_rule})'
+
+        if min_items == 0 and separator_rule:
+            result = f'({item_rule} {item_operator}*)?'
+        else:
+            result += f'{item_operator}*'
+
+    return result
+
+
+class BuiltinRule:
+    def __init__(self, content: str, deps: list = None):
+        self.content = content
+        self.deps = deps or []
+
+_up_to_15_digits = _build_repetition('[0-9]', 0, 15)
+
+# whitespace is constrained to a single space char to prevent model "running away" in
+# whitespace. Also maybe improves generation quality?
+SPACE_RULE = '" "?'
+
+PRIMITIVE_RULES = {
+    'boolean'      : BuiltinRule('("true" | "false") space', []),
+    'decimal-part' : BuiltinRule('[0-9] ' + _up_to_15_digits, []),
+    'integral-part': BuiltinRule('[0-9] | [1-9] ' + _up_to_15_digits, []),
+    'number'       : BuiltinRule('("-"? integral-part) ("." decimal-part)? ([eE] [-+]? integral-part)? space', ['integral-part', 'decimal-part']),
+    'integer'      : BuiltinRule('("-"? integral-part) space', ['integral-part']),
+    'value'        : BuiltinRule('object | array | string | number | boolean | null', ['object', 'array', 'string', 'number', 'boolean', 'null']),
+    'object'       : BuiltinRule('"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space', ['string', 'value']),
+    'array'        : BuiltinRule('"[" space ( value ("," space value)* )? "]" space', ['value']),
+    'uuid'         : BuiltinRule(r'"\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + r' "\"" space', []),
+    'char'         : BuiltinRule(r'[^"\\] | "\\" (["\\/bfnrt] | "u" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])', []),
+    'string'       : BuiltinRule(r'"\"" char* "\"" space', ['char']),
+    'null'         : BuiltinRule('"null" space', []),
+}
+
+# TODO: support "uri", "email" string formats
+STRING_FORMAT_RULES = {
+    'date'            : BuiltinRule('[0-9] [0-9] [0-9] [0-9] "-" ( "0" [1-9] | "1" [0-2] ) "-" ( \"0\" [1-9] | [1-2] [0-9] | "3" [0-1] )', []),
+    'time'            : BuiltinRule('([01] [0-9] | "2" [0-3]) ":" [0-5] [0-9] ":" [0-5] [0-9] ( "." [0-9] [0-9] [0-9] )? ( "Z" | ( "+" | "-" ) ( [01] [0-9] | "2" [0-3] ) ":" [0-5] [0-9] )', []),
+    'date-time'       : BuiltinRule('date "T" time', ['date', 'time']),
+    'date-string'     : BuiltinRule('"\\"" date "\\"" space', ['date']),
+    'time-string'     : BuiltinRule('"\\"" time "\\"" space', ['time']),
+    'date-time-string': BuiltinRule('"\\"" date-time "\\"" space', ['date-time']),
+}
+
+DOTALL = '[\\U00000000-\\U0010FFFF]'
+DOT = '[^\\x0A\\x0D]'
+
+RESERVED_NAMES = set(["root", "dot", *PRIMITIVE_RULES.keys(), *STRING_FORMAT_RULES.keys()])
+
+INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
+GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
+GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
+GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
+
+NON_LITERAL_SET = set('|.()[]{}*+?')
+ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
+
+
+class SchemaConverter:
+    def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
+        self._prop_order = prop_order
+        self._allow_fetch = allow_fetch
+        self._dotall = dotall
+        self._raw_pattern = raw_pattern
+        self._rules = {
+            'space': SPACE_RULE,
+        }
+        self._refs = {}
+        self._refs_being_resolved = set()
+
+    def _format_literal(self, literal):
+        escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal
+        )
+        return f'"{escaped}"'
+
+    def not_literal(self, literal: str, dotall: bool = True, maybe_escaped_underscores = False) -> str:
+        '''
+            not_literal('a') -> '[^a]'
+            not_literal('abc') -> '([^a] | "a" ([^b] | "b" ([^c])?)?)?'
+        '''
+        assert len(literal) > 0, 'Empty literal not supported'
+        def recurse(i: int):
+            c = literal[i]
+            if maybe_escaped_underscores and c == '_':
+                yield f'[^{c}\\\\]'
+                yield ' | '
+                yield f'"\\\\"? "{c}"'
+            else:
+                yield f'[^{c}]'
+            if i < len(literal) - 1:
+                yield ' | '
+                yield self._format_literal(c)
+                yield ' ('
+                yield from recurse(i + 1)
+                yield ')?'
+
+        return ''.join(('(', *recurse(0), ')'))
+
+    def _add_rule(self, name, rule):
+        esc_name = INVALID_RULE_CHARS_RE.sub('-', name)
+        if esc_name not in self._rules or self._rules[esc_name] == rule:
+            key = esc_name
+        else:
+            i = 0
+            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
+                i += 1
+            key = f'{esc_name}{i}'
+        self._rules[key] = rule
+        return key
+
+    def resolve_refs(self, schema: dict, url: str):
+        '''
+            Resolves all $ref fields in the given schema, fetching any remote schemas,
+            replacing $ref with absolute reference URL and populating self._refs with the
+            respective referenced (sub)schema dictionaries.
+        '''
+        def visit(n: dict):
+            if isinstance(n, list):
+                return [visit(x) for x in n]
+            elif isinstance(n, dict):
+                ref = n.get('$ref')
+                if ref is not None and ref not in self._refs:
+                    if ref.startswith('https://'):
+                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
+                        import requests
+
+                        frag_split = ref.split('#')
+                        base_url = frag_split[0]
+
+                        target = self._refs.get(base_url)
+                        if target is None:
+                            target = self.resolve_refs(requests.get(ref).json(), base_url)
+                            self._refs[base_url] = target
+
+                        if len(frag_split) == 1 or frag_split[-1] == '':
+                            return target
+                    elif ref.startswith('#/'):
+                        target = schema
+                        ref = f'{url}{ref}'
+                        n['$ref'] = ref
+                    else:
+                        raise ValueError(f'Unsupported ref {ref}')
+
+                    for sel in ref.split('#')[-1].split('/')[1:]:
+                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
+                        target = target[sel]
+
+                    self._refs[ref] = target
+                else:
+                    for v in n.values():
+                        visit(v)
+
+            return n
+        return visit(schema)
+
+    def _generate_union_rule(self, name, alt_schemas):
+        return ' | '.join((
+            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
+            for i, alt_schema in enumerate(alt_schemas)
+        ))
+
+    def _visit_pattern(self, pattern, name):
+        '''
+            Transforms a regular expression pattern into a GBNF rule.
+
+            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
+            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
+
+            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
+
+            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
+            we define sub-rules to keep the output lean.
+        '''
+
+        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
+        pattern = pattern[1:-1]
+        sub_rule_ids = {}
+
+        i = 0
+        length = len(pattern)
+
+        def to_rule(s: Tuple[str, bool]) -> str:
+            (txt, is_literal) = s
+            return "\"" + txt + "\"" if is_literal else txt
+
+        def transform() -> Tuple[str, bool]:
+            '''
+                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
+            '''
+            nonlocal i
+            nonlocal pattern
+            nonlocal sub_rule_ids
+
+            start = i
+            # For each component of this sequence, store its string representation and whether it's a literal.
+            # We only need a flat structure here to apply repetition operators to the last item, and
+            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
+            # (GBNF's syntax is luckily very close to regular expressions!)
+            seq: list[Tuple[str, bool]] = []
+
+            def get_dot():
+                if self._dotall:
+                    rule = DOTALL
+                else:
+                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
+                    rule = DOT
+                return self._add_rule(f'dot', rule)
+
+            def join_seq():
+                nonlocal seq
+                ret = []
+                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
+                    if is_literal:
+                        ret.append((''.join(x[0] for x in g), True))
+                    else:
+                        ret.extend(g)
+                if len(ret) == 1:
+                    return ret[0]
+                return (' '.join(to_rule(x) for x in seq), False)
+
+            while i < length:
+                c = pattern[i]
+                if c == '.':
+                    seq.append((get_dot(), False))
+                    i += 1
+                elif c == '(':
+                    i += 1
+                    if i < length:
+                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
+                    seq.append((f'({to_rule(transform())})', False))
+                elif c == ')':
+                    i += 1
+                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
+                    return join_seq()
+                elif c == '[':
+                    square_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != ']':
+                        if pattern[i] == '\\':
+                            square_brackets += pattern[i:i+2]
+                            i += 2
+                        else:
+                            square_brackets += pattern[i]
+                            i += 1
+                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    square_brackets += ']'
+                    i += 1
+                    seq.append((square_brackets, False))
+                elif c == '|':
+                    seq.append(('|', False))
+                    i += 1
+                elif c in ('*', '+', '?'):
+                    seq[-1] = (to_rule(seq[-1]) + c, False)
+                    i += 1
+                elif c == '{':
+                    curly_brackets = c
+                    i += 1
+                    while i < length and pattern[i] != '}':
+                        curly_brackets += pattern[i]
+                        i += 1
+                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
+                    curly_brackets += '}'
+                    i += 1
+                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
+                    min_times = 0
+                    max_times = None
+                    try:
+                        if len(nums) == 1:
+                            min_times = int(nums[0])
+                            max_times = min_times
+                        else:
+                            assert len(nums) == 2
+                            min_times = int(nums[0]) if nums[0] else 0
+                            max_times = int(nums[1]) if nums[1] else None
+                    except ValueError:
+                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
+
+                    (sub, sub_is_literal) = seq[-1]
+
+                    if not sub_is_literal:
+                        id = sub_rule_ids.get(sub)
+                        if id is None:
+                            id = self._add_rule(f'{name}-{len(sub_rule_ids) + 1}', sub)
+                            sub_rule_ids[sub] = id
+                        sub = id
+
+                    seq[-1] = (_build_repetition(f'"{sub}"' if sub_is_literal else sub, min_times, max_times, item_rule_is_literal=sub_is_literal), False)
+                else:
+                    literal = ''
+                    while i < length:
+                        if pattern[i] == '\\' and i < length - 1:
+                            next = pattern[i + 1]
+                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
+                                i += 1
+                                literal += pattern[i]
+                                i += 1
+                            else:
+                                literal += pattern[i:i+2]
+                                i += 2
+                        elif pattern[i] == '"' and not self._raw_pattern:
+                            literal += '\\"'
+                            i += 1
+                        elif pattern[i] not in NON_LITERAL_SET and \
+                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
+                            literal += pattern[i]
+                            i += 1
+                        else:
+                            break
+                    if literal:
+                        seq.append((literal, True))
+
+            return join_seq()
+
+        return self._add_rule(
+            name,
+            to_rule(transform()) if self._raw_pattern \
+                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+
+
+    def _resolve_ref(self, ref):
+        ref_name = ref.split('/')[-1]
+        if ref_name not in self._rules and ref not in self._refs_being_resolved:
+            self._refs_being_resolved.add(ref)
+            resolved = self._refs[ref]
+            ref_name = self.visit(resolved, ref_name)
+            self._refs_being_resolved.remove(ref)
+        return ref_name
+
+    def _generate_constant_rule(self, value):
+        return self._format_literal(json.dumps(value))
+
+    def visit(self, schema, name):
+        schema_type = schema.get('type')
+        schema_format = schema.get('format')
+        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
+
+        if (ref := schema.get('$ref')) is not None:
+            return self._add_rule(rule_name, self._resolve_ref(ref))
+
+        elif 'oneOf' in schema or 'anyOf' in schema:
+            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
+
+        elif isinstance(schema_type, list):
+            return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type]))
+
+        elif 'const' in schema:
+            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
+
+        elif 'enum' in schema:
+            rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum']))
+            return self._add_rule(rule_name, rule)
+
+        elif schema_type in (None, 'object') and \
+             ('properties' in schema or \
+              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
+            required = set(schema.get('required', []))
+            properties = list(schema.get('properties', {}).items())
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
+
+        elif schema_type in (None, 'object') and 'allOf' in schema:
+            required = set()
+            properties = []
+            hybrid_name = name
+            def add_component(comp_schema, is_required):
+                if (ref := comp_schema.get('$ref')) is not None:
+                    comp_schema = self._refs[ref]
+
+                if 'properties' in comp_schema:
+                    for prop_name, prop_schema in comp_schema['properties'].items():
+                        properties.append((prop_name, prop_schema))
+                        if is_required:
+                            required.add(prop_name)
+
+            for t in schema['allOf']:
+                if 'anyOf' in t:
+                    for tt in t['anyOf']:
+                        add_component(tt, is_required=False)
+                else:
+                    add_component(t, is_required=True)
+
+            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[]))
+
+        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
+            items = schema.get('items') or schema['prefixItems']
+            if isinstance(items, list):
+                return self._add_rule(
+                    rule_name,
+                    '"[" space ' +
+                    ' "," space '.join(
+                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
+                        for i, item in enumerate(items)) +
+                    ' "]" space')
+            else:
+                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
+                min_items = schema.get("minItems", 0)
+                max_items = schema.get("maxItems")
+                return self._add_rule(rule_name, '"[" space ' + _build_repetition(item_rule_name, min_items, max_items, separator_rule='"," space') + ' "]" space')
+
+        elif schema_type in (None, 'string') and 'pattern' in schema:
+            return self._visit_pattern(schema['pattern'], rule_name)
+
+        elif schema_type in (None, 'string') and re.match(r'^uuid[1-5]?$', schema_format or ''):
+            return self._add_primitive(
+                'root' if rule_name == 'root' else schema_format,
+                PRIMITIVE_RULES['uuid']
+            )
+
+        elif schema_type in (None, 'string') and f'{schema_format}-string' in STRING_FORMAT_RULES:
+            prim_name = f'{schema_format}-string'
+            return self._add_rule(rule_name, self._add_primitive(prim_name, STRING_FORMAT_RULES[prim_name]))
+
+        elif schema_type == 'string' and ('minLength' in schema or 'maxLength' in schema):
+            char_rule = self._add_primitive('char', PRIMITIVE_RULES['char'])
+            min_len = schema.get('minLength', 0)
+            max_len = schema.get('maxLength')
+
+            return self._add_rule(rule_name, r'"\"" ' + _build_repetition(char_rule, min_len, max_len) + r' "\"" space')
+
+        elif (schema_type == 'object') or (len(schema) == 0):
+            return self._add_rule(rule_name, self._add_primitive('object', PRIMITIVE_RULES['object']))
+
+        else:
+            assert schema_type in PRIMITIVE_RULES, f'Unrecognized schema: {schema}'
+            # TODO: support minimum, maximum, exclusiveMinimum, exclusiveMaximum at least for zero
+            return self._add_primitive('root' if rule_name == 'root' else schema_type, PRIMITIVE_RULES[schema_type])
+
+    def _add_primitive(self, name: str, rule: BuiltinRule):
+        n = self._add_rule(name, rule.content)
+
+        for dep in rule.deps:
+            dep_rule = PRIMITIVE_RULES.get(dep) or STRING_FORMAT_RULES.get(dep)
+            assert dep_rule, f'Rule {dep} not known'
+            if dep not in self._rules:
+                self._add_primitive(dep, dep_rule)
+        return n
+
+    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
+        prop_order = self._prop_order
+        # sort by position in prop_order (if specified) then by original order
+        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
+
+        prop_kv_rule_names = {}
+        for prop_name, prop_schema in properties:
+            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+            prop_kv_rule_names[prop_name] = self._add_rule(
+                f'{name}{"-" if name else ""}{prop_name}-kv',
+                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
+            )
+        required_props = [k for k in sorted_props if k in required]
+        optional_props = [k for k in sorted_props if k not in required]
+
+        if additional_properties == True or isinstance(additional_properties, dict):
+            sub_name = f'{name}{"-" if name else ""}additional'
+            value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
+            prop_kv_rule_names["*"] = self._add_rule(
+                f'{sub_name}-kv',
+                self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+            )
+            optional_props.append("*")
+
+        rule = '"{" space '
+        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
+
+        if optional_props:
+            rule += ' ('
+            if required_props:
+                rule += ' "," space ( '
+
+            def get_recursive_refs(ks, first_is_optional):
+                [k, *rest] = ks
+                kv_rule_name = prop_kv_rule_names[k]
+                if k == '*':
+                    res = self._add_rule(
+                        f'{name}{"-" if name else ""}additional-kvs',
+                        f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*'
+                    )
+                elif first_is_optional:
+                    res = f'( "," space {kv_rule_name} )?'
+                else:
+                    res = kv_rule_name
+                if len(rest) > 0:
+                    res += ' ' + self._add_rule(
+                        f'{name}{"-" if name else ""}{k}-rest',
+                        get_recursive_refs(rest, first_is_optional=True)
+                    )
+                return res
+
+            rule += ' | '.join(
+                get_recursive_refs(optional_props[i:], first_is_optional=False)
+                for i in range(len(optional_props))
+            )
+            if required_props:
+                rule += ' )'
+            rule += ' )?'
+
+        rule += ' "}" space'
+
+        return rule
+
+    def format_grammar(self):
+        return '\n'.join(
+            f'{name} ::= {rule}'
+            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
+        )
+
+
+def main(args_in = None):
+    parser = argparse.ArgumentParser(
+        description='''
+            Generates a grammar (suitable for use in ./main) that produces JSON conforming to a
+            given JSON schema. Only a subset of JSON schema features are supported; more may be
+            added in the future.
+        ''',
+    )
+    parser.add_argument(
+        '--prop-order',
+        default=[],
+        type=lambda s: s.split(','),
+        help='''
+            comma-separated property names defining the order of precedence for object properties;
+            properties not specified here are given lower precedence than those that are, and
+            are kept in their original order from the schema. Required properties are always
+            given precedence over optional properties.
+        '''
+    )
+    parser.add_argument(
+        '--allow-fetch',
+        action='store_true',
+        default=False,
+        help='Whether to allow fetching referenced schemas over HTTPS')
+    parser.add_argument(
+        '--dotall',
+        action='store_true',
+        default=False,
+        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
+    parser.add_argument(
+        '--raw-pattern',
+        action='store_true',
+        default=False,
+        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
+
+    parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
+    args = parser.parse_args(args_in)
+
+    if args.schema.startswith('https://'):
+        url = args.schema
+        import requests
+        schema = requests.get(url).json()
+    elif args.schema == '-':
+        url = 'stdin'
+        schema = json.load(sys.stdin)
+    else:
+        url = f'file://{args.schema}'
+        with open(args.schema) as f:
+            schema = json.load(f)
+    converter = SchemaConverter(
+        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
+        allow_fetch=args.allow_fetch,
+        dotall=args.dotall,
+        raw_pattern=args.raw_pattern)
+    schema = converter.resolve_refs(schema, url)
+    converter.visit(schema, '')
+    print(converter.format_grammar())
+
+
+if __name__ == '__main__':
+    main()

