llama : fix BPE pre-tokenization (#6920)

* merged the changes from deepseeker models to main branch

* Moved regex patterns to unicode.cpp and updated unicode.h

* Moved header files

* Resolved issues

* added and refactored unicode_regex_split and related functions

* Updated/merged the deepseek coder pr

* Refactored code

* Adding unicode regex mappings

* Adding unicode regex function

* Added needed functionality, testing remains

* Fixed issues

* Fixed issue with gpt2 regex custom preprocessor

* unicode : fix? unicode_wstring_to_utf8

* lint : fix whitespaces

* tests : add tokenizer tests for numbers

* unicode : remove redundant headers

* tests : remove and rename tokenizer test scripts

* tests : add sample usage

* gguf-py : reader prints warnings on duplicate keys

* llama : towards llama3 tokenization support (wip)

* unicode : shot in the dark to fix tests on Windows

* unicode : first try custom implementations

* convert : add "tokenizer.ggml.pre" GGUF KV (wip)

* llama : use new pre-tokenizer type

* convert : fix pre-tokenizer type writing

* lint : fix

* make : add test-tokenizer-0-llama-v3

* wip

* models : add llama v3 vocab file

* llama : adapt punctuation regex + add llama 3 regex

* minor

* unicode : set bomb

* unicode : set bomb

* unicode : always use std::wregex

* unicode : support \p{N}, \p{L} and \p{P} natively

* unicode : try fix windows

* unicode : category support via std::regex

* unicode : clean-up

* unicode : simplify

* convert : add convert-hf-to-gguf-update.py

ggml-ci

* lint : update

* convert : add falcon

ggml-ci

* unicode : normalize signatures

* lint : fix

* lint : fix

* convert : remove unused functions

* convert : add comments

* convert : exercise contractions

ggml-ci

* lint : fix

* cmake : refactor test targets

* tests : refactor vocab tests

ggml-ci

* tests : add more vocabs and tests

ggml-ci

* unicode : cleanup

* scripts : ignore new update script in check-requirements.sh

* models : add phi-3, mpt, gpt-2, starcoder

* tests : disable obsolete

ggml-ci

* tests : use faster bpe test

ggml-ci

* llama : more prominent warning for old BPE models

* tests : disable test-tokenizer-1-bpe due to slowness

ggml-ci

---------

Co-authored-by: Jaggzh <jaggz.h@gmail.com>
Co-authored-by: Kazim Abrar Mahi <kazimabrarmahi135@gmail.com>

.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.ref }}-${{ github.head_ref || github.run_id }}-${{ 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

@@ -16,7 +16,7 @@ on:
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 env:
@@ -31,7 +31,9 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -52,7 +54,7 @@ jobs:
         id: cmake_test
         run: |
           cd build
-          ctest -L main --verbose --timeout 900
+          ctest -L 'main|curl' --verbose --timeout 900
 
       - name: Determine tag name
         id: tag
@@ -76,10 +78,10 @@ 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-macos-arm64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
+          name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
     runs-on: macos-latest
@@ -87,7 +89,9 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - name: Dependencies
         id: depends
@@ -101,7 +105,9 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
+          # 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
@@ -132,10 +138,10 @@ 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-macos-x64.zip
+          path: llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
+          name: llama-bin-macos-x64.zip
 
   ubuntu-focal-make:
     runs-on: ubuntu-20.04
@@ -146,7 +152,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -158,7 +164,7 @@ jobs:
         with:
           node-version: "20"
 
-      - uses: actions/setup-python@v4
+      - uses: actions/setup-python@v5
         with:
           python-version: "3.11"
 
@@ -181,7 +187,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -203,27 +209,29 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
 
       - 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
@@ -236,6 +244,33 @@ jobs:
           ./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
 
+      - 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-ubuntu-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-ubuntu-x64.zip
+          name: llama-bin-ubuntu-x64.zip
+
 #  ubuntu-latest-cmake-sanitizer:
 #    runs-on: ubuntu-latest
 #
@@ -249,7 +284,7 @@ jobs:
 #    steps:
 #      - name: Clone
 #        id: checkout
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        id: depends
@@ -283,7 +318,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -311,7 +346,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -357,7 +392,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -398,7 +433,7 @@ jobs:
 
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Build
         id: cmake_build
@@ -418,7 +453,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -449,7 +484,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         id: depends
@@ -558,6 +593,63 @@ jobs:
         run: |
             make swift
 
+  windows-msys2:
+    runs-on: windows-latest
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - { sys: UCRT64,  env: ucrt-x86_64,  build: Release }
+          - { sys: CLANG64, env: clang-x86_64, build: Release }
+
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+
+      - name: Setup ${{ matrix.sys }}
+        uses: msys2/setup-msys2@v2
+        with:
+          update: true
+          msystem: ${{matrix.sys}}
+          install: >-
+            base-devel
+            mingw-w64-${{matrix.env}}-toolchain
+            mingw-w64-${{matrix.env}}-cmake
+            mingw-w64-${{matrix.env}}-openblas
+
+      - name: Build using make
+        shell: msys2 {0}
+        run: |
+            make -j $(nproc)
+
+      - name: Clean after building using make
+        shell: msys2 {0}
+        run: |
+            make clean
+
+      - name: Build using make w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            make LLAMA_OPENBLAS=1 -j $(nproc)
+
+      - name: Build using CMake
+        shell: msys2 {0}
+        run: |
+            cmake -B build
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
+      - name: Clean after building using CMake
+        shell: msys2 {0}
+        run: |
+            rm -rf build
+
+      - name: Build using CMake w/ OpenBLAS
+        shell: msys2 {0}
+        run: |
+            cmake -B build -DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS
+            cmake --build build --config ${{ matrix.build }} -j $(nproc)
+
   windows-latest-cmake:
     runs-on: windows-latest
 
@@ -593,7 +685,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -723,23 +815,23 @@ 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
 
@@ -755,7 +847,7 @@ 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
@@ -779,10 +871,10 @@ 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: |
@@ -793,10 +885,10 @@ 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
@@ -812,7 +904,7 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
         with:
           fetch-depth: 0
 
@@ -840,21 +932,21 @@ jobs:
         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\*
+          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
+        uses: actions/upload-artifact@v4
         with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
+          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
@@ -864,7 +956,7 @@ jobs:
 
     steps:
       - name: Clone
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Set up JDK
         uses: actions/setup-java@v3
@@ -887,7 +979,7 @@ 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
@@ -911,14 +1003,14 @@ 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
 
@@ -937,7 +1029,13 @@ 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
@@ -956,7 +1054,7 @@ jobs:
             const path = require('path');
             const fs = require('fs');
             const release_id = '${{ steps.create_release.outputs.id }}';
-            for (let file of await fs.readdirSync('./artifact')) {
+            for (let file of await fs.readdirSync('./artifact/release')) {
               if (path.extname(file) === '.zip') {
                 console.log('uploadReleaseAsset', file);
                 await github.repos.uploadReleaseAsset({
@@ -964,7 +1062,7 @@ jobs:
                   repo: context.repo.repo,
                   release_id: release_id,
                   name: file,
-                  data: await fs.readFileSync(`./artifact/${file}`)
+                  data: await fs.readFileSync(`./artifact/release/${file}`)
                 });
               }
             }
@@ -978,7 +1076,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1002,7 +1100,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1026,7 +1124,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Dependencies
 #        run: |
@@ -1056,7 +1154,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1072,7 +1170,7 @@ jobs:
 #          msbuild ALL_BUILD.vcxproj -t:build -p:configuration=${{ 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/${{ matrix.build }}
@@ -1095,7 +1193,7 @@ jobs:
 #
 #    steps:
 #      - name: Clone
-#        uses: actions/checkout@v3
+#        uses: actions/checkout@v4
 #
 #      - name: Add msbuild to PATH
 #        uses: microsoft/setup-msbuild@v1
@@ -1127,7 +1225,7 @@ jobs:
 #
 #      - name: Upload binaries
 #        if: matrix.blas == 'ON'
-#        uses: actions/upload-artifact@v1
+#        uses: actions/upload-artifact@v4
 #        with:
 #          name: llama-blas-bin-${{ matrix.arch }}
 #          path: build/bin/${{ matrix.build }}
@@ -1141,7 +1239,7 @@ 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

@@ -6,7 +6,7 @@ env:
   GGML_N_THREADS: 1
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -14,7 +14,7 @@ jobs:
     runs-on: ubuntu-20.04
     steps:
       - name: Checkout
-        uses: actions/checkout@v3
+        uses: actions/checkout@v4
 
       - name: Dependencies
         run: |

.github/workflows/docker.yml

@@ -16,7 +16,7 @@ on:
       - master
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -46,7 +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
@@ -91,6 +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
@@ -98,7 +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)
@@ -107,5 +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

@@ -15,13 +15,13 @@ on:
       - master
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  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

@@ -18,7 +18,7 @@ on:
     paths: ['**/*.nix', 'flake.lock']
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:

.github/workflows/nix-ci.yml

@@ -9,7 +9,7 @@ on:
     types: [opened, synchronize, reopened]
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:

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

@@ -17,7 +17,7 @@ on:
       - 'requirements/*.txt'
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -26,9 +26,9 @@ jobs:
     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

@@ -3,7 +3,7 @@ name: flake8 Lint
 on: [push, pull_request]
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -12,13 +12,13 @@ jobs:
     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
         uses: py-actions/flake8@v2
         with:
             ignore: "E203,E211,E221,E225,E231,E241,E251,E261,E266,E501,E701,E704,W503"
-            exclude: "examples/*,examples/*/**,*/**/__init__.py"
+            exclude: "examples/*,examples/*/**,*/**/__init__.py,convert-hf-to-gguf-update.py"

.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
@@ -11,15 +15,15 @@ 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.ref }}
+  group: ${{ github.workflow }}-${{ github.ref }}-${{ github.head_ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -37,32 +41,56 @@ jobs:
             sanitizer: ""
       fail-fast: false # While -DLLAMA_SANITIZE_THREAD=ON is broken
 
-    container:
-      image: ubuntu:latest
-      ports:
-        - 8888
-      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 \
+          sudo apt-get update
+          sudo 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: Python setup
+        id: setup_python
+        uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+
+      - name: Tests dependencies
+        id: test_dependencies
+        run: |
+          pip install -r examples/server/tests/requirements.txt
+
+      - 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 "${not_ignored_files}"
+            exit 1
+          fi
+
       - name: Build
         id: cmake_build
         run: |
@@ -76,10 +104,6 @@ jobs:
               -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON ;
           cmake --build . --config ${{ matrix.build_type }} -j $(nproc) --target server
 
-      - name: Tests dependencies
-        id: test_dependencies
-        run: |
-          pip install -r examples/server/tests/requirements.txt
 
       - name: Tests
         id: server_integration_tests
@@ -102,9 +126,10 @@ jobs:
     steps:
       - name: Clone
         id: checkout
-        uses: actions/checkout@v3
+        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: libCURL
         id: get_libcurl

.github/workflows/zig-build.yml

@@ -7,7 +7,7 @@ on:
       - master
 
 concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
   cancel-in-progress: true
 
 jobs:
@@ -18,7 +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

@@ -34,6 +34,7 @@ lcov-report/
 gcovr-report/
 
 build*
+!build.zig
 cmake-build-*
 out/
 tmp/
@@ -48,6 +49,7 @@ models-mnt
 /convert-llama2c-to-ggml
 /embd-input-test
 /embedding
+/eval-callback
 /gguf
 /gguf-llama-simple
 /gguf-split
@@ -77,6 +79,7 @@ models-mnt
 /batched-bench
 /export-lora
 /finetune
+/retrieval
 /speculative
 /parallel
 /train-text-from-scratch
@@ -98,7 +101,25 @@ qnt-*.txt
 perf-*.txt
 
 examples/jeopardy/results.txt
+examples/server/*.html.hpp
+examples/server/*.js.hpp
+examples/server/*.mjs.hpp
 
 poetry.lock
 poetry.toml
 nppBackup
+
+# Test binaries
+/tests/test-grammar-parser
+/tests/test-llama-grammar
+/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
+/tests/test-tokenizer-1-spm
+/tests/test-tokenizer-1-bpe
+/tests/test-rope
+/tests/test-backend-ops

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

@@ -43,6 +43,8 @@ else()
     set(LLAMA_METAL_DEFAULT OFF)
 endif()
 
+set(LLAMA_LLAMAFILE_DEFAULT ON)
+
 # general
 option(BUILD_SHARED_LIBS                "build shared libraries"                                OFF)
 option(LLAMA_STATIC                     "llama: static link libraries"                          OFF)
@@ -88,9 +90,10 @@ endif()
 # 3rd party libs
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
+option(LLAMA_LLAMAFILE                       "llama: use llamafile SGEMM"                       ${LLAMA_LLAMAFILE_DEFAULT})
 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")
@@ -113,6 +116,9 @@ 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)
@@ -250,6 +256,16 @@ 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
@@ -273,6 +289,7 @@ if (LLAMA_METAL)
         ${METALKIT_FRAMEWORK}
         )
 endif()
+
 if (LLAMA_BLAS)
     if (LLAMA_STATIC)
         set(BLA_STATIC ON)
@@ -355,23 +372,37 @@ if (LLAMA_BLAS)
     endif()
 endif()
 
+if (LLAMA_LLAMAFILE)
+    add_compile_definitions(GGML_USE_LLAMAFILE)
+
+    set(GGML_HEADERS_LLAMAFILE sgemm.h)
+    set(GGML_SOURCES_LLAMAFILE sgemm.cpp)
+endif()
+
 if (LLAMA_QKK_64)
     add_compile_definitions(GGML_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()
@@ -420,7 +451,7 @@ if (LLAMA_CUBLAS)
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     else()
-        message(WARNING "cuBLAS not found")
+        message(WARNING "CUDA not found")
     endif()
 endif()
 
@@ -519,9 +550,11 @@ 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)
@@ -543,7 +576,7 @@ if (LLAMA_HIPBLAS)
     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")
@@ -826,7 +859,7 @@ endif()
 
 set(CUDA_CXX_FLAGS "")
 
-if (LLAMA_CUBLAS)
+if (LLAMA_CUDA)
     set(CUDA_FLAGS -use_fast_math)
 
     if (LLAMA_FATAL_WARNINGS)
@@ -1051,7 +1084,7 @@ 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 "")
@@ -1129,15 +1162,16 @@ add_library(ggml OBJECT
             ggml-backend.h
             ggml-quants.c
             ggml-quants.h
-            ${GGML_SOURCES_CUDA}    ${GGML_HEADERS_CUDA}
-            ${GGML_SOURCES_OPENCL}  ${GGML_HEADERS_OPENCL}
-            ${GGML_SOURCES_METAL}   ${GGML_HEADERS_METAL}
-            ${GGML_SOURCES_MPI}     ${GGML_HEADERS_MPI}
-            ${GGML_SOURCES_EXTRA}   ${GGML_HEADERS_EXTRA}
-            ${GGML_SOURCES_SYCL}    ${GGML_HEADERS_SYCL}
-            ${GGML_SOURCES_KOMPUTE} ${GGML_HEADERS_KOMPUTE}
-            ${GGML_SOURCES_VULKAN}  ${GGML_HEADERS_VULKAN}
-            ${GGML_SOURCES_ROCM}    ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_CUDA}      ${GGML_HEADERS_CUDA}
+            ${GGML_SOURCES_OPENCL}    ${GGML_HEADERS_OPENCL}
+            ${GGML_SOURCES_METAL}     ${GGML_HEADERS_METAL}
+            ${GGML_SOURCES_MPI}       ${GGML_HEADERS_MPI}
+            ${GGML_SOURCES_EXTRA}     ${GGML_HEADERS_EXTRA}
+            ${GGML_SOURCES_SYCL}      ${GGML_HEADERS_SYCL}
+            ${GGML_SOURCES_KOMPUTE}   ${GGML_HEADERS_KOMPUTE}
+            ${GGML_SOURCES_VULKAN}    ${GGML_HEADERS_VULKAN}
+            ${GGML_SOURCES_ROCM}      ${GGML_HEADERS_ROCM}
+            ${GGML_SOURCES_LLAMAFILE} ${GGML_HEADERS_LLAMAFILE}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
@@ -1161,6 +1195,7 @@ add_library(llama
             llama.h
             unicode.h
             unicode.cpp
+            unicode-data.cpp
             )
 
 target_include_directories(llama PUBLIC .)
@@ -1256,6 +1291,12 @@ 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

@@ -1,16 +1,28 @@
 # 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 gguf-split 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-json-schema-to-grammar
+	tests/test-autorelease \
+	tests/test-backend-ops \
+	tests/test-double-float \
+	tests/test-grad0 \
+	tests/test-grammar-integration \
+	tests/test-grammar-parser \
+	tests/test-json-schema-to-grammar \
+	tests/test-llama-grammar \
+	tests/test-model-load-cancel \
+	tests/test-opt \
+	tests/test-quantize-fns \
+	tests/test-quantize-perf \
+	tests/test-rope \
+	tests/test-sampling \
+	tests/test-tokenizer-0 \
+	tests/test-tokenizer-1-bpe \
+	tests/test-tokenizer-1-spm
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -27,6 +39,17 @@ ifndef UNAME_M
 UNAME_M := $(shell uname -m)
 endif
 
