Fix compiler warnings (shadow variable)

.clang-tidy

@@ -12,7 +12,6 @@ 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

@@ -26,8 +26,8 @@ COPY . .
 
 # Set nvcc architecture
 ENV CUDA_DOCKER_ARCH=${CUDA_DOCKER_ARCH}
-# Enable CUDA
-ENV LLAMA_CUDA=1
+# Enable cuBLAS
+ENV LLAMA_CUBLAS=1
 
 RUN make
 

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

@@ -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-cuda
+Name:           llama.cpp-cublas
 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_CUDA=1
+make -j LLAMA_CUBLAS=1
 
 %install
 mkdir -p %{buildroot}%{_bindir}/
-cp -p main %{buildroot}%{_bindir}/llamacppcuda
-cp -p server %{buildroot}%{_bindir}/llamacppcudaserver
-cp -p simple %{buildroot}%{_bindir}/llamacppcudasimple
+cp -p main %{buildroot}%{_bindir}/llamacppcublas
+cp -p server %{buildroot}%{_bindir}/llamacppcublasserver
+cp -p simple %{buildroot}%{_bindir}/llamacppcublassimple
 
 mkdir -p %{buildroot}/usr/lib/systemd/system
-%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacuda.service
+%{__cat} <<EOF  > %{buildroot}/usr/lib/systemd/system/llamacublas.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/llamacppcudaserver $LLAMA_ARGS
+ExecStart=/usr/bin/llamacppcublasserver $LLAMA_ARGS
 ExecReload=/bin/kill -s HUP $MAINPID
 Restart=never
 
@@ -67,10 +67,10 @@ rm -rf %{buildroot}
 rm -rf %{_builddir}/*
 
 %files
-%{_bindir}/llamacppcuda
-%{_bindir}/llamacppcudaserver
-%{_bindir}/llamacppcudasimple
-/usr/lib/systemd/system/llamacuda.service
+%{_bindir}/llamacppcublas
+%{_bindir}/llamacppcublasserver
+%{_bindir}/llamacppcublassimple
+/usr/lib/systemd/system/llamacublas.service
 %config /etc/sysconfig/llama
 
 %pre

.devops/main-cuda.Dockerfile

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

.devops/nix/docker.nix

@@ -1,37 +0,0 @@
-{
-  lib,
-  dockerTools,
-  buildEnv,
-  llama-cpp,
-  interactive ? true,
-  coreutils,
-}:
-
-# A tar that can be fed into `docker load`:
-#
-# $ nix build .#llamaPackages.docker
-# $ docker load < result
-
-# For details and variations cf.
-# - https://nixos.org/manual/nixpkgs/unstable/#ssec-pkgs-dockerTools-buildLayeredImage
-# - https://discourse.nixos.org/t/a-faster-dockertools-buildimage-prototype/16922
-# - https://nixery.dev/
-
-# Approximate (compressed) sizes, at the time of writing, are:
-#
-# .#llamaPackages.docker: 125M;
-# .#llamaPackagesCuda.docker: 537M;
-# .#legacyPackages.aarch64-linux.llamaPackagesXavier.docker: 415M.
-
-dockerTools.buildLayeredImage {
-  name = llama-cpp.pname;
-  tag = "latest";
-
-  contents =
-    [ llama-cpp ]
-    ++ lib.optionals interactive [
-      coreutils
-      dockerTools.binSh
-      dockerTools.caCertificates
-    ];
-}

.devops/nix/package.nix

@@ -1,17 +1,15 @@
 {
   lib,
-  glibc,
   config,
   stdenv,
   mkShell,
-  runCommand,
   cmake,
   ninja,
   pkg-config,
   git,
   python3,
   mpi,
-  blas,
+  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
   cudaPackages,
   darwin,
   rocmPackages,
@@ -24,7 +22,7 @@
     useOpenCL
     useRocm
     useVulkan
-  ] && blas.meta.available,
+  ],
   useCuda ? config.cudaSupport,
   useMetalKit ? stdenv.isAarch64 && stdenv.isDarwin && !useOpenCL,
   useMpi ? false, # Increases the runtime closure size by ~700M
@@ -32,12 +30,6 @@
   useRocm ? config.rocmSupport,
   useVulkan ? false,
   llamaVersion ? "0.0.0", # Arbitrary version, substituted by the flake
-
-  # 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,
-  precompileMetalShaders ? false
 }@inputs:
 
 let
@@ -49,7 +41,10 @@ let
     versionOlder
     ;
 
+  # It's necessary to consistently use backendStdenv when building with CUDA support,
+  # otherwise we get libstdc++ errors downstream.
   stdenv = throw "Use effectiveStdenv instead";
+  effectiveStdenv = if useCuda then cudaPackages.backendStdenv else inputs.stdenv;
 
   suffices =
     lib.optionals useBlas [ "BLAS" ]
@@ -67,15 +62,10 @@ 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
@@ -94,11 +84,6 @@ let
     ]
   );
 
-  xcrunHost = runCommand "xcrunHost" {} ''
-    mkdir -p $out/bin
-    ln -s /usr/bin/xcrun $out/bin
-  '';
-
   # apple_sdk is supposed to choose sane defaults, no need to handle isAarch64
   # separately
   darwinBuildInputs =
@@ -162,17 +147,12 @@ effectiveStdenv.mkDerivation (
     postPatch = ''
       substituteInPlace ./ggml-metal.m \
         --replace '[bundle pathForResource:@"ggml-metal" ofType:@"metal"];' "@\"$out/bin/ggml-metal.metal\";"
-      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 $PATH
-    # see https://github.com/ggerganov/llama.cpp/pull/6118 for discussion
-    __noChroot = effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders;
+      # TODO: Package up each Python script or service appropriately.
+      # If we were to migrate to buildPythonPackage and prepare the `pyproject.toml`,
+      # we could make those *.py into setuptools' entrypoints
+      substituteInPlace ./*.py --replace "/usr/bin/env python" "${llama-python}/bin/python"
+    '';
 
     nativeBuildInputs =
       [
@@ -187,11 +167,6 @@ effectiveStdenv.mkDerivation (
         # TODO: Replace with autoAddDriverRunpath
         # once https://github.com/NixOS/nixpkgs/pull/275241 has been merged
         cudaPackages.autoAddOpenGLRunpathHook
-      ]
-      ++ optionals (effectiveStdenv.hostPlatform.isGnu && enableStatic) [
-        glibc.static
-      ] ++ optionals (effectiveStdenv.isDarwin && useMetalKit && precompileMetalShaders) [
-        xcrunHost
       ];
 
     buildInputs =
@@ -200,23 +175,21 @@ effectiveStdenv.mkDerivation (
       ++ optionals useMpi [ mpi ]
       ++ optionals useOpenCL [ clblast ]
       ++ optionals useRocm rocmBuildInputs
-      ++ optionals useBlas [ blas ]
       ++ optionals useVulkan vulkanBuildInputs;
 
     cmakeFlags =
       [
         (cmakeBool "LLAMA_NATIVE" false)
         (cmakeBool "LLAMA_BUILD_SERVER" true)
-        (cmakeBool "BUILD_SHARED_LIBS" (!enableStatic))
+        (cmakeBool "BUILD_SHARED_LIBS" true)
         (cmakeBool "CMAKE_SKIP_BUILD_RPATH" true)
         (cmakeBool "LLAMA_BLAS" useBlas)
         (cmakeBool "LLAMA_CLBLAST" useOpenCL)
-        (cmakeBool "LLAMA_CUDA" useCuda)
+        (cmakeBool "LLAMA_CUBLAS" useCuda)
         (cmakeBool "LLAMA_HIPBLAS" useRocm)
         (cmakeBool "LLAMA_METAL" useMetalKit)
         (cmakeBool "LLAMA_MPI" useMpi)
         (cmakeBool "LLAMA_VULKAN" useVulkan)
-        (cmakeBool "LLAMA_STATIC" enableStatic)
       ]
       ++ optionals useCuda [
         (
@@ -236,16 +209,14 @@ 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")
-        (cmakeBool "LLAMA_METAL_EMBED_LIBRARY" (!precompileMetalShaders))
-      ];
+      ++ optionals useMetalKit [ (lib.cmakeFeature "CMAKE_C_FLAGS" "-D__ARM_FEATURE_DOTPROD=1") ]
+      ++ optionals useBlas [ (lib.cmakeFeature "LLAMA_BLAS_VENDOR" "OpenBLAS") ];
 
     # TODO(SomeoneSerge): It's better to add proper install targets at the CMake level,
     # if they haven't been added yet.
     postInstall = ''
-      mv $out/bin/main${executableSuffix} $out/bin/llama${executableSuffix}
-      mv $out/bin/server${executableSuffix} $out/bin/llama-server${executableSuffix}
+      mv $out/bin/main $out/bin/llama
+      mv $out/bin/server $out/bin/llama-server
       mkdir -p $out/include
       cp $src/llama.h $out/include/
     '';
@@ -284,11 +255,11 @@ effectiveStdenv.mkDerivation (
       # Configurations we don't want even the CI to evaluate. Results in the
       # "unsupported platform" messages. This is mostly a no-op, because
       # cudaPackages would've refused to evaluate anyway.
-      badPlatforms = optionals (useCuda || useOpenCL) lib.platforms.darwin;
+      badPlatforms = optionals (useCuda || useOpenCL || useVulkan) lib.platforms.darwin;
 
       # Configurations that are known to result in build failures. Can be
       # overridden by importing Nixpkgs with `allowBroken = true`.
-      broken = (useMetalKit && !effectiveStdenv.isDarwin);
+      broken = (useMetalKit && !effectiveStdenv.isDarwin) || (useVulkan && effectiveStdenv.isDarwin);
 
       description = "Inference of LLaMA model in pure C/C++${descriptionSuffix}";
       homepage = "https://github.com/ggerganov/llama.cpp/";

.devops/nix/scope.nix

@@ -12,8 +12,5 @@ lib.makeScope newScope (
   self: {
     inherit llamaVersion;
     llama-cpp = self.callPackage ./package.nix { };
-    docker = self.callPackage ./docker.nix { };
-    docker-min = self.callPackage ./docker.nix { interactive = false; };
-    sif = self.callPackage ./sif.nix { };
   }
 )

.devops/nix/sif.nix

@@ -1,27 +0,0 @@
-{
-  lib,
-  singularity-tools,
-  llama-cpp,
-  bashInteractive,
-  interactive ? false,
-}:
-
-let
-  optionalInt = cond: x: if cond then x else 0;
-in
-singularity-tools.buildImage rec {
-  inherit (llama-cpp) name;
-  contents = [ llama-cpp ] ++ lib.optionals interactive [ bashInteractive ];
-
-  # These are excessive (but safe) for most variants. Building singularity
-  # images requires superuser privileges, so we build them inside a VM in a
-  # writable image of pre-determined size.
-  #
-  # ROCm is currently affected by https://github.com/NixOS/nixpkgs/issues/276846
-  #
-  # Expected image sizes:
-  # - cpu/blas: 150M,
-  # - cuda, all gencodes: 560M,
-  diskSize = 4096 + optionalInt llama-cpp.useRocm 16384;
-  memSize = diskSize;
-}

.devops/server-cuda.Dockerfile

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

.github/ISSUE_TEMPLATE/bug.md

@@ -7,5 +7,3 @@ assignees: ''
 ---
 
 Please include information about your system, the steps to reproduce the bug, and the version of llama.cpp that you are using. If possible, please provide a minimal code example that reproduces the bug.
-
-If the bug concerns the server, please try to reproduce it first using the [server test scenario framework](https://github.com/ggerganov/llama.cpp/tree/master/examples/server/tests).

.github/workflows/bench.yml

@@ -1,280 +0,0 @@
-# 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: ['.github/workflows/bench.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/bench/**.*']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/bench.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/bench/**.*']
-  schedule:
-    -  cron: '04 2 * * *'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  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
-    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@v3
-        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: Install k6
-        id: k6_installation
-        run: |
-          cd examples/server/bench
-          wget --quiet https://github.com/grafana/k6/releases/download/v0.49.0/k6-v0.49.0-linux-amd64.tar.gz
-          tar xzf k6*.tar.gz --strip-components=1
-
-      - 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
-          BENCH_K6_BIN_PATH=./k6 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 phi-2/ggml-model-q4_0.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: benchmark-results
-          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
-        continue-on-error: true # If not authorized on external repo
-        with:
-          authToken: ${{secrets.GITHUB_TOKEN}}
-          sha: ${{ inputs.sha || github.event.pull_request.head.sha || github.sha }}
-          context: bench-server-baseline
-          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: ${{ github.event.pull_request != '' }}
-        with:
-          message-id: bench-${{ github.job }}-${{ env.RUNNER_LABEL }}
-          message: |
-            📈 **llama.cpp server** for _${{ github.job }}_ on _${{ env.RUNNER_LABEL }}_: **${{ env.BENCH_ITERATIONS}} iterations** 🚀
-
-            - Concurrent users: ${{ env.N_USERS }}, duration: ${{ github.event.inputs.duration || env.DURATION }}
-            - HTTP request          : avg=${{ env.HTTP_REQ_DURATION_AVG }}ms        p(90)=${{ env.HTTP_REQ_DURATION_P_90_ }}ms fails=${{ 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=${{ env.LLAMACPP_PROMPT_TOKENS_AVG }}tk/s p(90)=${{ env.LLAMACPP_PROMPT_TOKENS_P_90_ }}tk/s **total=${{ env.LLAMACPP_PROMPT_TOKENS_TOTAL_COUNTER_RATE }}tk/s**
-            - Token generation  (tg): avg=${{ env.LLAMACPP_TOKENS_SECOND_AVG }}tk/s p(90)=${{ env.LLAMACPP_TOKENS_SECOND_P_90_ }}tk/s **total=${{ env.LLAMACPP_COMPLETION_TOKENS_TOTAL_COUNTER_RATE }}tk/s**
-            - ${{ env.BENCH_GRAPH_XLABEL }}
-
-            <details>
-
-            <summary>Time series</summary>
-
-            <p align="center">
-
-            <img width="100%" height="100%" src="${{ env.IMAGE_O }}" alt="prompt_tokens_seconds" />
-
-            <details>
-
-            <summary>More</summary>
-
-            ```mermaid
-            ${{ env.PROMPT_TOKENS_SECONDS }}
-            ```
-
-            </details>
-
-            <img width="100%" height="100%" src="${{ env.IMAGE_1 }}" alt="predicted_tokens_seconds"/>
-
-            <details>
-                <summary>More</summary>
-
-            ```mermaid
-            ${{ env.PREDICTED_TOKENS_SECONDS }}
-            ```
-
-            </details>
-
-            </p>
-
-            <details>
-
-            <summary>Details</summary>
-
-            <p align="center">
-
-            <img width="100%" height="100%" src="${{ env.IMAGE_2 }}" alt="kv_cache_usage_ratio" />
-
-            <details>
-                <summary>More</summary>
-
-            ```mermaid
-            ${{ env.KV_CACHE_USAGE_RATIO }}
-            ```
-
-            </details>
-
-            <img width="100%" height="100%" src="${{ env.IMAGE_3 }}" alt="requests_processing"/>
-
-            <details>
-                <summary>More</summary>
-
-            ```mermaid
-            ${{ env.REQUESTS_PROCESSING }}
-            ```
-
-            </details>
-
-            </p>
-            </details>
-            </details>

.github/workflows/build.yml

@@ -15,133 +15,14 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m']
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 env:
   BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
 jobs:
-  macOS-latest-cmake-arm64:
-    runs-on: macos-14
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          mkdir build
-          cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
-          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
-        with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-macos-arm64.zip
-
-  macOS-latest-cmake-x64:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        continue-on-error: true
-        run: |
-          brew update
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          mkdir build
-          cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON ..
-          cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
-
-      - name: Test
-        id: cmake_test
-        run: |
-          cd build
-          ctest -L main --verbose --timeout 900
-
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp LICENSE ./build/bin/
-          zip -r llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
-        with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-macos-x64.zip
-
   ubuntu-focal-make:
     runs-on: ubuntu-20.04
-    env:
-      LLAMA_NODE_AVAILABLE: true
-      LLAMA_PYTHON_AVAILABLE: true
 
     steps:
       - name: Clone
@@ -154,18 +35,8 @@ jobs:
           sudo apt-get update
           sudo apt-get install build-essential gcc-8
 
-      - uses: actions/setup-node@v4
-        with:
-          node-version: "20"
-
-      - uses: actions/setup-python@v4
-        with:
-          python-version: "3.11"
-
       - name: Build
         id: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
         run: |
           CC=gcc-8 make -j $(nproc)
 
@@ -175,28 +46,6 @@ jobs:
           CC=gcc-8 make tests -j $(nproc)
           make test -j $(nproc)
 