examples/llama-bench/llama-bench.cpp

@@ -113,7 +113,7 @@ static std::string get_cpu_info() {
 
 static std::string get_gpu_info() {
     std::string id;
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
     int count = ggml_backend_cuda_get_device_count();
     for (int i = 0; i < count; i++) {
         char buf[128];
@@ -249,6 +249,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "q5_1") {
         return GGML_TYPE_Q5_1;
     }
+    if (s == "iq4_nl") {
+        return GGML_TYPE_IQ4_NL;
+    }
 
     return GGML_TYPE_COUNT;
 }
@@ -805,7 +808,7 @@ struct test {
 
 const std::string test::build_commit = LLAMA_COMMIT;
 const int         test::build_number = LLAMA_BUILD_NUMBER;
-const bool        test::cuda         = !!ggml_cpu_has_cublas();
+const bool        test::cuda         = !!ggml_cpu_has_cuda();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::kompute      = !!ggml_cpu_has_kompute();

examples/llava/MobileVLM-README.md

@@ -1,11 +1,13 @@
 # MobileVLM
 
-Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants.
+Currently this implementation supports [MobileVLM-1.7B](https://huggingface.co/mtgv/MobileVLM-1.7B) / [MobileVLM_V2-1.7B](https://huggingface.co/mtgv/MobileVLM_V2-1.7B) variants.
 
 for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM)
 
 The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
 
+Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
+
 ## Usage
 Build with cmake or run `make llava-cli` to build it.
 
@@ -20,7 +22,7 @@ After building, run: `./llava-cli` to see the usage. For example:
 
 ## Model conversion
 
-- Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally:
+1. Clone `mobileVLM-1.7B` and `clip-vit-large-patch14-336` locally:
 
 ```sh
 git clone https://huggingface.co/mtgv/MobileVLM-1.7B
@@ -34,7 +36,7 @@ git clone https://huggingface.co/openai/clip-vit-large-patch14-336
 python ./examples/llava/llava-surgery.py -m path/to/MobileVLM-1.7B
 ```
 
-3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF:
+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 \
@@ -44,6 +46,14 @@ python ./examples/llava/convert-image-encoder-to-gguf \
     --projector-type ldp
 ```
 
+```sh
+python ./examples/llava/convert-image-encoder-to-gguf \
+    -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 \
+    --projector-type ldpv2
+```
+
 4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
@@ -68,7 +78,7 @@ cd examples/llava/android/build_64
 ### run on Android
 refer to `android/adb_run.sh`, modify resources' `name` and `path`
 
-## some result on Android with `Snapdragon 888` chip
+## Some result on Android with `Snapdragon 888` chip
 ### case 1
 **input**
 ```sh
@@ -99,7 +109,6 @@ llama_print_timings:       total time =   34731.93 ms
     --image /data/local/tmp/cat.jpeg \
     -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat is in the image? ASSISTANT:"
 ```
-
 **output**
 ```sh
 encode_image_with_clip: image encoded in 21149.51 ms by CLIP (  146.87 ms per image patch)
@@ -111,12 +120,82 @@ llama_print_timings:        eval time =    1279.03 ms /    18 runs   (   71.06 m
 llama_print_timings:       total time =   34570.79 ms
 ```
 
+
+## Some result on Android with `Snapdragon 778G` chip
+### MobileVLM-1.7B case
+#### llava-cli release-b2005
+**input**
+```sh
+/data/local/tmp/llava-cli \
+    -m /data/local/tmp/ggml-model-q4_k.gguf \
+    --mmproj /data/local/tmp/mmproj-model-f16.gguf \
+    -t 4 \
+    --image /data/local/tmp/many_llamas.jpeg \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:"
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in 18728.52 ms by CLIP (  130.06 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ A group of llamas are standing in a green pasture.
+
+llama_print_timings:        load time =   20357.33 ms
+llama_print_timings:      sample time =       2.96 ms /    14 runs   (    0.21 ms per token,  4734.53 tokens per second)
+llama_print_timings: prompt eval time =    8119.49 ms /   191 tokens (   42.51 ms per token,    23.52 tokens per second)
+llama_print_timings:        eval time =    1005.75 ms /    14 runs   (   71.84 ms per token,    13.92 tokens per second)
+llama_print_timings:       total time =   28038.34 ms /   205 tokens
+```
+#### llava-cli latest-version
+**input**
+
+Just the same as above.
+
+**output**(seems to be much slower)
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in 288268.88 ms by CLIP ( 2001.87 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ It is a group of sheep standing together in a grass field.
+
+llama_print_timings:        load time =  818120.91 ms
+llama_print_timings:      sample time =       3.44 ms /    14 runs   (    0.25 ms per token,  4067.40 tokens per second)
+llama_print_timings: prompt eval time =  529274.69 ms /   191 tokens ( 2771.07 ms per token,     0.36 tokens per second)
+llama_print_timings:        eval time =   43894.02 ms /    13 runs   ( 3376.46 ms per token,     0.30 tokens per second)
+llama_print_timings:       total time =  865441.76 ms /   204 tokens
+```
+### MobileVLM_V2-1.7B case
+#### llava-cli release-2005b
+**input**
+
+Just the same as above.
+
+**output**
+```sh
+encode_image_with_clip: image encoded in 20609.61 ms by CLIP (  143.12 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ This image captures a lively scene of 20 llamas in motion on an expansive, grassy field. The llama is scattered across the landscape with some standing and others sitting down as if taking rest or observing their surroundings from different vantage points within this verdant setting.
+
+The background offers glimpses into a picturesque town nestled amidst hills under an overcast sky, adding depth to the scene while also emphasizing that distance between these llama and human-made structures like houses or roads in which they roam freely without any barriers around them. The image is framed by text at both right angles on white backgrounds against a contrasting blue backdrop with green foliage, further drawing attention to the llamas amidst their natural habitat while also inviting viewers into this picturesque landscape within town limits of Alta Llama
+
+llama_print_timings:        load time =   22406.77 ms
+llama_print_timings:      sample time =      49.26 ms /   186 runs   (    0.26 ms per token,  3776.27 tokens per second)
+llama_print_timings: prompt eval time =    9044.54 ms /   191 tokens (   47.35 ms per token,    21.12 tokens per second)
+llama_print_timings:        eval time =   14497.49 ms /   186 runs   (   77.94 ms per token,    12.83 tokens per second)
+llama_print_timings:       total time =   44411.01 ms /   377 tokens
+```
+
 ## Orin compile and run
 ### compile
 ```sh
-make LLAMA_CUBLAS=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
+make LLAMA_CUDA=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
 ```
-
 ### run on Orin
 ### case 1
 **input**
@@ -165,8 +244,121 @@ llama_print_timings:        eval time =     166.65 ms /    11 runs   (   15.15 m
 llama_print_timings:       total time =    1365.47 ms /   243 tokens
 ```
 