+# In GNU make default CXX is g++ instead of c++.  Let's fix that so that users
+# of non-gcc compilers don't have to provide g++ alias or wrapper.
+DEFCC  := cc
+DEFCXX := c++
+ifeq ($(origin CC),default)
+CC  := $(DEFCC)
+endif
+ifeq ($(origin CXX),default)
+CXX := $(DEFCXX)
+endif
+
 # Mac OS + Arm can report x86_64
 # ref: https://github.com/ggerganov/whisper.cpp/issues/66#issuecomment-1282546789
 ifeq ($(UNAME_S),Darwin)
@@ -49,11 +72,17 @@ default: $(BUILD_TARGETS)
 test: $(TEST_TARGETS)
 	@failures=0; \
 	for test_target in $(TEST_TARGETS); do \
-		if [ "$$test_target" = "tests/test-tokenizer-0-llama" ]; then \
-			./$$test_target $(CURDIR)/models/ggml-vocab-llama.gguf; \
-		elif [ "$$test_target" = "tests/test-tokenizer-0-falcon" ]; then \
+		if [ "$$test_target" = "tests/test-tokenizer-0" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-llama-spm.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-llama-bpe.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-phi-3.gguf; \
 			./$$test_target $(CURDIR)/models/ggml-vocab-falcon.gguf; \
-		elif [ "$$test_target" = "tests/test-tokenizer-1-llama" ]; then \
+			./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-coder.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-deepseek-llm.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-bert-bge.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-starcoder.gguf; \
+			./$$test_target $(CURDIR)/models/ggml-vocab-gpt-2.gguf; \
+		elif [ "$$test_target" = "tests/test-tokenizer-1-spm" ]; then \
 			continue; \
 		elif [ "$$test_target" = "tests/test-tokenizer-1-bpe" ]; then \
 			continue; \
@@ -384,20 +413,31 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
+ifndef LLAMA_NO_LLAMAFILE
+	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
+	OBJS        += sgemm.o
+endif
+
 ifdef LLAMA_BLIS
 	MK_CPPFLAGS += -DGGML_USE_OPENBLAS -I/usr/local/include/blis -I/usr/include/blis
 	MK_LDFLAGS  += -lblis -L/usr/local/lib
 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
@@ -458,16 +498,25 @@ 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
-
 	MK_CPPFLAGS += -DGGML_USE_CLBLAST $(shell pkg-config --cflags-only-I clblast OpenCL)
 	MK_CFLAGS   += $(shell pkg-config --cflags-only-other clblast OpenCL)
 	MK_CXXFLAGS += $(shell pkg-config --cflags-only-other clblast OpenCL)
@@ -510,7 +559,6 @@ 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))
@@ -522,7 +570,7 @@ 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
@@ -539,8 +587,14 @@ 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
@@ -581,6 +635,11 @@ ggml-mpi.o: ggml-mpi.c ggml-mpi.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
+ifndef LLAMA_NO_LLAMAFILE
+sgemm.o: sgemm.cpp sgemm.h ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 GF_CC := $(CC)
 include scripts/get-flags.mk
 
@@ -593,7 +652,7 @@ 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
@@ -618,19 +677,26 @@ $(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
 #
@@ -650,13 +716,16 @@ 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 $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@
@@ -687,6 +756,7 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 
 clean:
 	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
 
 #
@@ -723,11 +793,11 @@ batched: examples/batched/batched.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)
 
-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o 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)
 
-quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.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)
 
@@ -755,10 +825,19 @@ 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 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)
+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 examples/server/json-schema-to-grammar.mjs.hpp 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)
 
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% Makefile
+	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
+		echo "unsigned char $${NAME}[] = {" && \
+		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
+		echo "};" && \
+		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
+	) > $@
+
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -767,6 +846,10 @@ 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)
@@ -804,6 +887,10 @@ 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)
@@ -830,6 +917,10 @@ 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)
 
+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 $@ $(LDFLAGS)
+
 ifeq ($(UNAME_S),Darwin)
 swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
@@ -877,6 +968,10 @@ 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)
@@ -905,11 +1000,7 @@ tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-0: tests/test-tokenizer-0.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
@@ -917,7 +1008,7 @@ tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMM
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
+tests/test-tokenizer-1-spm: tests/test-tokenizer-1-spm.cpp ggml.o llama.o $(COMMON_DEPS) console.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 

Package.swift

@@ -2,6 +2,45 @@
 
 import PackageDescription
 
+var sources = [
+    "ggml.c",
+    "sgemm.cpp",
+    "llama.cpp",
+    "unicode.cpp",
+    "unicode-data.cpp",
+    "ggml-alloc.c",
+    "ggml-backend.c",
+    "ggml-quants.c",
+]
+
+var resources: [Resource] = []
+var linkerSettings: [LinkerSetting] = []
+var cSettings: [CSetting] =  [
+    .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
+    .unsafeFlags(["-fno-objc-arc"]),
+    // NOTE: NEW_LAPACK will required iOS version 16.4+
+    // We should consider add this in the future when we drop support for iOS 14
+    // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
+    // .define("ACCELERATE_NEW_LAPACK"),
+    // .define("ACCELERATE_LAPACK_ILP64")
+]
+
+#if canImport(Darwin)
+sources.append("ggml-metal.m")
+resources.append(.process("ggml-metal.metal"))
+linkerSettings.append(.linkedFramework("Accelerate"))
+cSettings.append(
+    contentsOf: [
+        .define("GGML_USE_ACCELERATE"),
+        .define("GGML_USE_METAL")
+    ]
+)
+#endif
+
+#if os(Linux)
+    cSettings.append(.define("_GNU_SOURCE"))
+#endif
+
 let package = Package(
     name: "llama",
     platforms: [
@@ -28,33 +67,11 @@ let package = Package(
                "ggml-cuda.h",
                "Makefile"
             ],
-            sources: [
-                "ggml.c",
-                "llama.cpp",
-                "unicode.cpp",
-                "ggml-alloc.c",
-                "ggml-backend.c",
-                "ggml-quants.c",
-                "ggml-metal.m",
-            ],
-            resources: [
-                .process("ggml-metal.metal")
-            ],
+            sources: sources,
+            resources: resources,
             publicHeadersPath: "spm-headers",
-            cSettings: [
-                .unsafeFlags(["-Wno-shorten-64-to-32", "-O3", "-DNDEBUG"]),
-                .define("GGML_USE_ACCELERATE"),
-                .unsafeFlags(["-fno-objc-arc"]),
-                .define("GGML_USE_METAL"),
-                // NOTE: NEW_LAPACK will required iOS version 16.4+
-                // We should consider add this in the future when we drop support for iOS 14
-                // (ref: ref: https://developer.apple.com/documentation/accelerate/1513264-cblas_sgemm?language=objc)
-                // .define("ACCELERATE_NEW_LAPACK"),
-                // .define("ACCELERATE_LAPACK_ILP64")
-            ],
-            linkerSettings: [
-                .linkedFramework("Accelerate")
-            ]
+            cSettings: cSettings,
+            linkerSettings: linkerSettings
         )
     ],
     cxxLanguageStandard: .cxx11

README-sycl.md

@@ -3,32 +3,41 @@
 - [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.
@@ -45,216 +54,233 @@ 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/server-intel.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:
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
-# For Nvidia GPUs
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
+
+#build all binary
+cmake --build . --config Release -j -v
+```
+
+#### Nvidia GPU
+```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 FP32 (recommended for better performance in most cases)
 cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
 
-# Build example/main only
-#cmake --build . --config Release --target main
+# Option 2: Use FP16
+cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
 
-# Or, build all binary
-cmake --build . --config Release -v
+#build all binary
+cmake --build . --config Release -j -v
 
-cd ..
 ```
 
-or
+### III. Run the inference
 
-```sh
-./examples/sycl/build.sh
-```
+1. Retrieve and prepare model
 
-### Run
-
-1. Put model file to folder **models**
-
-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|               |
@@ -268,22 +294,22 @@ 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:
 
@@ -303,74 +329,63 @@ 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):
 ```
@@ -380,7 +395,7 @@ 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/
 
@@ -390,76 +405,57 @@ 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
-
-::  for FP32
+# Option 1: Use FP32 (recommended for better performance in most cases)
 cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
 
+# Option 2: Or FP16
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
 
-::  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|               |
@@ -474,23 +470,23 @@ 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:
 