-  ubuntu-focal-make-curl:
-    runs-on: ubuntu-20.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential gcc-8 libcurl4-openssl-dev
-
-      - name: Build
-        id: make_build
-        env:
-          LLAMA_FATAL_WARNINGS: 1
-          LLAMA_CURL: 1
-        run: |
-          CC=gcc-8 make -j $(nproc)
-
   ubuntu-latest-cmake:
     runs-on: ubuntu-latest
 
@@ -216,7 +65,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_FATAL_WARNINGS=ON
+          cmake ..
           cmake --build . --config Release -j $(nproc)
 
       - name: Test
@@ -225,51 +74,40 @@ jobs:
           cd build
           ctest -L main --verbose --timeout 900
 
-      - name: Test llama2c conversion
-        id: llama2c_test
+  ubuntu-latest-cmake-sanitizer:
+    runs-on: ubuntu-latest
+
+    continue-on-error: true
+
+    strategy:
+      matrix:
+        sanitizer: [ADDRESS, THREAD, UNDEFINED]
+        build_type: [Debug, Release]
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt-get update
+          sudo apt-get install build-essential
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
+          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
+
+      - name: Test
+        id: cmake_test
         run: |
           cd build
-          echo "Fetch tokenizer"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/tok512.bin
-          echo "Fetch llama2c model"
-          wget https://huggingface.co/karpathy/tinyllamas/resolve/main/stories260K/stories260K.bin
-          ./bin/convert-llama2c-to-ggml --copy-vocab-from-model ./tok512.bin --llama2c-model stories260K.bin --llama2c-output-model stories260K.gguf
-          ./bin/main -m stories260K.gguf -p "One day, Lily met a Shoggoth" -n 500 -c 256
-
-#  ubuntu-latest-cmake-sanitizer:
-#    runs-on: ubuntu-latest
-#
-#    continue-on-error: true
-#
-#    strategy:
-#      matrix:
-#        sanitizer: [ADDRESS, THREAD, UNDEFINED]
-#        build_type: [Debug, Release]
-#
-#    steps:
-#      - name: Clone
-#        id: checkout
-#        uses: actions/checkout@v3
-#
-#      - name: Dependencies
-#        id: depends
-#        run: |
-#          sudo apt-get update
-#          sudo apt-get install build-essential
-#
-#      - name: Build
-#        id: cmake_build
-#        run: |
-#          mkdir build
-#          cd build
-#          cmake .. -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
-#          cmake --build . --config ${{ matrix.build_type }} -j $(nproc)
-#
-#      - name: Test
-#        id: cmake_test
-#        run: |
-#          cd build
-#          ctest -L main --verbose --timeout 900
+          ctest -L main --verbose --timeout 900
 
   ubuntu-latest-cmake-mpi:
     runs-on: ubuntu-latest
@@ -305,28 +143,6 @@ jobs:
           cd build
           ctest -L main --verbose
 
-  ubuntu-22-cmake-vulkan:
-    runs-on: ubuntu-22.04
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt-get install build-essential libvulkan-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake -DLLAMA_VULKAN=ON ..
-          cmake --build . --config Release -j $(nproc)
-
   ubuntu-22-cmake-sycl:
     runs-on: ubuntu-22.04
 
@@ -428,8 +244,6 @@ jobs:
 
       - name: Build
         id: make_build
-        env:
-            LLAMA_FATAL_WARNINGS: 1
         run: |
           LLAMA_NO_METAL=1 make -j $(sysctl -n hw.logicalcpu)
 
@@ -463,7 +277,7 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DLLAMA_METAL=OFF ..
+          cmake -DLLAMA_METAL=OFF ..
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -493,7 +307,6 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
-            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -522,7 +335,6 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
-            -DLLAMA_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
             -DLLAMA_BUILD_SERVER=OFF \
@@ -587,8 +399,6 @@ jobs:
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_KOMPUTE=ON -DKOMPUTE_OPT_DISABLE_VULKAN_VERSION_CHECK=ON -DBUILD_SHARED_LIBS=ON'
           - build: 'vulkan'
             defines: '-DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_VULKAN=ON -DBUILD_SHARED_LIBS=ON'
-          - build: 'arm64'
-            defines: '-A ARM64 -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DBUILD_SHARED_LIBS=ON'
 
     steps:
       - name: Clone
@@ -684,7 +494,7 @@ jobs:
       - name: Test
         id: cmake_test
         # not all machines have native AVX-512
-        if: ${{ matrix.build != 'arm64' && matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
+        if: ${{ matrix.build != 'clblast' && matrix.build != 'kompute' && matrix.build != 'vulkan' && (matrix.build != 'avx512' || env.HAS_AVX512F == '1') }}
         run: |
           cd build
           ctest -L main -C Release --verbose --timeout 900
@@ -728,13 +538,13 @@ jobs:
           path: |
             llama-${{ steps.tag.outputs.name }}-bin-win-${{ matrix.build }}-x64.zip
 
-  windows-latest-cmake-cuda:
+  windows-latest-cmake-cublas:
     runs-on: windows-latest
 
     strategy:
       matrix:
         cuda: ['12.2.0', '11.7.1']
-        build: ['cuda']
+        build: ['cublas']
 
     steps:
       - name: Clone
@@ -755,7 +565,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUDA=ON -DBUILD_SHARED_LIBS=ON
+          cmake .. -DLLAMA_NATIVE=OFF -DLLAMA_BUILD_SERVER=ON -DLLAMA_CUBLAS=ON -DBUILD_SHARED_LIBS=ON
           cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
 
       - name: Determine tag name
@@ -800,7 +610,6 @@ jobs:
 
   windows-latest-cmake-sycl:
     runs-on: windows-latest
-
     defaults:
       run:
         shell: bash
@@ -809,6 +618,7 @@ jobs:
       WINDOWS_BASEKIT_URL: https://registrationcenter-download.intel.com/akdlm/IRC_NAS/62641e01-1e8d-4ace-91d6-ae03f7f8a71f/w_BaseKit_p_2024.0.0.49563_offline.exe
       WINDOWS_DPCPP_MKL: intel.oneapi.win.cpp-dpcpp-common:intel.oneapi.win.mkl.devel
 
+
     steps:
       - name: Clone
         id: checkout
@@ -823,32 +633,6 @@ jobs:
         id: cmake_build
         run:  examples/sycl/win-build-sycl.bat
 
-      - name: Determine tag name
-        id: tag
-        shell: bash
-        run: |
-          BUILD_NUMBER="$(git rev-list --count HEAD)"
-          SHORT_HASH="$(git rev-parse --short=7 HEAD)"
-          if [[ "${{ env.BRANCH_NAME }}" == "master" ]]; then
-            echo "name=b${BUILD_NUMBER}" >> $GITHUB_OUTPUT
-          else
-            SAFE_NAME=$(echo "${{ env.BRANCH_NAME }}" | tr '/' '-')
-            echo "name=${SAFE_NAME}-b${BUILD_NUMBER}-${SHORT_HASH}" >> $GITHUB_OUTPUT
-          fi
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          7z a llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip ./build/bin/*
-
-      - name: Upload artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v3
-        with:
-          path: |
-            llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip
-
   ios-xcode-build:
     runs-on: macos-latest
 
@@ -881,7 +665,8 @@ jobs:
         run: |
           cd examples/llama.android
 
-          ./gradlew build --no-daemon
+          # Skip armeabi-v7a for now (https://github.com/llvm/llvm-project/issues/65820).
+          ./gradlew build --no-daemon -Pskip-armeabi-v7a
 
 #  freeBSD-latest:
 #    runs-on: macos-12
@@ -911,9 +696,7 @@ jobs:
       - macOS-latest-make
       - macOS-latest-cmake
       - windows-latest-cmake
-      - windows-latest-cmake-cuda
-      - macOS-latest-cmake-arm64
-      - macOS-latest-cmake-x64
+      - windows-latest-cmake-cublas
 
     steps:
       - name: Clone

.github/workflows/close-issue.yml

@@ -1,23 +0,0 @@
-name: Close inactive issues
-on:
-  schedule:
-    - cron: "42 0 * * *"
-
-jobs:
-  close-issues:
-    runs-on: ubuntu-latest
-    permissions:
-      issues: write
-      pull-requests: write
-    steps:
-      - uses: actions/stale@v5
-        with:
-          exempt-issue-labels: "refactor,help wanted,good first issue,research"
-          days-before-issue-stale: 30
-          days-before-issue-close: 14
-          stale-issue-label: "stale"
-          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: 10000
-          repo-token: ${{ secrets.GITHUB_TOKEN }}

.github/workflows/code-coverage.yml

@@ -5,10 +5,6 @@ env:
   GGML_NLOOP: 3
   GGML_N_THREADS: 1
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   run:
     runs-on: ubuntu-20.04

.github/workflows/docker.yml

@@ -15,10 +15,6 @@ on:
     branches:
       - master
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   push_to_registry:
     name: Push Docker image to Docker Hub

.github/workflows/editorconfig.yml

@@ -14,10 +14,6 @@ on:
     branches:
       - master
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   editorconfig:
     runs-on: ubuntu-latest

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

@@ -17,12 +17,9 @@ on:
     types: [opened, synchronize, reopened]
     paths: ['**/*.nix', 'flake.lock']
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   nix-build-aarch64:
+    if: ${{ vars.CACHIX_NAME != '' }}
     runs-on: ubuntu-latest
     steps:
     - name: Checkout repository
@@ -40,8 +37,8 @@ jobs:
         extra-conf: |
           extra-platforms = aarch64-linux
           extra-system-features = nixos-test kvm
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -49,7 +46,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Show all output paths
       run: >
           nix run github:nix-community/nix-eval-jobs

.github/workflows/nix-ci.yml

@@ -8,10 +8,6 @@ on:
   pull_request:
     types: [opened, synchronize, reopened]
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   nix-eval:
     strategy:
@@ -27,8 +23,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -41,6 +37,7 @@ jobs:
           --flake
           ".#packages.$(nix eval --raw --impure --expr builtins.currentSystem)"
   nix-build:
+    if: ${{ vars.CACHIX_NAME != '' }}
     strategy:
       fail-fast: false
       matrix:
@@ -54,8 +51,8 @@ jobs:
       with:
         github-token: ${{ secrets.GITHUB_TOKEN }}
         extra-conf: |
-          extra-substituters = https://llama-cpp.cachix.org https://cuda-maintainers.cachix.org
-          extra-trusted-public-keys = llama-cpp.cachix.org-1:H75X+w83wUKTIPSO1KWy9ADUrzThyGs8P5tmAbkWhQc= cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
+          extra-substituters = https://${{ vars.CACHIX_NAME }}.cachix.org https://cuda-maintainers.cachix.org
+          extra-trusted-public-keys = ${{ vars.CACHIX_PUBLIC_KEY }} cuda-maintainers.cachix.org-1:0dq3bujKpuEPMCX6U4WylrUDZ9JyUG0VpVZa7CNfq5E=
     - uses: DeterminateSystems/magic-nix-cache-action@v2
       with:
         upstream-cache: https://${{ matrix.cachixName }}.cachix.org
@@ -63,7 +60,7 @@ jobs:
       uses: cachix/cachix-action@v13
       with:
         authToken: '${{ secrets.CACHIX_AUTH_TOKEN }}'
-        name: llama-cpp
+        name: ${{ vars.CACHIX_NAME }}
     - name: Build
       run: >
           nix run github:Mic92/nix-fast-build

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

@@ -3,23 +3,17 @@ name: Python check requirements.txt
 on:
   push:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
       - 'requirements/*.txt'
   pull_request:
     paths:
-      - '.github/workflows/python-check-requirements.yml'
       - 'scripts/check-requirements.sh'
       - 'convert*.py'
       - 'requirements.txt'
       - 'requirements/*.txt'
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   python-check-requirements:
     runs-on: ubuntu-latest
@@ -32,4 +26,4 @@ jobs:
         with:
           python-version: "3.11"
       - name: Run check-requirements.sh script
-        run:  bash scripts/check-requirements.sh
+        run:  bash scripts/check-requirements.sh nocleanup

.github/workflows/python-lint.yml

@@ -2,10 +2,6 @@ name: flake8 Lint
 
 on: [push, pull_request]
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   flake8-lint:
     runs-on: ubuntu-latest

.github/workflows/server.yml

@@ -1,154 +0,0 @@
-# Server build and tests
-name: Server
-
-on:
-  workflow_dispatch: # allows manual triggering
-    inputs:
-      slow_tests:
-        description: 'Run slow tests'
-        required: true
-        type: boolean
-  push:
-    branches:
-      - master
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  pull_request:
-    types: [opened, synchronize, reopened]
-    paths: ['.github/workflows/server.yml', '**/CMakeLists.txt', '**/Makefile', '**/*.h', '**/*.hpp', '**/*.c', '**/*.cpp', '**/*.cu', '**/*.swift', '**/*.m', 'examples/server/tests/**.*']
-  schedule:
-    -  cron: '0 0 * * *'
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
-jobs:
-  server:
-    runs-on: ubuntu-latest
-
-    strategy:
-      matrix:
-        # TODO: temporary disabled due to linux kernel issues
-        #sanitizer: [ADDRESS, THREAD, UNDEFINED]
-        sanitizer: [UNDEFINED]
-        build_type: [Debug]
-        include:
-          - build_type: Release
-            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 \
-            build-essential \
-            git \
-            cmake \
-            python3-pip \
-            wget \
-            language-pack-en \
-            libcurl4-openssl-dev
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. \
-              -DLLAMA_NATIVE=OFF \
-              -DLLAMA_BUILD_SERVER=ON \
-              -DLLAMA_CURL=ON \
-              -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} \
-              -DLLAMA_SANITIZE_${{ 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
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd examples/server/tests
-          PORT=8888 ./tests.sh --stop --no-skipped --no-capture --tags slow
-
-
-  server-windows:
-    runs-on: windows-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v3
-        with:
-          fetch-depth: 0
-
-      - name: libCURL
-        id: get_libcurl
-        env:
-          CURL_VERSION: 8.6.0_6
-        run: |
-          curl.exe -o $env:RUNNER_TEMP/curl.zip -L "https://curl.se/windows/dl-${env:CURL_VERSION}/curl-${env:CURL_VERSION}-win64-mingw.zip"
-          mkdir $env:RUNNER_TEMP/libcurl
-          tar.exe -xvf $env:RUNNER_TEMP/curl.zip --strip-components=1 -C $env:RUNNER_TEMP/libcurl
-
-      - name: Build
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -DLLAMA_CURL=ON -DCURL_LIBRARY="$env:RUNNER_TEMP/libcurl/lib/libcurl.dll.a" -DCURL_INCLUDE_DIR="$env:RUNNER_TEMP/libcurl/include"
-          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS} --target server
-
-      - 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: Copy Libcurl
-        id: prepare_libcurl
-        run: |
-          cp $env:RUNNER_TEMP/libcurl/bin/libcurl-x64.dll ./build/bin/Release/libcurl-x64.dll
-
-      - name: Tests
-        id: server_integration_tests
-        if: ${{ !matrix.disabled_on_pr || !github.event.pull_request }}
-        run: |
-          cd examples/server/tests
-          behave.exe --summary --stop --no-capture --exclude 'issues|wrong_usages|passkey' --tags llama.cpp
-
-      - name: Slow tests
-        id: server_integration_tests_slow
-        if: ${{ (github.event.schedule || github.event.inputs.slow_tests == 'true') && matrix.build_type == 'Release' }}
-        run: |
-          cd examples/server/tests
-          behave.exe --stop --no-skipped --no-capture --tags slow

.github/workflows/tidy-post.yml

@@ -0,0 +1,20 @@
+name: clang-tidy review post comments
+
+on:
+  workflow_dispatch:
+    workflows: ["clang-tidy-review"]
+    types:
+      - completed
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+
+    steps:
+      - uses: ZedThree/clang-tidy-review/post@v0.13.0
+        # lgtm_comment_body, max_comments, and annotations need to be set on the posting workflow in a split setup
+        with:
+          # adjust options as necessary
+          lgtm_comment_body: ''
+          annotations: false
+          max_comments: 25

.github/workflows/tidy-review.yml

@@ -0,0 +1,23 @@
+name: clang-tidy-review
+
+on:
+  pull_request:
+    branches:
+      - master
+
+jobs:
+  clang-tidy-review:
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v3
+
+    - uses: ZedThree/clang-tidy-review@v0.13.0
+      id: review
+      with:
+        lgtm_comment_body: ''
+        build_dir: build
+        cmake_command: cmake . -B build -DCMAKE_EXPORT_COMPILE_COMMANDS=on
+        split_workflow: true
+
+    - uses: ZedThree/clang-tidy-review/upload@v0.13.0

.github/workflows/zig-build.yml

@@ -6,10 +6,6 @@ on:
     branches:
       - master
 
-concurrency:
-  group: ${{ github.workflow }}-${{ github.ref }}
-  cancel-in-progress: true
-
 jobs:
   build:
     strategy:

.gitignore

@@ -11,10 +11,7 @@
 *.gcda
 *.dot
 *.bat
-*.tmp
 *.metallib
-*.etag
-*.lastModified
 .DS_Store
 .build/
 .cache/
@@ -26,15 +23,11 @@
 .clang-tidy
 .vs/
 .vscode/
-.idea/
-
-ggml-metal-embed.metal
 
 lcov-report/
 gcovr-report/
 
 build*
-cmake-build-*
 out/
 tmp/
 
@@ -50,8 +43,6 @@ models-mnt
 /embedding
 /gguf
 /gguf-llama-simple
-/gguf-split
-/gritlm
 /imatrix
 /infill
 /libllama.so
@@ -59,9 +50,6 @@ models-mnt
 /llava-cli
 /lookahead
 /lookup
-/lookup-create
-/lookup-merge
-/lookup-stats
 /main
 /metal
 /passkey
@@ -77,7 +65,6 @@ models-mnt
 /batched-bench
 /export-lora
 /finetune
-/retrieval
 /speculative
 /parallel
 /train-text-from-scratch

CMakeLists.txt

@@ -55,9 +55,6 @@ option(LLAMA_ALL_WARNINGS               "llama: enable all compiler warnings"
 option(LLAMA_ALL_WARNINGS_3RD_PARTY     "llama: enable all compiler warnings in 3rd party libs" OFF)
 option(LLAMA_GPROF                      "llama: enable gprof"                                   OFF)
 
-# build
-option(LLAMA_FATAL_WARNINGS             "llama: enable -Werror flag"                            OFF)
-
 # sanitizers
 option(LLAMA_SANITIZE_THREAD            "llama: enable thread sanitizer"                        OFF)
 option(LLAMA_SANITIZE_ADDRESS           "llama: enable address sanitizer"                       OFF)
@@ -89,8 +86,8 @@ endif()
 option(LLAMA_ACCELERATE                      "llama: enable Accelerate framework"               ON)
 option(LLAMA_BLAS                            "llama: use BLAS"                                  OFF)
 set(LLAMA_BLAS_VENDOR "Generic" CACHE STRING "llama: BLAS library vendor")
-option(LLAMA_CUDA                            "llama: use CUDA"                                  OFF)
-option(LLAMA_CUBLAS                          "llama: use CUDA (deprecated, use LLAMA_CUDA)"     OFF)
+option(LLAMA_CUBLAS                          "llama: use CUDA"                                  OFF)
+#option(LLAMA_CUDA_CUBLAS                     "llama: use cuBLAS for prompt processing"          OFF)
 option(LLAMA_CUDA_FORCE_DMMV                 "llama: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(LLAMA_CUDA_FORCE_MMQ                  "llama: use mmq kernels instead of cuBLAS"         OFF)
 set(LLAMA_CUDA_DMMV_X      "32" CACHE STRING "llama: x stride for dmmv CUDA kernels")
@@ -99,8 +96,6 @@ option(LLAMA_CUDA_F16                        "llama: use 16 bit floats for some
 set(LLAMA_CUDA_KQUANTS_ITER "2" CACHE STRING "llama: iters./thread per block for Q2_K/Q6_K")
 set(LLAMA_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                              "llama: max. batch size for using peer access")
-option(LLAMA_CUDA_NO_PEER_COPY               "llama: do not use peer to peer copies"            OFF)
-option(LLAMA_CURL                            "llama: use libcurl to download model from an URL" OFF)
 option(LLAMA_HIPBLAS                         "llama: use hipBLAS"                               OFF)
 option(LLAMA_HIP_UMA                         "llama: use HIP unified memory architecture"       OFF)
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
@@ -112,25 +107,22 @@ option(LLAMA_VULKAN_RUN_TESTS                "llama: run Vulkan tests"
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 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)
 option(LLAMA_SYCL                            "llama: use SYCL"                                  OFF)
 option(LLAMA_SYCL_F16                        "llama: use 16 bit floats for sycl calculations"   OFF)
-set(LLAMA_SYCL_TARGET   "INTEL" CACHE STRING "llama: sycl target device")
-option(LLAMA_CPU_HBM                         "llama: use memkind for CPU HBM"                   OFF)
-set(LLAMA_SCHED_MAX_COPIES  "4" CACHE STRING "llama: max input copies for pipeline parallelism")
 
 option(LLAMA_BUILD_TESTS                     "llama: build tests"    ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_EXAMPLES                  "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER                    "llama: build server example"                      ON)
 
+
 # add perf arguments
 option(LLAMA_PERF                            "llama: enable perf"                               OFF)
+if (LLAMA_PERF)
+    add_definitions(-DGGML_PERF)
+endif()
 
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
@@ -138,7 +130,6 @@ include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
 #
 # Compile flags
 #
-
 if (LLAMA_SYCL)
     set(CMAKE_CXX_STANDARD 17)
 else()
@@ -149,12 +140,9 @@ set(CMAKE_CXX_STANDARD_REQUIRED true)
 set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
-
 find_package(Threads REQUIRED)
 include(CheckCXXCompilerFlag)
 
-add_compile_definitions(GGML_SCHED_MAX_COPIES=${LLAMA_SCHED_MAX_COPIES})
-
 # enable libstdc++ assertions for debug builds
 if (CMAKE_SYSTEM_NAME MATCHES "Linux")
     add_compile_definitions($<$<CONFIG:Debug>:_GLIBCXX_ASSERTIONS>)
@@ -163,17 +151,17 @@ endif()
 if (NOT MSVC)
     if (LLAMA_SANITIZE_THREAD)
         add_compile_options(-fsanitize=thread)
-        link_libraries     (-fsanitize=thread)
+        link_libraries(-fsanitize=thread)
     endif()
 
     if (LLAMA_SANITIZE_ADDRESS)
         add_compile_options(-fsanitize=address -fno-omit-frame-pointer)
-        link_libraries     (-fsanitize=address)
+        link_libraries(-fsanitize=address)
     endif()
 
     if (LLAMA_SANITIZE_UNDEFINED)
         add_compile_options(-fsanitize=undefined)
-        link_libraries     (-fsanitize=undefined)
+        link_libraries(-fsanitize=undefined)
     endif()
 endif()
 
@@ -205,80 +193,40 @@ if (LLAMA_METAL)
         add_compile_definitions(GGML_METAL_NDEBUG)
     endif()
 
-    # copy ggml-common.h and ggml-metal.metal to bin directory
-    configure_file(ggml-common.h    ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h    COPYONLY)
+    # get full path to the file
+    #add_compile_definitions(GGML_METAL_DIR_KERNELS="${CMAKE_CURRENT_SOURCE_DIR}/")
+
+    # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
-    if (LLAMA_METAL_EMBED_LIBRARY)
-        enable_language(ASM)
-        add_compile_definitions(GGML_METAL_EMBED_LIBRARY)
-
-        set(METALLIB_COMMON "${CMAKE_CURRENT_SOURCE_DIR}/ggml-common.h")
-        set(METALLIB_SOURCE "${CMAKE_CURRENT_SOURCE_DIR}/ggml-metal.metal")
-
-        file(MAKE_DIRECTORY "${CMAKE_BINARY_DIR}/autogenerated")
-
-        # merge ggml-common.h and ggml-metal.metal into a single file
-        set(METALLIB_EMBED_ASM    "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.s")
-        set(METALLIB_SOURCE_EMBED "${CMAKE_BINARY_DIR}/autogenerated/ggml-metal-embed.metal")
-
-        add_custom_command(
-            OUTPUT ${METALLIB_EMBED_ASM}
-            COMMAND echo "Embedding Metal library"
-            COMMAND sed -e '/\#include \"ggml-common.h\"/r ${METALLIB_COMMON}' -e '/\#include \"ggml-common.h\"/d' < ${METALLIB_SOURCE} > ${METALLIB_SOURCE_EMBED}
-            COMMAND echo ".section __DATA,__ggml_metallib"          >  ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_start"              >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_start:"                    >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".incbin \\\"${METALLIB_SOURCE_EMBED}\\\"" >> ${METALLIB_EMBED_ASM}
-            COMMAND echo ".globl _ggml_metallib_end"                >> ${METALLIB_EMBED_ASM}
-            COMMAND echo "_ggml_metallib_end:"                      >> ${METALLIB_EMBED_ASM}
-            DEPENDS ggml-metal.metal ggml-common.h
-            COMMENT "Generate assembly for embedded Metal library"
-        )
-
-        set(GGML_SOURCES_METAL ${GGML_SOURCES_METAL} ${METALLIB_EMBED_ASM})
-    else()
-        if (LLAMA_METAL_SHADER_DEBUG)
-            # custom command to do the following:
-            #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
-            #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
-            #
-            # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
-            #       disabling fast math is needed in order to pass tests/test-backend-ops
-            # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
-            # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
-            #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
-            set(XC_FLAGS -fno-fast-math -fno-inline -g)
-        else()
-            set(XC_FLAGS -O3)
-        endif()
-
-        # Append macOS metal versioning flags
-        if (LLAMA_METAL_MACOSX_VERSION_MIN)
-            message(STATUS "Adding -mmacosx-version-min=${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})
+    if (LLAMA_METAL_SHADER_DEBUG)
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
+        set(XC_FLAGS -fno-fast-math -fno-inline -g)
+        if (LLAMA_QKK_64)
+            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
         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
             COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
-            COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal
-            DEPENDS ggml-metal.metal ggml-common.h
+            DEPENDS ggml-metal.metal
             COMMENT "Compiling Metal kernels"
-            )
+        )
 
         add_custom_target(
             ggml-metal ALL
             DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-            )
-    endif() # LLAMA_METAL_EMBED_LIBRARY
+        )
+    endif()
 
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
@@ -350,17 +298,14 @@ if (LLAMA_BLAS)
         endif()
 
         message(STATUS "BLAS found, Includes: ${BLAS_INCLUDE_DIRS}")
-
         add_compile_options(${BLAS_LINKER_FLAGS})
-
         add_compile_definitions(GGML_USE_OPENBLAS)
-
         if (${BLAS_INCLUDE_DIRS} MATCHES "mkl" AND (${LLAMA_BLAS_VENDOR} MATCHES "Generic" OR ${LLAMA_BLAS_VENDOR} MATCHES "Intel"))
             add_compile_definitions(GGML_BLAS_USE_MKL)
         endif()
-
-        set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${BLAS_LIBRARIES})
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ${BLAS_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${BLAS_INCLUDE_DIRS})
+
     else()
         message(WARNING "BLAS not found, please refer to "
         "https://cmake.org/cmake/help/latest/module/FindBLAS.html#blas-lapack-vendors"
@@ -373,25 +318,21 @@ if (LLAMA_QKK_64)
 endif()
 
 if (LLAMA_CUBLAS)
-    message(WARNING "LLAMA_CUBLAS is deprecated and will be removed in the future.\nUse LLAMA_CUDA instead")
-    set(LLAMA_CUDA ON)
-endif()
-
-if (LLAMA_CUDA)
     cmake_minimum_required(VERSION 3.17)
 
     find_package(CUDAToolkit)
     if (CUDAToolkit_FOUND)
-        message(STATUS "CUDA found")
+        message(STATUS "cuBLAS found")
 
         enable_language(CUDA)
 
         set(GGML_HEADERS_CUDA ggml-cuda.h)
+        set(GGML_SOURCES_CUDA ggml-cuda.cu)
 
-        file(GLOB GGML_SOURCES_CUDA "ggml-cuda/*.cu")
-        list(APPEND GGML_SOURCES_CUDA "ggml-cuda.cu")
-
-        add_compile_definitions(GGML_USE_CUDA)
+        add_compile_definitions(GGML_USE_CUBLAS)
+#        if (LLAMA_CUDA_CUBLAS)
+#            add_compile_definitions(GGML_CUDA_CUBLAS)
+#        endif()
         if (LLAMA_CUDA_FORCE_DMMV)
             add_compile_definitions(GGML_CUDA_FORCE_DMMV)
         endif()
@@ -408,9 +349,6 @@ if (LLAMA_CUDA)
         endif()
         add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
-        if (LLAMA_CUDA_NO_PEER_COPY)
-            add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-        endif()
 
         if (LLAMA_STATIC)
             if (WIN32)
@@ -440,7 +378,7 @@ if (LLAMA_CUDA)
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     else()
-        message(WARNING "CUDA not found")
+        message(WARNING "cuBLAS not found")
     endif()
 endif()
 
@@ -449,20 +387,15 @@ if (LLAMA_MPI)
     find_package(MPI)
     if (MPI_C_FOUND)
         message(STATUS "MPI found")
-
         set(GGML_HEADERS_MPI ggml-mpi.h)
-        set(GGML_SOURCES_MPI ggml-mpi.c)
-
+        set(GGML_SOURCES_MPI ggml-mpi.c ggml-mpi.h)
         add_compile_definitions(GGML_USE_MPI)
         add_compile_definitions(${MPI_C_COMPILE_DEFINITIONS})
-
         if (NOT MSVC)
             add_compile_options(-Wno-cast-qual)
         endif()
-
         set(LLAMA_EXTRA_LIBS     ${LLAMA_EXTRA_LIBS}     ${MPI_C_LIBRARIES})
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${MPI_C_INCLUDE_DIRS})
-
         # Even if you're only using the C header, C++ programs may bring in MPI
         # C++ functions, so more linkage is needed
         if (MPI_CXX_FOUND)
@@ -494,28 +427,31 @@ if (LLAMA_VULKAN)
     if (Vulkan_FOUND)
         message(STATUS "Vulkan found")
 
-        set(GGML_HEADERS_VULKAN ggml-vulkan.h)
-        set(GGML_SOURCES_VULKAN ggml-vulkan.cpp)
+        add_library(ggml-vulkan OBJECT ggml-vulkan.cpp ggml-vulkan.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-vulkan PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
 
         add_compile_definitions(GGML_USE_VULKAN)
 
         if (LLAMA_VULKAN_CHECK_RESULTS)
-            add_compile_definitions(GGML_VULKAN_CHECK_RESULTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_CHECK_RESULTS)
         endif()
 
         if (LLAMA_VULKAN_DEBUG)
-            add_compile_definitions(GGML_VULKAN_DEBUG)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_DEBUG)
         endif()
 
         if (LLAMA_VULKAN_VALIDATE)
-            add_compile_definitions(GGML_VULKAN_VALIDATE)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_VALIDATE)
         endif()
 
         if (LLAMA_VULKAN_RUN_TESTS)
-            add_compile_definitions(GGML_VULKAN_RUN_TESTS)
+            target_compile_definitions(ggml-vulkan PRIVATE GGML_VULKAN_RUN_TESTS)
         endif()
 
-        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} Vulkan::Vulkan)
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-vulkan)
     else()
         message(WARNING "Vulkan not found")
     endif()
@@ -527,72 +463,56 @@ if (LLAMA_HIPBLAS)
     if (NOT ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CC=/opt/rocm/llvm/bin/clang")
     endif()
-
     if (NOT ${CMAKE_CXX_COMPILER_ID} MATCHES "Clang")
         message(WARNING "Only LLVM is supported for HIP, hint: CXX=/opt/rocm/llvm/bin/clang++")
     endif()
 
-    find_package(hip     REQUIRED)
-    find_package(hipblas REQUIRED)
-    find_package(rocblas REQUIRED)
+    find_package(hip)
+    find_package(hipblas)
+    find_package(rocblas)
 