-## Minor shortcomings
-The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quickly, we uniformly modified `clip_n_patches` function to a quarter. when counting the time consumption, the calculated time will be 4 times bigger than the real cost.
+## Running on Intel(R) Core(TM) i7-10750H
+### Operating system
+Ubuntu22.04
+### compile
+```sh
+make -j32
+```
+### MobileVLM-1.7B case
+**input**
+```sh
+-m /path/to/ggml-model-q4_k.gguf \
+    --mmproj /path/to/mmproj-model-f16.gguf \
+    --image /path/to/many_llamas.jpeg
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:" \
+```
+**output**
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in  2730.94 ms by CLIP (   18.96 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that?ASSISTANT:
+
+ A group of llamas are walking together in a field.
+
+llama_print_timings:        load time =    5506.60 ms
+llama_print_timings:      sample time =       0.44 ms /    13 runs   (    0.03 ms per token, 29545.45 tokens per second)
+llama_print_timings: prompt eval time =    2031.58 ms /   190 tokens (   10.69 ms per token,    93.52 tokens per second)
+llama_print_timings:        eval time =     438.92 ms /    12 runs   (   36.58 ms per token,    27.34 tokens per second)
+llama_print_timings:       total time =    5990.25 ms /   202 tokens
+```
+
+### MobileVLM_V2-1.7B case
+**input**
+
+Just the same as above.
+
+**ouput**
+```sh
+encode_image_with_clip: image embedding created: 144 tokens
+
+encode_image_with_clip: image encoded in  3223.89 ms by CLIP (   22.39 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that?ASSISTANT:
+
+ The image captures a tranquil scene in a park, where a group of approximately 20 llamas are gathered. The llamas, a mix of white and black, are standing in a line, their black and white patterns contrasting with the lush green grass of the park. The lamas are arranged in a line, suggesting a social order.
+
+The park itself is lush and green, with trees dotting the landscape in the background. A sign reading "Llamas Tico  Ana" is also visible in the image, possibly indicating the location or the breed of the llamas. The image seems to be taken from a distance, providing a wide view of the scene and the surrounding environment.
+
+The llamas' positions relative to each other, the sign, and the trees create a harmonious composition. The image does not contain any discernible text. The overall scene is one of peace and natural beauty, with the llamas in their natural habitat, surrounded by the vibrant colors and lush greenery of the park.
+
+llama_print_timings:        load time =    6642.61 ms
+llama_print_timings:      sample time =       8.15 ms /   223 runs   (    0.04 ms per token, 27358.61 tokens per second)
+llama_print_timings: prompt eval time =    2475.07 ms /   190 tokens (   13.03 ms per token,    76.77 tokens per second)
+llama_print_timings:        eval time =    8760.60 ms /   222 runs   (   39.46 ms per token,    25.34 tokens per second)
+llama_print_timings:       total time =   15513.95 ms /   412 tokens
+```
+
+## Run on Intel(R) Core(TM) Ultra7 115H
+### operation system
+Windows11
+### comiple
+```sh
+make -j32
+```
+### MobileVLM-1.7B case
+**input**
+```sh
+-m /path/to/ggml-model-q4_k.gguf \
+    --mmproj /path/to/tmp/mmproj-model-f16.gguf \
+    -p "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat's that? ASSISTANT:" \
+```
+**output**
+```sh
+encode_image_with_clip: image encoded in  4902.81 ms by CLIP (   34.05 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ The image features a group of brown and white llamas standing in a grassy field.
+
+llama_print_timings:        load time =    7441.06 ms
+llama_print_timings:      sample time =       0.72 ms /    19 runs   (    0.04 ms per token, 26279.39 tokens per second)
+llama_print_timings: prompt eval time =    2090.71 ms /   191 tokens (   10.95 ms per token,    91.36 tokens per second)
+llama_print_timings:        eval time =     512.35 ms /    18 runs   (   28.46 ms per token,    35.13 tokens per second)
+llama_print_timings:       total time =    7987.23 ms /   209 tokens
+```
+
+### MobileVLM_V2-1.7B case
+**input**
+
+Just the same as above.
+
+**output**
+```sh
+encode_image_with_clip: image encoded in  4682.44 ms by CLIP (   32.52 ms per image patch)
+system_prompt: A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER:
+user_prompt: \nWhat's that? ASSISTANT:
+
+ This image captures a lively scene of a group of 14 llamas in a grassy field. The llamas, with their distinctive black and white coats, are standing and walking in a line, seemingly engaged in a social activity. One
+ of them, possibly the first in the line, has its back turned, perhaps observing something in the distance.
+
+The llama in the front of the line stands out due to its black and white coloring, which is quite unusual for llama patterns. The llama in the front also seems to be more aware of its surroundings, as it faces the camera, giving a sense of engagement with the viewer.
+
+The image is taken from the side of the llama, providing a clear view of the llama in the front and its companions. The lameness in the llama in
+ front is not visible, indicating that it might not be the main focus of the photo.
+
+The background of the image features a grassy field, with a fence and a tree visible in the distance. The tree appears to be bare, suggesting that it might be during a time of year when most trees are dormant or have shed their leaves.
+
+
+llama_print_timings:        load time =    7015.35 ms
+llama_print_timings:      sample time =      10.61 ms /   256 runs   (    0.04 ms per token, 24119.09 tokens per second)
+llama_print_timings: prompt eval time =    2052.45 ms /   191 tokens (   10.75 ms per token,    93.06 tokens per second)
+llama_print_timings:        eval time =    7259.43 ms /   255 runs   (   28.47 ms per token,    35.13 tokens per second)
+llama_print_timings:       total time =   14371.19 ms /   446 tokens
+```
 
 ## TODO
 
@@ -181,5 +373,5 @@ The `n_patch` of output in `ldp` is 1/4 of the input. In order to implement quic
 
 ## contributor
 ```sh
-zhangjidong05, yangyang260, huyiming03, chenxiaotao03
+zhangjidong05, yangyang260, huyiming03, chenxiaotao03, ZiangWu-77
 ```

examples/llava/README.md

@@ -24,7 +24,7 @@ After building, run: `./llava-cli` to see the usage. For example:
 
 ## LLaVA 1.5
 
-- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
+1. Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b

examples/llava/clip.cpp

@@ -7,7 +7,7 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
@@ -119,6 +119,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_LLAVA_PROJ      "mm.%d.%s"
 #define TN_MVLM_PROJ_MLP   "mm.model.mlp.%d.%s"
 #define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
+#define TN_MVLM_PROJ_PEG   "mm.model.peg.%d.%s"
 #define TN_IMAGE_NEWLINE   "model.image_newline"
 
 
@@ -126,12 +127,14 @@ enum projector_type {
     PROJECTOR_TYPE_MLP,
     PROJECTOR_TYPE_MLP_NORM,
     PROJECTOR_TYPE_LDP,
+    PROJECTOR_TYPE_LDPV2,
     PROJECTOR_TYPE_UNKNOWN,
 };
 
 static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
     { PROJECTOR_TYPE_MLP, "mlp" },
     { PROJECTOR_TYPE_LDP, "ldp" },
+    { PROJECTOR_TYPE_LDPV2, "ldpv2"},
 };
 
 
@@ -475,6 +478,14 @@ struct clip_vision_model {
     struct ggml_tensor * mm_model_block_2_block_2_0_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_w;
     struct ggml_tensor * mm_model_block_2_block_2_1_b;
+
+    // MobileVLM_V2 projection
+    struct ggml_tensor * mm_model_mlp_0_w;
+    struct ggml_tensor * mm_model_mlp_0_b;
+    struct ggml_tensor * mm_model_mlp_2_w;
+    struct ggml_tensor * mm_model_mlp_2_b;
+    struct ggml_tensor * mm_model_peg_0_w;
+    struct ggml_tensor * mm_model_peg_0_b;
 };
 
 struct clip_ctx {
@@ -807,6 +818,30 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             }
             embeddings = block_1;
         }
+        else if (ctx->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            int n_patch = 24;
+            struct ggml_tensor * mlp_0 = ggml_mul_mat(ctx0, model.mm_model_mlp_0_w, embeddings);
+            mlp_0 = ggml_add(ctx0, mlp_0, model.mm_model_mlp_0_b);
+            mlp_0 = ggml_gelu(ctx0, mlp_0);
+            struct ggml_tensor * mlp_2 = ggml_mul_mat(ctx0, model.mm_model_mlp_2_w, mlp_0);
+            mlp_2 = ggml_add(ctx0, mlp_2, model.mm_model_mlp_2_b);
+            // mlp_2 ne = [2048, 576, 1, 1]
+            // // AVG Pool Layer 2*2, strides = 2
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 0, 2, 3));
+            // mlp_2 ne = [576, 2048, 1, 1]
+            mlp_2 = ggml_reshape_4d(ctx0, mlp_2, n_patch, n_patch, mlp_2->ne[1], mlp_2->ne[2]);
+            // mlp_2 ne [24, 24, 2048, 1]
+            mlp_2 = ggml_pool_2d(ctx0, mlp_2, GGML_OP_POOL_AVG, 2, 2, 2, 2, 0, 0);
+            // weight ne = [3, 3, 2048, 1]
+            struct ggml_tensor * peg_0 = ggml_conv_depthwise_2d(ctx0, model.mm_model_peg_0_w, mlp_2, 1, 1, 1, 1, 1, 1);
+            peg_0 = ggml_cont(ctx0, ggml_permute(ctx0, peg_0, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, model.mm_model_peg_0_b);
+            mlp_2 = ggml_cont(ctx0, ggml_permute(ctx0, mlp_2, 1, 2, 0, 3));
+            peg_0 = ggml_add(ctx0, peg_0, mlp_2);
+            peg_0 = ggml_reshape_3d(ctx0, peg_0, peg_0->ne[0], peg_0->ne[1] * peg_0->ne[2], peg_0->ne[3]);
+            embeddings = peg_0;
+        }
         else {
             GGML_ASSERT(false);
         }
@@ -934,7 +969,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
     }
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
     new_clip->backend = ggml_backend_cuda_init(0);
     printf("%s: CLIP using CUDA backend\n", __func__);
 #endif
@@ -1177,7 +1212,18 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             vision_model.mm_model_block_2_block_2_0_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "0.weight"));
             vision_model.mm_model_block_2_block_2_1_w   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.weight"));
             vision_model.mm_model_block_2_block_2_1_b   = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_BLOCK, 2, 2, "1.bias"));
-        } else {
+        }
+        else if (new_clip->proj_type == PROJECTOR_TYPE_LDPV2)
+        {
+            // MobilVLM_V2 projection
+            vision_model.mm_model_mlp_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "weight"));
+            vision_model.mm_model_mlp_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 0, "bias"));
+            vision_model.mm_model_mlp_2_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "weight"));
+            vision_model.mm_model_mlp_2_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_MLP, 2, "bias"));
+            vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
+            vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
+        }
+        else {
             std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
             throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
         }
@@ -1710,7 +1756,7 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
     int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);
 
-    if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
+    if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
         n_patches /= 4;
     }
 
@@ -1966,6 +2012,9 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
     if (ctx->proj_type == PROJECTOR_TYPE_LDP) {
         return ctx->vision_model.mm_model_block_1_block_2_1_b->ne[0];
     }
+    if (ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
+        return ctx->vision_model.mm_model_peg_0_b->ne[0];
+    }
     if (ctx->proj_type == PROJECTOR_TYPE_MLP) {
         return ctx->vision_model.mm_2_b->ne[0];
     }

examples/llava/convert-image-encoder-to-gguf.py

@@ -1,6 +1,7 @@
 import argparse
 import os
 import json
+import re
 
 import torch
 import numpy as np
@@ -38,9 +39,11 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b
 def get_tensor_name(name: str) -> str:
     if "projection" in name:
         return name
-
     if "mm_projector" in name:
-        return name.replace("model.mm_projector", "mm")
+        name = name.replace("model.mm_projector", "mm")
+        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
+        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
+        return name
 
     return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
 
@@ -83,7 +86,7 @@ ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
 ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
                 help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
-ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
 # Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5

examples/llava/llava-cli.cpp

@@ -146,7 +146,6 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     int n_past = 0;
 
     const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx_llava->ctx_llama));
 
     std::string system_prompt, user_prompt;
     size_t image_pos = prompt.find("<image>");
@@ -180,7 +179,7 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         }
     }
 
-    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
+    eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
     llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
     eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
 

examples/lookahead/lookahead.cpp

@@ -64,13 +64,10 @@ int main(int argc, char ** argv) {
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // Tokenize the prompt
-    const bool add_bos = llama_should_add_bos_token(model);
-    LOG("add_bos tgt: %d\n", add_bos);
-
     std::vector<llama_token> inp;
     std::vector<llama_token> all;
 
-    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
     all = inp;
 
     const int max_context_size     = llama_n_ctx(ctx);

examples/lookup/CMakeLists.txt

@@ -3,3 +3,21 @@ add_executable(${TARGET} lookup.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-create)
+add_executable(${TARGET} lookup-create.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-merge)
+add_executable(${TARGET} lookup-merge.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET lookup-stats)
+add_executable(${TARGET} lookup-stats.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/lookup/lookup-create.cpp