@@ -505,7 +501,7 @@ 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
@@ -513,19 +509,13 @@ 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
 ```
 
@@ -533,67 +523,51 @@ 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 +10,9 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Recent API changes
 
+- [2024 Apr 21] `llama_token_to_piece` can now optionally render special tokens https://github.com/ggerganov/llama.cpp/pull/6807
+- [2024 Apr 4] State and session file functions reorganized under `llama_state_*` https://github.com/ggerganov/llama.cpp/pull/6341
+- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
 - [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
 - [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
 - [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
@@ -17,12 +20,13 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ### Hot topics
 
+- **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 parallelizm support https://github.com/ggerganov/llama.cpp/pull/6017
+- 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
-- Support loading sharded model, using `gguf-split` CLI https://github.com/ggerganov/llama.cpp/pull/6187
 
 ----
 
@@ -89,9 +93,11 @@ Typically finetunes of the base models below are supported as well.
 
 - [X] LLaMA 🦙
 - [x] LLaMA 2 🦙🦙
+- [x] LLaMA 3 🦙🦙🦙
 - [X] [Mistral 7B](https://huggingface.co/mistralai/Mistral-7B-v0.1)
 - [x] [Mixtral MoE](https://huggingface.co/models?search=mistral-ai/Mixtral)
-- [X] Falcon
+- [x] [DBRX](https://huggingface.co/databricks/dbrx-instruct)
+- [X] [Falcon](https://huggingface.co/models?search=tiiuae/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)
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
@@ -114,7 +120,14 @@ 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] [Command-R](https://huggingface.co/CohereForAI/c4ai-command-r-v01)
+- [x] [Grok-1](https://huggingface.co/keyfan/grok-1-hf)
+- [x] [Xverse](https://huggingface.co/models?search=xverse)
+- [x] [Command-R models](https://huggingface.co/models?search=CohereForAI/c4ai-command-r)
+- [x] [SEA-LION](https://huggingface.co/models?search=sea-lion)
+- [x] [GritLM-7B](https://huggingface.co/GritLM/GritLM-7B) + [GritLM-8x7B](https://huggingface.co/GritLM/GritLM-8x7B)
+- [x] [OLMo](https://allenai.org/olmo)
+
+(instructions for supporting more models: [HOWTO-add-model.md](./docs/HOWTO-add-model.md))
 
 **Multimodal models:**
 
@@ -124,6 +137,8 @@ Typically finetunes of the base models below are supported as well.
 - [x] [ShareGPT4V](https://huggingface.co/models?search=Lin-Chen/ShareGPT4V)
 - [x] [MobileVLM 1.7B/3B models](https://huggingface.co/models?search=mobileVLM)
 - [x] [Yi-VL](https://huggingface.co/models?search=Yi-VL)
+- [x] [Mini CPM](https://huggingface.co/models?search=MiniCPM)
+- [x] [Moondream](https://huggingface.co/vikhyatk/moondream2)
 
 **HTTP server**
 
@@ -138,6 +153,7 @@ 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)
@@ -147,6 +163,7 @@ 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:**
 
@@ -173,6 +190,14 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [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)
+- [AI Sublime Text plugin](https://github.com/yaroslavyaroslav/OpenAI-sublime-text) (MIT)
+
+*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---
 
@@ -448,44 +473,41 @@ 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
@@ -502,7 +524,7 @@ 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):
@@ -510,7 +532,7 @@ 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)
@@ -521,18 +543,18 @@ 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
 
   OpenCL acceleration is provided by the matrix multiplication kernels from the [CLBlast](https://github.com/CNugteren/CLBlast) project and custom kernels for ggml that can generate tokens on the GPU.
 
   You will need the [OpenCL SDK](https://github.com/KhronosGroup/OpenCL-SDK).
-    - For Ubuntu or Debian, the packages `opencl-headers`, `ocl-icd` may be needed.
+    - For Ubuntu, Debian, and Fedora the packages `opencl-headers`, `ocl-icd` may be needed.
 
     - For Windows, a pre-built SDK is available on the [OpenCL Releases](https://github.com/KhronosGroup/OpenCL-SDK/releases) page.
 
@@ -557,6 +579,12 @@ Building the program with BLAS support may lead to some performance improvements
 
   Pre-built CLBlast binaries may be found on the [CLBlast Releases](https://github.com/CNugteren/CLBlast/releases) page. For Unix variants, it may also be found in your operating system's packages.
 
+  Linux packaging:
+  Fedora Linux:
+  ```bash
+  sudo dnf install clblast
+  ```
+
   Alternatively, they may be built from source.
 
   - <details>
@@ -733,11 +761,11 @@ 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
 
@@ -745,7 +773,7 @@ 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 |
@@ -1093,7 +1121,9 @@ docker run --gpus all -v /path/to/models:/models local/llama.cpp:server-cuda -m
 - Clean-up any trailing whitespaces, use 4 spaces for indentation, brackets on the same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 - Tensors store data in row-major order. We refer to dimension 0 as columns, 1 as rows, 2 as matrices
-- Matrix multiplication is unconventional: [`z = ggml_mul_mat(ctx, x, y)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means `zT = x @ yT`
+- Matrix multiplication is unconventional: [`C = ggml_mul_mat(ctx, A, B)`](https://github.com/ggerganov/llama.cpp/blob/880e352277fc017df4d5794f0c21c44e1eae2b84/ggml.h#L1058-L1064) means $C^T = A B^T \Leftrightarrow C = B A^T.$
+
+![matmul](media/matmul.png)
 
 ### Docs
 

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

@@ -112,10 +112,12 @@ pub fn build(b: *std.build.Builder) !void {
     make.enable_lto = b.option(bool, "lto", "Enable LTO optimization, (default: false)") orelse false;
 
     const ggml = make.obj("ggml", "ggml.c");
+    const sgemm = make.obj("sgemm", "sgemm.cpp");
     const ggml_alloc = make.obj("ggml-alloc", "ggml-alloc.c");
     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");
@@ -127,15 +129,44 @@ pub fn build(b: *std.build.Builder) !void {
     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, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, 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, json_schema_to_grammar, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, sgemm, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, json_schema_to_grammar, buildinfo, sampling, grammar_parser, clip, llava });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }
+
+    const server_assets = [_][]const u8{ "index.html", "index.js", "completion.js", "json-schema-to-grammar.mjs" };
+    for (server_assets) |asset| {
+        const input_path = b.fmt("examples/server/public/{s}", .{asset});
+        const output_path = b.fmt("examples/server/{s}.hpp", .{asset});
+
+        // Portable equivalent of `b.addSystemCommand(&.{ "xxd", "-n", asset, "-i", input_path, output_path }) })`:
+
+        const input = try std.fs.cwd().readFileAlloc(b.allocator, input_path, std.math.maxInt(usize));
+        defer b.allocator.free(input);
+
+        var buf = std.ArrayList(u8).init(b.allocator);
+        defer buf.deinit();
+
+        for (input) |byte| {
+            try std.fmt.format(buf.writer(), "0x{X:0>2}, ", .{byte});
+        }
+
+        var name = try std.mem.replaceOwned(u8, b.allocator, asset, "-", "_");
+        defer b.allocator.free(name);
+        std.mem.replaceScalar(u8, name, '.', '_');
+
+        try std.fs.cwd().writeFile(output_path, b.fmt(
+            "unsigned char {s}[] = {{{s}}};\nunsigned int {s}_len = {d};\n",
+            .{ name, buf.items, name, input.len },
+        ));
+
+        std.debug.print("Dumped hex of \"{s}\" ({s}) to {s}\n", .{ input_path, name, output_path });
+    }
 }

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
@@ -153,6 +153,55 @@ function gg_sum_ctest_release {
     gg_printf '```\n'
 }
 
+# test_scripts_debug
+
+function gg_run_test_scripts_debug {
+    cd ${SRC}
+
+    set -e
+
+    # TODO: too slow, run on dedicated node
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+   #(cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-debug/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+
+    set +e
+}
+
+function gg_sum_test_scripts_debug {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test scripts in debug mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
+# test_scripts_release
+
+function gg_run_test_scripts_release {
+    cd ${SRC}
+
+    set -e
+
+    (cd ./examples/gguf-split && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+    (cd ./examples/quantize   && time bash tests.sh "$SRC/build-ci-release/bin" "$MNT/models") 2>&1 | tee -a $OUT/${ci}-scripts.log
+
+    set +e
+}
+
+function gg_sum_test_scripts_release {
+    gg_printf '### %s\n\n' "${ci}"
+
+    gg_printf 'Runs test scripts in release mode\n'
+    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
+    gg_printf '```\n'
+    gg_printf '%s\n' "$(cat $OUT/${ci}-scripts.log)"
+    gg_printf '```\n'
+    gg_printf '\n'
+}
+
 function gg_get_model {
     local gguf_3b="$MNT/models/open-llama/3B-v2/ggml-model-f16.gguf"
     local gguf_7b="$MNT/models/open-llama/7B-v2/ggml-model-f16.gguf"
@@ -412,8 +461,8 @@ 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 +624,7 @@ 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
@@ -642,6 +691,9 @@ test $ret -eq 0 && gg_run ctest_release
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
     test $ret -eq 0 && gg_run embd_bge_small
 
+    test $ret -eq 0 && gg_run test_scripts_debug
+    test $ret -eq 0 && gg_run test_scripts_release
+
     if [ -z ${GG_BUILD_VRAM_GB} ] || [ ${GG_BUILD_VRAM_GB} -ge 8 ]; then
         if [ -z ${GG_BUILD_CUDA} ]; then
             test $ret -eq 0 && gg_run open_llama_3b_v2

common/CMakeLists.txt

@@ -47,9 +47,6 @@ 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)
 
 add_library(${TARGET} STATIC
@@ -63,6 +60,7 @@ add_library(${TARGET} STATIC
     grammar-parser.h
     grammar-parser.cpp
     json.hpp
+    json-schema-to-grammar.cpp
     train.h
     train.cpp
     ngram-cache.h

common/common.cpp

@@ -1,4 +1,6 @@
 #include "common.h"
+#include "json.hpp"
+#include "json-schema-to-grammar.h"
 #include "llama.h"
 
 #include <algorithm>
@@ -16,6 +18,7 @@
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
+#include <codecvt>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -27,7 +30,6 @@
 #ifndef NOMINMAX
 #   define NOMINMAX
 #endif
-#include <codecvt>
 #include <locale>
 #include <windows.h>
 #include <fcntl.h>
@@ -48,12 +50,12 @@
 #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)
@@ -68,6 +70,8 @@
 #define LLAMA_CURL_MAX_HEADER_LENGTH 256
 #endif // LLAMA_USE_CURL
 
+using json = nlohmann::ordered_json;
+
 int32_t get_num_physical_cores() {
 #ifdef __linux__
     // enumerate the set of thread siblings, num entries is num cores
@@ -104,6 +108,79 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+#include <pthread.h>
+
+static void cpuid(unsigned leaf, unsigned subleaf,
+                  unsigned *eax, unsigned *ebx, unsigned *ecx, unsigned *edx) {
+    __asm__("movq\t%%rbx,%%rsi\n\t"
+            "cpuid\n\t"
+            "xchgq\t%%rbx,%%rsi"
+            : "=a"(*eax), "=S"(*ebx), "=c"(*ecx), "=d"(*edx)
+            : "0"(leaf), "2"(subleaf));
+}
+
+static int pin_cpu(int cpu) {
+    cpu_set_t mask;
+    CPU_ZERO(&mask);
+    CPU_SET(cpu, &mask);
+    return pthread_setaffinity_np(pthread_self(), sizeof(mask), &mask);
+}
+
+static bool is_hybrid_cpu(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(7, 0, &eax, &ebx, &ecx, &edx);
+    return !!(edx & (1u << 15));
+}
+
+static bool is_running_on_efficiency_core(void) {
+    unsigned eax, ebx, ecx, edx;
+    cpuid(0x1a, 0, &eax, &ebx, &ecx, &edx);
+    int intel_atom = 0x20;
+    int core_type = (eax & 0xff000000u) >> 24;
+    return core_type == intel_atom;
+}
+
+static int count_math_cpus(int cpu_count) {
+    int result = 0;
+    for (int cpu = 0; cpu < cpu_count; ++cpu) {
+        if (pin_cpu(cpu)) {
+            return -1;
+        }
+        if (is_running_on_efficiency_core()) {
+            continue; // efficiency cores harm lockstep threading
+        }
+        ++cpu; // hyperthreading isn't useful for linear algebra
+        ++result;
+    }
+    return result;
+}
+
+#endif // __x86_64__ && __linux__
+
+/**
+ * Returns number of CPUs on system that are useful for math.
+ */
+int get_math_cpu_count() {
+#if defined(__x86_64__) && defined(__linux__) && !defined(__ANDROID__)
+    int cpu_count = sysconf(_SC_NPROCESSORS_ONLN);
+    if (cpu_count < 1) {
+        return get_num_physical_cores();
+    }
+    if (is_hybrid_cpu()) {
+        cpu_set_t affinity;
+        if (!pthread_getaffinity_np(pthread_self(), sizeof(affinity), &affinity)) {
+            int result = count_math_cpus(cpu_count);
+            pthread_setaffinity_np(pthread_self(), sizeof(affinity), &affinity);
+            if (result > 0) {
+                return result;
+            }
+        }
+    }
+#endif
+    return get_num_physical_cores();
+}
+
 void process_escapes(std::string & input) {
     std::size_t input_len = input.length();
     std::size_t output_idx = 0;
@@ -157,15 +234,63 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     return result;
 }
 
-static 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;
+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.val_i64 = std::atol(sep);
+    } else if (strncmp(sep, "float:", 6) == 0) {
+        sep += 6;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_FLOAT;
+        kvo.val_f64 = 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.val_bool = true;
+        } else if (std::strcmp(sep, "false") == 0) {
+            kvo.val_bool = false;
+        } else {
+            fprintf(stderr, "%s: invalid boolean value for KV override '%s'\n", __func__, data);
+            return false;
+        }
+    } else if (strncmp(sep, "str:", 4) == 0) {
+        sep += 4;
+        kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+        if (strlen(sep) > 127) {
+            fprintf(stderr, "%s: malformed KV override '%s', value cannot exceed 127 chars\n", __func__, data);
+            return false;
+        }
+        strncpy(kvo.val_str, sep, 127);
+        kvo.val_str[127] = '\0';
+    } else {
+        fprintf(stderr, "%s: invalid type for KV override '%s'\n", __func__, data);
+        return false;
+    }
+    overrides.emplace_back(std::move(kvo));
+    return true;
+}
+
+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") {
         if (++i >= argc) {
             invalid_param = true;
             return true;
         }
+        // This is temporary, in the future the samplign state will be moved fully to llama_sampling_context.
         params.seed = std::stoul(argv[i]);
+        sparams.seed = std::stoul(argv[i]);
         return true;
     }
     if (arg == "-t" || arg == "--threads") {
@@ -861,9 +986,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
             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") {
@@ -889,9 +1014,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
             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") {
@@ -917,9 +1042,9 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
                 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") {
@@ -1010,6 +1135,10 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
         params.n_print = std::stoi(argv[i]);
         return true;
     }
+    if (arg == "--check-tensors") {
+        params.check_tensors = true;
+        return true;
+    }
     if (arg == "--ppl-output-type") {
         if (++i >= argc) {
             invalid_param = true;
@@ -1062,8 +1191,8 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
         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") {
@@ -1148,52 +1277,24 @@ static bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg,
         );
         return true;
     }
+    if (arg == "-j" || arg == "--json-schema") {
+        if (++i >= argc) {
+            invalid_param = true;
+            return true;
+        }
+        sparams.grammar = json_schema_to_grammar(json::parse(argv[i]));
+        return true;
+    }
     if (arg == "--override-kv") {
         if (++i >= argc) {
             invalid_param = true;
             return true;
         }
-        char* sep = strchr(argv[i], '=');
-        if (sep == nullptr || sep - argv[i] >= 128) {
-            fprintf(stderr, "error: Malformed KV override: %s\n", argv[i]);
-            invalid_param = true;
-            return true;
-        }
-        struct llama_model_kv_override kvo;
-        std::strncpy(kvo.key, argv[i], sep - argv[i]);
-        kvo.key[sep - argv[i]] = 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, "error: Invalid boolean value for KV override: %s\n", argv[i]);
-                invalid_param = true;
-                return true;
-            }
-        }
-        else {
+        if (!parse_kv_override(argv[i], params.kv_overrides)) {
             fprintf(stderr, "error: Invalid type for KV override: %s\n", argv[i]);
             invalid_param = true;
             return true;
         }
-        params.kv_overrides.push_back(kvo);
         return true;
     }
 #ifndef LOG_DISABLE_LOGS
@@ -1353,6 +1454,9 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        or `--logit-bias 15043-1` to decrease likelihood of token ' Hello'\n");
     printf("  --grammar GRAMMAR     BNF-like grammar to constrain generations (see samples in grammars/ dir)\n");
     printf("  --grammar-file FNAME  file to read grammar from\n");
+    printf("  -j SCHEMA, --json-schema SCHEMA\n");
+    printf("                        JSON schema to constrain generations (https://json-schema.org/), e.g. `{}` for any JSON object.\n");
+    printf("                        For schemas w/ external $refs, use --grammar + example/json_schema_to_grammar.py instead\n");
     printf("  --cfg-negative-prompt PROMPT\n");
     printf("                        negative prompt to use for guidance. (default: empty)\n");
     printf("  --cfg-negative-prompt-file FNAME\n");
@@ -1373,7 +1477,7 @@ 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");
@@ -1461,9 +1565,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     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");
+    printf("                        types: int, float, bool, str. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
     printf("  -ptc N, --print-token-count N\n");
     printf("                        print token count every N tokens (default: %d)\n", params.n_print);
+    printf("  --check-tensors       check model tensor data for invalid values\n");
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -1500,6 +1605,77 @@ 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
 //
@@ -1517,6 +1693,18 @@ std::vector<std::string> string_split(std::string input, char separator) {
     return parts;
 }
 
+std::string string_strip(const std::string & str) {
+    size_t start = 0;
+    size_t end = str.size();
+    while (start < end && std::isspace(str[start])) {
+        start++;
+    }
+    while (end > start && std::isspace(str[end - 1])) {
+        end--;
+    }
+    return str.substr(start, end - start);
+}
+
 std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
     std::unordered_map<std::string, llama_sampler_type> sampler_canonical_name_map {
         {"top_k",       llama_sampler_type::TOP_K},
@@ -1613,6 +1801,7 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
     mparams.tensor_split    = params.tensor_split;
     mparams.use_mmap        = params.use_mmap;
     mparams.use_mlock       = params.use_mlock;
+    mparams.check_tensors   = params.check_tensors;
     if (params.kv_overrides.empty()) {
         mparams.kv_overrides = NULL;
     } else {
@@ -1674,6 +1863,8 @@ 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);
@@ -1928,11 +2119,6 @@ struct llama_model * llama_load_model_from_url(
         return NULL;
     }
 
-    if (!curl) {
-        fprintf(stderr, "%s: error initializing libcurl\n", __func__);
-        return NULL;
-    }
-
     if (!llama_download_file(curl, model_url, path_model)) {
         return NULL;
     }
@@ -2126,7 +2312,7 @@ 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), };
@@ -2146,23 +2332,23 @@ 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);
@@ -2170,12 +2356,12 @@ std::vector<llama_token> llama_tokenize(
     return result;
 }
 
-std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token) {
+std::string llama_token_to_piece(const struct llama_context * ctx, llama_token token, bool special) {
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
     if (n_tokens < 0) {
         result.resize(-n_tokens);
-        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size());
+        int check = llama_token_to_piece(llama_get_model(ctx), token, result.data(), result.size(), special);
         GGML_ASSERT(check == -n_tokens);
     } else {
         result.resize(n_tokens);
@@ -2387,7 +2573,7 @@ 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");
@@ -2489,7 +2675,7 @@ 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

@@ -39,6 +39,7 @@ extern char const *LLAMA_BUILD_TARGET;
 
 struct llama_control_vector_load_info;
 
+int get_math_cpu_count();
 int32_t get_num_physical_cores();
 
 //
@@ -48,7 +49,7 @@ int32_t get_num_physical_cores();
 struct gpt_params {
     uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
 
-    int32_t n_threads             = get_num_physical_cores();
+    int32_t n_threads             = get_math_cpu_count();
     int32_t n_threads_draft       = -1;
     int32_t n_threads_batch       = -1;    // number of threads to use for batch processing (-1 = use n_threads)
     int32_t n_threads_batch_draft = -1;
@@ -80,10 +81,13 @@ 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;
-    llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
+    enum llama_rope_scaling_type rope_scaling_type = LLAMA_ROPE_SCALING_TYPE_UNSPECIFIED;
+    enum llama_pooling_type      pooling_type      = LLAMA_POOLING_TYPE_UNSPECIFIED; // pooling type for embeddings
 
     // // sampling parameters
     struct llama_sampling_params sparams;
@@ -156,6 +160,8 @@ 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
+    bool check_tensors     = false; // validate tensor data
 
     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
@@ -165,18 +171,24 @@ struct gpt_params {
     std::string image  = ""; // path to an image file
 };
 
+bool parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
+
 bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params);
 
 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
 //
@@ -184,6 +196,7 @@ void process_escapes(std::string& input);
 std::vector<llama_sampler_type> sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names);
 std::vector<llama_sampler_type> sampler_types_from_chars(const std::string & names_string);
 std::vector<std::string> string_split(std::string input, char separator);
+std::string string_strip(const std::string & str);
 std::string sampler_type_to_name_string(llama_sampler_type sampler_type);
 
 //
@@ -219,20 +232,21 @@ 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
+// tokenizes a token into a piece, optionally renders special/control tokens
 // should work similar to Python's `tokenizer.id_to_piece`
 std::string llama_token_to_piece(
         const struct llama_context * ctx,
-                       llama_token   token);
+                       llama_token   token,
+                       bool          special = true);
 
 // TODO: these should be moved in llama.h C-style API under single `llama_detokenize` function
 //       that takes into account the tokenizer type and decides how to handle the leading space

common/json-schema-to-grammar.cpp

@@ -11,35 +11,101 @@
 
 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 = "\" \"?";
 
-std::unordered_map<std::string, std::string> 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"},
-    {"value", "object | array | string | number | boolean"},
-    {"object", "\"{\" space ( string \":\" space value (\",\" space string \":\" space value)* )? \"}\" space"},
-    {"array", "\"[\" space ( value (\",\" space value)* )? \"]\" space"},
-    {"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"},
-    {"string", " \"\\\"\" (\n"
-               "        [^\"\\\\] |\n"
-               "        \"\\\\\" ([\"\\\\/bfnrt] | \"u\" [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F] [0-9a-fA-F])\n"
-               "      )* \"\\\"\" space"},
-    {"null", "\"null\" space"}
+struct BuiltinRule {
+    std::string content;
+    std::vector<std::string> deps;
 };
-std::vector<std::string> OBJECT_RULE_NAMES = {"object", "array", "string", "number", "boolean", "null", "value"};
 
-std::unordered_map<std::string, std::string> DATE_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-string", "\"\\\"\" date \"\\\"\" space"},
-    {"time-string", "\"\\\"\" time \"\\\"\" space"},
-    {"date-time-string", "\"\\\"\" date-time \"\\\"\" space"}
+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) {
@@ -47,7 +113,7 @@ static bool is_reserved_name(const std::string & name) {
     if (RESERVED_NAMES.empty()) {
         RESERVED_NAMES.insert("root");
         for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : DATE_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();
 }
@@ -192,7 +258,7 @@ private:
                 if (_dotall) {
                     rule = "[\\U00000000-\\U0010FFFF]";
                 } else {
-                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
+                    rule = "[^\\x0A\\x0D]";
                 }
                 return _add_rule("dot", rule);
             };
@@ -308,47 +374,21 @@ private:
                     auto &sub = last.first;
                     auto sub_is_literal = last.second;
 
-                    if (min_times == 0 && max_times == std::numeric_limits<int>::max()) {
-                        sub += "*";
-                    } else if (min_times == 0 && max_times == 1) {
-                        sub += "?";
-                    } else if (min_times == 1 && max_times == std::numeric_limits<int>::max()) {
-                        sub += "+";
-                    } else {
-                        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;
+                    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);
                         }
-                        std::string result;
-                        if (sub_is_literal && min_times > 0) {
-                            result = "\"" + repeat(sub.substr(1, sub.length() - 2), min_times) + "\"";
-                        } else {
-                            for (int j = 0; j < min_times; j++) {
-                                if (j > 0) {
-                                    result += " ";
-                                }
-                                result += sub;
-                            }
-                        }
-                        if (min_times > 0 && min_times < max_times) {
-                            result += " ";
-                        }
-                        if (max_times == std::numeric_limits<int>::max()) {
-                            result += sub + "*";
-                        } else {
-                            for (int j = min_times; j < max_times; j++) {
-                                if (j > min_times) {
-                                    result += " ";
-                                }
-                                result += sub + "?";
-                            }
-                        }
-                        seq.back().first = result;
-                        seq.back().second = false;
+                        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) {
@@ -424,7 +464,7 @@ private:
         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_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
+            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("*");
         }
@@ -486,6 +526,25 @@ private:
         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,
@@ -647,49 +706,33 @@ public:
                 return _add_rule(rule_name, rule);
             } else {
                 std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
-                std::string list_item_operator = "( \",\" space " + item_rule_name + " )";
-                std::string successive_items;
                 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>() : -1;
-                if (min_items > 0) {
-                    successive_items += repeat(list_item_operator, min_items - 1);
-                    min_items--;
-                }
-                if (max_items >= 0 && max_items > min_items) {
-                    successive_items += repeat(list_item_operator + "?", max_items - min_items - 1);
-                } else {