-    message(STATUS "HIP and hipBLAS found")
+    if (${hipblas_FOUND} AND ${hip_FOUND})
+        message(STATUS "HIP and hipBLAS found")
+        add_compile_definitions(GGML_USE_HIPBLAS GGML_USE_CUBLAS)
+        if (LLAMA_HIP_UMA)
+            add_compile_definitions(GGML_HIP_UMA)
+        endif()
+        add_library(ggml-rocm OBJECT ggml-cuda.cu ggml-cuda.h)
+        if (BUILD_SHARED_LIBS)
+            set_target_properties(ggml-rocm PROPERTIES POSITION_INDEPENDENT_CODE ON)
+        endif()
+        if (LLAMA_CUDA_FORCE_DMMV)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_DMMV)
+        endif()
+        if (LLAMA_CUDA_FORCE_MMQ)
+            target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_FORCE_MMQ)
+        endif()
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
+        target_compile_definitions(ggml-rocm PRIVATE GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
+        target_compile_definitions(ggml-rocm PRIVATE K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
+        set_source_files_properties(ggml-cuda.cu PROPERTIES LANGUAGE CXX)
+        target_link_libraries(ggml-rocm PRIVATE hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 
-    set(GGML_HEADERS_ROCM ggml-cuda.h)
-
-    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)
+        if (LLAMA_STATIC)
+            message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
+        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} ggml-rocm)
+    else()
+        message(WARNING "hipBLAS or HIP not found. Try setting CMAKE_PREFIX_PATH=/opt/rocm")
     endif()
-
-    if (LLAMA_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
-    if (LLAMA_CUDA_FORCE_MMQ)
-        add_compile_definitions(GGML_CUDA_FORCE_MMQ)
-    endif()
-
-    if (LLAMA_CUDA_NO_PEER_COPY)
-        add_compile_definitions(GGML_CUDA_NO_PEER_COPY)
-    endif()
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${LLAMA_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${LLAMA_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${LLAMA_CUDA_KQUANTS_ITER})
-
-    set_source_files_properties(${GGML_SOURCES_ROCM} PROPERTIES LANGUAGE CXX)
-
-    if (LLAMA_STATIC)
-        message(FATAL_ERROR "Static linking not supported for HIP/ROCm")
-    endif()
-
-    set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} hip::device PUBLIC hip::host roc::rocblas roc::hipblas)
 endif()
 
 if (LLAMA_SYCL)
-    if (NOT LLAMA_SYCL_TARGET MATCHES "^(INTEL|NVIDIA)$")
-        message(FATAL_ERROR "Invalid backend chosen, supported options are INTEL or NVIDIA")
-    endif()
-
     if ( NOT DEFINED ENV{ONEAPI_ROOT})
         message(FATAL_ERROR "Not detect ENV {ONEAPI_ROOT}, please install oneAPI & source it, like: source /opt/intel/oneapi/setvars.sh")
     endif()
     #todo: AOT
 
     find_package(IntelSYCL REQUIRED)
-
-    message(STATUS "SYCL found")
-
-    add_compile_definitions(GGML_USE_SYCL)
-
     if (LLAMA_SYCL_F16)
         add_compile_definitions(GGML_SYCL_F16)
     endif()
+    add_compile_definitions(GGML_USE_SYCL)
 
     add_compile_options(-I./) #include DPCT
     add_compile_options(-I/${SYCL_INCLUDE_DIR})
@@ -600,21 +520,14 @@ if (LLAMA_SYCL)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O3")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl -L${MKLROOT}/lib")
-    if (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-        set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsycl-targets=nvptx64-nvidia-cuda")
-    endif()
 
-    set(GGML_HEADERS_SYCL ggml-sycl.h)
+    set(GGML_HEADERS_SYCL ggml.h ggml-sycl.h)
     set(GGML_SOURCES_SYCL ggml-sycl.cpp)
 
     if (WIN32)
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl sycl7 OpenCL mkl_sycl_blas_dll.lib mkl_intel_ilp64_dll.lib mkl_sequential_dll.lib mkl_core_dll.lib)
     else()
-        if (LLAMA_SYCL_TARGET STREQUAL "INTEL")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
-        elseif (LLAMA_SYCL_TARGET STREQUAL "NVIDIA")
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl pthread m dl onemkl)
-        endif()
+        set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} -fsycl OpenCL mkl_core pthread m dl mkl_sycl_blas mkl_intel_ilp64 mkl_tbb_thread)
     endif()
 endif()
 
@@ -627,61 +540,61 @@ if (LLAMA_KOMPUTE)
     endif()
 
     function(compile_shader)
-        set(options)
-        set(oneValueArgs)
-        set(multiValueArgs SOURCES)
-        cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
-        foreach(source ${compile_shader_SOURCES})
-            get_filename_component(filename ${source} NAME)
-            set(spv_file ${filename}.spv)
-            add_custom_command(
-                OUTPUT ${spv_file}
-                DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
-                ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
-                COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
-                COMMENT "Compiling ${source} to ${spv_file}"
-                )
+      set(options)
+      set(oneValueArgs)
+      set(multiValueArgs SOURCES)
+      cmake_parse_arguments(compile_shader "${options}" "${oneValueArgs}" "${multiValueArgs}" ${ARGN})
+      foreach(source ${compile_shader_SOURCES})
+        get_filename_component(filename ${source} NAME)
+        set(spv_file ${filename}.spv)
+        add_custom_command(
+            OUTPUT ${spv_file}
+            DEPENDS ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/common.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_getrows.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n_pre.comp
+              ${CMAKE_CURRENT_SOURCE_DIR}/kompute-shaders/op_mul_mv_q_n.comp
+              COMMAND ${glslc_executable} --target-env=vulkan1.2 -o ${spv_file} ${CMAKE_CURRENT_SOURCE_DIR}/${source}
+            COMMENT "Compiling ${source} to ${spv_file}"
+        )
 
-            get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
-            set(FILE_NAME "shader${RAW_FILE_NAME}")
-            string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
-            string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
-            string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
-            set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
-            message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
-            if(CMAKE_GENERATOR MATCHES "Visual Studio")
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
-                    )
-            else()
-                add_custom_command(
-                    OUTPUT ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
-                    COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
-                    DEPENDS ${spv_file} xxd
-                    COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
-                    )
-            endif()
-        endforeach()
+        get_filename_component(RAW_FILE_NAME ${spv_file} NAME)
+        set(FILE_NAME "shader${RAW_FILE_NAME}")
+        string(REPLACE ".comp.spv" ".h" HEADER_FILE ${FILE_NAME})
+        string(TOUPPER ${HEADER_FILE} HEADER_FILE_DEFINE)
+        string(REPLACE "." "_" HEADER_FILE_DEFINE "${HEADER_FILE_DEFINE}")
+        set(OUTPUT_HEADER_FILE "${HEADER_FILE}")
+        message(STATUS "${HEADER_FILE} generating ${HEADER_FILE_DEFINE}")
+        if(CMAKE_GENERATOR MATCHES "Visual Studio")
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/$<CONFIG>/xxd"
+            )
+        else()
+            add_custom_command(
+              OUTPUT ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "/*THIS FILE HAS BEEN AUTOMATICALLY GENERATED - DO NOT EDIT*/" > ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#ifndef ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace kp {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "namespace shader_data {" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_BINARY_DIR}/bin/xxd -i ${RAW_FILE_NAME} >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo "}}" >> ${OUTPUT_HEADER_FILE}
+              COMMAND ${CMAKE_COMMAND} -E echo \"\#endif // define ${HEADER_FILE_DEFINE}\" >> ${OUTPUT_HEADER_FILE}
+              DEPENDS ${spv_file} xxd
+              COMMENT "Converting to hpp: ${FILE_NAME} ${CMAKE_BINARY_DIR}/bin/xxd"
+            )
+        endif()
+      endforeach()
     endfunction()
 
     if (EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/kompute/CMakeLists.txt")
@@ -691,66 +604,66 @@ if (LLAMA_KOMPUTE)
 
         # Compile our shaders
         compile_shader(SOURCES
-            kompute-shaders/op_scale.comp
-            kompute-shaders/op_scale_8.comp
-            kompute-shaders/op_add.comp
-            kompute-shaders/op_addrow.comp
-            kompute-shaders/op_mul.comp
-            kompute-shaders/op_silu.comp
-            kompute-shaders/op_relu.comp
-            kompute-shaders/op_gelu.comp
-            kompute-shaders/op_softmax.comp
-            kompute-shaders/op_norm.comp
-            kompute-shaders/op_rmsnorm.comp
-            kompute-shaders/op_diagmask.comp
-            kompute-shaders/op_mul_mat_mat_f32.comp
-            kompute-shaders/op_mul_mat_f16.comp
-            kompute-shaders/op_mul_mat_q8_0.comp
-            kompute-shaders/op_mul_mat_q4_0.comp
-            kompute-shaders/op_mul_mat_q4_1.comp
-            kompute-shaders/op_mul_mat_q6_k.comp
-            kompute-shaders/op_getrows_f16.comp
-            kompute-shaders/op_getrows_q4_0.comp
-            kompute-shaders/op_getrows_q4_1.comp
-            kompute-shaders/op_getrows_q6_k.comp
-            kompute-shaders/op_rope_f16.comp
-            kompute-shaders/op_rope_f32.comp
-            kompute-shaders/op_cpy_f16_f16.comp
-            kompute-shaders/op_cpy_f16_f32.comp
-            kompute-shaders/op_cpy_f32_f16.comp
-            kompute-shaders/op_cpy_f32_f32.comp
+          kompute-shaders/op_scale.comp
+          kompute-shaders/op_scale_8.comp
+          kompute-shaders/op_add.comp
+          kompute-shaders/op_addrow.comp
+          kompute-shaders/op_mul.comp
+          kompute-shaders/op_silu.comp
+          kompute-shaders/op_relu.comp
+          kompute-shaders/op_gelu.comp
+          kompute-shaders/op_softmax.comp
+          kompute-shaders/op_norm.comp
+          kompute-shaders/op_rmsnorm.comp
+          kompute-shaders/op_diagmask.comp
+          kompute-shaders/op_mul_mat_mat_f32.comp
+          kompute-shaders/op_mul_mat_f16.comp
+          kompute-shaders/op_mul_mat_q8_0.comp
+          kompute-shaders/op_mul_mat_q4_0.comp
+          kompute-shaders/op_mul_mat_q4_1.comp
+          kompute-shaders/op_mul_mat_q6_k.comp
+          kompute-shaders/op_getrows_f16.comp
+          kompute-shaders/op_getrows_q4_0.comp
+          kompute-shaders/op_getrows_q4_1.comp
+          kompute-shaders/op_getrows_q6_k.comp
+          kompute-shaders/op_rope_f16.comp
+          kompute-shaders/op_rope_f32.comp
+          kompute-shaders/op_cpy_f16_f16.comp
+          kompute-shaders/op_cpy_f16_f32.comp
+          kompute-shaders/op_cpy_f32_f16.comp
+          kompute-shaders/op_cpy_f32_f32.comp
         )
 
         # Create a custom target for our generated shaders
         add_custom_target(generated_shaders DEPENDS
-            shaderop_scale.h
-            shaderop_scale_8.h
-            shaderop_add.h
-            shaderop_addrow.h
-            shaderop_mul.h
-            shaderop_silu.h
-            shaderop_relu.h
-            shaderop_gelu.h
-            shaderop_softmax.h
-            shaderop_norm.h
-            shaderop_rmsnorm.h
-            shaderop_diagmask.h
-            shaderop_mul_mat_mat_f32.h
-            shaderop_mul_mat_f16.h
-            shaderop_mul_mat_q8_0.h
-            shaderop_mul_mat_q4_0.h
-            shaderop_mul_mat_q4_1.h
-            shaderop_mul_mat_q6_k.h
-            shaderop_getrows_f16.h
-            shaderop_getrows_q4_0.h
-            shaderop_getrows_q4_1.h
-            shaderop_getrows_q6_k.h
-            shaderop_rope_f16.h
-            shaderop_rope_f32.h
-            shaderop_cpy_f16_f16.h
-            shaderop_cpy_f16_f32.h
-            shaderop_cpy_f32_f16.h
-            shaderop_cpy_f32_f32.h
+          shaderop_scale.h
+          shaderop_scale_8.h
+          shaderop_add.h
+          shaderop_addrow.h
+          shaderop_mul.h
+          shaderop_silu.h
+          shaderop_relu.h
+          shaderop_gelu.h
+          shaderop_softmax.h
+          shaderop_norm.h
+          shaderop_rmsnorm.h
+          shaderop_diagmask.h
+          shaderop_mul_mat_mat_f32.h
+          shaderop_mul_mat_f16.h
+          shaderop_mul_mat_q8_0.h
+          shaderop_mul_mat_q4_0.h
+          shaderop_mul_mat_q4_1.h
+          shaderop_mul_mat_q6_k.h
+          shaderop_getrows_f16.h
+          shaderop_getrows_q4_0.h
+          shaderop_getrows_q4_1.h
+          shaderop_getrows_q6_k.h
+          shaderop_rope_f16.h
+          shaderop_rope_f32.h
+          shaderop_cpy_f16_f16.h
+          shaderop_cpy_f16_f32.h
+          shaderop_cpy_f32_f16.h
+          shaderop_cpy_f32_f32.h
         )
 
         # Create a custom command that depends on the generated_shaders
@@ -763,10 +676,8 @@ if (LLAMA_KOMPUTE)
 
         # Add the stamp to the main sources to ensure dependency tracking
         set(GGML_SOURCES_KOMPUTE ggml-kompute.cpp ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-        set(GGML_HEADERS_KOMPUTE ggml-kompute.h   ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
-
+        set(GGML_HEADERS_KOMPUTE ggml-kompute.h ${CMAKE_CURRENT_BINARY_DIR}/ggml-kompute.stamp)
         add_compile_definitions(GGML_USE_KOMPUTE)
-
         set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} kompute)
         set(LLAMA_EXTRA_INCLUDES ${LLAMA_EXTRA_INCLUDES} ${CMAKE_BINARY_DIR})
     else()
@@ -774,18 +685,6 @@ if (LLAMA_KOMPUTE)
     endif()
 endif()
 
-if (LLAMA_CPU_HBM)
-    find_library(memkind memkind REQUIRED)
-
-    add_compile_definitions(GGML_USE_CPU_HBM)
-
-    target_link_libraries(ggml PUBLIC memkind)
-endif()
-
-if (LLAMA_PERF)
-    add_compile_definitions(GGML_PERF)
-endif()
-
 function(get_flags CCID CCVER)
     set(C_FLAGS "")
     set(CXX_FLAGS "")
@@ -810,30 +709,28 @@ function(get_flags CCID CCVER)
         if (CCVER VERSION_GREATER_EQUAL 8.1.0)
             list(APPEND CXX_FLAGS -Wextra-semi)
         endif()
+    elseif (CCID MATCHES "Intel")
+        if (NOT LLAMA_SYCL)
+            # enable max optimization level when using Intel compiler
+            set(C_FLAGS   -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            set(CXX_FLAGS -ipo -O3 -static -fp-model=fast -flto -fno-stack-protector)
+            add_link_options(-fuse-ld=lld -static-intel)
+        endif()
     endif()
 
     set(GF_C_FLAGS   ${C_FLAGS}   PARENT_SCOPE)
     set(GF_CXX_FLAGS ${CXX_FLAGS} PARENT_SCOPE)
 endfunction()
 
-if (LLAMA_FATAL_WARNINGS)
-    if (CMAKE_CXX_COMPILER_ID MATCHES "GNU" OR CMAKE_CXX_COMPILER_ID MATCHES "Clang")
-        list(APPEND C_FLAGS   -Werror)
-        list(APPEND CXX_FLAGS -Werror)
-    elseif (CMAKE_CXX_COMPILER_ID STREQUAL "MSVC")
-        add_compile_options(/WX)
-    endif()
-endif()
-
 if (LLAMA_ALL_WARNINGS)
     if (NOT MSVC)
-        list(APPEND WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
-        list(APPEND C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
-                                  -Werror=implicit-int -Werror=implicit-function-declaration)
-        list(APPEND CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
+        set(WARNING_FLAGS -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function)
+        set(C_FLAGS       -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmissing-prototypes
+                          -Werror=implicit-int -Werror=implicit-function-declaration)
+        set(CXX_FLAGS     -Wmissing-declarations -Wmissing-noreturn)
 
-        list(APPEND C_FLAGS   ${WARNING_FLAGS})
-        list(APPEND CXX_FLAGS ${WARNING_FLAGS})
+        set(C_FLAGS   ${WARNING_FLAGS} ${C_FLAGS})
+        set(CXX_FLAGS ${WARNING_FLAGS} ${CXX_FLAGS})
 
         get_flags(${CMAKE_CXX_COMPILER_ID} ${CMAKE_CXX_COMPILER_VERSION})
 
@@ -848,11 +745,10 @@ endif()
 
 set(CUDA_CXX_FLAGS "")
 
-if (LLAMA_CUDA)
-    set(CUDA_FLAGS -use_fast_math)
-
-    if (LLAMA_FATAL_WARNINGS)
-        list(APPEND CUDA_FLAGS -Werror all-warnings)
+if (LLAMA_CUBLAS)
+    set(CUDA_FLAGS ${CXX_FLAGS} -use_fast_math)
+    if (NOT MSVC)
+        list(APPEND CUDA_FLAGS -Wno-pedantic)
     endif()
 
     if (LLAMA_ALL_WARNINGS AND NOT MSVC)
@@ -886,11 +782,7 @@ if (LLAMA_CUDA)
         message("-- CUDA host compiler is ${CUDA_CCID} ${CUDA_CCVER}")
 
         get_flags(${CUDA_CCID} ${CUDA_CCVER})
-        list(APPEND CUDA_CXX_FLAGS ${CXX_FLAGS} ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
-    endif()
-
-    if (NOT MSVC)
-        list(APPEND CUDA_CXX_FLAGS -Wno-pedantic)
+        list(APPEND CUDA_CXX_FLAGS ${GF_CXX_FLAGS})  # This is passed to -Xcompiler later
     endif()
 endif()
 