@@ -0,0 +1,41 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    GGML_ASSERT(model != nullptr);
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+    fprintf(stderr, "%s: tokenization done\n", __func__);
+
+
+    llama_ngram_cache ngram_cache;
+    llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
+    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
+
+    llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
+}

examples/lookup/lookup-merge.cpp

@@ -0,0 +1,47 @@
+#include "ggml.h"
+#include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
+
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <iostream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+static void print_usage() {
+    fprintf(stderr, "Merges multiple lookup cache files into a single one.\n");
+    fprintf(stderr, "Usage: lookup-merge [--help] lookup_part_1.bin lookup_part_2.bin ... lookup_merged.bin\n");
+}
+
+int main(int argc, char ** argv){
+    if (argc < 3) {
+        print_usage();
+        exit(1);
+    }
+
+    std::vector<std::string> args;
+    args.resize(argc-1);
+    for (int i = 0; i < argc-1; ++i) {
+        args[i] = argv[i+1];
+        if (args[i] == "-h" || args[i] == "--help") {
+            print_usage();
+            exit(0);
+        }
+    }
+
+    fprintf(stderr, "lookup-merge: loading file %s\n", args[0].c_str());
+    llama_ngram_cache ngram_cache_merged = llama_ngram_cache_load(args[0]);
+
+    for (size_t i = 1; i < args.size()-1; ++i) {
+        fprintf(stderr, "lookup-merge: loading file %s\n", args[i].c_str());
+        llama_ngram_cache ngram_cache = llama_ngram_cache_load(args[i]);
+
+        llama_ngram_cache_merge(ngram_cache_merged, ngram_cache);
+    }
+
+    fprintf(stderr, "lookup-merge: saving file %s\n", args.back().c_str());
+    llama_ngram_cache_save(ngram_cache_merged, args.back());
+}

examples/lookup/lookup-stats.cpp

@@ -0,0 +1,160 @@
+#include "ggml.h"
+#include "common.h"
+#include "llama.h"
+#include "log.h"
+#include "ngram-cache.h"
+
+#include <cmath>
+#include <cstdint>
+#include <cstdio>
+#include <fstream>
+#include <string>
+#include <vector>
+#include <unordered_map>
+
+int main(int argc, char ** argv){
+    gpt_params params;
+
+    if (!gpt_params_parse(argc, argv, params)) {
+        return 1;
+    }
+
+    const int n_draft = params.n_draft;
+
+    // init llama.cpp
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model = NULL;
+    llama_context * ctx = NULL;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
+
+    // tokenize the prompt
+    std::vector<llama_token> inp;
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        const int64_t t_start_draft_us = ggml_time_us();
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
+
+    const int n_input = inp.size();
+    const int n_ctx = params.n_ctx;
+
+    int n_drafted = 0;
+    int n_accept  = 0;
+
+    const int64_t t_start_ms = ggml_time_ms();
+
+    // Iterate over input tokens in chunks of size n_ctx.
+    // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
+    for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
+        const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
+        std::vector<llama_token> pseudo_output;
+        pseudo_output.push_back(inp_slice[0]);
+
+        while ((int) pseudo_output.size() < n_ctx) {
+            // Simulate drafting and decoding from draft:
+            std::vector<llama_token> draft;
+            draft.push_back(pseudo_output.back());
+
+            {
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
+
+            n_drafted += draft.size() - 1;
+
+            for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
+                const llama_token ground_truth = inp_slice[pseudo_output.size()];
+                const llama_token drafted = draft[j];
+
+                if (ground_truth != drafted) {
+                    break;
+                }
+
+                ++n_accept;
+                pseudo_output.push_back(ground_truth);
+
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            // After each simulated batch decoding simulate the sampling of a single token:
+            if ((int) pseudo_output.size() < n_ctx) {
+                pseudo_output.push_back(inp_slice[pseudo_output.size()]);
+                {
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
+            }
+
+            draft.erase(draft.begin());
+
+        }
+        if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
+            const int64_t t_now_ms = ggml_time_ms();
+            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
+            const int64_t eta_min  = eta_ms / (60*1000);
+            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
+
+            LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
+        }
+
+        // After each chunk, update the dynamic ngram cache with the context ngram cache:
+        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+        ngram_cache_context.clear();
+    }
+
+    LOG_TEE("\n");
+
+    LOG_TEE("\n");
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_input - n_input % n_ctx);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    fprintf(stderr, "\n\n");
+
+    return 0;
+}

examples/lookup/lookup.cpp

@@ -1,12 +1,15 @@
-#include "common.h"
 #include "ggml.h"
 #include "llama.h"
+#include "common.h"
+#include "ngram-cache.h"
 
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
+#include <fstream>
 #include <string>
 #include <vector>
+#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
@@ -15,11 +18,7 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    // max/min n-grams size to search for in prompt
-    const int ngram_max = 4;
-    const int ngram_min = 1;
-
-    // length of the candidate / draft sequence, if match is found
+    // max. number of additional tokens to draft if match is found
     const int n_draft = params.n_draft;
 
     const bool dump_kv_cache = params.dump_kv_cache;
@@ -39,13 +38,41 @@ int main(int argc, char ** argv){
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    llama_set_rng_seed(ctx, params.seed);
+    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
-    const bool add_bos = llama_should_add_bos_token(model);
-    LOG("add_bos tgt: %d\n", add_bos);
-
     std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
+
+    llama_ngram_cache ngram_cache_context;
+    llama_ngram_cache ngram_cache_dynamic;
+    llama_ngram_cache ngram_cache_static;
+    int64_t t_draft_flat_us = 0;
+    int64_t t_draft_us = 0;
+
+    {
+        // Fill up context ngram cache with tokens from user input:
+        const int64_t t_start_draft_us = ggml_time_us();
+        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
+
+        if (!params.lookup_cache_static.empty()) {
+            try {
+                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
+            } catch (std::ifstream::failure const &) {
+                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
+                exit(1);
+            }
+        }
+
+        if (!params.lookup_cache_dynamic.empty()) {
+            try {
+                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
+            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
+        }
+
+        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
+    }
 
     const int max_context_size     = llama_n_ctx(ctx);
     const int max_tokens_list_size = max_context_size - 4;
@@ -76,8 +103,6 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
-    int64_t t_draft_us = 0;
-
     int n_past = inp.size();
 
     bool has_eos = false;
@@ -129,6 +154,12 @@ int main(int argc, char ** argv){
                 ++n_past;
                 ++i_dft;
                 inp.push_back(id);
+                {
+                    // Update context ngram cache with the newly accepted token:
+                    const int64_t t_start_draft_us = ggml_time_us();
+                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                    t_draft_us += ggml_time_us() - t_start_draft_us;
+                }
 
                 if (params.use_color) {
                     // color accepted draft token
@@ -149,6 +180,12 @@ int main(int argc, char ** argv){
             draft.clear();
             draft.push_back(id);
             inp.push_back(id);
+            {
+                // Update context ngram cache with the newly accepted token:
+                const int64_t t_start_draft_us = ggml_time_us();
+                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
+                t_draft_us += ggml_time_us() - t_start_draft_us;
+            }
             break;
         }
 
@@ -163,44 +200,19 @@ int main(int argc, char ** argv){
         llama_batch_clear(batch_tgt);
         llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
 
-        // generate n_pred tokens through prompt lookup
-        auto prompt_lookup = [&]() -> void {
-            const int inp_size = inp.size();
-            for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
-                const llama_token * ngram = &inp[inp_size - ngram_size];
-
-                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
-                    bool match = true;
-                    for (int j = 0; j < ngram_size; ++j) {
-                        if (inp[i + j] != ngram[j]) {
-                            match = false;
-                            break;
-                        }
-                    }
-
-                    if (match) {
-                        const int startIdx = i + ngram_size;
-                        const int endIdx = startIdx + n_draft;
-                        if (endIdx < inp_size) {
-                            for (int j = startIdx; j < endIdx; ++j) {
-                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
-                                draft.push_back(inp[j]);
-                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
-                                ++n_drafted;
-                            }
-                            return;
-                        }
-                    }
-                }
-            }
-            return;
-        };
-
+        // Draft already contains a single token sampled from the model:
+        GGML_ASSERT(draft.size() == 1);
+        GGML_ASSERT(draft[0] == inp.back());
         const int64_t t_start_draft_us = ggml_time_us();
 
-        prompt_lookup();
+        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
+
+        for (size_t i = 1; i < draft.size(); ++i) {
+            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
+        }
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
+        n_drafted += draft.size() - 1;
 
         llama_decode(ctx, batch_tgt);
         ++n_past;
@@ -210,19 +222,24 @@ int main(int argc, char ** argv){
 
     auto t_dec_end = ggml_time_us();
 
+    // Update dynamic ngram cache with context ngram cache and save it to disk:
+    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
+    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
+
     LOG_TEE("\n\n");
 
     LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
     LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_TEE("\n");
-    LOG_TEE("n_draft   = %d\n", n_draft);
-    LOG_TEE("n_predict = %d\n", n_predict);
-    LOG_TEE("n_drafted = %d\n", n_drafted);
-    LOG_TEE("t_draft   = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+    LOG_TEE("n_draft      = %d\n", n_draft);
+    LOG_TEE("n_predict    = %d\n", n_predict);
+    LOG_TEE("n_drafted    = %d\n", n_drafted);
+    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
+    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
             t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
-    LOG_TEE("n_accept  = %d\n", n_accept);
-    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
+    LOG_TEE("n_accept     = %d\n", n_accept);
+    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx);

examples/main-cmake-pkg/README.md

@@ -8,7 +8,7 @@ Because this example is "outside of the source tree", it is important to first b
 
 ### Considerations
 
-When hardware acceleration libraries are used (e.g. CUBlas, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
+When hardware acceleration libraries are used (e.g. CUDA, Metal, CLBlast, etc.), CMake must be able to locate the associated CMake package. In the example below, when building _main-cmake-pkg_ notice the `CMAKE_PREFIX_PATH` includes the Llama CMake package location _in addition to_ the CLBlast package—which was used when compiling _llama.cpp_.
 
 ### Build llama.cpp and install to C:\LlamaCPP directory
 

examples/main/README.md

@@ -296,7 +296,9 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Batch Size
 
--   `-b N, --batch-size N`: Set the batch size for prompt processing (default: 512). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+-   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
+
+- `-ub N`, `--ubatch-size N`: physical maximum batch size. This is for pipeline parallelization. Default: `512`.
 
 ### Prompt Caching
 
@@ -308,7 +310,7 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 ### Quantization
 
-For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-data--run).
+For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
 ## Additional Options
 
@@ -316,8 +318,8 @@ These options provide extra functionality and customization when running the LLa
 
 -   `-h, --help`: Display a help message showing all available options and their default values. This is particularly useful for checking the latest options and default values, as they can change frequently, and the information in this document may become outdated.
 -   `--verbose-prompt`: Print the prompt before generating text.
--   `-ngl N, --n-gpu-layers N`: When compiled with appropriate support (currently CLBlast or cuBLAS), this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
--   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
--   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
+-   `-ngl N, --n-gpu-layers N`: When compiled with GPU support, this option allows offloading some layers to the GPU for computation. Generally results in increased performance.
+-   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
+-   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
 -   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -235,7 +235,7 @@ int main(int argc, char ** argv) {
             // The file exists and is not empty
             session_tokens.resize(n_ctx);
             size_t n_token_count_out = 0;
-            if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
+            if (!llama_state_load_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
                 LOG_TEE("%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
                 return 1;
             }
@@ -246,6 +246,7 @@ int main(int argc, char ** argv) {
     }
 
     const bool add_bos = llama_should_add_bos_token(model);
+    GGML_ASSERT(llama_add_eos_token(model) != 1);
     LOG("add_bos: %d\n", add_bos);
 
     std::vector<llama_token> embd_inp;
@@ -255,7 +256,7 @@ int main(int argc, char ** argv) {
         if (params.chatml) {
             params.prompt = "<|im_start|>system\n" + params.prompt + "<|im_end|>";
         }
-        embd_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+        embd_inp = ::llama_tokenize(ctx, params.prompt, true, true);
     } else {
         LOG("use session tokens\n");
         embd_inp = session_tokens;
@@ -277,10 +278,10 @@ int main(int argc, char ** argv) {
     if (ctx_guidance) {
         LOG("cfg_negative_prompt: \"%s\"\n", log_tostr(sparams.cfg_negative_prompt));
 
-        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, add_bos, true);
+        guidance_inp = ::llama_tokenize(ctx_guidance, sparams.cfg_negative_prompt, true, true);
         LOG("guidance_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx_guidance, guidance_inp).c_str());
 
-        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+        std::vector<llama_token> original_inp = ::llama_tokenize(ctx, params.prompt, true, true);
         LOG("original_inp tokenized: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, original_inp).c_str());
 
         original_prompt_len = original_inp.size();
@@ -339,14 +340,14 @@ int main(int argc, char ** argv) {
     }
 
     // prefix & suffix for instruct mode
-    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", add_bos, true);
-    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false,   true);
+    const auto inp_pfx = ::llama_tokenize(ctx, "\n\n### Instruction:\n\n", true,  true);
+    const auto inp_sfx = ::llama_tokenize(ctx, "\n\n### Response:\n\n",    false, true);
 