-                    successive_items += list_item_operator + "*";
-                }
-                std::string rule;
-                if (min_items == 0) {
-                    rule =  "\"[\" space ( " + item_rule_name + " " + successive_items + " )? \"]\" space";
-                } else {
-                    rule =  "\"[\" space " + item_rule_name + " " + successive_items + " \"]\" space";
-                }
-                return _add_rule(rule_name, rule);
+                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_rule(rule_name == "root" ? "root" : schema_format, PRIMITIVE_RULES.at("uuid"));
-        } else if ((schema_type.is_null() || schema_type == "string") && DATE_RULES.find(schema_format) != DATE_RULES.end()) {
-            for (const auto & kv : DATE_RULES) {
-                _add_rule(kv.first, kv.second);
-            }
-            return schema_format + "-string";
+            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") {
-            for (const auto & n : OBJECT_RULE_NAMES) {
-                _add_rule(n, PRIMITIVE_RULES.at(n));
-            }
-            return _add_rule(rule_name, "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_rule(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
+            return _add_primitive(rule_name == "root" ? "root" : schema_type.get<std::string>(), PRIMITIVE_RULES.at(schema_type.get<std::string>()));
         }
     }
 

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) || 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) || 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

@@ -1,6 +1,8 @@
 #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,

common/sampling.cpp

@@ -1,4 +1,6 @@
+#define LLAMA_API_INTERNAL
 #include "sampling.h"
+#include <random>
 
 struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_params & params) {
     struct llama_sampling_context * result = new llama_sampling_context();
@@ -33,6 +35,8 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
 
     result->prev.resize(params.n_prev);
 
+    llama_sampling_set_rng_seed(result, params.seed);
+
     return result;
 }
 
@@ -62,6 +66,13 @@ void llama_sampling_reset(llama_sampling_context * ctx) {
     ctx->cur.clear();
 }
 
+void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed) {
+    if (seed == LLAMA_DEFAULT_SEED) {
+        seed = std::random_device{}();
+    }
+    ctx->rng.seed(seed);
+}
+
 void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst) {
     if (dst->grammar) {
         llama_grammar_free(dst->grammar);
@@ -168,77 +179,20 @@ 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);
@@ -260,7 +214,7 @@ static llama_token llama_sampling_sample_impl(
 
             sampler_queue(ctx_main, params, cur_p, min_keep);
 
-            id = llama_sample_token(ctx_main, &cur_p);
+            id = llama_sample_token_with_rng(ctx_main, &cur_p, ctx_sampling->rng);
 
             //{
             //    const int n_top = 10;
@@ -302,11 +256,13 @@ 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 +271,7 @@ 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 +280,10 @@ 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 +323,11 @@ 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 +340,14 @@ 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

@@ -4,9 +4,10 @@
 
 #include "grammar-parser.h"
 
+#include <random>
 #include <string>
-#include <vector>
 #include <unordered_map>
+#include <vector>
 
 // sampler types
 enum class llama_sampler_type : char {
@@ -20,25 +21,26 @@ enum class llama_sampler_type : char {
 
 // sampling parameters
 typedef struct llama_sampling_params {
-    int32_t     n_prev                = 64;       // number of previous tokens to remember
-    int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t     top_k                 = 40;       // <= 0 to use vocab size
-    float       top_p                 = 0.95f;    // 1.0 = disabled
-    float       min_p                 = 0.05f;    // 0.0 = disabled
-    float       tfs_z                 = 1.00f;    // 1.0 = disabled
-    float       typical_p             = 1.00f;    // 1.0 = disabled
-    float       temp                  = 0.80f;    // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float       dynatemp_range        = 0.00f;    // 0.0 = disabled
-    float       dynatemp_exponent     = 1.00f;    // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t     penalty_last_n        = 64;       // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float       penalty_repeat        = 1.00f;    // 1.0 = disabled
-    float       penalty_freq          = 0.00f;    // 0.0 = disabled
-    float       penalty_present       = 0.00f;    // 0.0 = disabled
-    int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float       mirostat_tau          = 5.00f;    // target entropy
-    float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = false;     // consider newlines as a repeatable token
+    int32_t     n_prev                = 64;                 // number of previous tokens to remember
+    int32_t     n_probs               = 0;                  // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t     min_keep              = 0;                  // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t     top_k                 = 40;                 // <= 0 to use vocab size
+    float       top_p                 = 0.95f;              // 1.0 = disabled
+    float       min_p                 = 0.05f;              // 0.0 = disabled
+    float       tfs_z                 = 1.00f;              // 1.0 = disabled
+    float       typical_p             = 1.00f;              // 1.0 = disabled
+    float       temp                  = 0.80f;              // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float       dynatemp_range        = 0.00f;              // 0.0 = disabled
+    float       dynatemp_exponent     = 1.00f;              // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t     penalty_last_n        = 64;                 // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float       penalty_repeat        = 1.00f;              // 1.0 = disabled
+    float       penalty_freq          = 0.00f;              // 0.0 = disabled
+    float       penalty_present       = 0.00f;              // 0.0 = disabled
+    int32_t     mirostat              = 0;                  // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float       mirostat_tau          = 5.00f;              // target entropy
+    float       mirostat_eta          = 0.10f;              // learning rate
+    bool        penalize_nl           = false;              // consider newlines as a repeatable token
+    uint32_t    seed                  = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampling_context
 
     std::vector<llama_sampler_type> samplers_sequence = {
         llama_sampler_type::TOP_K,
@@ -79,6 +81,8 @@ struct llama_sampling_context {
     // TODO: replace with ring-buffer
     std::vector<llama_token>      prev;
     std::vector<llama_token_data> cur;
+
+    std::mt19937 rng;
 };
 
 #include "common.h"
@@ -93,6 +97,9 @@ void llama_sampling_free(struct llama_sampling_context * ctx);
 // - reset grammar
 void llama_sampling_reset(llama_sampling_context * ctx);
 
+// Set the sampler seed
+void llama_sampling_set_rng_seed(struct llama_sampling_context * ctx, uint32_t seed);
+
 // Copy the sampler context
 void llama_sampling_cp(llama_sampling_context * src, llama_sampling_context * dst);
 
@@ -129,14 +136,16 @@ 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-update.py

@@ -0,0 +1,275 @@
+# This script downloads the tokenizer models of the specified models from Huggingface and
+# generates the get_vocab_base_pre() function for convert-hf-to-gguf.py
+#
+# This is necessary in order to analyze the type of pre-tokenizer used by the model and
+# provide the necessary information to llama.cpp via the GGUF header in order to implement
+# the same pre-tokenizer.
+#
+# ref: https://github.com/ggerganov/llama.cpp/pull/6920
+#
+# Instructions:
+#
+# - Add a new model to the "models" list
+# - Run the script with your huggingface token:
+#
+#   python3 convert-hf-to-gguf-update.py <huggingface_token>
+#
+# - Copy-paste the generated get_vocab_base_pre() function into convert-hf-to-gguf.py
+# - Update llama.cpp with the new pre-tokenizer if necessary
+#
+# TODO: generate tokenizer tests for llama.cpp
+# TODO: automate the update of convert-hf-to-gguf.py
+#
+
+import os
+import requests
+import sys
+import json
+
+from hashlib import sha256
+from enum import IntEnum, auto
+
+class TOKENIZER_TYPE(IntEnum):
+    SPM = auto()
+    BPE = auto()
+    WPM = auto()
+
+# TODO: this string has to exercise as much pre-tokenizer functionality as possible
+#       will be updated with time - contributions welcome
+chktxt = '\n \n\n \n\n\n \t \t\t \t\n  \n   \n    \n     \n🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ 🦙🦙 3 33 333 3333 33333 333333 3333333 33333333 3.3 3..3 3...3 កាន់តែពិសេសអាច😁 ?我想在apple工作1314151天～ ------======= нещо на Български \'\'\'\'\'\'```````\"\"\"\"......!!!!!!?????? I\'ve been \'told he\'s there, \'RE you sure? \'M not sure I\'ll make it, \'D you like some tea? We\'Ve a\'lL'
+
+if len(sys.argv) == 2:
+    token = sys.argv[1]
+else:
+    print("Usage: python convert-hf-to-gguf-update.py <huggingface_token>")
+    sys.exit(1)
+
+# TODO: add models here, base models preferred
+models = [
+        { "name": "llama-spm",      "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/meta-llama/Llama-2-7b-hf", },
+        { "name": "llama-bpe",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/meta-llama/Meta-Llama-3-8B", },
+        { "name": "phi-3",          "tokt": TOKENIZER_TYPE.SPM, "repo": "https://huggingface.co/microsoft/Phi-3-mini-4k-instruct", },
+        { "name": "deepseek-llm",   "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-llm-7b-base", },
+        { "name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
+        { "name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
+        { "name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+        { "name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
+        { "name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
+        { "name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },
+        ]
+
+# make directory "models/tokenizers" if it doesn't exist
+if not os.path.exists("models/tokenizers"):
+    os.makedirs("models/tokenizers")
+
+def download_file_with_auth(url, token, save_path):
+    headers = {"Authorization": f"Bearer {token}"}
+    response = requests.get(url, headers=headers)
+    if response.status_code == 200:
+        with open(save_path, 'wb') as f:
+            f.write(response.content)
+        print(f"File {save_path} downloaded successfully")
+    else:
+        print(f"Failed to download file. Status code: {response.status_code}")
+
+# download the tokenizer models
+for model in models:
+    name = model["name"]
+    repo = model["repo"]
+    tokt = model["tokt"]
+
+    if not os.path.exists(f"models/tokenizers/{name}"):
+        os.makedirs(f"models/tokenizers/{name}")
+    else:
+        print(f"Directory models/tokenizers/{name} already exists - skipping")
+        continue
+
+    print(f"Downloading {name} to models/tokenizers/{name}")
+
+    url = f"{repo}/raw/main/config.json"
+    save_path = f"models/tokenizers/{name}/config.json"
+    download_file_with_auth(url, token, save_path)
+
+    url = f"{repo}/raw/main/tokenizer.json"
+    save_path = f"models/tokenizers/{name}/tokenizer.json"
+    download_file_with_auth(url, token, save_path)
+
+    if tokt == TOKENIZER_TYPE.SPM:
+        url = f"{repo}/resolve/main/tokenizer.model"
+        save_path = f"models/tokenizers/{name}/tokenizer.model"
+        download_file_with_auth(url, token, save_path)
+
+    url = f"{repo}/raw/main/tokenizer_config.json"
+    save_path = f"models/tokenizers/{name}/tokenizer_config.json"
+    download_file_with_auth(url, token, save_path)
+
+# generate the source code for the convert-hf-to-gguf.py:get_vocab_base_pre() function:
+# TODO: auto-update convert-hf-to-gguf.py with the generated function
+
+src_ifs = ""
+for model in models:
+    name = model["name"]
+    tokt = model["tokt"]
+
+    if tokt == TOKENIZER_TYPE.SPM:
+        continue
+
+    # create the tokenizer
+    from transformers import AutoTokenizer
+    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+
+    chktok = tokenizer.encode(chktxt)
+    chkhsh = sha256(str(chktok).encode()).hexdigest()
+
+    print(f"model: {name}")
+    print(f"tokt: {tokt}")
+    print(f"repo: {model['repo']}")
+    print(f"chktok: {chktok}")
+    print(f"chkhsh: {chkhsh}")
+
+    # print the "pre_tokenizer" content from the tokenizer.json
+    with open(f"models/tokenizers/{name}/tokenizer.json", "r") as f:
+        cfg = json.load(f)
+        pre_tokenizer = cfg["pre_tokenizer"]
+        print("pre_tokenizer: " + json.dumps(pre_tokenizer, indent=4))
+
+    print(f"\n")
+
+    src_ifs += f"        if chkhsh == \"{chkhsh}\":\n"
+    src_ifs += f"            # ref: {model['repo']}\n"
+    src_ifs += f"            res = \"{name}\"\n"
+
+src_func = ""
+src_func +=  "    def get_vocab_base_pre(self, tokenizer) -> str:\n"
+src_func +=  "        # encoding this string and hashing the resulting tokens would (hopefully) give us a unique identifier that\n"
+src_func +=  "        # is specific for the BPE pre-tokenizer used by the model\n"
+src_func +=  "        # we will use this unique identifier to write a \"tokenizer.ggml.pre\" entry in the GGUF file which we can\n"
+src_func +=  "        # use in llama.cpp to implement the same pre-tokenizer\n"
+src_func +=  "\n"
+src_func += f"        chktxt = {repr(chktxt)}\n"
+src_func +=  "\n"
+src_func +=  "        chktok = tokenizer.encode(chktxt)\n"
+src_func +=  "        chkhsh = sha256(str(chktok).encode()).hexdigest()\n"
+src_func +=  "\n"
+src_func +=  "        print(f\"chktok: {chktok}\")\n"
+src_func +=  "        print(f\"chkhsh: {chkhsh}\")\n"
+src_func +=  "\n"
+src_func +=  "        res = None\n"
+src_func +=  "\n"
+src_func +=  "        # NOTE: if you get an error here, you need to add the model to the if-elif chain below\n"
+src_func +=  "        #       don't do this manually - use the convert-hf-to-gguf-update.py script!\n"
+src_func += f"{src_ifs}\n"
+src_func +=  "        if res is None:\n"
+src_func +=  "            print(\"\\n\")\n"
+src_func +=  "            print(\"**************************************************************************************\")\n"
+src_func +=  "            print(\"** WARNING: The BPE pre-tokenizer was not recognized!\")\n"
+src_func +=  "            print(\"**          This means that it was not added yet or you are using an older version.\")\n"
+src_func +=  "            print(\"**          Check convert-hf-to-gguf-update.py and update it accordingly.\")\n"
+src_func +=  "            print(\"**\")\n"
+src_func +=  "            print(f\"** chkhsh:  {chkhsh}\")\n"
+src_func +=  "            print(\"**************************************************************************************\")\n"
+src_func +=  "            print(\"\\n\")\n"
+src_func +=  "            raise NotImplementedError(\"BPE pre-tokenizer was not recognized - update get_vocab_base_pre()\")\n"
+src_func +=  "\n"
+src_func +=  "        print(f\"tokenizer.ggml.pre: {res}\")\n"
+src_func +=  "        print(f\"chkhsh: {chkhsh}\")\n"
+src_func +=  "\n"
+src_func +=  "        return res\n"
+
+print(src_func)
+
+print("\n")
+print("!!! Copy-paste the function above into convert-hf-to-gguf.py !!!")
+print("\n")
+
+# generate tests for each tokenizer model
+
+tests = [
+    "",
+    " ",
+    "  ",
+    "   ",
+    "\t",
+    "\n",
+    "\n\n",
+    "\n\n\n",
+    "\t\n",
+    "Hello world",
+    " Hello world",
+    "Hello World",
+    " Hello World",
+    " Hello World!",
+    "Hello, world!",
+    " Hello, world!",
+    " this is 🦙.cpp",
+    "w048 7tuijk dsdfhu",
+    "нещо на Български",
+    "កាន់តែពិសេសអាចខលចេញ",
+    "🚀 (normal) 😶‍🌫️ (multiple emojis concatenated) ✅ (only emoji that has its own token)",
+    "Hello",
+    " Hello",
+    "  Hello",
+    "   Hello",
+    "    Hello",
+    "    Hello\n    Hello",
+    " (",
+    "\n =",
+    "' era",
+    "Hello, y'all! How are you 😁 ?我想在apple工作1314151天～",
+    "3",
+    "33",
+    "333",
+    "3333",
+    "33333",
+    "333333",
+    "3333333",
+    "33333333",
+    "333333333",
+    chktxt,
+]
+
+# write the tests to ./models/ggml-vocab-{name}.gguf.inp
+# the format is:
+#
+# test0
+# __ggml_vocab_test__
+# test1
+# __ggml_vocab_test__
+# ...
+#
+
+# with each model, encode all tests and write the results in ./models/ggml-vocab-{name}.gguf.out
+# for each test, write the resulting tokens on a separate line
+
+for model in models:
+    name = model["name"]
+    tokt = model["tokt"]
+
+    # create the tokenizer
+    from transformers import AutoTokenizer
+    tokenizer = AutoTokenizer.from_pretrained(f"models/tokenizers/{name}")
+
+    with open(f"models/ggml-vocab-{name}.gguf.inp", "w") as f:
+        for text in tests:
+            f.write(f"{text}")
+            f.write("\n__ggml_vocab_test__\n")
+
+    with open(f"models/ggml-vocab-{name}.gguf.out", "w") as f:
+        for text in tests:
+            res = tokenizer.encode(text, add_special_tokens=False)
+            for r in res:
+                f.write(f" {r}")
+            f.write("\n")
+
+    print(f"Tests for {name} written in ./models/ggml-vocab-{name}.gguf.*")
+
+# generate commands for creating vocab files
+
+print("\nRun the following commands to generate the vocab files for testing:\n")
+
+for model in models:
+    name = model["name"]
+
+    print(f"python3 convert-hf-to-gguf.py models/tokenizers/{name}/ --outfile models/ggml-vocab-{name}.gguf --vocab-only")
+
+print("\n")

convert-llama-ggml-to-gguf.py

@@ -281,6 +281,7 @@ class GGMLToGGUF:
     def add_vocab(self, gguf_writer):
         hp = self.model.hyperparameters
         gguf_writer.add_tokenizer_model('llama')
+        gguf_writer.add_tokenizer_pre('default')
         tokens = []
         scores = []
         toktypes = []

convert-persimmon-to-gguf.py

@@ -1,4 +1,6 @@
 #!/usr/bin/env python3
+from __future__ import annotations
+
 import argparse
 import os
 import sys
@@ -97,6 +99,7 @@ def main():
 
     tokens, scores, toktypes = _get_sentencepiece_tokenizer_info(args.model_dir)
     gguf_writer.add_tokenizer_model('llama')
+    gguf_writer.add_tokenizer_pre('default')
     gguf_writer.add_token_list(tokens)
     gguf_writer.add_token_scores(scores)
     gguf_writer.add_token_types(toktypes)
@@ -106,12 +109,12 @@ 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 +513,47 @@ 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']
+        is_llama3 = (
+            tokenizer_model['type'] == 'BPE' and tokenizer_model.get('ignore_merges', False)
+            and not tokenizer_model.get('byte_fallback', True)
+        )
+        if is_llama3:
+            raise TypeError('Llama 3 must be converted with BpeVocab')
+
+        if not is_llama3 and (
+            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 +581,7 @@ 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 +633,7 @@ 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 +651,18 @@ 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 +674,18 @@ 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 +753,7 @@ 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 +762,7 @@ 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 +783,7 @@ 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 +835,15 @@ 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 +880,10 @@ 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 +904,7 @@ 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 +963,7 @@ 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 +1021,14 @@ 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 +1053,11 @@ 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 +1108,6 @@ 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 +1207,7 @@ 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 +1217,11 @@ 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 +1248,7 @@ 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 +1258,26 @@ 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 +1301,7 @@ 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 +1318,7 @@ 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 +1357,16 @@ 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 +1380,14 @@ 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 +1396,29 @@ 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 +1475,8 @@ 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 +1498,12 @@ 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 +1518,11 @@ 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.swift/Sources/main.swift

@@ -153,7 +153,7 @@ while n_cur <= n_len {
         // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
-        if new_token_id == llama_token_eos(model) || n_cur == n_len {
+        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {
@@ -229,7 +229,7 @@ private func tokenize(text: String, add_bos: Bool) -> [llama_token] {
 
 private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String? {
     var result = [CChar](repeating: 0, count: 8)
-    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count))
+    let nTokens = llama_token_to_piece(model, token, &result, Int32(result.count), false)
     if nTokens < 0 {
         let actualTokensCount = -Int(nTokens)
         result = .init(repeating: 0, count: actualTokensCount)
@@ -237,7 +237,8 @@ private func token_to_piece(token: llama_token, buffer: inout [CChar]) -> String
             model,
             token,
             &result,
-            Int32(result.count)
+            Int32(result.count),
+            false
         )
         assert(check == actualTokensCount)
     } else {

examples/batched/batched.cpp

@@ -191,8 +191,8 @@ int main(int argc, char ** argv) {
 
             //const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
-            // is it an end of stream? -> mark the stream as finished
-            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+            // is it an end of generation? -> mark the stream as finished
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
                 i_batch[i] = -1;
                 LOG_TEE("\n");
                 if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -47,7 +47,7 @@ struct beam_search_callback_data {
 // In this case, end-of-beam (eob) is equivalent to end-of-sentence (eos) but this need not always be the same.
 // For example, eob can be flagged due to maximum token length, stop words, etc.
 static bool is_at_eob(const beam_search_callback_data & callback_data, const llama_token * tokens, size_t n_tokens) {
-    return n_tokens && tokens[n_tokens-1] == llama_token_eos(llama_get_model(callback_data.ctx));
+    return n_tokens && llama_token_is_eog(llama_get_model(callback_data.ctx), tokens[n_tokens-1]);
 }
 
 // Function matching type llama_beam_search_callback_fn_t.

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

@@ -880,7 +880,7 @@ int main(int argc, char ** argv) {
     TransformerWeights weights = {};
     {
         LOG("%s: Loading llama2c model from %s\n", __func__, params.fn_llama2c_model);
-        FILE *file = fopen(params.fn_llama2c_model, "r");
+        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;

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/README.md

@@ -5,5 +5,6 @@ CLI to split / merge GGUF files.
 **Command line options:**
 
 - `--split`: split GGUF to multiple GGUF, default operation.
+- `--split-max-size`: max size per split in `M` or `G`, f.ex. `500M` or `2G`.
 - `--split-max-tensors`: maximum tensors in each split: default(128)
 - `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -28,9 +28,11 @@ enum split_operation : uint8_t {
 
 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) {
@@ -41,15 +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 = (size_t)n * 1024 * 1024; // megabytes
+    } else if (str.back() == 'G') {
+        sscanf(str.c_str(), "%d", &n);
+        n_bytes = (size_t)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;
@@ -62,6 +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);
@@ -71,23 +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);
@@ -99,24 +147,17 @@ 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(EXIT_FAILURE);
-        }
+        split_params_parse_ex(argc, argv, params);
     }
     catch (const std::invalid_argument & ex) {
         fprintf(stderr, "%s\n", ex.what());
@@ -140,15 +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,
@@ -158,79 +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_u16(ctx_out, LLM_KV_SPLIT_NO, i_split);
-        gguf_set_val_u16(ctx_out, LLM_KV_SPLIT_COUNT, n_split);
-        gguf_set_val_i32(ctx_out, LLM_KV_SPLIT_TENSORS_COUNT, n_tensors);
-
-        // 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);
-        }
-
-        char split_path[PATH_MAX] = {0};
-        llama_split_path(split_path, sizeof(split_path), params.output.c_str(), i_split, n_split);
-
-        fprintf(stderr, "%s: %s ...", __func__, split_path);
-        fout = std::ofstream(split_path, 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);
     }
 };
 
@@ -254,32 +353,22 @@ static void gguf_split(const split_params & split_params) {
         exit(EXIT_FAILURE);
     }
 
+    // prepare the strategy
     split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
+    int n_split = strategy.ctx_outs.size();
+    strategy.print_info();
 
-    char first_split_path[PATH_MAX] = {0};
-    llama_split_path(first_split_path, sizeof(first_split_path),
-                     split_params.output.c_str(), strategy.i_split, strategy.n_split);
-    fprintf(stderr, "%s: %s -> %s (%d tensors per file)\n",
-            __func__, split_params.input.c_str(),
-            first_split_path,
-            split_params.n_split_tensors);
-
-    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) {
@@ -448,10 +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);
 

examples/gguf-split/tests.sh

@@ -0,0 +1,89 @@
+#!/bin/bash
+
+set -eu
+
+if [ $# -lt 1 ]
+then
+    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
+    echo "example: $0 ../../build/bin ../../tmp"
+    exit 1
+fi
+
+if [ $# -gt 1 ]
+then
+    TMP_DIR=$2
+else
+    TMP_DIR=/tmp
+fi
+
+set -x
+
+SPLIT=$1/gguf-split
+MAIN=$1/main
+WORK_PATH=$TMP_DIR/gguf-split
+ROOT_DIR=$(realpath $(dirname $0)/../../)
+
+mkdir -p "$WORK_PATH"
+
+# Clean up in case of previously failed test
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf
+
+# 1. Get a model
+(
+cd $WORK_PATH
+"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/gemma-1.1-2b-it-Q8_0-GGUF --file gemma-1.1-2b-it.Q8_0.gguf
+)
+echo PASS
+
+# 2. Split with max tensors strategy
+$SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/ggml-model-split
+echo PASS
+echo
+
+# 2b. Test the sharded model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-split-00001-of-00006.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# 3. Merge
+$SPLIT --merge $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-merge.gguf
+echo PASS
+echo
+
+# 3b. Test the merged model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-merge.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# 4. Split with no tensor in metadata
+#$SPLIT --split-max-tensors 32 --no-tensor-in-metadata $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-32-tensors
+#echo PASS
+#echo
+
+# 4b. Test the sharded model is loading properly
+#$MAIN --model $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf --random-prompt --n-predict 32
+#echo PASS
+#echo
+
+# 5. Merge
+#$SPLIT --merge $WORK_PATH/ggml-model-split-32-tensors-00001-of-00006.gguf $WORK_PATH/ggml-model-merge-2.gguf
+#echo PASS
+#echo
+
+# 5b. Test the merged model is loading properly
+#$MAIN --model $WORK_PATH/ggml-model-merge-2.gguf --random-prompt --n-predict 32
+#echo PASS
+#echo
+
+# 6. Split with size strategy
+$SPLIT --split-max-size 2G $WORK_PATH/ggml-model-merge.gguf $WORK_PATH/ggml-model-split-2G
+echo PASS
+echo
+
+# 6b. Test the sharded model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-split-2G-00001-of-00002.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# Clean up
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-merge*.gguf

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/gritlm/README.md

@@ -21,12 +21,12 @@ not have to be performed at all.
 ### Running the example
 Download a Grit model:
 ```console
-$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf
+$ scripts/hf.sh --repo cohesionet/GritLM-7B_gguf --file gritlm-7b_q4_1.gguf --outdir models
 ```
 
 Run the example using the downloaded model:
 ```console
-$ ./gritlm -m gritlm-7b_q4_1.gguf
+$ ./gritlm -m models/gritlm-7b_q4_1.gguf
 