@@ -929,7 +821,6 @@ execute_process(
     ERROR_VARIABLE output
     OUTPUT_QUIET
 )
-
 if (output MATCHES "dyld-1015\.7")
     add_compile_definitions(HAVE_BUGGY_APPLE_LINKER)
 endif()
@@ -939,10 +830,10 @@ endif()
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 message(STATUS "CMAKE_SYSTEM_PROCESSOR: ${CMAKE_SYSTEM_PROCESSOR}")
 if (MSVC)
-    string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
-    message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
+  string(TOLOWER "${CMAKE_GENERATOR_PLATFORM}" CMAKE_GENERATOR_PLATFORM_LWR)
+  message(STATUS "CMAKE_GENERATOR_PLATFORM: ${CMAKE_GENERATOR_PLATFORM}")
 else ()
-    set(CMAKE_GENERATOR_PLATFORM_LWR "")
+  set(CMAKE_GENERATOR_PLATFORM_LWR "")
 endif ()
 
 if (NOT MSVC)
@@ -964,21 +855,11 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
      CMAKE_SYSTEM_PROCESSOR MATCHES "^(aarch64|arm.*|ARM64)$"))
     message(STATUS "ARM detected")
     if (MSVC)
-        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
         add_compile_definitions(__ARM_NEON)
         add_compile_definitions(__ARM_FEATURE_FMA)
-
-        set(CMAKE_REQUIRED_FLAGS_PREV ${CMAKE_REQUIRED_FLAGS})
-        string(JOIN " " CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS} "/arch:armv8.2")
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { int8x16_t _a, _b; int32x4_t _s = vdotq_s32(_s, _a, _b); return 0; }" GGML_COMPILER_SUPPORT_DOTPROD)
-        if (GGML_COMPILER_SUPPORT_DOTPROD)
-            add_compile_definitions(__ARM_FEATURE_DOTPROD)
-        endif ()
-        check_cxx_source_compiles("#include <arm_neon.h>\nint main() { float16_t _a; float16x8_t _s = vdupq_n_f16(_a); return 0; }" GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-        if (GGML_COMPILER_SUPPORT_FP16_VECTOR_ARITHMETIC)
-            add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
-        endif ()
-        set(CMAKE_REQUIRED_FLAGS ${CMAKE_REQUIRED_FLAGS_PREV})
+        add_compile_definitions(__ARM_FEATURE_DOTPROD)
+        # add_compile_definitions(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) # MSVC doesn't support vdupq_n_f16, vld1q_f16, vst1q_f16
+        add_compile_definitions(__aarch64__) # MSVC defines _M_ARM64 instead
     else()
         check_cxx_compiler_flag(-mfp16-format=ieee COMPILER_SUPPORTS_FP16_FORMAT_I3E)
         if (NOT "${COMPILER_SUPPORTS_FP16_FORMAT_I3E}" STREQUAL "")
@@ -989,16 +870,10 @@ if (CMAKE_OSX_ARCHITECTURES STREQUAL "arm64" OR CMAKE_GENERATOR_PLATFORM_LWR STR
             list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
-            if ("${CMAKE_SYSTEM_NAME}" STREQUAL "Android")
-                # Android armeabi-v7a
-                list(APPEND ARCH_FLAGS -mfpu=neon-vfpv4 -mno-unaligned-access -funsafe-math-optimizations)
-            else()
-                # Raspberry Pi 2
-                list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
-            endif()
+            # Raspberry Pi 2
+            list(APPEND ARCH_FLAGS -mfpu=neon-fp-armv8 -mno-unaligned-access -funsafe-math-optimizations)
         endif()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
-            # Android arm64-v8a
             # Raspberry Pi 3, 4, Zero 2 (32-bit)
             list(APPEND ARCH_FLAGS -mno-unaligned-access)
         endif()
@@ -1073,7 +948,7 @@ endif()
 add_compile_options("$<$<COMPILE_LANGUAGE:CXX>:${ARCH_FLAGS}>")
 add_compile_options("$<$<COMPILE_LANGUAGE:C>:${ARCH_FLAGS}>")
 
-if (LLAMA_CUDA)
+if (LLAMA_CUBLAS)
     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 "")
@@ -1142,6 +1017,11 @@ endif()
 
 # ggml
 
+if (GGML_USE_CPU_HBM)
+    add_definitions(-DGGML_USE_CPU_HBM)
+    find_library(memkind memkind REQUIRED)
+endif()
+
 add_library(ggml OBJECT
             ggml.c
             ggml.h
@@ -1158,17 +1038,16 @@ add_library(ggml OBJECT
             ${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}
             )
 
 target_include_directories(ggml PUBLIC . ${LLAMA_EXTRA_INCLUDES})
-target_compile_features   (ggml PUBLIC c_std_11) # don't bump
-
+target_compile_features(ggml PUBLIC c_std_11) # don't bump
 target_link_libraries(ggml PUBLIC Threads::Threads ${LLAMA_EXTRA_LIBS})
+if (GGML_USE_CPU_HBM)
+    target_link_libraries(ggml PUBLIC memkind)
+endif()
 
 add_library(ggml_static STATIC $<TARGET_OBJECTS:ggml>)
-
 if (BUILD_SHARED_LIBS)
     set_target_properties(ggml PROPERTIES POSITION_INDEPENDENT_CODE ON)
     add_library(ggml_shared SHARED $<TARGET_OBJECTS:ggml>)
@@ -1181,14 +1060,10 @@ endif()
 add_library(llama
             llama.cpp
             llama.h
-            unicode.h
-            unicode.cpp
-            unicode-data.cpp
             )
 
 target_include_directories(llama PUBLIC .)
-target_compile_features   (llama PUBLIC cxx_std_11) # don't bump
-
+target_compile_features(llama PUBLIC cxx_std_11) # don't bump
 target_link_libraries(llama PRIVATE
     ggml
     ${LLAMA_EXTRA_LIBS}
@@ -1239,7 +1114,7 @@ install(FILES ${CMAKE_CURRENT_BINARY_DIR}/LlamaConfig.cmake
         DESTINATION ${CMAKE_INSTALL_LIBDIR}/cmake/Llama)
 
 set(GGML_PUBLIC_HEADERS "ggml.h" "ggml-alloc.h" "ggml-backend.h"
-        "${GGML_HEADERS_CUDA}"  "${GGML_HEADERS_OPENCL}"
+        "${GGML_HEADERS_CUDA}" "${GGML_HEADERS_OPENCL}"
         "${GGML_HEADERS_METAL}" "${GGML_HEADERS_MPI}" "${GGML_HEADERS_EXTRA}")
 
 set_target_properties(ggml PROPERTIES PUBLIC_HEADER "${GGML_PUBLIC_HEADERS}")
@@ -1279,12 +1154,6 @@ 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()
 
 #

Makefile

@@ -1,16 +1,15 @@
 # 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  \
-	retrieval 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 llama-bench libllava.a llava-cli baby-llama beam-search  \
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey 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-backend-ops tests/test-model-load-cancel tests/test-autorelease
 
 # Code coverage output files
 COV_TARGETS = *.gcno tests/*.gcno *.gcda tests/*.gcda *.gcov tests/*.gcov lcov-report gcovr-report
@@ -98,10 +97,9 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS  = -I. -Icommon
-MK_CFLAGS    = -std=c11   -fPIC
-MK_CXXFLAGS  = -std=c++11 -fPIC
-MK_NVCCFLAGS = -std=c++11
+MK_CPPFLAGS = -I. -Icommon
+MK_CFLAGS   = -std=c11   -fPIC
+MK_CXXFLAGS = -std=c++11 -fPIC
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -168,17 +166,13 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
-ifdef LLAMA_SCHED_MAX_COPIES
-	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
-endif
-
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
+		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -206,10 +200,6 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
-ifdef LLAMA_SERVER_SSL
-	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
-	MK_LDFLAGS += -lssl -lcrypto
-endif
 
 ifdef LLAMA_CODE_COVERAGE
 	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
@@ -225,11 +215,6 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
-ifeq ($(LLAMA_FATAL_WARNINGS),1)
-	MK_CFLAGS   += -Werror
-	MK_CXXFLAGS += -Werror
-endif
-
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -390,24 +375,10 @@ ifdef LLAMA_BLIS
 endif # LLAMA_BLIS
 
 ifdef LLAMA_CUBLAS
-# LLAMA_CUBLAS is deprecated and will be removed in the future
-	LLAMA_CUDA := 1
-endif
-
-ifdef LLAMA_CUDA
-	ifneq ('', '$(wildcard /opt/cuda)')
-		CUDA_PATH ?= /opt/cuda
-	else
-		CUDA_PATH ?= /usr/local/cuda
-	endif
-	MK_CPPFLAGS  += -DGGML_USE_CUDA -I$(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/$(UNAME_M)-linux/lib -L/usr/lib/wsl/lib
+	MK_CPPFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I$(CUDA_PATH)/targets/x86_64-linux/include -I/usr/local/cuda/targets/aarch64-linux/include
+	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-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
-endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -458,30 +429,19 @@ ifdef LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 else
 	MK_NVCCFLAGS += -DGGML_CUDA_PEER_MAX_BATCH_SIZE=128
 endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
-ifdef LLAMA_CUDA_NO_PEER_COPY
-	MK_NVCCFLAGS += -DGGML_CUDA_NO_PEER_COPY
-endif # LLAMA_CUDA_NO_PEER_COPY
+#ifdef LLAMA_CUDA_CUBLAS
+#	MK_NVCCFLAGS += -DGGML_CUDA_CUBLAS
+#endif # LLAMA_CUDA_CUBLAS
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 ifdef JETSON_EOL_MODULE_DETECT
-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
+	$(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 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-define NVCC_COMPILE
-	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
-endef # NVCC_COMPILE
+	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
-
-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
+endif # LLAMA_CUBLAS
 
 ifdef LLAMA_CLBLAST
 
@@ -527,6 +487,7 @@ 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))
@@ -538,7 +499,7 @@ 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_CUDA
+	MK_CPPFLAGS += -DGGML_USE_HIPBLAS -DGGML_USE_CUBLAS
 ifdef LLAMA_HIP_UMA
 	MK_CPPFLAGS += -DGGML_HIP_UMA
 endif # LLAMA_HIP_UMA
@@ -551,18 +512,9 @@ endif # LLAMA_HIP_UMA
 ifdef LLAMA_CUDA_FORCE_DMMV
 	HIPFLAGS 	+= -DGGML_CUDA_FORCE_DMMV
 endif # LLAMA_CUDA_FORCE_DMMV
-ifdef LLAMA_CUDA_NO_PEER_COPY
-	HIPFLAGS 	+= -DGGML_CUDA_NO_PEER_COPY
-endif # LLAMA_CUDA_NO_PEER_COPY
 	OBJS        += ggml-cuda.o
-	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)
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	$(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
@@ -572,30 +524,11 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
-ifdef LLAMA_METAL_EMBED_LIBRARY
-	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
-	OBJS        += ggml-metal-embed.o
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
-ggml-metal.o: ggml-metal.m ggml-metal.h ggml.h
+ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
-
-ifdef LLAMA_METAL_EMBED_LIBRARY
-ggml-metal-embed.o: ggml-metal.metal ggml-common.h
-	@echo "Embedding Metal library"
-	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
-	$(eval TEMP_ASSEMBLY=$(shell mktemp))
-	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
-	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
-	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
-	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
-	@$(AS) $(TEMP_ASSEMBLY) -o $@
-	@rm -f ${TEMP_ASSEMBLY}
-endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -607,23 +540,17 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
-override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
 override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
 # identify CUDA host compiler
-ifdef LLAMA_CUDA
+ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
-endif
-
-ifdef LLAMA_CURL
-override CXXFLAGS := $(CXXFLAGS) -DLLAMA_USE_CURL
-override LDFLAGS  := $(LDFLAGS) -lcurl
+CUDA_CXXFLAGS := $(GF_CXXFLAGS)
 endif
 
 #
@@ -640,25 +567,10 @@ $(info I NVCCFLAGS: $(NVCCFLAGS))
 $(info I LDFLAGS:   $(LDFLAGS))
 $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
-ifdef LLAMA_CUDA
-$(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)
-endif # CUDA_POWER_ARCH
-endif # CUDA_DOCKER_ARCH
-endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
-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 I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+endif # LLAMA_CUBLAS
 $(info )
-endif
 
 #
 # Build library
@@ -673,18 +585,12 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-unicode.o: unicode.cpp unicode.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.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
+llama.o: llama.cpp 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
@@ -702,15 +608,9 @@ console.o: common/console.cpp common/console.h
 grammar-parser.o: common/grammar-parser.cpp common/grammar-parser.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-json-schema-to-grammar.o: common/json-schema-to-grammar.cpp common/json-schema-to-grammar.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 train.o: common/train.cpp common/train.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-ngram-cache.o: common/ngram-cache.cpp common/ngram-cache.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)
 
@@ -718,8 +618,7 @@ libllama.a: llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 	ar rcs libllama.a llama.o ggml.o $(OBJS) $(COMMON_DEPS)
 
 clean:
-	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult lookup-create lookup-merge lookup-stats common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
-	rm -vrf ggml-cuda/*.o
+	rm -vrf *.o tests/*.o *.so *.a *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
 	find examples pocs -type f -name "*.o" -delete
 
 #
@@ -780,26 +679,19 @@ embedding: examples/embedding/embedding.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)
 
-gritlm: examples/gritlm/gritlm.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)
-
 save-load-state: examples/save-load-state/save-load-state.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)
 
-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/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h 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)
+	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
+	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
 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)
 
-gguf-split: examples/gguf-split/gguf-split.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)
@@ -837,10 +729,6 @@ 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)
@@ -853,15 +741,9 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-lookup: examples/lookup/lookup.cpp ggml.o llama.o ngram-cache.o $(COMMON_DEPS) $(OBJS)
+lookup: examples/lookup/lookup.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)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-create.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-create.cpp) -o lookup-create $(LDFLAGS)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-merge.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-merge.cpp) -o lookup-merge $(LDFLAGS)
-	$(CXX) $(CXXFLAGS) -c examples/lookup/lookup-stats.cpp -o $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, examples/lookup/lookup-stats.cpp) -o lookup-stats $(LDFLAGS)
 
 passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -918,10 +800,6 @@ tests/test-double-float: tests/test-double-float.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-tests/test-json-schema-to-grammar: tests/test-json-schema-to-grammar.cpp json-schema-to-grammar.o ggml.o llama.o grammar-parser.o $(OBJS)
-	$(CXX) $(CXXFLAGS) -Iexamples/server -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 tests/test-grad0: tests/test-grad0.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
@@ -976,7 +854,3 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
-tests/test-chat-template: tests/test-chat-template.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)

Package.swift

@@ -31,8 +31,6 @@ let package = Package(
             sources: [
                 "ggml.c",
                 "llama.cpp",
-                "unicode.cpp",
-                "unicode-data.cpp",
                 "ggml-alloc.c",
                 "ggml-backend.c",
                 "ggml-quants.c",

README-sycl.md

@@ -1,9 +1,8 @@
 # llama.cpp for SYCL
 
 - [Background](#background)
-- [News](#news)
 - [OS](#os)
-- [Supported Devices](#supported-devices)
+- [Intel GPU](#intel-gpu)
 - [Docker](#docker)
 - [Linux](#linux)
 - [Windows](#windows)
@@ -14,42 +13,17 @@
 
 ## Background
 
-**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.
+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.
 
-**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:
+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.
 
-- **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.
+Intel uses the SYCL as direct programming language to support CPU, GPUs and FPGAs.
 
-### Llama.cpp + SYCL
-This SYCL "backend" follows the same design found in other llama.cpp BLAS-based paths such as *OpenBLAS, cuBLAS, CLBlast etc..*. 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.
+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.
 
-The llama.cpp SYCL backend supports:
-- Intel GPUs.
-- Nvidia GPUs.
+The llama.cpp for SYCL is used to support Intel GPUs.
 
-*Upcoming support: AMD GPUs*.
-
-When targetting **Intel CPUs**, it is recommended to  use llama.cpp for [x86_64](README.md#intel-onemkl) approach.
-
-## News
-
-- 2024.3
-  - A blog is published: **Run LLM on all Intel GPUs Using llama.cpp**: [intel.com](https://www.intel.com/content/www/us/en/developer/articles/technical/run-llm-on-all-gpus-using-llama-cpp-artical.html) or [medium.com](https://medium.com/@jianyu_neo/run-llm-on-all-intel-gpus-using-llama-cpp-fd2e2dcbd9bd).
-  - New base line is ready: [tag b2437](https://github.com/ggerganov/llama.cpp/tree/b2437).
-  - Support multiple cards: **--split-mode**: [none|layer]; not support [row], it's on developing.
-  - Support to assign main GPU by **--main-gpu**, replace $GGML_SYCL_DEVICE.
-  - Support detecting all GPUs with level-zero and same top **Max compute units**.
-  - Support OPs
-    - hardsigmoid
-    - hardswish
-    - pool2d
-
-- 2024.1
-  - Create SYCL backend for Intel GPU.
-  - Support Windows build
+For Intel CPU, recommend to use llama.cpp for X86 (Intel MKL building).
 