     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());
 
     // chatml prefix & suffix
-    const auto cml_pfx = ::llama_tokenize(ctx, "\n<|im_start|>user\n", add_bos, true);
+    const auto cml_pfx = ::llama_tokenize(ctx, "\n<|im_start|>user\n", true, true);
     const auto cml_sfx = ::llama_tokenize(ctx, "<|im_end|>\n<|im_start|>assistant\n", false, true);
 
     LOG("cml_pfx: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, cml_pfx).c_str());
@@ -693,7 +694,7 @@ int main(int argc, char ** argv) {
             // optionally save the session on first sample (for faster prompt loading next time)
             if (!path_session.empty() && need_to_save_session && !params.prompt_cache_ro) {
                 need_to_save_session = false;
-                llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+                llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
 
                 LOG("saved session to %s\n", path_session.c_str());
             }
@@ -935,7 +936,7 @@ int main(int argc, char ** argv) {
 
     if (!path_session.empty() && params.prompt_cache_all && !params.prompt_cache_ro) {
         LOG_TEE("\n%s: saving final output to session file '%s'\n", __func__, path_session.c_str());
-        llama_save_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
+        llama_state_save_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.size());
     }
 
     llama_print_timings(ctx);

examples/parallel/parallel.cpp

@@ -132,7 +132,6 @@ int main(int argc, char ** argv) {
     llama_context * ctx = NULL;
 
     // load the target model
-    params.logits_all = true;
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
 
     // load the prompts from an external file if there are any

examples/perplexity/README.md

@@ -3,19 +3,18 @@
 TODO
 
 ## Llama 2 70B Scorechart
-Quantization | Model size (GiB) | Perplexity | Delta to fp16
--- | -- | -- | --
-Q4_0 | 36.20 | 3.5550 | 3.61%
-Q4_1 | 40.20 | 3.5125 | 2.37%
-Q5_0 | 44.20 | 3.4744 | 1.26%
-Q2_K | 27.27 | 3.7339 | 8.82%
-Q3_K_S | 27.86 | 3.7019 | 7.89%
-Q3_K_M | 30.83 | 3.5932 | 4.72%
-Q3_K_L | 33.67 | 3.5617 | 3.80%
-Q4_K_S | 36.39 | 3.4852 | 1.57%
-Q4_K_M | 38.54 | 3.4725 | 1.20%
-Q5_K_S | 44.20 | 3.4483 | 0.50%
-Q5_K_M | 45.41 | 3.4451 | 0.40%
-Q6_K | 52.70 | 3.4367 | 0.16%
-fp16 | 128.5 | 3.4313 | -
-
+| Quantization | Model size (GiB) | Perplexity | Delta to fp16 |
+|--------------|------------------|------------|---------------|
+| Q4_0         | 36.20            | 3.5550     | 3.61%         |
+| Q4_1         | 40.20            | 3.5125     | 2.37%         |
+| Q5_0         | 44.20            | 3.4744     | 1.26%         |
+| Q2_K         | 27.27            | 3.7339     | 8.82%         |
+| Q3_K_S       | 27.86            | 3.7019     | 7.89%         |
+| Q3_K_M       | 30.83            | 3.5932     | 4.72%         |
+| Q3_K_L       | 33.67            | 3.5617     | 3.80%         |
+| Q4_K_S       | 36.39            | 3.4852     | 1.57%         |
+| Q4_K_M       | 38.54            | 3.4725     | 1.20%         |
+| Q5_K_S       | 44.20            | 3.4483     | 0.50%         |
+| Q5_K_M       | 45.41            | 3.4451     | 0.40%         |
+| Q6_K         | 52.70            | 3.4367     | 0.16%         |
+| fp16         | 128.5            | 3.4313     | -             |

examples/perplexity/perplexity.cpp

@@ -315,10 +315,11 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, true);
 
     const int n_ctx = llama_n_ctx(ctx);
 
@@ -380,6 +381,7 @@ static results_perplexity perplexity_v2(llama_context * ctx, const gpt_params &
             const int batch_size  = std::min(end - batch_start, n_batch);
 
             //fprintf(stderr, "    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
+            // TODO: use llama_batch.logits 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__);
                 return {tokens, -1, logit_history, prob_history};
@@ -453,6 +455,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     // BOS tokens will be added for each chunk before eval
 
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     std::ofstream logits_stream;
     if (!params.logits_file.empty()) {
@@ -469,7 +472,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
     auto tim1 = std::chrono::high_resolution_clock::now();
     fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
 
-    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+    std::vector<llama_token> tokens = ::llama_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());
@@ -552,6 +555,8 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
+            int n_outputs = 0;
+
             batch.n_tokens = 0;
             for (int seq = 0; seq < n_seq_batch; seq++) {
                 int seq_start = batch_start + seq*n_ctx;
@@ -566,11 +571,13 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 
                 for (int k = 0; k < batch_size; ++k) {
                     const int idx = seq*n_ctx + k;
-                    batch.token[idx] = tokens[seq_start + k];
-                    batch.pos[idx] = j*n_batch + k;
-                    batch.n_seq_id[idx] = 1;
-                    batch.seq_id[idx][0] = seq;
-                    batch.logits[idx] = batch.pos[idx] >= first ? 1 : 0;
+                    batch.token   [idx]    = tokens[seq_start + k];
+                    batch.pos     [idx]    = j*n_batch + k;
+                    batch.n_seq_id[idx]    = 1;
+                    batch.seq_id  [idx][0] = seq;
+                    batch.logits  [idx]    = batch.pos[idx] >= first ? 1 : 0;
+
+                    n_outputs += batch.logits[idx] != 0;
                 }
                 batch.n_tokens += batch_size;
 
@@ -583,9 +590,9 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
                 return {tokens, -1, logit_history, prob_history};
             }
 
-            if (num_batches > 1) {
+            if (num_batches > 1 && n_outputs > 0) {
                 const auto * batch_logits = llama_get_logits(ctx);
-                logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+                logits.insert(logits.end(), batch_logits, batch_logits + n_outputs * n_vocab);
             }
         }
 
@@ -604,14 +611,15 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
         }
 
         for (int seq = 0; seq < n_seq_batch; seq++) {
-            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx);
+            const float * all_logits = num_batches > 1 ? logits.data() : llama_get_logits_ith(ctx, seq*n_ctx + first);
+
             llama_token * tokens_data = tokens.data() + start + seq*n_ctx + first;
             if (!params.logits_file.empty()) {
-                process_logits(logits_stream, n_vocab, all_logits + first*n_vocab,
+                process_logits(logits_stream, n_vocab, all_logits,
                         tokens_data, n_ctx - 1 - first,
                         workers, log_probs, nll, nll2);
             } else {
-                process_logits(n_vocab, all_logits + first*n_vocab,
+                process_logits(n_vocab, all_logits,
                         tokens_data, n_ctx - 1 - first,
                         workers, nll, nll2,
                         logit_history.data() + start + seq*n_ctx + first,
@@ -652,6 +660,7 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
 }
 
 static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<float> & batch_logits, int32_t n_batch, int32_t n_vocab) {
+    int prev_outputs = 0;
     for (int32_t i = 0; i < (int32_t) batch.n_tokens; i += n_batch) {
         const int32_t n_tokens = std::min(n_batch, (int32_t) (batch.n_tokens - i));
 
@@ -672,7 +681,14 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             return false;
         }
 
-        memcpy(batch_logits.data() + i*n_vocab, llama_get_logits(ctx), n_tokens*n_vocab*sizeof(float));
+        int n_outputs = 0;
+        for (int i = 0; i < n_tokens; ++i) {
+            n_outputs += batch_view.logits[i] != 0;
+        }
+
+        memcpy(batch_logits.data() + prev_outputs*n_vocab, llama_get_logits(ctx), n_outputs*n_vocab*sizeof(float));
+
+        prev_outputs += n_outputs;
     }
 
     return true;
@@ -757,9 +773,6 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     const bool is_spm = llama_vocab_type(llama_get_model(ctx)) == LLAMA_VOCAB_TYPE_SPM;
     fprintf(stderr, "================================= is_spm = %d\n", is_spm);
 
-    // This is needed as usual for LLaMA models
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-
     // The tasks should be randomized so the score stabilizes quickly.
     bool randomize_tasks = true;
 
@@ -779,7 +792,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         size_t ending_logprob_count[4];
         double ending_logprob[4];
 
-        size_t i_batch;         // starting index in the llama_batch
+        size_t i_logits;        // starting index of logits in the llama_batch
         size_t common_prefix;   // max number of initial tokens that are the same in all sentences
         size_t required_tokens; // needed number of tokens to evaluate all 4 endings
         std::vector<llama_token> seq_tokens[4];
@@ -804,7 +817,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         hs_cur.gold_ending_idx = std::stoi( prompt_lines[idx*6+1] );
         for (size_t j = 0; j < 4; j++) {
             hs_cur.ending[j] = prompt_lines[idx*6+2+j];
-            hs_cur.seq_tokens[j] = ::llama_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], add_bos);
+            hs_cur.seq_tokens[j] = ::llama_tokenize(ctx, hs_cur.context + " " + hs_cur.ending[j], true);
         }
 
         // determine the common prefix of the endings
@@ -823,7 +836,7 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             hs_cur.seq_tokens[2].size() - hs_cur.common_prefix +
             hs_cur.seq_tokens[3].size() - hs_cur.common_prefix;
 
-        //GGML_ASSERT(hs_cur.common_prefix >= ::llama_tokenize(ctx, hs_cur.context, add_bos).size());
+        //GGML_ASSERT(hs_cur.common_prefix >= ::llama_tokenize(ctx, hs_cur.context, true).size());
 
         // Delete the selected random example from the prompt
         if (randomize_tasks) {
@@ -844,9 +857,10 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
     const int max_tasks_per_batch = 32;
     const int max_seq = std::min(4*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 4);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -857,16 +871,17 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         while (n_cur + (int) hs_data[i1].required_tokens <= n_ctx) {
             auto & hs_cur = hs_data[i1];
+            int n_logits = 0;
 
             const int s0 = 4*(i1 - i0);
             if (s0 + 4 > max_seq) {
@@ -874,18 +889,23 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
             }
 
             for (size_t i = 0; i < hs_cur.common_prefix; ++i) {
-                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3}, false);
+                llama_batch_add(batch, hs_cur.seq_tokens[0][i], i, { s0 + 0, s0 + 1, s0 + 2, s0 + 3 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < 4; ++s) {
-                for (size_t i = hs_cur.common_prefix; i < hs_cur.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = hs_cur.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = hs_cur.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, hs_cur.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
-            hs_cur.i_batch = i_batch;
-            i_batch += hs_cur.required_tokens;
+            hs_cur.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += hs_data[i1].required_tokens;
             if (++i1 == hs_task_count) {
@@ -911,12 +931,11 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
-            size_t li = hs_cur.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             for (int s = 0; s < 4; ++s) {
                 for (size_t j = hs_cur.common_prefix; j < hs_cur.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.emplace_back(hs_cur.i_batch + li++, hs_cur.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(hs_cur.i_logits + li++, hs_cur.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -928,7 +947,8 @@ static void hellaswag_score(llama_context * ctx, const gpt_params & params) {
         for (size_t i = i0; i < i1; ++i) {
             auto & hs_cur = hs_data[i];
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(hs_cur.i_batch + hs_cur.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*hs_cur.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -978,7 +998,7 @@ struct winogrande_entry {
     std::array<std::string, 2> choices;
     int answer;
 
-    size_t i_batch;
+    size_t i_logits;
     size_t common_prefix;
     size_t required_tokens;
     size_t n_base1; // number of tokens for context + choice 1
@@ -1089,12 +1109,9 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
     fprintf(stderr, "%s : tokenizing selected tasks\n", __func__);
 
-    // This is needed as usual for LLaMA models
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-
     for (auto & task : data) {
-        task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, add_bos);
-        task.seq_tokens[1] = ::llama_tokenize(ctx, task.first + task.choices[1] + task.second, add_bos);
+        task.seq_tokens[0] = ::llama_tokenize(ctx, task.first + task.choices[0] + task.second, true);
+        task.seq_tokens[1] = ::llama_tokenize(ctx, task.first + task.choices[1] + task.second, true);
 
         task.common_prefix = 0;
         for (size_t k = 0; k < task.seq_tokens[0].size(); k++) {
@@ -1104,12 +1121,13 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
             task.common_prefix++;
         }
 
+        // TODO: the last token of each of the sequences don't need to be evaluated
         task.required_tokens = task.common_prefix +
             task.seq_tokens[0].size() - task.common_prefix +
             task.seq_tokens[1].size() - task.common_prefix;
 
-        task.n_base1 = ::llama_tokenize(ctx, task.first + task.choices[0], add_bos).size();
-        task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], add_bos).size();
+        task.n_base1 = ::llama_tokenize(ctx, task.first + task.choices[0], true).size();
+        task.n_base2 = ::llama_tokenize(ctx, task.first + task.choices[1], true).size();
     }
 
     fprintf(stderr, "%s : calculating winogrande score over selected tasks.\n", __func__);
@@ -1121,9 +1139,10 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
     const int max_tasks_per_batch = 128;
     const int max_seq = std::min(2*max_tasks_per_batch, (int) llama_n_seq_max(ctx));
 