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "A purely peer-to-peer version of electronic cash w" is: 0.605
 Cosine similarity between "Bitcoin: A Peer-to-Peer Electronic Cash System" and "All text-based language problems can be reduced to" is: 0.103

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

@@ -23,6 +23,7 @@ struct Stats {
 };
 
 struct StatParams {
+    std::string dataset;
     std::string ofile = "imatrix.dat";
     int         n_output_frequency = 10;
     int         verbosity = 1;
@@ -44,41 +45,44 @@ private:
     std::mutex                             m_mutex;
     int                                    m_last_call = 0;
     std::vector<float>                     m_src1_data;
-    std::vector<int>                       m_ids; // the expert ids from ggml_mul_mat_id
+    std::vector<char>                      m_ids; // the expert ids from ggml_mul_mat_id
                                                   //
-    void save_imatrix(const char * file_name) const;
+    void save_imatrix(const char * file_name, const char * dataset) const;
     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
     if (ask) {
         if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
         if (t->op != GGML_OP_MUL_MAT) return false;
+        // why are small batches ignored (<16 tokens)?
         if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return false;
         if (!(wname.substr(0, 4) == "blk." || (m_params.collect_output_weight && wname == "output.weight"))) return false;
         return true;
@@ -96,44 +100,60 @@ 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];
+        //   ids  -> [n_experts_used, n_tokens]
+        //   src1 -> [cols, n_expert_used, n_tokens]
+        const ggml_tensor * ids = t->src[2];
+        const int n_as = src0->ne[2];
+        const int n_ids = ids->ne[0];
 
-        // 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
+        // take into account that ids is not contiguous!
 
+        GGML_ASSERT(ids->ne[1] == src1->ne[2]);
+
+        m_ids.resize(ggml_nbytes(ids));
+        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;
+
+        if (e.values.empty()) {
+            e.values.resize(src1->ne[0]*n_as, 0);
+        }
+        else if (e.values.size() != (size_t)src1->ne[0]*n_as) {
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]*n_as);
+            exit(1); //GGML_ASSERT(false);
+        }
+        if (m_params.verbosity > 1) {
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, wname.c_str(), ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[2], (int)src1->type);
+        }
         // 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];
-            if (e.values.empty()) {
-                e.values.resize(src1->ne[0], 0);
-            }
-            else if (e.values.size() != (size_t)src1->ne[0]) {
-                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", wname.c_str(), (int)e.values.size(), (int)src1->ne[0]);
-                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);
-            }
-            for (int row = 0; row < (int)src1->ne[1]; ++row) {
-                const int excur = m_ids[row*n_as + idx];
-                GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
-                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];
+            size_t e_start = ex*src1->ne[0];
+
+            for (int idx = 0; idx < n_ids; ++idx) {
+                for (int row = 0; row < (int)src1->ne[2]; ++row) {
+                    const int excur = *(const int32_t *) (m_ids.data() + row*ids->nb[1] + idx*ids->nb[0]);
+
+                    GGML_ASSERT(excur >= 0 && excur < n_as); // sanity check
+
+                    if (excur != ex) continue;
+
+                    const int64_t i11 = idx % src1->ne[1];
+                    const int64_t i12 = row;
+                    const float * x = (const float *)((const char *)data + i11*src1->nb[1] + i12*src1->nb[2]);
+
+                    for (int j = 0; j < (int)src1->ne[0]; ++j) {
+                        e.values[e_start + j] += x[j]*x[j];
+                    }
                 }
             }
             if (e.ncall > m_last_call) {
@@ -180,7 +200,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
 }
 
 void IMatrixCollector::save_imatrix() const {
-    save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str());
+    save_imatrix(m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str(), m_params.dataset.c_str());
 }
 
 void IMatrixCollector::keep_imatrix(int ncall) const {
@@ -188,24 +208,33 @@ void IMatrixCollector::keep_imatrix(int ncall) const {
     if (file_name.empty()) file_name = "imatrix.dat";
     file_name += ".at_";
     file_name += std::to_string(ncall);
-    save_imatrix(file_name.c_str());
+    save_imatrix(file_name.c_str(), m_params.dataset.c_str());
 }
 
-void IMatrixCollector::save_imatrix(const char * fname) const {
+void IMatrixCollector::save_imatrix(const char * fname, const char * dataset) const {
     std::ofstream out(fname, std::ios::binary);
     int n_entries = m_stats.size();
-    out.write((const char*)&n_entries, sizeof(n_entries));
-    for (auto& p : m_stats) {
+    out.write((const char *) &n_entries, sizeof(n_entries));
+    for (const auto & p : m_stats) {
         int len = p.first.size();
-        out.write((const char*)&len, sizeof(len));
+        out.write((const char *) &len, sizeof(len));
         out.write(p.first.c_str(), len);
-        out.write((const char*)&p.second.ncall, sizeof(p.second.ncall));
+        out.write((const char *) &p.second.ncall, sizeof(p.second.ncall));
         int nval = p.second.values.size();
-        out.write((const char*)&nval, sizeof(nval));
-        if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float));
+        out.write((const char *) &nval, sizeof(nval));
+        if (nval > 0) out.write((const char *) p.second.values.data(), nval * sizeof(float));
     }
+
+    // Write the number of call the matrix was computed with
+    out.write((const char *) &m_last_call, sizeof(m_last_call));
+
+    // Write the dataset name at the end of the file to later on specify it in quantize
+    int n_dataset = strlen(dataset);
+    out.write((const char *) &n_dataset, sizeof(n_dataset));
+    out.write(dataset, n_dataset);
+
     if (m_params.verbosity > 0) {
-        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname);
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n", __func__, m_last_call, fname);
     }
 }
 
@@ -343,12 +372,13 @@ 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 +451,7 @@ 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;
@@ -526,6 +557,29 @@ int main(int argc, char ** argv) {
         }
     }
 
+    gpt_params params;
+    params.n_batch = 512;
+    if (!gpt_params_parse(args.size(), args.data(), params)) {
+        return 1;
+    }
+
+    params.logits_all = true;
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    print_build_info();
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    sparams.dataset = params.prompt_file;
     g_collector.set_parameters(std::move(sparams));
 
     if (!combine_files.empty()) {
@@ -564,49 +618,21 @@ int main(int argc, char ** argv) {
         }
     }
 
-    gpt_params params;
-    params.n_batch = 512;
-    if (!gpt_params_parse(args.size(), args.data(), params)) {
-        return 1;
-    }
-
-    params.logits_all = true;
-    params.n_batch = std::min(params.n_batch, params.n_ctx);
-
-    print_build_info();
-
-    if (params.seed == LLAMA_DEFAULT_SEED) {
-        params.seed = time(NULL);
-    }
-
-    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
-
-    std::mt19937 rng(params.seed);
-    if (params.random_prompt) {
-        params.prompt = gpt_random_prompt(rng);
-    }
-
     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();
@@ -585,7 +586,7 @@ int main(int argc, char ** argv) {
 
             // deal with eot token in infill mode
             if ((llama_sampling_last(ctx_sampling) == llama_token_eot(model) || is_interacting) && params.interactive){
-                if(is_interacting && !params.interactive_first) {
+                if (is_interacting && !params.interactive_first) {
                     // print an eot token
                     printf("%s", llama_token_to_piece(ctx, llama_token_eot(model)).c_str());
                 }
@@ -650,8 +651,8 @@ int main(int argc, char ** argv) {
                 // LOG_TEE("took new input\n");
                 is_interacting = false;
             }
-            // deal with end of text token in interactive mode
-            else if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
+            // deal with end of generation tokens in interactive mode
+            else if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -730,8 +731,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(model) && !params.interactive) {
+        // end of generation
+        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !params.interactive) {
             break;
         }
 

examples/json_schema_to_grammar.py

@@ -6,37 +6,94 @@ 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': '("true" | "false") space',
-    'number': '("-"? ([0-9] | [1-9] [0-9]*)) ("." [0-9]+)? ([eE] [-+]? [0-9]+)? space',
-    'integer': '("-"? ([0-9] | [1-9] [0-9]*)) space',
-    'value'  : 'object | array | string | number | boolean',
-    'object' : '"{" space ( string ":" space value ("," space string ":" space value)* )? "}" space',
-    'array'  : '"[" space ( value ("," space value)* )? "]" space',
-    'uuid'   : '"\\"" ' + ' "-" '.join('[0-9a-fA-F]' * n for n in [8, 4, 4, 4, 12]) + ' "\\"" 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',
+    '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', []),
 }
-OBJECT_RULE_NAMES = ['object', 'array', 'string', 'number', 'boolean', 'null', 'value']
 
 # TODO: support "uri", "email" string formats
-DATE_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-string': '"\\"" date "\\"" space',
-    'time-string': '"\\"" time "\\"" space',
-    'date-time-string': '"\\"" date-time "\\"" space',
+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']),
 }
 
-RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()])
+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"]')
@@ -46,8 +103,6 @@ GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']'
 NON_LITERAL_SET = set('|.()[]{}*+?')
 ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = set('[]()|{}*+?')
 