 ## OS
 
@@ -59,16 +33,9 @@ When targetting **Intel CPUs**, it is recommended to  use llama.cpp for [x86_64]
 |Windows|Support|Windows 11|
 
 
-## Supported devices
+## Intel GPU
 
-### Intel GPUs
-
-The oneAPI Math Kernel Library, which the oneAPI base-toolkit includes, supports intel GPUs. In order to make it "visible", simply run the following:
-```sh
-source /opt/intel/oneapi/setvars.sh
-```
-
-- **Tested devices**
+### Verified
 
 |Intel GPU| Status | Verified Model|
 |-|-|-|
@@ -78,335 +45,259 @@ source /opt/intel/oneapi/setvars.sh
 |Intel built-in Arc GPU| Support| built-in Arc GPU in Meteor Lake|
 |Intel iGPU| Support| iGPU in i5-1250P, i7-1260P, i7-1165G7|
 
-*Notes:*
+Note: If the EUs (Execution Unit) in iGPU is less than 80, the inference speed will be too slow to use.
 
-- Device memory can be a limitation when running a large model on an intel GPU. The loaded model size, *`llm_load_tensors: buffer_size`*, is displayed in the log when running `./bin/main`.
+### Memory
 
-- 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 GPUs and 4.0GB for discrete GPUs.
+The memory is a limitation to run LLM on GPUs.
 
-- If the iGPU has less than 80  EUs *(Execution Unit)*, the inference speed will likely be too slow for practical use.
+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`.
 
-### Nvidia GPUs
-The BLAS acceleration on Nvidia GPUs through oneAPI can be obtained using the Nvidia plugins for oneAPI and the cuBLAS backend of the upstream oneMKL library. Details and instructions on how to setup the runtime and library can be found in [this section](#i-setup-environment)
-
-- **Tested devices**
-
-|Nvidia GPU| Status | Verified Model|
-|-|-|-|
-|Ampere Series| Support| A100, A4000|
-|Ampere Series *(Mobile)*| Support| RTX 40 Series|
-
-*Notes:*
-  - Support for Nvidia targets through oneAPI is currently limited to Linux platforms.
-
-  - Please make sure the native oneAPI MKL *(dedicated to intel CPUs and GPUs)* is not "visible" at this stage to properly setup and use the built-from-source oneMKL with cuBLAS backend in llama.cpp for Nvidia GPUs.
+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+.
 
+For dGPU, please make sure the device memory is enough. For llama-2-7b.Q4_0, recommend the device memory is 4GB+.
 
 ## Docker
-The docker build option is currently limited to *intel GPU* targets.
-### Build image
+
+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.
+
+
 ```sh
-# Using FP16
-docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
+# For F16:
+#docker build -t llama-cpp-sycl --build-arg="LLAMA_SYCL_F16=ON" -f .devops/main-intel.Dockerfile .
+
+# Or, for F32:
+docker build -t llama-cpp-sycl -f .devops/main-intel.Dockerfile .
+
+# Note: you can also use the ".devops/main-server.Dockerfile", which compiles the "server" example
 ```
 
-*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
+### Run
 
 ```sh
-# First, find all the DRI cards
+# Firstly, find all the DRI cards:
 ls -la /dev/dri
-# Then, pick the card that you want to use (here for e.g. /dev/dri/card1).
+# Then, pick the card that you want to use.
+
+# For example with "/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
 
-### I. Setup Environment
+### Setup Environment
 
-1. **Install GPU drivers**
+1. Install Intel GPU driver.
 
-  - **Intel GPU**
+a. Please install Intel GPU driver by official guide: [Install GPU Drivers](https://dgpu-docs.intel.com/driver/installation.html).
 
-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).
+Note: for iGPU, please install the client GPU driver.
 
-*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.
+b. Add user to group: video, render.
 
 ```sh
-sudo usermod -aG render $USER
-sudo usermod -aG video $USER
+sudo usermod -aG render username
+sudo usermod -aG video username
 ```
 
-*Note*: logout/re-login for the changes to take effect.
+Note: re-login to enable it.
 
-Verify installation through `clinfo`:
+c. Check
 
 ```sh
 sudo apt install clinfo
 sudo clinfo -l
 ```
 
-Sample output:
+Output (example):
 
-```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]
 ```
 
-- **Nvidia GPU**
+2. Install Intel® oneAPI Base toolkit.
 
-In order to target Nvidia GPUs through SYCL, please make sure the CUDA/CUBLAS native requirements *-found [here](README.md#cublas)-* are installed.
-Installation can be verified by running the following:
-```sh
-nvidia-smi
-```
-Please make sure at least one CUDA device is available, which can be displayed like this *(here an A100-40GB Nvidia GPU)*:
-```
-+---------------------------------------------------------------------------------------+
-| NVIDIA-SMI 535.54.03              Driver Version: 535.54.03    CUDA Version: 12.2     |
-|-----------------------------------------+----------------------+----------------------+
-| GPU  Name                 Persistence-M | Bus-Id        Disp.A | Volatile Uncorr. ECC |
-| Fan  Temp   Perf          Pwr:Usage/Cap |         Memory-Usage | GPU-Util  Compute M. |
-|                                         |                      |               MIG M. |
-|=========================================+======================+======================|
-|   0  NVIDIA A100-PCIE-40GB          On  | 00000000:8D:00.0 Off |                    0 |
-| N/A   36C    P0              57W / 250W |      4MiB / 40960MiB |      0%      Default |
-|                                         |                      |             Disabled |
-+-----------------------------------------+----------------------+----------------------+
-```
+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).
 
+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.
 
-- **Base installation**
+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**: 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**: 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
-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
 ```
 
-- **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:
+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]**.
 
+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]
+
 ```
 
-- **Nvidia GPU**
+2. Build locally:
 
-Similarly, user targetting 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]
-```
+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.
 
-### II. Build llama.cpp
-
-#### Intel GPU
 ```sh
-# Export relevant ENV variables
+mkdir -p build
+cd build
 source /opt/intel/oneapi/setvars.sh
 
-# Build LLAMA with MKL BLAS acceleration for intel GPU
-mkdir -p build && cd build
+# For FP16:
+#cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
 
-# Option 1: Use FP16 for better performance in long-prompt  inference
-cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
+# Or, for FP32:
 cmake .. -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Build example/main only
+#cmake --build . --config Release --target main
+
+# Or, build all binary
+cmake --build . --config Release -v
+
+cd ..
 ```
 
-#### Nvidia GPU
+or
+
 ```sh
-# Export relevant ENV variables
-export LD_LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LD_LIBRARY_PATH
-export LIBRARY_PATH=/path/to/oneMKL/buildWithCublas/lib:$LIBRARY_PATH
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/buildWithCublas/include:$CPLUS_INCLUDE_DIR
-export CPLUS_INCLUDE_DIR=/path/to/oneMKL/include:$CPLUS_INCLUDE_DIR
-
-# Build LLAMA with Nvidia BLAS acceleration through SYCL
-mkdir -p build && cd build
-
-# Option 1: Use FP16 for better performance in long-prompt  inference
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL_F16=ON
-
-# Option 2: Use FP32 by default
-cmake .. -DLLAMA_SYCL=ON -DLLAMA_SYCL_TARGET=NVIDIA -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+./examples/sycl/build.sh
 ```
 
-### III. Run the inference
+### Run
 
-1. Retrieve and prepare model
+1. Put model file to folder **models**
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You 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.
 
 2. Enable oneAPI running environment
 
-```sh
+```
 source /opt/intel/oneapi/setvars.sh
 ```
 
-3. List devices information
+3. List device ID
 
-Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
+Run without parameter:
 
 ```sh
 ./build/bin/ls-sycl-device
+
+# or running the "main" executable and look at the output log:
+
+./build/bin/main
 ```
-A example of such log in a system with 1 *intel CPU* and 1 *intel GPU* can look like the following:
+
+Check the ID in startup log, like:
+
 ```
-found 6 SYCL devices:
-|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
-|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
-|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
-| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
-| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
-| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
-| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
-| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
-| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+
 ```
 
 |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|
+|compute capability 1.3|Level-zero running time, recommended |
+|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
 
-4. Launch inference
+4. Set device ID and execute llama.cpp
 
-There are two device selection modes:
-
-- 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)|
-
-Examples:
-
-- Use device 0:
+Set device ID = 0 by **GGML_SYCL_DEVICE=0**
 
 ```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 none -mg 0
+GGML_SYCL_DEVICE=0 ./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
 ```
 or run by script:
 
-```sh
-./examples/sycl/run_llama2.sh 0
-```
-
-- Use multiple devices:
-
-```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
-```
-
-Otherwise, you can run the script:
-
 ```sh
 ./examples/sycl/run_llama2.sh
 ```
 
-*Notes:*
+Note:
 
-- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `/bin/main` if faced with the issue.
-- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
+- 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.
 
-```sh
-detect 1 SYCL GPUs: [0] with top Max compute units:512
+
+5. Check the device ID in output
+
+Like:
 ```
-Or
-```sh
-use 1 SYCL GPUs: [0] with Max compute units:512
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 ```
 
 ## Windows
 
-### I. Setup Environment
+### Setup Environment
 
-1. Install GPU driver
+1. Install Intel GPU driver.
 
-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).
+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).
 
-2. Install Visual Studio
+Note: **The driver is mandatory for compute function**.
 
-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/).
+2. Install Visual Studio.
 
-3. Install Intel® oneAPI Base toolkit
+Please install [Visual Studio](https://visualstudio.microsoft.com/) which impact oneAPI environment enabling in Windows.
 
-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.
+3. Install Intel® oneAPI Base toolkit.
 
-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)*.
+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).
 
-Following guidelines/code snippets assume the default installation values. Otherwise, please make sure the necessary changes are reflected where applicable.
+Recommend to install to default folder: **/opt/intel/oneapi**.
+
+Following guide uses the default folder as example. If you use other folder, please modify the following guide info with your folder.
 
 b. Enable oneAPI running environment:
 
-- Type "oneAPI" in the search bar, then open the `Intel oneAPI command prompt for Intel 64 for Visual Studio 2022` App.
+- In Search, input 'oneAPI'.
 
-- On the command prompt, enable the runtime environment with the following:
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-c. Verify installation
+c. Check GPU
 
-In the oneAPI command line, run the following to print the available SYCL devices:
+In oneAPI command line:
 
 ```
 sycl-ls
 ```
 
-There should be one or more *level-zero* GPU devices displayed as **[ext_oneapi_level_zero:gpu]**. Below is example of such output detecting an *intel Iris Xe* GPU as a Level-zero SYCL device:
+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]**.
 
 Output (example):
 ```
@@ -416,74 +307,96 @@ 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 build tools
+4. Install cmake & make
 
 a. Download & install cmake for Windows: https://cmake.org/download/
 
 b. Download & install mingw-w64 make for Windows provided by w64devkit
 
-- Download the 1.19.0 version of [w64devkit](https://github.com/skeeto/w64devkit/releases/download/v1.19.0/w64devkit-1.19.0.zip).
+- Download the latest fortran version of [w64devkit](https://github.com/skeeto/w64devkit/releases).
 
 - Extract `w64devkit` on your pc.
 
-- Add the **bin** folder path in the Windows system PATH environment (for e.g. `C:\xxx\w64devkit\bin\`).
+- Add the **bin** folder path in the Windows system PATH environment, like `C:\xxx\w64devkit\bin\`.
 
-### II. Build llama.cpp
+### Build locally:
 
-On the oneAPI command line window, step into the llama.cpp main directory and run the following:
+In oneAPI command line window:
 
 ```
 mkdir -p build
 cd build
 @call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
 
-cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release -DLLAMA_SYCL_F16=ON
+::  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
 
-make
+::  for FP32
+cmake -G "MinGW Makefiles" ..  -DLLAMA_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icx  -DCMAKE_BUILD_TYPE=Release
+
+
+::  build example/main only
+::  make main
+
+::  build all binary
+make -j
+cd ..
 ```
 
-Otherwise, run the `win-build-sycl.bat` wrapper which encapsulates the former instructions:
-```sh
+or
+
+```
 .\examples\sycl\win-build-sycl.bat
 ```
 
-*Notes:*
+Note:
 
-- 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`.
+- 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.
 
-### III. Run the inference
+### Run
 
-1. Retrieve and prepare model
+1. Put model file to folder **models**
 
-You can refer to the general [*Prepare and Quantize*](README#prepare-and-quantize) guide for model prepration, or simply download [llama-2-7b.Q4_0.gguf](https://huggingface.co/TheBloke/Llama-2-7B-GGUF/blob/main/llama-2-7b.Q4_0.gguf) model as example.
+You 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.
 
 2. Enable oneAPI running environment
 
-On the oneAPI command line window, run the following and step into the llama.cpp directory:
+- In Search, input 'oneAPI'.
+
+Search & open "Intel oneAPI command prompt for Intel 64 for Visual Studio 2022"
+
+- In Run:
+
+In CMD:
 ```
 "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64
 ```
 
-3. List devices information
+3. List device ID
 
-Similar to the native `sycl-ls`, available SYCL devices can be queried as follow:
+Run without parameter:
 
 ```
 build\bin\ls-sycl-device.exe
+
+or
+
+build\bin\main.exe
 ```
 
-The output of this command in a system with 1 *intel CPU* and 1 *intel GPU* would look like the following:
+Check the ID in startup log, like:
+
 ```
-found 6 SYCL devices:
-|  |                  |                                             |Compute   |Max compute|Max work|Max sub|               |
-|ID|       Device Type|                                         Name|capability|units      |group   |group  |Global mem size|
-|--|------------------|---------------------------------------------|----------|-----------|--------|-------|---------------|
-| 0|[level_zero:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       1.3|        512|    1024|     32|    16225243136|
-| 1|[level_zero:gpu:1]|                    Intel(R) UHD Graphics 770|       1.3|         32|     512|     32|    53651849216|
-| 2|    [opencl:gpu:0]|               Intel(R) Arc(TM) A770 Graphics|       3.0|        512|    1024|     32|    16225243136|
-| 3|    [opencl:gpu:1]|                    Intel(R) UHD Graphics 770|       3.0|         32|     512|     32|    53651849216|
-| 4|    [opencl:cpu:0]|         13th Gen Intel(R) Core(TM) i7-13700K|       3.0|         24|    8192|     64|    67064815616|
-| 5|    [opencl:acc:0]|               Intel(R) FPGA Emulation Device|       1.2|         24|67108864|     64|    67064815616|
+found 4 SYCL devices:
+  Device 0: Intel(R) Arc(TM) A770 Graphics,	compute capability 1.3,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
+  Device 1: Intel(R) FPGA Emulation Device,	compute capability 1.2,
+    max compute_units 24,	max work group size 67108864,	max sub group size 64,	global mem size 67065057280
+  Device 2: 13th Gen Intel(R) Core(TM) i7-13700K,	compute capability 3.0,
+    max compute_units 24,	max work group size 8192,	max sub group size 64,	global mem size 67065057280
+  Device 3: Intel(R) Arc(TM) A770 Graphics,	compute capability 3.0,
+    max compute_units 512,	max work group size 1024,	max sub group size 32,	global mem size 16225243136
 
 ```
 
@@ -492,33 +405,15 @@ found 6 SYCL devices:
 |compute capability 1.3|Level-zero running time, recommended |
 |compute capability 3.0|OpenCL running time, slower than level-zero in most cases|
 