-    llama_batch batch = llama_batch_init(n_ctx, 0, max_seq);
+    llama_batch batch = llama_batch_init(n_ctx, 0, 2);
 
     std::vector<float> tok_logits(n_vocab);
+    // TODO: this could be made smaller; it's currently the worst-case size
     std::vector<float> batch_logits(n_vocab*n_ctx);
 
     std::vector<std::pair<size_t, llama_token>> eval_pairs;
@@ -1137,29 +1156,33 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0;
+        size_t i_logits = 0;
 
         llama_batch_clear(batch);
 
         while (n_cur + (int) data[i1].required_tokens <= n_ctx) {
+            int n_logits = 0;
             const int s0 = 2*(i1 - i0);
             if (s0 + 2 > max_seq) {
                 break;
             }
 
             for (size_t i = 0; i < data[i1].common_prefix; ++i) {
-                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1}, false);
+                llama_batch_add(batch, data[i1].seq_tokens[0][i], i, { s0 + 0, s0 + 1 }, false);
             }
             batch.logits[batch.n_tokens - 1] = true;
+            n_logits += 1;
 
             for (int s = 0; s < 2; ++s) {
+                // TODO: end before the last token, no need to predict past the end of the sequences
                 for (size_t i = data[i1].common_prefix; i < data[i1].seq_tokens[s].size(); ++i) {
                     llama_batch_add(batch, data[i1].seq_tokens[s][i], i, { s0 + s }, true);
+                    n_logits += 1;
                 }
             }
 
-            data[i1].i_batch = i_batch;
-            i_batch += data[i1].required_tokens;
+            data[i1].i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += data[i1].required_tokens;
             if (++i1 == data.size()) {
@@ -1190,15 +1213,16 @@ static void winogrande_score(llama_context * ctx, const gpt_params & params) {
 
             const auto& n_base1 = skip_choice ? task.n_base1 : task.common_prefix;
             const int last_1st = task.seq_tokens[0].size() - n_base1 > 1 ? 1 : 0;
-            size_t li = n_base1 - 1;
+            size_t li = n_base1 - task.common_prefix;
             for (size_t j = n_base1-1; j < task.seq_tokens[0].size()-1-last_1st; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[0][j+1]);
+                eval_pairs.emplace_back(task.i_logits + li++, task.seq_tokens[0][j+1]);
             }
             const auto& n_base2 = skip_choice ? task.n_base2 : task.common_prefix;
             const int last_2nd = task.seq_tokens[1].size() - n_base2 > 1 ? 1 : 0;
-            li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - 1;
+            // FIXME: this uses the wrong first logits when not skipping the choice word
+            li = task.seq_tokens[0].size() - task.common_prefix + n_base2 - task.common_prefix;
             for (size_t j = n_base2-1; j < task.seq_tokens[1].size()-1-last_2nd; ++j) {
-                eval_pairs.emplace_back(task.i_batch + li++, task.seq_tokens[1][j+1]);
+                eval_pairs.emplace_back(task.i_logits + li++, task.seq_tokens[1][j+1]);
             }
         }
         compute_logprobs(batch_logits.data(), n_vocab, workers, eval_pairs, eval_results);
@@ -1287,14 +1311,14 @@ struct multiple_choice_task {
     }
 
     // For evaluation
-    size_t i_batch;         // starting index in the llama_batch
+    size_t i_logits;        // starting index of logits in the llama_batch
     size_t common_prefix;   // max number of initial tokens that are the same in all sentences
     size_t required_tokens; // needed number of tokens to evaluate all answers
     std::vector<std::vector<llama_token>> seq_tokens;
     std::vector<float> log_probs;
 };
 
-static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos, multiple_choice_task& task, bool log_error) {
+static bool multiple_choice_prepare_one_task(llama_context * ctx, multiple_choice_task& task, bool log_error) {
     if (task.question.empty() || task.mc1.answers.empty()) {
         if (log_error) {
             printf("%s: found bad task with empty question and/or answers\n", __func__);
@@ -1309,7 +1333,7 @@ static bool multiple_choice_prepare_one_task(llama_context * ctx, bool add_bos,
             }
             return false;
         }
-        task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, add_bos));
+        task.seq_tokens.emplace_back(::llama_tokenize(ctx, task.question + " " + answer, true));
     }
     auto min_len = task.seq_tokens.front().size();
     for (auto& seq : task.seq_tokens) {
@@ -1366,7 +1390,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
     std::vector<uint32_t> task_pos(n_task);
     strstream.read((char *)task_pos.data(), task_pos.size()*sizeof(uint32_t));
     if (strstream.fail()) {
-        printf("%s: failed to raad task positions from prompt\n", __func__);
+        printf("%s: failed to read task positions from prompt\n", __func__);
         return;
     }
 
@@ -1408,9 +1432,6 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         n_task = params.multiple_choice_tasks;
     }
 
-    // This is needed as usual for LLaMA models
-    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
-
     printf("%s: preparing task data", __func__);
     fflush(stdout);
     if (n_task > 500) {
@@ -1418,7 +1439,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         fflush(stdout);
         std::atomic<int> counter(0);
         std::atomic<int> n_bad(0);
-        auto prepare = [&counter, &n_bad, &tasks, ctx, add_bos] () {
+        auto prepare = [&counter, &n_bad, &tasks, ctx] () {
             int num_tasks = tasks.size();
             int n_bad_local = 0;
             while (true) {
@@ -1429,7 +1450,7 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
                 }
                 int last = std::min(first + K_TOKEN_CHUNK, num_tasks);
                 for (int i = first; i < last; ++i) {
-                    if (!multiple_choice_prepare_one_task(ctx, add_bos, tasks[i], false)) ++n_bad_local;
+                    if (!multiple_choice_prepare_one_task(ctx, tasks[i], false)) ++n_bad_local;
                 }
             }
         };
@@ -1447,11 +1468,11 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             return;
         }
     } else {
-        int n_dot = n_task/100;
+        int n_dot = std::max((int) n_task/100, 1);
         int i_task = 0;
         for (auto& task : tasks) {
             ++i_task;
-            if (!multiple_choice_prepare_one_task(ctx, add_bos, task, true)) {
+            if (!multiple_choice_prepare_one_task(ctx, task, true)) {
                 return;
             }
             if (i_task%n_dot == 0) {
@@ -1491,17 +1512,18 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         int n_cur = 0;
 
         size_t i1 = i0;
-        size_t i_batch = 0; // this tells us where in `llama_batch` we are currently
+        size_t i_logits = 0; // this tells us how many logits were needed before this point in the batch
 
         llama_batch_clear(batch);
 
         // batch as much tasks as possible into the available context
-        // each task has 4 unique seuqnce ids - one for each ending
+        // each task has 4 unique sequence ids - one for each ending
         // the common prefix is shared among the 4 sequences to save tokens
         // we extract logits only from the last common token and from all ending tokens of each sequence
         int s0 = 0;
         while (n_cur + (int) tasks[i1].required_tokens <= n_ctx) {
             auto& cur_task = tasks[i1];
+            int n_logits = 0;
 
             int num_answers = cur_task.seq_tokens.size();
             if (s0 + num_answers > max_seq) {
@@ -1518,17 +1540,22 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
                 llama_batch_add(batch, cur_task.seq_tokens[0][i], i, batch_indeces, false);
             }
             batch.logits[batch.n_tokens - 1] = true; // we need logits for the last token of the common prefix
+            n_logits += 1;
 
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
-                for (size_t i = cur_task.common_prefix; i < cur_task.seq_tokens[s].size(); ++i) {
-                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, true);
+                const size_t seq_tokens_size = cur_task.seq_tokens[s].size();
+                // TODO: don't evaluate the last token of each sequence
+                for (size_t i = cur_task.common_prefix; i < seq_tokens_size; ++i) {
+                    const bool needs_logits = i < seq_tokens_size - 1;
+                    llama_batch_add(batch, cur_task.seq_tokens[s][i], i, { s0 + s }, needs_logits);
+                    n_logits += needs_logits;
                 }
             }
 
             s0 += num_answers;
 
-            cur_task.i_batch = i_batch;
-            i_batch += cur_task.required_tokens;
+            cur_task.i_logits = i_logits;
+            i_logits += n_logits;
 
             n_cur += cur_task.required_tokens;
             if (++i1 == tasks.size()) {
@@ -1554,12 +1581,11 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
         eval_pairs.clear();
         for (size_t i = i0; i < i1; ++i) {
             auto& cur_task = tasks[i];
-            size_t li = cur_task.common_prefix;
+            size_t li = 1; // skip the last logit of the common prefix (computed separately below)
             for (int s = 0; s < int(cur_task.seq_tokens.size()); ++s) {
                 for (size_t j = cur_task.common_prefix; j < cur_task.seq_tokens[s].size() - 1; j++) {
-                    eval_pairs.emplace_back(cur_task.i_batch + li++, cur_task.seq_tokens[s][j + 1]);
+                    eval_pairs.emplace_back(cur_task.i_logits + li++, cur_task.seq_tokens[s][j + 1]);
                 }
-                ++li;
             }
         }
         // Then we do the actual calculation
@@ -1578,7 +1604,8 @@ static void multiple_choice_score(llama_context * ctx, const gpt_params & params
             //}
             //printf("\n    common_prefix: %zu\n", cur_task.common_prefix);
 
-            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*(cur_task.i_batch + cur_task.common_prefix - 1), n_vocab*sizeof(float));
+            // get the logits of the last token of the common prefix
+            std::memcpy(tok_logits.data(), batch_logits.data() + n_vocab*cur_task.i_logits, n_vocab*sizeof(float));
 
             const auto first_probs = softmax(tok_logits);
 
@@ -1681,6 +1708,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
     const int num_batches = (n_ctx + n_batch - 1)/n_batch;
     const int nv = 2*((n_vocab + 1)/2) + 4;
     const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    GGML_ASSERT(llama_add_eos_token(llama_get_model(ctx)) != 1);
 
     std::vector<uint16_t> log_probs_uint16(size_t(n_ctx - 1 - n_ctx/2) * nv);
     std::vector<float> kld_values(size_t(n_ctx - 1 - n_ctx/2)*n_chunk);
@@ -1730,6 +1758,7 @@ static void kl_divergence(llama_context * ctx, const gpt_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
+            // TODO: use llama_batch.logits 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__);
                 return;

examples/quantize/README.md

@@ -4,17 +4,17 @@ TODO
 
 ## Llama 2 7B
 
-Quantization | Bits per Weight (BPW)
--- | --
-Q2_K | 3.35
-Q3_K_S | 3.50
-Q3_K_M | 3.91
-Q3_K_L | 4.27
-Q4_K_S | 4.58
-Q4_K_M | 4.84
-Q5_K_S | 5.52
-Q5_K_M | 5.68
-Q6_K | 6.56
+| Quantization | Bits per Weight (BPW) |
+|--------------|-----------------------|
+| Q2_K         | 3.35                  |
+| Q3_K_S       | 3.50                  |
+| Q3_K_M       | 3.91                  |
+| Q3_K_L       | 4.27                  |
+| Q4_K_S       | 4.58                  |
+| Q4_K_M       | 4.84                  |
+| Q5_K_S       | 5.52                  |
+| Q5_K_M       | 5.68                  |
+| Q6_K         | 6.56                  |
 
 ## Llama 2 13B
 Quantization | Bits per Weight (BPW)

examples/quantize/quantize.cpp

@@ -26,6 +26,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "IQ2_S",  LLAMA_FTYPE_MOSTLY_IQ2_S,  " 2.5  bpw quantization",            },
     { "IQ2_M",  LLAMA_FTYPE_MOSTLY_IQ2_M,  " 2.7  bpw quantization",            },
     { "IQ1_S",  LLAMA_FTYPE_MOSTLY_IQ1_S,  " 1.56 bpw quantization",            },
+    { "IQ1_M",  LLAMA_FTYPE_MOSTLY_IQ1_M,  " 1.75 bpw quantization",            },
     { "Q2_K",   LLAMA_FTYPE_MOSTLY_Q2_K,   " 2.63G, +0.6717 ppl @ LLaMA-v1-7B", },
     { "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.16G, +9.0634 ppl @ LLaMA-v1-7B", },
     { "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization",            },
@@ -87,13 +88,17 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
 //
 [[noreturn]]
 static void usage(const char * executable) {
-    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
+    printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] [--pure] [--imatrix] [--include-weights] [--exclude-weights] [--output-tensor-type] [--token-embedding-type] [--override-kv] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
     printf("  --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
     printf("  --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
     printf("  --pure: Disable k-quant mixtures and quantize all tensors to the same type\n");
     printf("  --imatrix file_name: use data in file_name as importance matrix for quant optimizations\n");
     printf("  --include-weights tensor_name: use importance matrix for this/these tensor(s)\n");
     printf("  --exclude-weights tensor_name: use importance matrix for this/these tensor(s)\n");
+    printf("  --output-tensor-type ggml_type: use this ggml_type for the output.weight tensor\n");
+    printf("  --token-embedding-type ggml_type: use this ggml_type for the token embeddings tensor\n");
+    printf("  --override-kv KEY=TYPE:VALUE\n");
+    printf("      Advanced option to override model metadata by key in the quantized model. May be specified multiple times.\n");
     printf("Note: --include-weights and --exclude-weights cannot be used together\n");
     printf("\nAllowed quantization types:\n");
     for (auto & it : QUANT_OPTIONS) {
@@ -107,56 +112,60 @@ static void usage(const char * executable) {
     exit(1);
 }
 