-DATE_PATTERN = '[0-9]{4}-(0[1-9]|1[0-2])-([0-2][0-9]|3[0-1])'
-TIME_PATTERN = '([01][0-9]|2[0-3])(:[0-5][0-9]){2}(\\.[0-9]{1,3})?(Z|[+-](([01][0-9]|2[0-3]):[0-5][0-9]))' # Cap millisecond precision w/ 3 digits
 
 class SchemaConverter:
     def __init__(self, *, prop_order, allow_fetch, dotall, raw_pattern):
@@ -55,7 +110,9 @@ class SchemaConverter:
         self._allow_fetch = allow_fetch
         self._dotall = dotall
         self._raw_pattern = raw_pattern
-        self._rules = {'space': SPACE_RULE}
+        self._rules = {
+            'space': SPACE_RULE,
+        }
         self._refs = {}
         self._refs_being_resolved = set()
 
@@ -65,6 +122,29 @@ class SchemaConverter:
         )
         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:
@@ -169,10 +249,10 @@ class SchemaConverter:
 
             def get_dot():
                 if self._dotall:
-                    rule = '[\\U00000000-\\U0010FFFF]'
+                    rule = DOTALL
                 else:
                     # Accept any character... except \n and \r line break chars (\x0A and \xOD)
-                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
+                    rule = DOT
                 return self._add_rule(f'dot', rule)
 
             def join_seq():
@@ -246,26 +326,14 @@ class SchemaConverter:
 
                     (sub, sub_is_literal) = seq[-1]
 
-                    if min_times == 0 and max_times is None:
-                        seq[-1] = (f'{sub}*', False)
-                    elif min_times == 0 and max_times == 1:
-                        seq[-1] = (f'{sub}?', False)
-                    elif min_times == 1 and max_times is None:
-                        seq[-1] = (f'{sub}+', False)
-                    else:
-                        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
+                    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] = (
-                            ' '.join(
-                                ([f'"{sub[1:-1] * min_times}"'] if sub_is_literal else [sub] * min_times) +
-                                ([f'{sub}?'] * (max_times - min_times) if max_times is not None else [f'{sub}*'])),
-                            False
-                        )
+                    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:
@@ -373,49 +441,47 @@ class SchemaConverter:
                     ' "]" space')
             else:
                 item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
-                list_item_operator = f'( "," space {item_rule_name} )'
-                successive_items = ""
                 min_items = schema.get("minItems", 0)
                 max_items = schema.get("maxItems")
-                if min_items > 0:
-                    successive_items = list_item_operator * (min_items - 1)
-                    min_items -= 1
-                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 + "*"
-                if min_items == 0:
-                    rule = f'"[" space ( {item_rule_name} {successive_items} )? "]" space'
-                else:
-                    rule = f'"[" space {item_rule_name} {successive_items} "]" space'
-                return self._add_rule(rule_name, rule)
+                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_rule(
+            return self._add_primitive(
                 'root' if rule_name == 'root' else schema_format,
                 PRIMITIVE_RULES['uuid']
             )
 
-        elif schema_type in (None, 'string') and schema_format in DATE_RULES:
-            for t, r in DATE_RULES.items():
-                self._add_rule(t, r)
-            return schema_format + '-string'
+        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):
-            for n in OBJECT_RULE_NAMES:
-                self._add_rule(n, PRIMITIVE_RULES[n])
-            return self._add_rule(rule_name, 'object')
+            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_rule(
-                'root' if rule_name == 'root' else schema_type,
-                PRIMITIVE_RULES[schema_type]
-            )
+            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
@@ -437,7 +503,7 @@ class SchemaConverter:
             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_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
+                self._add_primitive('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
             )
             optional_props.append("*")
 

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];
@@ -190,7 +190,7 @@ static const cmd_params cmd_params_defaults = {
     /* n_ubatch      */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
-    /* n_threads     */ {get_num_physical_cores()},
+    /* n_threads     */ {get_math_cpu_count()},
     /* n_gpu_layers  */ {99},
     /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
     /* main_gpu      */ {0},
@@ -808,7 +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/llama.android/app/src/main/cpp/llama-android.cpp

@@ -408,7 +408,7 @@ Java_com_example_llama_Llm_completion_1loop(
     const auto new_token_id = llama_sample_token_greedy(context, &candidates_p);
 
     const auto n_cur = env->CallIntMethod(intvar_ncur, la_int_var_value);
-    if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+    if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
         return env->NewStringUTF("");
     }
 

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

@@ -158,7 +158,7 @@ actor LlamaContext {
             new_token_id = llama_sample_token_greedy(context, &candidates_p)
         }
 
-        if new_token_id == llama_token_eos(model) || n_cur == n_len {
+        if llama_token_is_eog(model, new_token_id) || n_cur == n_len {
             print("\n")
             let new_token_str = String(cString: temporary_invalid_cchars + [0])
             temporary_invalid_cchars.removeAll()
@@ -322,7 +322,7 @@ actor LlamaContext {
         defer {
             result.deallocate()
         }
-        let nTokens = llama_token_to_piece(model, token, result, 8)
+        let nTokens = llama_token_to_piece(model, token, result, 8, false)
 
         if nTokens < 0 {
             let newResult = UnsafeMutablePointer<Int8>.allocate(capacity: Int(-nTokens))
@@ -330,7 +330,7 @@ actor LlamaContext {
             defer {
                 newResult.deallocate()
             }
-            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens)
+            let nNewTokens = llama_token_to_piece(model, token, newResult, -nTokens, false)
             let bufferPointer = UnsafeBufferPointer(start: newResult, count: Int(nNewTokens))
             return Array(bufferPointer)
         } else {

examples/llava/MobileVLM-README.md

@@ -6,7 +6,7 @@ for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com
 
 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 MobiVLM as an example, the different conversion step will be shown.
+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.
@@ -22,7 +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
@@ -36,7 +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` (for **V2** the arg is `--projector-type ldpv2`) 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 \
@@ -78,7 +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
@@ -109,7 +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)
@@ -121,12 +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**
@@ -175,8 +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
 
@@ -191,5 +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

@@ -3,11 +3,12 @@
 // I'll gradually clean and extend it
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
+#include "log.h"
 #include "ggml.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#ifdef GGML_USE_CUBLAS
+#ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
@@ -23,7 +24,6 @@
 #include <cstdlib>
 #include <cstring>
 #include <fstream>
-#include <iostream>
 #include <map>
 #include <regex>
 #include <stdexcept>
@@ -104,6 +104,7 @@ static std::string format(const char * fmt, ...) {
 #define TN_POS_EMBD        "%s.position_embd.weight"
 #define TN_CLASS_EMBD      "v.class_embd"
 #define TN_PATCH_EMBD      "v.patch_embd.weight"
+#define TN_PATCH_BIAS      "v.patch_embd.bias"
 #define TN_ATTN_K          "%s.blk.%d.attn_k.%s"
 #define TN_ATTN_Q          "%s.blk.%d.attn_q.%s"
 #define TN_ATTN_V          "%s.blk.%d.attn_v.%s"
@@ -145,7 +146,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
 static int get_key_idx(const gguf_context * ctx, const char * key) {
     int i = gguf_find_key(ctx, key);
     if (i == -1) {
-        fprintf(stderr, "key %s not found in file\n", key);
+        LOG_TEE("key %s not found in file\n", key);
         throw std::runtime_error(format("Missing required key: %s", key));
     }
 
@@ -247,7 +248,7 @@ static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
 
 static void print_tensor_info(const ggml_tensor * tensor, const char * prefix = "") {
     size_t tensor_size = ggml_nbytes(tensor);
-    printf("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
+    LOG_TEE("%s: n_dims = %d, name = %s, tensor_size=%zu, shape:[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "], type = %s\n",
             prefix, ggml_n_dims(tensor), tensor->name, tensor_size,
             tensor->ne[0], tensor->ne[1], tensor->ne[2], tensor->ne[3], ggml_type_name(tensor->type));
 }
@@ -265,7 +266,7 @@ static projector_type clip_projector_type_from_string(const std::string & name)
 static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        std::cerr << "Failed to open file for writing: " << filename << std::endl;
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
         return;
     }
 
@@ -284,7 +285,7 @@ static void clip_image_write_image_to_ppm(const clip_image_u8& img, const std::s
 static void clip_image_save_to_bmp(const clip_image_u8& img, const std::string& filename) {
     std::ofstream file(filename, std::ios::binary);
     if (!file.is_open()) {
-        std::cerr << "Failed to open file for writing: " << filename << std::endl;
+        LOG_TEE("Failed to open file for writing: %s\n", filename.c_str());
         return;
     }
 
@@ -425,6 +426,7 @@ struct clip_vision_model {
     // embeddings
     struct ggml_tensor * class_embedding;
     struct ggml_tensor * patch_embeddings;
+    struct ggml_tensor * patch_bias;
     struct ggml_tensor * position_embeddings;
 
     struct ggml_tensor * pre_ln_w;
@@ -501,6 +503,11 @@ struct clip_ctx {
     bool use_gelu = false;
     int32_t ftype = 1;
 
+    bool has_class_embedding = true;
+    bool has_pre_norm = true;
+    bool has_post_norm = false;
+    bool has_patch_bias = false;
+
     struct gguf_context * ctx_gguf;
     struct ggml_context * ctx_data;
 
@@ -515,7 +522,7 @@ struct clip_ctx {
 
 static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
     if (!ctx->has_vision_encoder) {
-        printf("This gguf file seems to have no vision encoder\n");
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
         return nullptr;
     }
 
@@ -526,7 +533,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     const int patch_size           = hparams.patch_size;
     const int num_patches          = ((image_size / patch_size) * (image_size / patch_size));
     const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
-    const int num_positions        = num_patches + 1;
+    const int num_positions        = num_patches + (ctx->has_class_embedding ? 1 : 0);
     const int hidden_size          = hparams.hidden_size;
     const int n_head               = hparams.n_head;
     const int d_head               = hidden_size / n_head;
@@ -557,16 +564,23 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
     inp = ggml_reshape_3d(ctx0, inp, num_patches, hidden_size, batch_size);
     inp = ggml_cont(ctx0, ggml_permute(ctx0, inp, 1, 0, 2, 3));
 
+    if (ctx->has_patch_bias) {
+        // inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
+        inp = ggml_add(ctx0, inp, model.patch_bias);
+    }
+
     // concat class_embeddings and patch_embeddings
-    struct ggml_tensor * embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+    struct ggml_tensor * embeddings = inp;
+    if (ctx->has_class_embedding) {
+        embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
+        embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
+        embeddings = ggml_acc(ctx0, embeddings, inp,
+                embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
+    }
     ggml_set_name(embeddings, "embeddings");
     ggml_set_input(embeddings);
 
-    embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
-            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
-
-    embeddings = ggml_acc(ctx0, embeddings, inp,
-            embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
 
     struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
     ggml_set_name(positions, "positions");
@@ -576,7 +590,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
 
     // pre-layernorm
-    {
+    if (ctx->has_pre_norm) {
         embeddings = ggml_norm(ctx0, embeddings, eps);
         ggml_set_name(embeddings, "pre_ln");
 
@@ -664,6 +678,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
         embeddings = cur;
     }
 
+    // post-layernorm
+    if (ctx->has_post_norm) {
+        embeddings = ggml_norm(ctx0, embeddings, eps);
+        ggml_set_name(embeddings, "post_ln");
+
+        embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.post_ln_w), model.post_ln_b);
+    }
+
     // llava projector
     {
         embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
@@ -835,9 +857,10 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
             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_add(ctx0, peg_0, mlp_2);
             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;
         }
@@ -878,21 +901,21 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         const int idx_name = gguf_find_key(ctx, KEY_NAME);
         if (idx_name != -1) { // make name optional temporarily as some of the uploaded models missing it due to a bug
             const std::string name = gguf_get_val_str(ctx, idx_name);
-            printf("%s: model name:   %s\n", __func__, name.c_str());
+            LOG_TEE("%s: model name:   %s\n", __func__, name.c_str());
         }
-        printf("%s: description:  %s\n", __func__, description.c_str());
-        printf("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
-        printf("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
-        printf("%s: n_tensors:    %d\n", __func__, n_tensors);
-        printf("%s: n_kv:         %d\n", __func__, n_kv);
-        printf("%s: ftype:        %s\n", __func__, ftype_str.c_str());
-        printf("\n");
+        LOG_TEE("%s: description:  %s\n", __func__, description.c_str());
+        LOG_TEE("%s: GGUF version: %d\n", __func__, gguf_get_version(ctx));
+        LOG_TEE("%s: alignment:    %zu\n", __func__, gguf_get_alignment(ctx));
+        LOG_TEE("%s: n_tensors:    %d\n", __func__, n_tensors);
+        LOG_TEE("%s: n_kv:         %d\n", __func__, n_kv);
+        LOG_TEE("%s: ftype:        %s\n", __func__, ftype_str.c_str());
+        LOG_TEE("\n");
     }
     const int n_tensors = gguf_get_n_tensors(ctx);
 
     // kv
     const int n_kv = gguf_get_n_kv(ctx);
-    printf("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
+    LOG_TEE("%s: loaded meta data with %d key-value pairs and %d tensors from %s\n",
         __func__, n_kv, n_tensors, fname);
     {
         std::map<enum ggml_type, uint32_t> n_type;
@@ -903,7 +926,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             n_type[type]++;
         }
 
-        printf("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
+        LOG_TEE("%s: Dumping metadata keys/values. Note: KV overrides do not apply in this output.\n", __func__);
         for (int i = 0; i < n_kv; i++) {
             const char * name           = gguf_get_key(ctx, i);
             const enum gguf_type type   = gguf_get_kv_type(ctx, i);
@@ -919,7 +942,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             }
             replace_all(value, "\n", "\\n");
 
-            printf("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
+            LOG_TEE("%s: - kv %3d: %42s %-16s = %s\n", __func__, i, name, type_name.c_str(), value.c_str());
         }
 
         // print type counts
@@ -928,7 +951,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
                 continue;
             }
 
-            printf("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
+            LOG_TEE("%s: - type %4s: %4d tensors\n", __func__, ggml_type_name(kv.first), kv.second);
         }
     }
 
@@ -943,7 +966,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             size_t tensor_size = ggml_nbytes(cur);
             model_size += tensor_size;
             if (verbosity >= 3) {
-                printf("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
+                LOG_TEE("%s: tensor[%d]: n_dims = %d, name = %s, tensor_size=%zu, offset=%zu, shape:[%" PRIu64 ", %" PRIu64 ", %" PRIu64 ", %" PRIu64 "], type = %s\n",
                        __func__, i, ggml_n_dims(cur), cur->name, tensor_size, offset, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3], ggml_type_name(type));
             }
         }
@@ -968,20 +991,20 @@ 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__);
+    LOG_TEE("%s: CLIP using CUDA backend\n", __func__);
 #endif
 
 #ifdef GGML_USE_METAL
     new_clip->backend = ggml_backend_metal_init();
-    printf("%s: CLIP using Metal backend\n", __func__);
+    LOG_TEE("%s: CLIP using Metal backend\n", __func__);
 #endif
 
 
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
-        printf("%s: CLIP using CPU backend\n", __func__);
+        LOG_TEE("%s: CLIP using CPU backend\n", __func__);
     }
 
     // model size and capabilities
@@ -1005,15 +1028,15 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         new_clip->use_gelu = gguf_get_val_bool(ctx, idx);
 
         if (verbosity >= 1) {
-            printf("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
-            printf("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
-            printf("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
-            printf("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
-            printf("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
+            LOG_TEE("%s: text_encoder:   %d\n", __func__, new_clip->has_text_encoder);
+            LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
+            LOG_TEE("%s: llava_projector:  %d\n", __func__, new_clip->has_llava_projector);
+            LOG_TEE("%s: model size:     %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
+            LOG_TEE("%s: metadata size:  %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
         }
     }
 
-    printf("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
+    LOG_TEE("%s: params backend buffer size = % 6.2f MB (%i tensors)\n", __func__, model_size / (1024.0 * 1024.0), n_tensors);
 