+4. Set device ID and execute 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)|
-
-Examples:
-
-- Use device 0:
+Set device ID = 0 by **set GGML_SYCL_DEVICE=0**
 
 ```
-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 none -mg 0
+set GGML_SYCL_DEVICE=0
+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
 ```
-
-- Use multiple devices:
-
-```
-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
-```
-Otherwise, run the following wrapper script:
+or run by script:
 
 ```
 .\examples\sycl\win-run-llama2.bat
@@ -526,15 +421,14 @@ Otherwise, run the following wrapper script:
 
 Note:
 
-- By default, `mmap` is used to read the model file. In some cases, it causes runtime hang issues. Please disable it by passing `--no-mmap` to the `main.exe` if faced with the issue.
-- Upon execution, verify the selected device(s) ID(s) in the output log, which can for instance be displayed as follow:
+- 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.
 
-```sh
-detect 1 SYCL GPUs: [0] with top Max compute units:512
+
+5. Check the device ID in output
+
+Like:
 ```
-Or
-```sh
-use 1 SYCL GPUs: [0] with Max compute units:512
+Using device **0** (Intel(R) Arc(TM) A770 Graphics) as main device
 ```
 
 ## Environment Variable
@@ -543,55 +437,58 @@ use 1 SYCL GPUs: [0] with Max compute units:512
 
 |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.|
+|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|
+
+#### Running
 
-#### Runtime
 
 |Name|Value|Function|
 |-|-|-|
+|GGML_SYCL_DEVICE|0 (default) or 1|Set the device id used. Check the device ids by default running output|
 |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
 
-- Hanging during startup
+- Hang during startup
 
-  llama.cpp uses *mmap* as the default mode for reading the model file and copying it to the GPU. In some systems, `memcpy` might behave abnormally and therefore hang.
+  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` flag to the `main` executable.
-
-- `Split-mode:[row]` is not supported.
+  Solution: add **--no-mmap** or **--mmap 0**.
 
 ## Q&A
 
 - Error:  `error while loading shared libraries: libsycl.so.7: cannot open shared object file: No such file or directory`.
 
-  - 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`.
+  Miss to enable oneAPI running environment.
 
-- General compiler error:
+  Install oneAPI base toolkit and enable it by: `source /opt/intel/oneapi/setvars.sh`.
 
-  - Remove build folder or try a clean-build.
+- In Windows, no result, not error.
 
-- I can **not** see `[ext_oneapi_level_zero:gpu]` afer installing the GPU driver on Linux.
+  Miss to enable oneAPI running environment.
 
-  Please double-check with `sudo sycl-ls`.
+- Meet compile error.
 
-  If it's present in the list, please add video/render group to your user then **logout/login** or restart your system:
+  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:
 
   ```
-  sudo usermod -aG render $USER
-  sudo usermod -aG video $USER
+  sudo usermod -aG render username
+  sudo usermod -aG video username
   ```
-  Otherwise, please double-check the GPU driver installation steps.
 
-### **GitHub contribution**:
-Please add the **[SYCL]** prefix/tag in issues/PRs titles to help the SYCL-team check/address them without delay.
+  Then **relogin**.
+
+  If you do not see it, please check the installation GPU steps again.
 
 ## Todo
 
-- Support row layer split for multiple card runs.
+- Support multiple cards.

README.md

@@ -8,22 +8,16 @@
 
 Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others) in pure C/C++
 
-### Recent API changes
-
-- [2024 Mar 26] Logits and embeddings API updated for compactness https://github.com/ggerganov/llama.cpp/pull/6122
-- [2024 Mar 13] Add `llama_synchronize()` + `llama_context_params.n_ubatch` https://github.com/ggerganov/llama.cpp/pull/6017
-- [2024 Mar 8] `llama_kv_cache_seq_rm()` returns a `bool` instead of `void`, and new `llama_n_seq_max()` returns the upper limit of acceptable `seq_id` in batches (relevant when dealing with multiple sequences) https://github.com/ggerganov/llama.cpp/pull/5328
-- [2024 Mar 4] Embeddings API updated https://github.com/ggerganov/llama.cpp/pull/5796
-- [2024 Mar 3] `struct llama_context_params` https://github.com/ggerganov/llama.cpp/pull/5849
-
 ### Hot topics
 
-- 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
-- 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
+- Remove LLAMA_MAX_DEVICES and LLAMA_SUPPORTS_GPU_OFFLOAD: https://github.com/ggerganov/llama.cpp/pull/5240
+- Incoming backends: https://github.com/ggerganov/llama.cpp/discussions/5138
+  - [SYCL backend](README-sycl.md) is ready (1/28/2024), support Linux/Windows in Intel GPUs (iGPU, Arc/Flex/Max series)
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
+- Collecting Apple Silicon performance stats:
+  - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
+  - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
+- Looking for contributions to improve and maintain the `server` example: https://github.com/ggerganov/llama.cpp/issues/4216
 
 ----
 
@@ -67,7 +61,7 @@ variety of hardware - locally and in the cloud.
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
 - AVX, AVX2 and AVX512 support for x86 architectures
-- 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
+- 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP)
 - Vulkan, SYCL, and (partial) OpenCL backend support
 - CPU+GPU hybrid inference to partially accelerate models larger than the total VRAM capacity
@@ -113,22 +107,16 @@ Typically finetunes of the base models below are supported as well.
 - [x] [Orion 14B](https://github.com/ggerganov/llama.cpp/pull/5118)
 - [x] [InternLM2](https://huggingface.co/models?search=internlm2)
 - [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)
 
 **Multimodal models:**
 
-- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e), [LLaVA 1.6 models](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2)
+- [x] [LLaVA 1.5 models](https://huggingface.co/collections/liuhaotian/llava-15-653aac15d994e992e2677a7e)
 - [x] [BakLLaVA](https://huggingface.co/models?search=SkunkworksAI/Bakllava)
 - [x] [Obsidian](https://huggingface.co/NousResearch/Obsidian-3B-V0.5)
 - [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)
 
-**HTTP server**
-
-[llama.cpp web server](./examples/server) is a lightweight [OpenAI API](https://github.com/openai/openai-openapi) compatible HTTP server that can be used to serve local models and easily connect them to existing clients.
 
 **Bindings:**
 
@@ -137,7 +125,6 @@ Typically finetunes of the base models below are supported as well.
 - Node.js: [withcatai/node-llama-cpp](https://github.com/withcatai/node-llama-cpp)
 - JS/TS (llama.cpp server client): [lgrammel/modelfusion](https://modelfusion.dev/integration/model-provider/llamacpp)
 - 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 (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)
@@ -148,7 +135,6 @@ 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:**
 
@@ -159,7 +145,6 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [nat/openplayground](https://github.com/nat/openplayground)
 - [Faraday](https://faraday.dev/) (proprietary)
 - [LMStudio](https://lmstudio.ai/) (proprietary)
-- [LocalAI](https://github.com/mudler/LocalAI) (MIT)
 - [LostRuins/koboldcpp](https://github.com/LostRuins/koboldcpp) (AGPL)
 - [Mozilla-Ocho/llamafile](https://github.com/Mozilla-Ocho/llamafile)
 - [nomic-ai/gpt4all](https://github.com/nomic-ai/gpt4all)
@@ -169,14 +154,8 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [cztomsik/ava](https://github.com/cztomsik/ava) (MIT)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
 - [pythops/tenere](https://github.com/pythops/tenere) (AGPL)
-- [RecurseChat](https://recurse.chat/) (proprietary)
 - [semperai/amica](https://github.com/semperai/amica)
 - [withcatai/catai](https://github.com/withcatai/catai)
-- [Mobile-Artificial-Intelligence/maid](https://github.com/Mobile-Artificial-Intelligence/maid) (MIT)
-- [Msty](https://msty.app) (proprietary)
-- [LLMFarm](https://github.com/guinmoon/LLMFarm?tab=readme-ov-file) (MIT)
-
-*(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
 ---
 
@@ -452,27 +431,30 @@ 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.
 
-- #### CUDA
+- #### cuBLAS
 
-  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).
+  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).
 
   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_CUDA=1
+    make LLAMA_CUBLAS=1
     ```
   - Using `CMake`:
 
     ```bash
     mkdir build
     cd build
-    cmake .. -DLLAMA_CUDA=ON
+    cmake .. -DLLAMA_CUBLAS=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. |
@@ -634,15 +616,6 @@ Building the program with BLAS support may lead to some performance improvements
 
 - #### Vulkan
 
-> [!WARNING]
->
-> Vulkan support has been broken in https://github.com/ggerganov/llama.cpp/pull/6122
-> due to relying on `GGML_OP_GET_ROWS` which is not yet properly supported by the Vulkan backend,
-> but should be fixed relatively soon (possibly in https://github.com/ggerganov/llama.cpp/pull/6155
-> (ref: https://github.com/ggerganov/llama.cpp/pull/6122#issuecomment-2015327635)).
->
-> Meanwhile, if you want to use the Vulkan backend, you should use the commit right before the breaking change, https://github.com/ggerganov/llama.cpp/commit/55c1b2a3bbd470e9e2a3a0618b92cf64a885f806
-
   **With docker**:
 
   You don't need to install Vulkan SDK. It will be installed inside the container.
@@ -795,7 +768,7 @@ The time per token is measured on a MacBook M1 Pro 32GB RAM using 4 and 8 thread
 
 #### How to run
 
-1. Download/extract: https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+1. Download/extract: https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip?ref=salesforce-research
 2. Run `./perplexity -m models/7B/ggml-model-q4_0.gguf -f wiki.test.raw`
 3. Output:
 ```
@@ -808,7 +781,7 @@ And after 4.45 hours, you will have the final perplexity.
 ### Interactive mode
 
 If you want a more ChatGPT-like experience, you can run in interactive mode by passing `-i` as a parameter.
-In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMA emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
+In this mode, you can always interrupt generation by pressing Ctrl+C and entering one or more lines of text, which will be converted into tokens and appended to the current context. You can also specify a *reverse prompt* with the parameter `-r "reverse prompt string"`. This will result in user input being prompted whenever the exact tokens of the reverse prompt string are encountered in the generation. A typical use is to use a prompt that makes LLaMa emulate a chat between multiple users, say Alice and Bob, and pass `-r "Alice:"`.
 
 Here is an example of a few-shot interaction, invoked with the command
 
@@ -872,7 +845,7 @@ Sample run:
 ```
 == Running in interactive mode. ==
  - Press Ctrl+C to interject at any time.
- - Press Return to return control to LLaMA.
+ - Press Return to return control to LLaMa.
  - If you want to submit another line, end your input in '\'.
 
  Below is an instruction that describes a task. Write a response that appropriately completes the request.
@@ -919,9 +892,6 @@ First, install the essential packages for termux:
 pkg install clang wget git cmake
 ```
 Second, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
-
-You can execute the following commands on your computer to avoid downloading the NDK to your mobile. Of course, you can also do this in Termux.
-
 ```
 $ mkdir build-android
 $ cd build-android
@@ -930,28 +900,7 @@ $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROI
 $ make
 ```
 Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
-Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
-
-(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
-```
-$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
-$cd /data/data/com.termux/files/home/bin
-$chmod +x ./*
-```
-
-Download model [llama-2-7b-chat.Q4_K_M.gguf](https://huggingface.co/TheBloke/Llama-2-7B-Chat-GGUF/blob/main/llama-2-7b-chat.Q4_K_M.gguf), and push it to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
-
-```
-$mv /sdcard/llama.cpp/llama-2-7b-chat.Q4_K_M.gguf /data/data/com.termux/files/home/model/
-```
-
-Now, you can start chatting:
-```
-$cd /data/data/com.termux/files/home/bin
-$./main -m ../model/llama-2-7b-chat.Q4_K_M.gguf -n 128 -cml
-```
-
-Here is a demo of an interactive session running on Pixel 5 phone:
+Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
 
@@ -1009,7 +958,7 @@ We have three Docker images available for this project:
 
 1. `ghcr.io/ggerganov/llama.cpp:full`: This image includes both the main executable file and the tools to convert LLaMA models into ggml and convert into 4-bit quantization. (platforms: `linux/amd64`, `linux/arm64`)
 2. `ghcr.io/ggerganov/llama.cpp:light`: This image only includes the main executable file. (platforms: `linux/amd64`, `linux/arm64`)
-3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executable file. (platforms: `linux/amd64`, `linux/arm64`)
+3. `ghcr.io/ggerganov/llama.cpp:server`: This image only includes the server executabhle file. (platforms: `linux/amd64`, `linux/arm64`)
 
 Additionally, there the following images, similar to the above:
 

awq-py/README.md

@@ -0,0 +1,116 @@
+# AWQ: Activation-aware Weight Quantization for LLM - version apply to llamacpp
+[[Paper](https://arxiv.org/abs/2306.00978)][[Original Repo](https://github.com/mit-han-lab/llm-awq)][[Easy-to-use Repo](https://github.com/casper-hansen/AutoAWQ)]
+
+**Supported models:**
+
+- [X] LLaMA
+- [x] LLaMA 2
+- [X] MPT
+- [X] Mistral AI v0.1
+- [ ] Bloom
+- [ ] Mixtral MoE
+
+**TODO:**
+- [x] Update version work with both MPT and MPT-AWQ model
+- [ ] Add OPT model
+- [ ] Add Bloom model
+- [ ] Add Mixtral MoE
+- [ ] Support w3, w2
+
+
+## Contents
+
+- [Install](##Install)
+- [Convert](##Convert)
+- [Quantize](##Quantize)
+- [Test](##Test)
+- [Benchmark](##Benchmark)
+- [Results](##Results)
+
+## Install
+Install requirements
+```bash
+pip install -r requirements.txt
+```
+Get the pre-computed AWQ search results for multiple model families, including LLaMA, LLaMA2, MPT, OPT
+```bash
+git clone https://huggingface.co/datasets/mit-han-lab/awq-model-zoo awq_cache
+```
+
+## Convert
+Example for llama model
+```bash
+# For llama7b and llama2 models
+python convert.py models/llama-7b/ --awq-path awq_cache/llama-7b-w4-g128.pt --outfile models/llama_7b_fp16.gguf
+# For mistral and mpt models
+python convert-hf-to-gguf.py models/mpt-7b/ --awq-path awq_cache/mpt-7b-w4-g128.pt --outfile models/mpt_7b_fp16.gguf
+```
+
+## Quantize
+```bash
+# We only benchmark and confirm the results on q4_0, q4_1, and q2_k types.
+./quantize models/llama_7b_fp16.gguf models/llama_7b_q4_0.gguf q4_0
+```
+
+## Test
+```bash
+# For all models.
+./build/bin/main -m models/llama_7b_q4_0.gguf -n 128 --prompt "Once upon a time"
+```
+
+## Benchmark
+The perplexity measurements in table above are done against the `wikitext2` test dataset (https://paperswithcode.com/dataset/wikitext-2), with context length of 512.
+```bash
+# For llama and llama2, and mistral models.
+./perplexity -m models/llama_7b_q4_0.gguf -f datasets/wikitext-2-raw/wiki.test.raw
+```
+
+## Results
+Results are run on OpenBLAS (CPU) and CuBLAS (GPU) for fair comparison
+We use three types of llamacpp quantization methods to work with our version, including q4_0, q4_1, and q2_k
+
+### Llama 7B (Build with OpenBLAS)
+
+| Model      | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-----------:|--------------|-------:|-------:|-------:|-------:|
+|Llama 7B    | perplexity   | 5.9066 | 6.1214 | 6.0643 | 6.5808 |
+|Llama 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama 7B| perplexity   | 5.9175 | 6.0252 | 5.9987 | 6.3692 |
+|AWQ-LLama 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Llama2 7B (Build with CuBLAS)
+
+| Model       | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|------------:|--------------|-------:|-------:|-------:|-------:|
+|Llama2 7B    | perplexity   | 5.8664 | 6.0260 | 6.0656 | 6.4496 |
+|Llama2 7B    | file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|Llama2 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-LLama2 7B| perplexity   | 5.8801 | 6.0054 | 5.9849 | 6.3650 |
+|AWQ-LLama2 7B| file size    |  12.9G  |   3.5G |   3.9G |   2.7G |
+|AWQ-LLama2 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+
+### Mistral 7B v0.1 (Build with CuBLAS)
+
+| Model        | Measure      | F16    | Q4_0   | Q4_1   | Q2_K   |
+|-------------:|--------------|-------:|-------:|-------:|-------:|
+|Mistral 7B    | perplexity   | 5.6931 | 5.8202 | 5.8268 | 6.1645 |
+|Mistral 7B    | file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|Mistral 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+|AWQ-Mistral 7B| perplexity   | 5.6934 | 5.8020 | 5.7691 | 6.0426 |
+|AWQ-Mistral 7B| file size     |  14.5G |   4.1G |   4.5G |   3.1G |
+|AWQ-Mistral 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6 |
+
+### MPT 7B (Build with OpenBLAS)
+
+| Model    | Measure      | F16    | Q4_0   | Q4_1   | Q2_K    |
+|---------:|--------------|-------:|-------:|-------:|--------:|
+|MPT 7B    | perplexity   | 8.4369 | 8.7956 | 8.6265 | 11.4913 |
+|MPT 7B    | file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|MPT 7B    | bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |
+|AWQ-MPT 7B| perplexity   | 8.4944 | 8.7053 |  8.6750 | 10.2873|
+|AWQ-MPT 7B| file size    |  13.7G  |   3.9G |   4.3G |   2.8G  |
+|AWQ-MPT 7B| bits/weight  |   16.0 |    4.5 |    5.0 |    2.6  |