-static void load_imatrix(const std::string& imatrix_file, std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void load_imatrix(const std::string & imatrix_file, std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     std::ifstream in(imatrix_file.c_str(), std::ios::binary);
     if (!in) {
-        printf("%s: failed to open %s\n",__func__,imatrix_file.c_str());
-        return;
+        printf("%s: failed to open %s\n",__func__, imatrix_file.c_str());
+        exit(1);
     }
     int n_entries;
-    in.read((char*)&n_entries, sizeof(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__, imatrix_file.c_str());
-        return;
+        exit(1);
     }
     for (int i = 0; i < n_entries; ++i) {
         int len; in.read((char *)&len, sizeof(len));
         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,imatrix_file.c_str());
-            return;
+            printf("%s: failed reading name for entry %d from %s\n", __func__, i+1, imatrix_file.c_str());
+            exit(1);
         }
         name_as_vec[len] = 0;
         std::string name{name_as_vec.data()};
-        auto& e = imatrix_data[std::move(name)];
+        auto & e = imatrix_data[name];
         int ncall;
-        in.read((char*)&ncall, sizeof(ncall));
+        in.read((char *)&ncall, sizeof(ncall));
         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);
+            printf("%s: failed reading number of values for entry %d\n", __func__, i);
             imatrix_data = {};
-            return;
+            exit(1);
         }
         e.resize(nval);
-        in.read((char*)e.data(), nval*sizeof(float));
+        in.read((char *)e.data(), nval*sizeof(float));
         if (in.fail()) {
-            printf("%s: failed reading data for entry %d\n",__func__,i);
+            printf("%s: failed reading data for entry %d\n", __func__, i);
             imatrix_data = {};
-            return;
+            exit(1);
         }
         if (ncall > 0) {
             for (auto& v : e) v /= ncall;
         }
+
+        if (getenv("LLAMA_TRACE")) {
+            printf("%s: loaded data (size = %6d, ncall = %6d) for '%s'\n", __func__, int(e.size()), ncall, name.c_str());
+        }
     }
-    printf("%s: loaded %d importance matrix entries from %s\n",__func__,int(imatrix_data.size()),imatrix_file.c_str());
+    printf("%s: loaded %d importance matrix entries from %s\n", __func__, int(imatrix_data.size()), imatrix_file.c_str());
 }
 
-static void prepare_imatrix(const std::string& imatrix_file,
-        const std::vector<std::string>& included_weights,
-        const std::vector<std::string>& excluded_weights,
-        std::unordered_map<std::string, std::vector<float>>& imatrix_data) {
+static void prepare_imatrix(const std::string & imatrix_file,
+        const std::vector<std::string> & included_weights,
+        const std::vector<std::string> & excluded_weights,
+        std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
     if (!imatrix_file.empty()) {
         load_imatrix(imatrix_file, imatrix_data);
     }
@@ -189,6 +198,55 @@ static void prepare_imatrix(const std::string& imatrix_file,
     }
 }
 
+static ggml_type parse_ggml_type(const char * arg) {
+    ggml_type result = GGML_TYPE_COUNT;
+    for (int j = 0; j < GGML_TYPE_COUNT; ++j) {
+        auto type = ggml_type(j);
+        const auto * name = ggml_type_name(type);
+        if (name && strcmp(arg, name) == 0) {
+            result = type; break;
+        }
+    }
+    return result;
+}
+
+static bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides) {
+    const char* sep = strchr(data, '=');
+    if (sep == nullptr || sep - data >= 128) {
+        fprintf(stderr, "%s: malformed KV override '%s'\n", __func__, data);
+        return false;
+    }
+    llama_model_kv_override kvo;
+    std::strncpy(kvo.key, data, sep - data);
+    kvo.key[sep - data] = 0;
+    sep++;
+    if (strncmp(sep, "int:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+        kvo.int_value = std::atol(sep);
+    } else if (strncmp(sep, "float:", 6) == 0) {
+        sep += 6;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+        kvo.float_value = std::atof(sep);
+    } else if (strncmp(sep, "bool:", 5) == 0) {
+        sep += 5;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_BOOL;
+        if (std::strcmp(sep, "true") == 0) {
+            kvo.bool_value = true;
+        } else if (std::strcmp(sep, "false") == 0) {
+            kvo.bool_value = false;
+        } else {
+            fprintf(stderr, "%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            return false;
+        }
+    } else {
+        fprintf(stderr, "%s: invalid type for KV override '%s'\n", __func__, data);
+        return false;
+    }
+    overrides.emplace_back(std::move(kvo));
+    return true;
+}
+
 int main(int argc, char ** argv) {
     if (argc < 3) {
         usage(argv[0]);
@@ -199,10 +257,27 @@ int main(int argc, char ** argv) {
     int arg_idx = 1;
     std::string imatrix_file;
     std::vector<std::string> included_weights, excluded_weights;
+    std::vector<llama_model_kv_override> kv_overrides;
 
     for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
         if (strcmp(argv[arg_idx], "--leave-output-tensor") == 0) {
             params.quantize_output_tensor = false;
+        } else if (strcmp(argv[arg_idx], "--output-tensor-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.output_tensor_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--token-embedding-type") == 0) {
+            if (arg_idx < argc-1) {
+                params.token_embedding_type = parse_ggml_type(argv[++arg_idx]);
+            } else {
+                usage(argv[0]);
+            }
+        } else if (strcmp(argv[arg_idx], "--override-kv") == 0) {
+            if (arg_idx == argc-1 || !parse_kv_override(argv[++arg_idx], kv_overrides)) {
+                usage(argv[0]);
+            }
         } else if (strcmp(argv[arg_idx], "--allow-requantize") == 0) {
             params.allow_requantize = true;
         } else if (strcmp(argv[arg_idx], "--pure") == 0) {
@@ -243,6 +318,11 @@ int main(int argc, char ** argv) {
     if (!imatrix_data.empty()) {
         params.imatrix = &imatrix_data;
     }
+    if (!kv_overrides.empty()) {
+        kv_overrides.emplace_back();
+        kv_overrides.back().key[0] = 0;
+        params.kv_overrides = &kv_overrides;
+    }
 
     llama_backend_init();
 
@@ -264,8 +344,7 @@ int main(int argc, char ** argv) {
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
-    }
-    else {
+    } else {
         fname_out = argv[arg_idx];
         arg_idx++;
 
@@ -296,10 +375,12 @@ int main(int argc, char ** argv) {
 
     if ((params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XS || params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_XXS ||
          params.ftype == LLAMA_FTYPE_MOSTLY_IQ2_S  ||
-         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S || params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S) && imatrix_data.empty()) {
-        fprintf(stderr, "\n===============================================================================================\n");
-        fprintf(stderr, "Please do not use IQ1_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
-        fprintf(stderr, "===============================================================================================\n\n\n");
+         params.ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_S  ||
+         params.ftype == LLAMA_FTYPE_MOSTLY_IQ1_M) && imatrix_data.empty()) {
+        fprintf(stderr, "\n==========================================================================================================\n");
+        fprintf(stderr, "Please do not use IQ1_S, IQ1_M, IQ2_S, IQ2_XXS, IQ2_XS or Q2_K_S quantization without an importance matrix\n");
+        fprintf(stderr, "==========================================================================================================\n\n\n");
         return 1;
     }
 

examples/regex-to-grammar.py

@@ -0,0 +1,20 @@
+import json, subprocess, sys, os
+
+assert len(sys.argv) >= 2
+[_, pattern, *rest] = sys.argv
+
+print(subprocess.check_output(
+    [
+        "python",
+        os.path.join(
+        os.path.dirname(os.path.realpath(__file__)),
+        "json_schema_to_grammar.py"),
+        *rest,
+        "-",
+        "--raw-pattern",
+    ],
+    text=True,
+    input=json.dumps({
+        "type": "string",
+        "pattern": pattern,
+    }, indent=2)))

examples/retrieval/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET retrieval)
+add_executable(${TARGET} retrieval.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/retrieval/README.md

@@ -0,0 +1,69 @@
+# llama.cpp/examples/retrieval
+
+Demonstration of simple retrieval technique based on cosine similarity
+
+More info:
+https://github.com/ggerganov/llama.cpp/pull/6193
+
+### How to use
+
+`retieval.cpp` has parameters of its own:
+- `--context-file`: file to be embedded - state this option multiple times to embed multiple files
+- `--chunk-size`: minimum size of each text chunk to be embedded
+- `--chunk-separator`: STRING to divide chunks by. newline by default
+
+`retrieval` example can be tested as follows:
+
+```bash
+make -j && ./retrieval --model ./models/bge-base-en-v1.5-f16.gguf --top-k 3 --context-file README.md --context-file License --chunk-size 100 --chunk-separator .
+```
+
+This chunks and embeds all given files and starts a loop requesting query inputs:
+
+```
+Enter query:
+```
+
+On each query input, top k chunks are shown along with file name, chunk position within file and original text:
+
+```
+Enter query: describe the mit license
+batch_decode: n_tokens = 6, n_seq = 1
+Top 3 similar chunks:
+filename: README.md
+filepos: 119
+similarity: 0.762334
+textdata:
+png)
+
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
+
+[Roadmap](https://github.
+--------------------
+filename: License
+filepos: 0
+similarity: 0.725146
+textdata:
+MIT License
+
+Copyright (c) 2023 Georgi Gerganov
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+--------------------
+filename: README.md
+filepos: 9178
+similarity: 0.621722
+textdata:
+com/cztomsik/ava) (MIT)
+- [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [pythops/tenere](https://github.
+--------------------
+```