     // load tensors
     {
@@ -1026,7 +1049,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         new_clip->ctx_data = ggml_init(params);
         if (!new_clip->ctx_data) {
-            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            LOG_TEE("%s: ggml_init() failed\n", __func__);
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1034,7 +1057,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
 
         auto fin = std::ifstream(fname, std::ios::binary);
         if (!fin) {
-            printf("cannot open model file for loading tensors\n");
+            LOG_TEE("cannot open model file for loading tensors\n");
             clip_free(new_clip);
             gguf_free(ctx);
             return nullptr;
@@ -1056,7 +1079,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             const size_t offset = gguf_get_data_offset(ctx) + gguf_get_tensor_offset(ctx, i);
             fin.seekg(offset, std::ios::beg);
             if (!fin) {
-                printf("%s: failed to seek for tensor %s\n", __func__, name);
+                LOG_TEE("%s: failed to seek for tensor %s\n", __func__, name);
                 clip_free(new_clip);
                 gguf_free(ctx);
                 return nullptr;
@@ -1127,34 +1150,61 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         }
 
         if (verbosity >= 2) {
-            printf("\n%s: vision model hparams\n", __func__);
-            printf("image_size         %d\n", hparams.image_size);
-            printf("patch_size         %d\n", hparams.patch_size);
-            printf("v_hidden_size      %d\n", hparams.hidden_size);
-            printf("v_n_intermediate   %d\n", hparams.n_intermediate);
-            printf("v_projection_dim   %d\n", hparams.projection_dim);
-            printf("v_n_head           %d\n", hparams.n_head);
-            printf("v_n_layer          %d\n", hparams.n_layer);
-            printf("v_eps              %f\n", hparams.eps);
-            printf("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
-            printf("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
-            printf("v_image_grid_pinpoints: ");
+            LOG_TEE("\n%s: vision model hparams\n", __func__);
+            LOG_TEE("image_size         %d\n", hparams.image_size);
+            LOG_TEE("patch_size         %d\n", hparams.patch_size);
+            LOG_TEE("v_hidden_size      %d\n", hparams.hidden_size);
+            LOG_TEE("v_n_intermediate   %d\n", hparams.n_intermediate);
+            LOG_TEE("v_projection_dim   %d\n", hparams.projection_dim);
+            LOG_TEE("v_n_head           %d\n", hparams.n_head);
+            LOG_TEE("v_n_layer          %d\n", hparams.n_layer);
+            LOG_TEE("v_eps              %f\n", hparams.eps);
+            LOG_TEE("v_image_mean       %f %f %f\n", new_clip->image_mean[0], new_clip->image_mean[1], new_clip->image_mean[2]);
+            LOG_TEE("v_image_std        %f %f %f\n", new_clip->image_std[0], new_clip->image_std[1], new_clip->image_std[2]);
+            LOG_TEE("v_image_grid_pinpoints: ");
             for (int i = 0; i < 32 && (hparams.image_grid_pinpoints[i] != 0); ++i) {
-                printf("%d ", hparams.image_grid_pinpoints[i]);
+                LOG_TEE("%d ", hparams.image_grid_pinpoints[i]);
             }
-            printf("\n");
-            printf("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
+            LOG_TEE("\n");
+            LOG_TEE("v_mm_patch_merge_type: %s\n", hparams.mm_patch_merge_type);
 
         }
 
+        try {
+            vision_model.class_embedding  = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
+            new_clip->has_class_embedding = true;
+        } catch (const std::exception& e) {
+            new_clip->has_class_embedding = false;
+        }
+
+        try {
+            vision_model.pre_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
+            vision_model.pre_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
+            new_clip->has_pre_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_pre_norm = false;
+        }
+
+        try {
+            vision_model.post_ln_w  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "weight"));
+            vision_model.post_ln_b  = get_tensor(new_clip->ctx_data, format(TN_LN_POST, "v", "bias"));
+            new_clip->has_post_norm = true;
+        } catch (std::exception & e) {
+            new_clip->has_post_norm = false;
+        }
+
+        try {
+            vision_model.patch_bias = get_tensor(new_clip->ctx_data, TN_PATCH_BIAS);
+            new_clip->has_patch_bias = true;
+        } catch (std::exception & e) {
+            new_clip->has_patch_bias = false;
+        }
+
         try {
             vision_model.patch_embeddings    = get_tensor(new_clip->ctx_data, TN_PATCH_EMBD);
-            vision_model.class_embedding     = get_tensor(new_clip->ctx_data, TN_CLASS_EMBD);
             vision_model.position_embeddings = get_tensor(new_clip->ctx_data, format(TN_POS_EMBD, "v"));
-            vision_model.pre_ln_w            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "weight"));
-            vision_model.pre_ln_b            = get_tensor(new_clip->ctx_data, format(TN_LN_PRE, "v", "bias"));
         } catch(const std::exception& e) {
-            fprintf(stderr, "%s: failed to load vision model tensors\n", __func__);
+            LOG_TEE("%s: failed to load vision model tensors\n", __func__);
         }
 
         // LLaVA projection
@@ -1183,7 +1233,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
             } catch (std::runtime_error & e) {  }
             try {
                 vision_model.image_newline = get_tensor(new_clip->ctx_data, TN_IMAGE_NEWLINE);
-                // fprintf(stderr, "%s: image_newline tensor (llava-1.6) found\n", __func__);
+                // LOG_TEE("%s: image_newline tensor (llava-1.6) found\n", __func__);
             } catch (std::runtime_error & e) {  }
         } else if (new_clip->proj_type == PROJECTOR_TYPE_LDP) {
             // MobileVLM projection
@@ -1263,7 +1313,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
         ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
         ggml_gallocr_reserve(new_clip->compute_alloc, gf);
         size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
-        printf("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
+        LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
     }
 
     return new_clip;
@@ -1303,7 +1353,7 @@ bool clip_image_load_from_file(const char * fname, clip_image_u8 * img) {
     int nx, ny, nc;
     auto * data = stbi_load(fname, &nx, &ny, &nc, 3);
     if (!data) {
-        fprintf(stderr, "%s: failed to load image '%s'\n", __func__, fname);
+        LOG_TEE("%s: failed to load image '%s'\n", __func__, fname);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1315,7 +1365,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
     int nx, ny, nc;
     auto * data = stbi_load_from_memory(bytes, bytes_length, &nx, &ny, &nc, 3);
     if (!data) {
-        fprintf(stderr, "%s: failed to decode image bytes\n", __func__);
+        LOG_TEE("%s: failed to decode image bytes\n", __func__);
         return false;
     }
     build_clip_img_from_data(data, nx, ny, img);
@@ -1324,7 +1374,7 @@ bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length
 }
 
 // Linear interpolation between two points
-inline float lerp(float s, float e, float t) {
+inline float clip_lerp(float s, float e, float t) {
     return s + (e - s) * t;
 }
 // Bilinear resize function
@@ -1346,17 +1396,17 @@ static void bilinear_resize(const clip_image_u8& src, clip_image_u8& dst, int ta
             float y_lerp = py - y_floor;
 
             for (int c = 0; c < 3; c++) {
-                float top = lerp(
+                float top = clip_lerp(
                     static_cast<float>(src.buf[3 * (y_floor * src.nx + x_floor) + c]),
                     static_cast<float>(src.buf[3 * (y_floor * src.nx + (x_floor + 1)) + c]),
                     x_lerp
                 );
-                float bottom = lerp(
+                float bottom = clip_lerp(
                     static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + x_floor) + c]),
                     static_cast<float>(src.buf[3 * ((y_floor + 1) * src.nx + (x_floor + 1)) + c]),
                     x_lerp
                 );
-                dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(lerp(top, bottom, y_lerp));
+                dst.buf[3 * (y * target_width + x) + c] = static_cast<uint8_t>(clip_lerp(top, bottom, y_lerp));
             }
         }
     }
@@ -1505,7 +1555,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int> & or
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -1544,7 +1594,7 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
 bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
     bool pad_to_square = true;
     if (!ctx->has_vision_encoder) {
-        printf("This gguf file seems to have no vision encoder\n");
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
         return false;
     }
     auto & params = ctx->vision_model.hparams;
@@ -1621,7 +1671,7 @@ bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, cli
             }
 
             for (size_t i = 0; i < patches.size(); i++) {
-                // printf("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
+                // LOG_TEE("patch %d: %d %d\n", i, patches[i]->nx, patches[i]->ny);
                 clip_image_u8_free(patches[i]);
             }
 
@@ -1755,7 +1805,7 @@ 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;
     }
 
@@ -1764,7 +1814,7 @@ int clip_n_patches(const struct clip_ctx * ctx) {
 
 bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
     if (!ctx->has_vision_encoder) {
-        printf("This gguf file seems to have no vision encoder\n");
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -1776,7 +1826,7 @@ bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f3
 
 bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_image_f32_batch * imgs, float * vec) {
     if (!ctx->has_vision_encoder) {
-        printf("This gguf file seems to have no vision encoder\n");
+        LOG_TEE("This gguf file seems to have no vision encoder\n");
         return false;
     }
 
@@ -1938,7 +1988,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             new_type = type;
             if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
                 new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
-                // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+                // LOG_TEE("%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
             }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
@@ -1957,7 +2007,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 f32_data = (float *)conv_buf.data();
                 break;
             default:
-                printf("Please use an input file in f32 or f16\n");
+                LOG_TEE("Please use an input file in f32 or f16\n");
                 gguf_free(ctx_out);
                 return false;
             }
@@ -1984,7 +2034,7 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
             fout.put(0);
         }
 
-        printf("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
+        LOG_TEE("%s: n_dims = %d | quantize=%d | size = %f MB -> %f MB\n", name.c_str(), ggml_n_dims(cur), quantize,
                orig_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
     }
 
@@ -2000,8 +2050,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
     gguf_free(ctx_out);
 
     {
-        printf("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
-        printf("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
+        LOG_TEE("%s: original  size = %8.2f MB\n", __func__, total_size_org / 1024.0 / 1024.0);
+        LOG_TEE("%s: quantized size = %8.2f MB\n", __func__, total_size_new / 1024.0 / 1024.0);
     }
 
     return true;

examples/llava/llava-cli.cpp

@@ -1,4 +1,5 @@
 #include "ggml.h"
+#include "log.h"
 #include "common.h"
 #include "clip.h"
 #include "llava.h"
@@ -18,7 +19,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
             n_eval = n_batch;
         }
         if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
-            fprintf(stderr, "%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
+            LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }
         *n_past += n_eval;
@@ -45,7 +46,7 @@ static const char * sample(struct llama_sampling_context * ctx_sampling,
     const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
     llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
     static std::string ret;
-    if (id == llama_token_eos(llama_get_model(ctx_llama))) {
+    if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
         ret = "</s>";
     } else {
         ret = llama_token_to_piece(ctx_llama, id);
@@ -73,7 +74,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
     size_t img_base64_str_start, img_base64_str_end;
     find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
     if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
-        fprintf(stderr, "%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
+        LOG_TEE("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
         return NULL;
     }
 
@@ -87,7 +88,7 @@ static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip
 
     auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
     if (!embed) {
-        fprintf(stderr, "%s: could not load image from base64 string.\n", __func__);
+        LOG_TEE("%s: could not load image from base64 string.\n", __func__);
         return NULL;
     }
 
@@ -112,8 +113,8 @@ struct llava_context {
 };
 
 static void show_additional_info(int /*argc*/, char ** argv) {
-    fprintf(stderr, "\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
-    fprintf(stderr, "  note: a lower temperature value like 0.1 is recommended for better quality.\n");
+    LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
+    LOG_TEE("  note: a lower temperature value like 0.1 is recommended for better quality.\n");
 }
 
 static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_params * params) {
@@ -123,18 +124,18 @@ static struct llava_image_embed * load_image(llava_context * ctx_llava, gpt_para
     auto prompt = params->prompt;
     if (prompt_contains_image(prompt)) {
         if (!params->image.empty()) {
-            fprintf(stderr, "using base64 encoded image instead of command line image path\n");
+            LOG_TEE("using base64 encoded image instead of command line image path\n");
         }
         embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->n_threads, prompt);
         if (!embed) {
-            fprintf(stderr, "%s: can't load image from prompt\n", __func__);
+            LOG_TEE("%s: can't load image from prompt\n", __func__);
             return NULL;
         }
         params->prompt = remove_image_from_prompt(prompt);
     } else {
         embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->n_threads, params->image.c_str());
         if (!embed) {
-            fprintf(stderr, "%s: is %s really an image file?\n", __func__, params->image.c_str());
+            LOG_TEE("%s: is %s really an image file?\n", __func__, params->image.c_str());
             return NULL;
         }
     }
@@ -146,7 +147,6 @@ 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>");
@@ -154,18 +154,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         // new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
         system_prompt = prompt.substr(0, image_pos);
         user_prompt = prompt.substr(image_pos + std::string("<image>").length());
-        printf("system_prompt: %s\n", system_prompt.c_str());
+        LOG_TEE("system_prompt: %s\n", system_prompt.c_str());
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
-        printf("user_prompt: %s\n", user_prompt.c_str());
+        LOG_TEE("user_prompt: %s\n", user_prompt.c_str());
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     } else {
@@ -175,18 +175,18 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
         if (params->verbose_prompt) {
             auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
             for (int i = 0; i < (int) tmp.size(); i++) {
-                printf("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
+                LOG_TEE("%6d -> '%s'\n", tmp[i], llama_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
             }
         }
     }
 
-    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);
 
     // generate the response
 
-    fprintf(stderr, "\n");
+    LOG_TEE("\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
     std::string response = "";
@@ -225,7 +225,7 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
     if (model == NULL) {
-        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        LOG_TEE("%s: error: unable to load model\n" , __func__);
         return NULL;
     }
 
@@ -235,7 +235,7 @@ static struct llava_context * llava_init(gpt_params * params) {
     llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
 
     if (ctx_llama == NULL) {
-        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
         return NULL;
     }
 
@@ -258,6 +258,12 @@ static void llava_free(struct llava_context * ctx_llava) {
     llama_backend_free();
 }
 
+static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
+    (void) level;
+    (void) user_data;
+    LOG_TEE("%s", text);
+}
+
 int main(int argc, char ** argv) {
     ggml_time_init();
 
@@ -267,6 +273,14 @@ int main(int argc, char ** argv) {
         show_additional_info(argc, argv);
         return 1;
     }
+
+#ifndef LOG_DISABLE_LOGS
+    log_set_target(log_filename_generator("llava", "log"));
+    LOG_TEE("Log start\n");
+    log_dump_cmdline(argc, argv);
+    llama_log_set(llama_log_callback_logTee, nullptr);
+#endif // LOG_DISABLE_LOGS
+
     if (params.mmproj.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
         gpt_print_usage(argc, argv, params);
         show_additional_info(argc, argv);
@@ -275,7 +289,7 @@ int main(int argc, char ** argv) {
 
     auto ctx_llava = llava_init(&params);
     if (ctx_llava == NULL) {
-        fprintf(stderr, "%s: error: failed to init llava\n", __func__);
+        LOG_TEE("%s: error: failed to init llava\n", __func__);
         return 1;
     }
 

examples/llava/llava.cpp

@@ -54,7 +54,7 @@ static std::pair<int, int> select_best_resolution(const std::pair<int, int>& ori
         int downscaled_height = static_cast<int>(original_height * scale);
         int effective_resolution = std::min(downscaled_width * downscaled_height, original_width * original_height);
         int wasted_resolution = (width * height) - effective_resolution;
-        // fprintf(stderr, "resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
+        // LOG_TEE("resolution: %d %d, scale: %f, downscaled: %d %d, effective: %d, wasted: %d\n", width, height, scale, downscaled_width, downscaled_height, effective_resolution, wasted_resolution);
         if (effective_resolution > max_effective_resolution || (effective_resolution == max_effective_resolution && wasted_resolution < min_wasted_resolution)) {
             max_effective_resolution = effective_resolution;
             min_wasted_resolution = wasted_resolution;
@@ -154,13 +154,13 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
     model.newline = ggml_new_tensor_1d(model.ctx, GGML_TYPE_F32, newline_tmp->ne[0]);
     if (newline_tmp->backend != GGML_BACKEND_TYPE_CPU) {
         if (newline_tmp->buffer == NULL) {
-            printf("newline_tmp tensor buffer is NULL\n");
+            LOG_TEE("newline_tmp tensor buffer is NULL\n");
         }
         ggml_backend_tensor_get(newline_tmp, model.newline->data, 0, ggml_nbytes(newline_tmp));
     } else {
         model.newline->data = newline_tmp->data;
         if (model.newline->data == NULL) {
-            printf("newline_tmp tensor data is NULL\n");
+            LOG_TEE("newline_tmp tensor data is NULL\n");
         }
     }
 
@@ -224,7 +224,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
     img_res_v.size = 0;
     img_res_v.data = nullptr;
     if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
-        fprintf(stderr, "%s: unable to preprocess image\n", __func__);
+        LOG_TEE("%s: unable to preprocess image\n", __func__);
         delete[] img_res_v.data;
         return false;
     }
@@ -239,7 +239,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
         delete[] img_res_v.data;
         if (!encoded) {
-            fprintf(stderr, "Unable to encode image\n");
+            LOG_TEE("Unable to encode image\n");
 
             return false;
         }
@@ -252,12 +252,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
             image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip)); // 576 patches * 4096 embeddings * 4 bytes = 9437184
             const bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]); // image data is in 3x336x336 format and will be converted to 336x336x3 inside
             if (!encoded) {
-                fprintf(stderr, "Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
+                LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
                 return false;
             }
         }
         const int64_t t_img_enc_batch_us = ggml_time_us();
-        printf("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
+        LOG_TEE("%s: %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
 
         const int32_t * image_grid = clip_image_grid(ctx_clip);
 
@@ -290,12 +290,12 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
         // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
     }
 
-    printf("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
+    LOG_TEE("%s: image embedding created: %d tokens\n", __func__, *n_img_pos);
 
     const int64_t t_img_enc_end_us = ggml_time_us();
     float t_img_enc_ms = (t_img_enc_end_us - t_img_enc_start_us) / 1000.0;
 