examples/retrieval/retrieval.cpp

@@ -0,0 +1,350 @@
+#include "common.h"
+#include "llama.h"
+
+#include <algorithm>
+#include <fstream>
+
+struct retrieval_params {
+    std::vector<std::string> context_files; // context files to embed
+    int32_t chunk_size            = 64;     // chunk size for context embedding
+    std::string chunk_separator   = "\n";   // chunk separator for context embedding
+};
+
+static void retrieval_params_print_usage(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & params) {
+    gpt_print_usage(argc, argv, gpt_params);
+    printf("retrieval options:\n");
+    printf("  --context-file FNAME  file containing context to embed.\n");
+    printf("                        specify multiple files by providing --context-file option multiple times.\n");
+    printf("  --chunk-size N        minimum length of embedded text chunk (default:%d)\n", params.chunk_size);
+    printf("  --chunk-separator STRING\n");
+    printf("                        string to separate chunks (default: \"\\n\")\n");
+    printf("\n");
+}
+
+static void retrieval_params_parse(int argc, char ** argv, gpt_params & gpt_params, retrieval_params & retrieval_params) {
+    int i = 1;
+    std::string arg;
+    while (i < argc) {
+        arg = argv[i];
+        bool invalid_gpt_param = false;
+        if(gpt_params_find_arg(argc, argv, argv[i], gpt_params, i, invalid_gpt_param)) {
+            if (invalid_gpt_param) {
+                fprintf(stderr, "error: invalid argument: %s\n", arg.c_str());
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            // option was parsed by gpt_params_find_arg
+        } else if (arg == "--context-file") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --context-file\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            // store the external file name in params
+            retrieval_params.context_files.push_back(argv[i]);
+        } else if (arg == "--chunk-size") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --chunk-size\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            retrieval_params.chunk_size = std::stoi(argv[i]);
+        } else if (arg == "--chunk-separator") {
+            if (++i >= argc) {
+                fprintf(stderr, "error: missing argument for --chunk-separator\n");
+                retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+                exit(1);
+            }
+            retrieval_params.chunk_separator = argv[i];
+        } else {
+            // unknown argument
+            fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
+            retrieval_params_print_usage(argc, argv, gpt_params, retrieval_params);
+            exit(1);
+        }
+        i++;
+    }
+}
+
+struct chunk {
+    // filename
+    std::string filename;
+    // original file position
+    size_t filepos;
+    // original text data
+    std::string textdata = "";
+    // tokenized text data
+    std::vector<llama_token> tokens;
+    // embedding
+    std::vector<float> embedding;
+};
+
+// chunk file data to chunks of size >= chunk_size
+// chunk_separator is the separator between chunks
+static std::vector<chunk> chunk_file(const std::string & filename, int chunk_size, const std::string & chunk_separator) {
+    std::vector<chunk> chunks;
+    std::ifstream f(filename.c_str());
+
+    if (!f.is_open()) {
+        fprintf(stderr, "Error: could not open file %s\n", filename.c_str());
+        return chunks;
+    }
+
+    chunk current_chunk;
+    char buffer[1024];
+    int64_t filepos = 0;
+    std::string current = "";
+    while (f.read(buffer, 1024)) {
+        current += std::string(buffer, f.gcount());
+        size_t pos;
+        while ((pos = current.find(chunk_separator)) != std::string::npos) {
+            current_chunk.textdata += current.substr(0, pos + chunk_separator.size());
+            if ((int) current_chunk.textdata.size() > chunk_size) {
+                // save chunk
+                current_chunk.filepos = filepos;
+                current_chunk.filename = filename;
+                chunks.push_back(current_chunk);
+                // update filepos
+                filepos += (int) current_chunk.textdata.size();
+                // reset current_chunk
+                current_chunk = chunk();
+            }
+            current = current.substr(pos + chunk_separator.size());
+        }
+
+    }
+    // add leftover data to last chunk
+    if (current_chunk.textdata.size() > 0) {
+        if (chunks.empty()) {
+            current_chunk.filepos = filepos;
+            current_chunk.filename = filename;
+            chunks.push_back(current_chunk);
+        } else {
+            chunks.back().textdata += current_chunk.textdata;
+        }
+    }
+    f.close();
+    return chunks;
+}
+
+static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
+    for (size_t i = 0; i < tokens.size(); i++) {
+        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
+    }
+}
+
+static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
+    // 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__);
+    }
+
+    for (int i = 0; i < batch.n_tokens; i++) {
+        if (!batch.logits[i]) {
+            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]);
+        if (embd == NULL) {
+            embd = llama_get_embeddings_ith(ctx, i);
+            if (embd == NULL) {
+                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
+                continue;
+            }
+        }
+
+        float * out = output + batch.seq_id[i][0] * n_embd;
+        llama_embd_normalize(embd, out, n_embd);
+    }
+}
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+    retrieval_params retrieval_params;
+
+    retrieval_params_parse(argc, argv, params, retrieval_params);
+
+    // For BERT models, batch size must be equal to ubatch size
+    params.n_ubatch = params.n_batch;
+
+    if (retrieval_params.chunk_size <= 0) {
+        fprintf(stderr, "chunk_size must be positive\n");
+        return 1;
+    }
+    if (retrieval_params.context_files.empty()) {
+        fprintf(stderr, "context_files must be specified\n");
+        return 1;
+    }
+    params.embedding = true;
+
+    print_build_info();
+
+    printf("processing files:\n");
+    for (auto & context_file : retrieval_params.context_files) {
+        printf("%s\n", context_file.c_str());
+    }
+
+    std::vector<chunk> chunks;
+    for (auto & context_file : retrieval_params.context_files) {
+        std::vector<chunk> file_chunk = chunk_file(context_file, retrieval_params.chunk_size, retrieval_params.chunk_separator);
+        chunks.insert(chunks.end(), file_chunk.begin(), file_chunk.end());
+    }
+    printf("Number of chunks: %ld\n", chunks.size());
+
+    llama_backend_init();
+    llama_numa_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    // load the model
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    const int n_ctx = llama_n_ctx(ctx);
+
+    if (n_ctx > n_ctx_train) {
+        fprintf(stderr, "%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", get_system_info(params).c_str());
+    }
+
+    // max batch size
+    const uint64_t n_batch = params.n_batch;
+    GGML_ASSERT(params.n_batch >= params.n_ctx);
+
+    // tokenize the prompts and trim
+    for (auto & chunk : chunks) {
+        auto inp = ::llama_tokenize(ctx, chunk.textdata, true, false);
+        if (inp.size() > n_batch) {
+            fprintf(stderr, "%s: error: chunk size (%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;
+        }
+        // add eos if not present
+        if (inp.empty() || inp.back() != llama_token_eos(model)) {
+            inp.push_back(llama_token_eos(model));
+        }
+        chunk.tokens = inp;
+    }
+
+    // tokenization stats
+    if (params.verbose_prompt) {
+        for (int i = 0; i < (int) chunks.size(); i++) {
+            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, chunks[i].textdata.c_str());
+            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, chunks[i].tokens.size());
+            for (int j = 0; j < (int) chunks[i].tokens.size(); j++) {
+                fprintf(stderr, "%6d -> '%s'\n", chunks[i].tokens[j], llama_token_to_piece(ctx, chunks[i].tokens[j]).c_str());
+            }
+            fprintf(stderr, "\n\n");
+        }
+    }
+
+    // initialize batch
+    const int n_chunks = chunks.size();
+    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
+
+    // allocate output
+    const int n_embd = llama_n_embd(model);
+    std::vector<float> embeddings(n_chunks * n_embd, 0);
+    float * emb = embeddings.data();
+
+    // break into batches
+    int p = 0; // number of prompts processed already
+    int s = 0; // number of prompts in current batch
+    for (int k = 0; k < n_chunks; k++) {
+        // clamp to n_batch tokens
+        auto & inp = chunks[k].tokens;
+
+        const uint64_t n_toks = inp.size();
+
+        // encode if at capacity
+        if (batch.n_tokens + n_toks > n_batch) {
+            float * out = emb + p * n_embd;
+            batch_decode(ctx, batch, out, s, n_embd);
+            llama_batch_clear(batch);
+            p += s;
+            s = 0;
+        }
+
+        // add to batch
+        batch_add_seq(batch, inp, s);
+        s += 1;
+    }
+
+    // final batch
+    float * out = emb + p * n_embd;
+    batch_decode(ctx, batch, out, s, n_embd);
+
+    // save embeddings to chunks
+    for (int i = 0; i < n_chunks; i++) {
+        chunks[i].embedding = std::vector<float>(emb + i * n_embd, emb + (i + 1) * n_embd);
+        // clear tokens as they are no longer needed
+        chunks[i].tokens.clear();
+    }
+
+    // start loop, receive query and return top k similar chunks based on cosine similarity
+    std::string query;
+    while (true) {
+        printf("Enter query: ");
+        std::getline(std::cin, query);
+        std::vector<int32_t> query_tokens = llama_tokenize(ctx, query, true);
+
+        struct llama_batch query_batch = llama_batch_init(n_batch, 0, 1);
+        batch_add_seq(query_batch, query_tokens, 0);
+
+        std::vector<float> query_emb(n_embd, 0);
+        batch_decode(ctx, query_batch, query_emb.data(), 1, n_embd);
+
+        llama_batch_clear(query_batch);
+
+        // compute cosine similarities
+        {
+            std::vector<std::pair<int, float>> similarities;
+            for (int i = 0; i < n_chunks; i++) {
+                float sim = llama_embd_similarity_cos(chunks[i].embedding.data(), query_emb.data(), n_embd);
+                similarities.push_back(std::make_pair(i, sim));
+            }
+
+            // sort similarities
+            std::sort(similarities.begin(), similarities.end(), [](const std::pair<int, float> & a, const std::pair<int, float> & b) {
+                return a.second > b.second;
+            });
+
+            printf("Top %d similar chunks:\n", params.sparams.top_k);
+            for (int i = 0; i < std::min(params.sparams.top_k, (int) chunks.size()); i++) {
+                printf("filename: %s\n", chunks[similarities[i].first].filename.c_str());
+                printf("filepos: %lld\n", (long long int) chunks[similarities[i].first].filepos);
+                printf("similarity: %f\n", similarities[i].second);
+                printf("textdata:\n%s\n", chunks[similarities[i].first].textdata.c_str());
+                printf("--------------------\n");
+            }
+        }
+    }
+
+    // clean up
+    llama_print_timings(ctx);
+    llama_free(ctx);
+    llama_free_model(model);
+    llama_backend_free();
+}

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

@@ -24,6 +24,7 @@ int main(int argc, char ** argv) {
 
     std::string result0;
     std::string result1;
+    std::string result2;
 
     // init
     llama_model * model;
@@ -44,8 +45,8 @@ int main(int argc, char ** argv) {
 
     // save state (rng, logits, embedding and kv_cache) to file
     {
-        std::vector<uint8_t> state_mem(llama_get_state_size(ctx));
-        const size_t written = llama_copy_state_data(ctx, state_mem.data());
+        std::vector<uint8_t> state_mem(llama_state_get_size(ctx));
+        const size_t written = llama_state_get_data(ctx, state_mem.data());
 
         FILE *fp_write = fopen("dump_state.bin", "wb");
         fwrite(state_mem.data(), 1, written, fp_write);
@@ -97,13 +98,13 @@ int main(int argc, char ** argv) {
 
     // load state (rng, logits, embedding and kv_cache) from file
     {
-        std::vector<uint8_t> state_mem(llama_get_state_size(ctx2));
+        std::vector<uint8_t> state_mem(llama_state_get_size(ctx2));
 
         FILE * fp_read = fopen("dump_state.bin", "rb");
         const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
         fclose(fp_read);
 
-        if (read != llama_set_state_data(ctx2, state_mem.data())) {
+        if (read != llama_state_set_data(ctx2, state_mem.data())) {
             fprintf(stderr, "\n%s : failed to read state\n", __func__);
             llama_free(ctx2);
             llama_free_model(model);
@@ -141,16 +142,104 @@ int main(int argc, char ** argv) {
         n_past += 1;
     }
 
-    printf("\n");
+    printf("\n\n");
 
     llama_free(ctx2);
-    llama_free_model(model);
 
     if (result0 != result1) {
         fprintf(stderr, "\n%s : error : the 2 generations are different\n", __func__);
         return 1;
     }
 
+    // make new context
+    auto* ctx3 = llama_new_context_with_model(model, llama_context_params_from_gpt_params(params));
+
+    printf("\nsingle seq run: %s", params.prompt.c_str());
+
+    // load state (rng, logits, embedding and kv_cache) from file
+    {
+        std::vector<uint8_t> state_mem(llama_state_get_size(ctx3));
+
+        FILE * fp_read = fopen("dump_state.bin", "rb");
+        const size_t read = fread(state_mem.data(), 1, state_mem.size(), fp_read);
+        fclose(fp_read);
+
+        if (read != llama_state_set_data(ctx3, state_mem.data())) {
+            fprintf(stderr, "\n%s : failed to read state\n", __func__);
+            llama_free(ctx3);
+            llama_free_model(model);
+            return 1;
+        }
+
+        fprintf(stderr, "%s : deserialized state from %zd out of a maximum of %zd bytes\n", __func__, read, state_mem.size());
+    }
+
+    // restore state (last tokens)
+    n_past = n_past_saved;
+
+    // save seq 0 and load into seq 1
+    {
+        // save kv of seq 0
+        std::vector<uint8_t> seq_store(llama_state_seq_get_size(ctx3, 0));
+        const size_t ncopy = llama_state_seq_get_data(ctx3, seq_store.data(), 0);
+        if (ncopy != seq_store.size()) {
+            fprintf(stderr, "\n%s : seq copy data length %zd does not match expected length %zd\n", __func__, ncopy, seq_store.size());
+            llama_free(ctx3);
+            llama_free_model(model);
+            return 1;
+        }
+        fprintf(stderr, "%s : seq 0 copied, %zd bytes\n", __func__, ncopy);
+
+        // erase whole kv
+        llama_kv_cache_clear(ctx3);
+        fprintf(stderr, "%s : kv cache cleared\n", __func__);
+
+        // restore kv into seq 1
+        const size_t nset = llama_state_seq_set_data(ctx3, seq_store.data(), 1);
+        if (nset != seq_store.size()) {
+            fprintf(stderr, "\n%s : seq set data length %zd does not match expected length %zd\n", __func__, nset, seq_store.size());
+            llama_free(ctx3);
+            llama_free_model(model);
+            return 1;
+        }
+        fprintf(stderr, "%s : seq 1 restored, %zd bytes\n", __func__, nset);
+    }
+
+    // third run with seq 1 instead of 0
+    for (auto i = 0; i < params.n_predict; i++) {
+        auto * logits = llama_get_logits(ctx3);
+        auto n_vocab = llama_n_vocab(model);
+        std::vector<llama_token_data> candidates;
+        candidates.reserve(n_vocab);
+        for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+            candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
+        }
+        llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+        auto next_token = llama_sample_token(ctx3, &candidates_p);
+        auto next_token_str = llama_token_to_piece(ctx3, next_token);
+
+        printf("%s", next_token_str.c_str());
+        result2 += next_token_str;
+
+        if (llama_decode(ctx3, llama_batch_get_one(&next_token, 1, n_past, 1))) {
+            fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_free(ctx3);
+            llama_free_model(model);
+            return 1;
+        }
+        n_past += 1;
+    }
+
+    printf("\n");
+
+    llama_free(ctx3);
+    llama_free_model(model);
+
+    if (result0 != result2) {
+        fprintf(stderr, "\n%s : error : the seq restore generation is different\n", __func__);
+        return 1;
+    }
+
     fprintf(stderr, "\n%s : success\n", __func__);
 
     return 0;