-    printf("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
+    LOG_TEE("\n%s: image encoded in %8.2f ms by CLIP (%8.2f ms per image patch)\n", __func__, t_img_enc_ms, t_img_enc_ms / *n_img_pos);
 
     return true;
 }
@@ -305,7 +305,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     int n_llama_embd = llama_n_embd(llama_get_model(ctx_llama));
     auto n_image_embd = clip_n_mmproj_embd(ctx_clip);
     if (n_image_embd != n_llama_embd) {
-        printf("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
+        LOG_TEE("%s: embedding dim of the multimodal projector (%d) is not equal to that of LLaMA (%d). Make sure that you use the correct mmproj file.\n", __func__, n_image_embd, n_llama_embd);
         return false;
     }
     return true;
@@ -314,13 +314,13 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
 bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
     float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
     if (!image_embd) {
-        fprintf(stderr, "Unable to allocate memory for image embeddings\n");
+        LOG_TEE("Unable to allocate memory for image embeddings\n");
         return false;
     }
 
     int n_img_pos;
     if (!encode_image_with_clip(ctx_clip, n_threads, img, image_embd, &n_img_pos)) {
-        fprintf(stderr, "%s: cannot encode image, aborting\n", __func__);
+        LOG_TEE("%s: cannot encode image, aborting\n", __func__);
         free(image_embd);
         return false;
     }
@@ -340,7 +340,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         }
         llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
         if (llama_decode(ctx_llama, batch)) {
-            fprintf(stderr, "%s : failed to eval\n", __func__);
+            LOG_TEE("%s : failed to eval\n", __func__);
             return false;
         }
         *n_past += n_eval;
@@ -352,7 +352,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     clip_image_u8 * img = clip_image_u8_init();
     if (!clip_image_load_from_bytes(image_bytes, image_bytes_length, img)) {
         clip_image_u8_free(img);
-        fprintf(stderr, "%s: can't load image from bytes, is it a valid image?", __func__);
+        LOG_TEE("%s: can't load image from bytes, is it a valid image?", __func__);
         return NULL;
     }
 
@@ -361,7 +361,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        fprintf(stderr, "%s: coulnd't embed the image\n", __func__);
+        LOG_TEE("%s: coulnd't embed the image\n", __func__);
         return NULL;
     }
 
@@ -375,7 +375,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
 static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long *sizeOut) {
     auto file = fopen(path, "rb");
     if (file == NULL) {
-        fprintf(stderr, "%s: can't read file %s\n", __func__, path);
+        LOG_TEE("%s: can't read file %s\n", __func__, path);
         return false;
     }
 
@@ -385,7 +385,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
 
     auto buffer = (unsigned char *)malloc(fileSize); // Allocate memory to hold the file data
     if (buffer == NULL) {
-        fprintf(stderr, "%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
+        LOG_TEE("%s: failed to alloc %ld bytes for file %s\n", __func__, fileSize, path);
         perror("Memory allocation error");
         fclose(file);
         return false;
@@ -410,7 +410,7 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
     if (!loaded) {
-        fprintf(stderr, "%s: failed to load %s\n", __func__, image_path);
+        LOG_TEE("%s: failed to load %s\n", __func__, image_path);
         return NULL;
     }
 

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);
@@ -302,7 +299,7 @@ int main(int argc, char ** argv) {
                 }
                 fflush(stdout);
 
-                if (id == llama_token_eos(model)) {
+                if (llama_token_is_eog(model, id)) {
                     has_eos = true;
                 }
 

examples/lookup/lookup-create.cpp

@@ -28,10 +28,8 @@ int main(int argc, char ** argv){
     GGML_ASSERT(model != nullptr);
 
     // tokenize the prompt
-    const bool add_bos = llama_should_add_bos_token(model);
-
     std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
+    inp = ::llama_tokenize(ctx, params.prompt, true, true);
     fprintf(stderr, "%s: tokenization done\n", __func__);
 
 

examples/lookup/lookup-stats.cpp

@@ -30,15 +30,11 @@ 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;

examples/lookup/lookup.cpp

@@ -38,15 +38,11 @@ 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;
@@ -144,7 +140,7 @@ int main(int argc, char ** argv){
                 printf("%s", token_str.c_str());
             }
 
-            if (id == llama_token_eos(model)) {
+            if (llama_token_is_eog(model, id)) {
                 has_eos = true;
             }
 

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

@@ -296,19 +296,23 @@ 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
 
 -   `--prompt-cache FNAME`: Specify a file to cache the model state after the initial prompt. This can significantly speed up the startup time when you're using longer prompts. The file is created during the first run and is reused and updated in subsequent runs. **Note**: Restoring a cached prompt does not imply restoring the exact state of the session at the point it was saved. So even when specifying a specific seed, you are not guaranteed to get the same sequence of tokens as the original generation.
 
-### Grammars
+### Grammars & JSON schemas
 
 -   `--grammar GRAMMAR`, `--grammar-file FILE`: Specify a grammar (defined inline or in a file) to constrain model output to a specific format. For example, you could force the model to output JSON or to speak only in emojis. See the [GBNF guide](../../grammars/README.md) for details on the syntax.
 
+-   `--json-schema SCHEMA`: Specify a [JSON schema](https://json-schema.org/) to constrain model output to (e.g. `{}` for any JSON object, or `{"items": {"type": "string", "minLength": 10, "maxLength": 100}, "minItems": 10}` for a JSON array of strings with size constraints). If a schema uses external `$ref`s, you should use `--grammar "$( python examples/json_schema_to_grammar.py myschema.json )"` instead.
+
 ### 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 +320,8 @@ 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

@@ -235,17 +235,17 @@ 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;
             }
             session_tokens.resize(n_token_count_out);
-            llama_set_rng_seed(ctx, params.seed);
             LOG_TEE("%s: loaded a session with prompt size of %d tokens\n", __func__, (int)session_tokens.size());
         }
     }
 
     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 +255,7 @@ 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 +277,10 @@ 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 +339,14 @@ 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 +693,7 @@ 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());
             }
@@ -794,8 +794,8 @@ int main(int argc, char ** argv) {
                 }
             }
 
-            // deal with end of text token in interactive mode
-            if (llama_sampling_last(ctx_sampling) == llama_token_eos(model)) {
+            // deal with end of generation tokens in interactive mode
+            if (llama_token_is_eog(model, llama_sampling_last(ctx_sampling))) {
                 LOG("found EOS token\n");
 
                 if (params.interactive) {
@@ -919,8 +919,8 @@ int main(int argc, char ** argv) {
             }
         }
 
-        // end of text token
-        if (!embd.empty() && embd.back() == llama_token_eos(model) && !(params.instruct || params.interactive || params.chatml)) {
+        // end of generation
+        if (!embd.empty() && llama_token_is_eog(model, embd.back()) && !(params.instruct || params.interactive || params.chatml)) {
             LOG_TEE(" [end of text]\n");
             break;
         }
@@ -935,7 +935,7 @@ 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
@@ -360,7 +359,7 @@ int main(int argc, char ** argv) {
                 //        client.id, client.seq_id, id, client.n_decoded, client.i_batch, token_str.c_str());
 
                 if (client.n_decoded > 2 &&
-                        (id == llama_token_eos(model) ||
+                        (llama_token_is_eog(model, id) ||
                          (params.n_predict > 0 && client.n_decoded + client.n_prompt >= params.n_predict) ||
                          client.response.find("User:") != std::string::npos ||
                          client.response.find('\n') != std::string::npos)) {

examples/passkey/passkey.cpp

@@ -252,8 +252,8 @@ int main(int argc, char ** argv) {
             // sample the most likely token
             const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
-            // is it an end of stream?
-            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+            // is it an end of generation?
+            if (llama_token_is_eog(model, new_token_id) || n_cur == n_len) {
                 LOG_TEE("\n");
 
                 break;

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;
@@ -1823,12 +1852,20 @@ int main(int argc, char ** argv) {
 
     const int32_t n_ctx = params.n_ctx;
 
+    if (n_ctx <= 0) {
+        fprintf(stderr, "%s: perplexity tool requires '--ctx-size' > 0\n", __func__);
+        return 1;
+    }
+
     const bool ppl = !params.hellaswag && !params.winogrande && !params.multiple_choice && !params.kl_divergence;
+
     if (ppl) {
-        int n_seq = std::max(1, params.n_batch / n_ctx);
-        int32_t n_kv = n_seq * n_ctx;
+        const int32_t n_seq = std::max(1, params.n_batch / n_ctx);
+        const int32_t n_kv = n_seq * n_ctx;
+
         params.n_parallel = n_seq;
-        params.n_ctx = n_kv;
+        params.n_ctx      = n_kv;
+
         params.n_batch = std::min(params.n_batch, n_kv);
     } else {
         params.n_batch = std::min(params.n_batch, params.n_ctx);

examples/quantize/CMakeLists.txt

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

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

@@ -8,7 +8,6 @@
 #include <unordered_map>
 #include <fstream>
 #include <cmath>
-#include <algorithm>
 
 struct quant_option {
     std::string name;
@@ -26,6 +25,7 @@ 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",            },
@@ -52,6 +52,10 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "COPY",   LLAMA_FTYPE_ALL_F32,       "only copy tensors, no quantizing", },
 };
 
+static const char * const LLM_KV_QUANTIZE_IMATRIX_FILE       = "quantize.imatrix.file";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_DATASET    = "quantize.imatrix.dataset";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES  = "quantize.imatrix.entries_count";
+static const char * const LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS   = "quantize.imatrix.chunks_count";
 
 static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std::string & ftype_str_out) {
     std::string ftype_str;
@@ -87,13 +91,18 @@ 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("  --keep-split: will generate quatized model in the same shards as input");
+    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) {
@@ -107,61 +116,80 @@ 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 int load_imatrix(const std::string & imatrix_file, std::string & imatrix_dataset, 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());
+
+    // latest imatrix version contains the dataset filename at the end of the file
+    int m_last_call = 0;
+    if (in.peek() != EOF) {
+        in.read((char *)&m_last_call, sizeof(m_last_call));
+        int dataset_len;
+        in.read((char *)&dataset_len, sizeof(dataset_len));
+        std::vector<char> dataset_as_vec(dataset_len);
+        in.read(dataset_as_vec.data(), dataset_len);
+        imatrix_dataset.assign(dataset_as_vec.begin(), dataset_as_vec.end());
+        printf("%s: imatrix dataset='%s'\n", __func__, imatrix_dataset.c_str());
+    }
+    printf("%s: loaded %d importance matrix entries from %s computed on %d chunks\n", __func__, int(imatrix_data.size()), imatrix_file.c_str(), m_last_call);
+    return m_last_call;
 }
 
-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 int prepare_imatrix(const std::string & imatrix_file,
+        std::string & imatrix_dataset,
+        const std::vector<std::string> & included_weights,
+        const std::vector<std::string> & excluded_weights,
+        std::unordered_map<std::string, std::vector<float>> & imatrix_data) {
+    int m_last_call = -1;
     if (!imatrix_file.empty()) {
-        load_imatrix(imatrix_file, imatrix_data);
+        m_last_call = load_imatrix(imatrix_file, imatrix_dataset, imatrix_data);
     }
     if (imatrix_data.empty()) {
-        return;
+        return m_last_call;
     }
     if (!excluded_weights.empty()) {
         for (auto& name : excluded_weights) {
@@ -187,6 +215,7 @@ static void prepare_imatrix(const std::string& imatrix_file,
     if (!imatrix_data.empty()) {
         printf("%s: have %d importance matrix entries\n", __func__, int(imatrix_data.size()));
     }
+    return m_last_call;
 }
 
 static ggml_type parse_ggml_type(const char * arg) {
@@ -211,6 +240,7 @@ 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) {
@@ -227,6 +257,10 @@ int main(int argc, char ** argv) {
             } 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) {
@@ -249,6 +283,8 @@ int main(int argc, char ** argv) {
             } else {
                 usage(argv[0]);
             }
+        } else if (strcmp(argv[arg_idx], "--keep-split")) {
+            params.keep_split = true;
         } else {
             usage(argv[0]);
         }
@@ -262,10 +298,48 @@ int main(int argc, char ** argv) {
         usage(argv[0]);
     }
 
+    std::string imatrix_dataset;
     std::unordered_map<std::string, std::vector<float>> imatrix_data;
-    prepare_imatrix(imatrix_file, included_weights, excluded_weights, imatrix_data);
+    int m_last_call = prepare_imatrix(imatrix_file, imatrix_dataset, included_weights, excluded_weights, imatrix_data);
     if (!imatrix_data.empty()) {
         params.imatrix = &imatrix_data;
+        {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_FILE);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+            strncpy(kvo.val_str, imatrix_file.c_str(), 127);
+            kvo.val_str[127] = '\0';
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+        if (!imatrix_dataset.empty()) {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_DATASET);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_STR;
+            strncpy(kvo.val_str, imatrix_dataset.c_str(), 127);
+            kvo.val_str[127] = '\0';
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+
+        {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_ENTRIES);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+            kvo.val_i64 = imatrix_data.size();
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+
+        if (m_last_call > 0) {
+            llama_model_kv_override kvo;
+            std::strcpy(kvo.key, LLM_KV_QUANTIZE_IMATRIX_N_CHUNKS);
+            kvo.tag = LLAMA_KV_OVERRIDE_TYPE_INT;
+            kvo.val_i64 = m_last_call;
+            kv_overrides.emplace_back(std::move(kvo));
+        }
+    }
+    if (!kv_overrides.empty()) {
+        kv_overrides.emplace_back();
+        kv_overrides.back().key[0] = 0;
+        params.kv_overrides = &kv_overrides;
     }
 
     llama_backend_init();
@@ -276,21 +350,28 @@ int main(int argc, char ** argv) {
     std::string fname_out;
 
     std::string ftype_str;
+    std::string suffix = ".gguf";
     if (try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
         std::string fpath;
         const size_t pos = fname_inp.find_last_of("/\\");
         if (pos != std::string::npos) {
             fpath = fname_inp.substr(0, pos + 1);
         }
-        // export as [inp path]/ggml-model-[ftype].gguf
-        fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
+
+        // export as [inp path]/ggml-model-[ftype]. Only add extension if there is no splitting
+        fname_out = fpath + "ggml-model-" + ftype_str;
+        if (!params.keep_split) {
+            fname_out += suffix;
+        }
         arg_idx++;
         if (ftype_str == "COPY") {
             params.only_copy = true;
         }
-    }
-    else {
+    } else {
         fname_out = argv[arg_idx];
+        if (params.keep_split && fname_out.find(suffix) != std::string::npos) {
+            fname_out = fname_out.substr(0, fname_out.length() - suffix.length());
+        }
         arg_idx++;
 
         if (argc <= arg_idx) {
@@ -320,10 +401,12 @@ 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/quantize/tests.sh

@@ -0,0 +1,65 @@
+#!/bin/bash
+
+set -eu
+
+if [ $# -lt 1 ]
+then
+    echo "usage:   $0 path_to_build_binary [path_to_temp_folder]"
+    echo "example: $0 ../../build/bin ../../tmp"
+    exit 1
+fi
+
+if [ $# -gt 1 ]
+then
+    TMP_DIR=$2
+else
+    TMP_DIR=/tmp
+fi
+
+set -x
+
+SPLIT=$1/gguf-split
+QUANTIZE=$1/quantize
+MAIN=$1/main
+WORK_PATH=$TMP_DIR/quantize
+ROOT_DIR=$(realpath $(dirname $0)/../../)
+
+mkdir -p "$WORK_PATH"
+
+# Clean up in case of previously failed test
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf
+
+# 1. Get a model
+(
+cd $WORK_PATH
+"$ROOT_DIR"/scripts/hf.sh --repo ggml-org/gemma-1.1-2b-it-Q8_0-GGUF --file gemma-1.1-2b-it.Q8_0.gguf
+)
+echo PASS
+
+# 2. Split model
+$SPLIT --split-max-tensors 28  $WORK_PATH/gemma-1.1-2b-it.Q8_0.gguf $WORK_PATH/ggml-model-split
+echo PASS
+echo
+
+# 3. Requant model with '--keep_split'
+$QUANTIZE --allow-requantize --keep_split $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant.gguf Q4_K
+echo PASS
+echo
+
+# 3a. Test the requanted model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-requant-00001-of-00006.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# 4. Requant mode without '--keep_split'
+$QUANTIZE --allow-requantize $WORK_PATH/ggml-model-split-00001-of-00006.gguf $WORK_PATH/ggml-model-requant-merge.gguf Q4_K
+echo PASS
+echo
+
+# 4b. Test the requanted model is loading properly
+$MAIN --model $WORK_PATH/ggml-model-requant-merge.gguf --random-prompt --n-predict 32
+echo PASS
+echo
+
+# Clean up
+rm -f $WORK_PATH/ggml-model-split*.gguf $WORK_PATH/ggml-model-requant*.gguf

examples/regex-to-grammar.py

@@ -8,7 +8,7 @@ print(subprocess.check_output(
         "python",
         os.path.join(
         os.path.dirname(os.path.realpath(__file__)),
-        "json-schema-to-grammar.py"),
+        "json_schema_to_grammar.py"),
         *rest,
         "-",
         "--raw-pattern",