awq-py/awq/apply_awq.py

@@ -0,0 +1,254 @@
+"""
+Implements the AWQ for llama.cpp use cases.
+Original paper: https://arxiv.org/abs/2306.00978
+
+This code is based on versions of the AWQ implementation found in the following repositories:
+* https://github.com/mit-han-lab/llm-awq
+* https://github.com/casper-hansen/AutoAWQ
+"""
+
+import os
+import torch
+import torch.nn as nn
+
+from transformers import AutoModelForCausalLM, AutoConfig
+from transformers.models.bloom.modeling_bloom import BloomGelu
+from transformers.models.llama.modeling_llama import LlamaRMSNorm
+from transformers.activations import GELUActivation
+
+
+class ScaledActivation(nn.Module):
+    """
+    ScaledActivation module wraps an existing activation function and applies a
+    scale factor to its output.
+
+    Args:
+        module (nn.Module): The activation function to be scaled.
+        scales (torch.Tensor): A tensor of size (num_features,) containing the initial
+            scale factors for each feature.
+
+    Returns:
+        torch.Tensor: The scaled output of the activation function.
+    """
+
+    def __init__(self, module, scales):
+        super().__init__()
+        self.act = module
+        self.scales = nn.Parameter(scales.data)
+
+    def forward(self, x):
+        return self.act(x) / self.scales.view(1, 1, -1).to(x.device)
+
+
+def set_op_by_name(layer, name, new_module):
+    """
+    Set the new module for given module's name.
+
+    Args:
+        layer (nn.Module): The layer in which to replace the submodule.
+        name (str): The path to the submodule to be replaced, using dot notation
+            to access nested modules.
+        new_module (nn.Module): The new module to replace the existing one.
+    """
+    levels = name.split(".")
+    if len(levels) > 1:
+        mod_ = layer
+        for l_idx in range(len(levels) - 1):
+            if levels[l_idx].isdigit():
+                mod_ = mod_[int(levels[l_idx])]
+            else:
+                mod_ = getattr(mod_, levels[l_idx])
+        setattr(mod_, levels[-1], new_module)
+    else:
+        setattr(layer, name, new_module)
+
+
+def get_op_by_name(module, op_name):
+    """
+    Retrieves a submodule within a given layer based on its name.
+
+    Args:
+        module (nn.Module): The layer containing the submodule to find.
+        op_name (str): The name of the submodule.
+
+    Returns:
+        nn.Module: The requested submodule found within the given layer.
+
+    Raises:
+        ValueError: If the specified submodule cannot be found within the layer.
+    """
+    for name, m in module.named_modules():
+        if name == op_name:
+            return m
+    raise ValueError(f"Cannot find op {op_name} in module {module}")
+
+
+@torch.no_grad()
+def scale_ln_fcs(ln, fcs, scales):
+    """
+    Scales the weights of a LayerNorm and a list of fully-connected layers proportionally.
+
+    Args:
+        ln (nn.LayerNorm): The LayerNorm module to be scaled.
+        fcs (List[nn.Linear]): A list of fully-connected layers to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+
+    if not isinstance(fcs, list):
+        fcs = [fcs]
+
+    scales = scales.to(ln.weight.device)
+
+    ln.weight.div_(scales)
+    if hasattr(ln, "bias") and ln.bias is not None:
+        ln.bias.div_(scales)
+
+    for fc in fcs:
+        fc.weight.mul_(scales.view(1, -1))
+
+    for p in ln.parameters():
+        assert torch.isnan(p).sum() == 0
+    for fc in fcs:
+        for p in fc.parameters():
+            assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_fc_fc(fc1, fc2, scales):
+    """
+    Scales the weights of two fully-connected layers in a specific pattern.
+
+    Args:
+        fc1 (nn.Linear): The first fully-connected layer to be scaled.
+        fc2 (nn.Linear): The second fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+    """
+    assert isinstance(fc1, nn.Linear)
+    assert isinstance(fc2, nn.Linear)
+
+    scales = scales.to(fc1.weight.device)
+
+    fc1.weight[-scales.size(0):].div_(scales.view(-1, 1))
+    if fc1.bias is not None:
+        fc1.bias.div_(scales.view(-1))
+
+    fc2.weight.mul_(scales.view(1, -1))
+
+    for p in fc1.parameters():
+        assert torch.isnan(p).sum() == 0
+    for p in fc2.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+@torch.no_grad()
+def scale_gelu_fc(gelu, fc, scales):
+    """
+    Scales the weight of a GELU activation and a fully-connected layer proportionally.
+
+    Args:
+        gelu (Union[nn.GELU, BloomGelu, GELUActivation]): The GELU activation module to be scaled.
+        fc (nn.Linear): The fully-connected layer to be scaled.
+        scales (torch.Tensor): A 1D tensor of size (num_features,).
+
+    Raises:
+        TypeError: If the `gelu` module is not of type `nn.GELU`, `BloomGelu`, or `GELUActivation`.
+        TypeError: If the `fc` module is not of type `nn.Linear`.
+    """
+    assert isinstance(gelu, (nn.GELU, BloomGelu, GELUActivation))
+    assert isinstance(fc, nn.Linear)
+
+    fc.weight.mul_(scales.view(1, -1).to(fc.weight.device))
+
+    for p in fc.parameters():
+        assert torch.isnan(p).sum() == 0
+
+
+def apply_scale(module, scales_list, input_feat_dict=None):
+    """
+    Applies different scaling strategies to layers based on their type and hierarchy within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layers to be scaled.
+        scales_list (List[Tuple[str, List[str], torch.Tensor]]): A list of tuples containing:
+            * prev_op_name (str): The name of the preceding operation or module,
+                relative to which the layers to be scaled are located.
+            * layer_names (List[str]): A list of names of the layers to be scaled, relative to the preceding operation.
+            * scales (torch.Tensor): A 1D tensor of size (num_features,) containing the scaling factors for each feature.
+        input_feat_dict (Optional[Dict[str, torch.Tensor]]): A dictionary mapping layer names to their corresponding
+            input features (optional).
+    """
+    for prev_op_name, layer_names, scales in scales_list:
+        prev_op = get_op_by_name(module, prev_op_name)
+        layers = [get_op_by_name(module, name) for name in layer_names]
+
+        prev_op.cuda()
+        for layer in layers:
+            layer.cuda()
+        scales.cuda()
+
+        if isinstance(prev_op, nn.Linear):
+            assert len(layers) == 1
+            scale_fc_fc(prev_op, layers[0], scales)
+        elif isinstance(prev_op, (nn.LayerNorm, LlamaRMSNorm)) or "rmsnorm" in str(prev_op.__class__).lower():
+            scale_ln_fcs(prev_op, layers, scales)
+        elif isinstance(prev_op, (nn.GELU, BloomGelu, GELUActivation)):
+            new_module = ScaledActivation(prev_op, scales)
+            set_op_by_name(module, prev_op_name, new_module)
+            scale_gelu_fc(prev_op, layers[0], scales)
+        else:
+            raise NotImplementedError(f"prev_op {type(prev_op)} not supported yet!")
+
+        # apply the scaling to input feat if given; prepare it for clipping
+        if input_feat_dict is not None:
+            for layer_name in layer_names:
+                inp = input_feat_dict[layer_name]
+                inp.div_(scales.view(1, -1).to(inp.device))
+
+        prev_op.cpu()
+        for layer in layers:
+            layer.cpu()
+        scales.cpu()
+
+
+@torch.no_grad()
+def apply_clip(module, clip_list):
+    """
+    Applies element-wise clipping to the weight of a specific layer within a given module.
+
+    Args:
+        module (nn.Module): The module containing the layer to be clipped.
+        clip_list (List[Tuple[str, torch.Tensor]]): A list of tuples containing:
+            * name (str): The name of the layer to be clipped, relative to the root of the module.
+            * max_val (torch.Tensor): A 1D or 2D tensor defining the upper bound for each element of the layer's weight.
+    """
+    for name, max_val in clip_list:
+        layer = get_op_by_name(module, name)
+        layer.cuda()
+        max_val = max_val.to(layer.weight.device)
+        org_shape = layer.weight.shape
+        layer.weight.data = layer.weight.data.reshape(*max_val.shape[:2], -1)
+        layer.weight.data = torch.clamp(layer.weight.data, -max_val, max_val)
+        layer.weight.data = layer.weight.data.reshape(org_shape)
+        layer.cpu()
+
+
+def add_scale_weights(model_path, scale_path, tmp_path):
+    """
+    Adds pre-computed Activation Weight Quantization (AWQ) results to a model,
+    including scaling factors and clipping bounds.
+
+    Args:
+        model_path (str): Path to the pre-trained model to be equipped with AWQ.
+        scale_path (str): Path to the AWQ scale factors (.pt file).
+        tmp_path (str): Path to the temporary directory where the equipped model will be saved.
+    """
+    config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, config=config, trust_remote_code=True
+    )
+    model.eval()
+    awq_results = torch.load(str(scale_path), map_location="cpu")
+    apply_scale(model, awq_results["scale"])
+    apply_clip(model, awq_results["clip"])
+    model.save_pretrained(str(tmp_path))
+    os.system(f"cp {str(model_path)}/tokenizer* {str(tmp_path)}")

awq-py/requirements.txt

@@ -0,0 +1,2 @@
+torch>=2.1.1
+transformers>=4.32.0

build.zig

@@ -115,27 +115,23 @@ pub fn build(b: *std.build.Builder) !void {
     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");
     const console = make.obj("console", "common/console.cpp");
     const sampling = make.obj("sampling", "common/sampling.cpp");
     const grammar_parser = make.obj("grammar-parser", "common/grammar-parser.cpp");
-    const json_schema_to_grammar = make.obj("json-schema-to-grammar", "common/json-schema-to-grammar.cpp");
     const train = make.obj("train", "common/train.cpp");
     const clip = make.obj("clip", "examples/llava/clip.cpp");
-    const llava = make.obj("llava", "examples/llava/llava.cpp");
 
-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, 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, common, buildinfo, train });
 
-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, unicode, unicode_data, common, buildinfo, sampling, grammar_parser, json_schema_to_grammar, clip, llava });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
     if (server.target.isWindows()) {
         server.linkSystemLibrary("ws2_32");
     }

ci/run.sh

@@ -33,20 +33,19 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
-CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
+CMAKE_EXTRA=""
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
     CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_SHADER_DEBUG=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUDA=1"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_CUBLAS=1"
 fi
 
 if [ ! -z ${GG_BUILD_SYCL} ]; then
     if [ -z ${ONEAPI_ROOT} ]; then
-        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:"
-        echo "source /opt/intel/oneapi/setvars.sh"
+        echo "Not detected ONEAPI_ROOT, please install oneAPI base toolkit and enable it by:\n source /opt/intel/oneapi/setvars.sh"
         exit 1
     fi
 
@@ -220,7 +219,7 @@ function gg_run_open_llama_3b_v2 {
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/resolve/main/pytorch_model.bin
     gg_wget models-mnt/open-llama/3B-v2/ https://huggingface.co/openlm-research/open_llama_3b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
     head -n 60 models-mnt/wikitext/wikitext-2-raw/wiki.test.raw > models-mnt/wikitext/wikitext-2-raw/wiki.test-60.raw
 
@@ -273,19 +272,19 @@ function gg_run_open_llama_3b_v2 {
     (time ./bin/main --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
     (time ./bin/main --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
     (time ./bin/save-load-state --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
@@ -344,17 +343,17 @@ function gg_run_open_llama_3b_v2 {
     python3 ../convert-lora-to-ggml.py ${path_lora}
 
     # f16
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
-    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-f16.log
+    (time ./bin/perplexity --model ${model_f16} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-f16.log
     compare_ppl "f16 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-f16.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare}                            -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-q8_0.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0.log
     compare_ppl "q8_0 shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     # q8_0 + f16 lora-base
-    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
+    (time ./bin/perplexity --model ${model_q8_0} -f ${shakespeare} --lora ${lora_shakespeare} --lora-base ${model_f16} -c 128 -b 128 --chunks 2 ) 2>&1 | tee -a $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log
     compare_ppl "q8_0 / f16 base shakespeare" "$(cat $OUT/${ci}-ppl-shakespeare-q8_0.log | grep "^\[1\]")" "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log | grep "^\[1\]")" | tee -a $OUT/${ci}-lora-ppl.log
 
     set +e
@@ -402,7 +401,7 @@ function gg_run_open_llama_7b_v2 {
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/resolve/main/pytorch_model-00002-of-00002.bin
     gg_wget models-mnt/open-llama/7B-v2/ https://huggingface.co/openlm-research/open_llama_7b_v2/raw/main/generation_config.json
 
-    gg_wget models-mnt/wikitext/ https://huggingface.co/datasets/ggml-org/ci/resolve/main/wikitext-2-raw-v1.zip
+    gg_wget models-mnt/wikitext/ https://s3.amazonaws.com/research.metamind.io/wikitext/wikitext-2-raw-v1.zip
     unzip -o models-mnt/wikitext/wikitext-2-raw-v1.zip -d models-mnt/wikitext/
 
     path_models="../models-mnt/open-llama/7B-v2"
@@ -412,8 +411,8 @@ function gg_run_open_llama_7b_v2 {
 
     set -e
 
-    (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
+    (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
 
     python3 ../convert.py ${path_models}
 
@@ -569,54 +568,6 @@ function gg_sum_open_llama_7b_v2 {
     #gg_printf '- shakespeare (q8_0 / f16 base lora):\n```\n%s\n```\n' "$(cat $OUT/${ci}-ppl-shakespeare-lora-q8_0-f16.log)"
 }
 
-# bge-small
-
-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_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
-    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/sentence_bert_config.json
-    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/vocab.txt
-    gg_wget models-mnt/bge-small/ https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/modules.json
-    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/1_Pooling https://huggingface.co/BAAI/bge-small-en-v1.5/raw/main/1_Pooling/config.json
-
-    path_models="../models-mnt/bge-small"
-
-    rm -rf build-ci-release && mkdir build-ci-release && cd build-ci-release
-
-    set -e
-
-    (time cmake -DCMAKE_BUILD_TYPE=Release ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
-
-    python3 ../convert-hf-to-gguf.py ${path_models}
-
-    model_f16="${path_models}/ggml-model-f16.gguf"
-    model_q8_0="${path_models}/ggml-model-q8_0.gguf"
-
-    ./bin/quantize ${model_f16} ${model_q8_0} q8_0
-
-    (time ./bin/embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-
-    set +e
-}
-
-function gg_sum_embd_bge_small {
-    gg_printf '### %s\n\n' "${ci}"
-
-    gg_printf 'BGE Small (BERT):\n'
-    gg_printf '- status: %s\n' "$(cat $OUT/${ci}.exit)"
-    gg_printf '- f16: \n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-f16.log)"
-    gg_printf '- q8_0:\n```\n%s\n```\n' "$(cat $OUT/${ci}-tg-q8_0.log)"
-}
-
 ## main
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
@@ -640,8 +591,6 @@ test $ret -eq 0 && gg_run ctest_debug
 test $ret -eq 0 && gg_run ctest_release
 
 if [ -z ${GG_BUILD_LOW_PERF} ]; then
-    test $ret -eq 0 && gg_run embd_bge_small
-
     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

@@ -19,12 +19,7 @@ if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
         endif()
     endif()
 
-    if(EXISTS "${GIT_DIR}/index")
-        set(GIT_INDEX "${GIT_DIR}/index")
-    else()
-        message(WARNING "Git index not found in git repository.")
-        set(GIT_INDEX "")
-    endif()
+    set(GIT_INDEX "${GIT_DIR}/index")
 else()
     message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
     set(GIT_INDEX "")
@@ -47,8 +42,6 @@ if (BUILD_SHARED_LIBS)
     set_target_properties(${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)
 
@@ -62,28 +55,14 @@ add_library(${TARGET} STATIC
     console.cpp
     grammar-parser.h
     grammar-parser.cpp
-    json.hpp
     train.h
     train.cpp
-    ngram-cache.h
-    ngram-cache.cpp
     )
 
 if (BUILD_SHARED_LIBS)
     set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
 endif()
 
-set(LLAMA_COMMON_EXTRA_LIBS build_info)
-
-# Use curl to download model url
-if (LLAMA_CURL)
-    find_package(CURL REQUIRED)
-    add_definitions(-DLLAMA_USE_CURL)
-    include_directories(${CURL_INCLUDE_DIRS})
-    find_library(CURL_LIBRARY curl REQUIRED)
-    set(LLAMA_COMMON_EXTRA_LIBS ${LLAMA_COMMON_EXTRA_LIBS} ${CURL_LIBRARY})
-endif ()
-
 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE ${LLAMA_COMMON_EXTRA_LIBS} PUBLIC llama)
+target_link_libraries(${TARGET} PRIVATE build_info PUBLIC llama)

common/common.h

@@ -37,16 +37,13 @@ extern char const *LLAMA_COMMIT;
 extern char const *LLAMA_COMPILER;
 extern char const *LLAMA_BUILD_TARGET;
 
-struct llama_control_vector_load_info;
-
-int32_t get_num_physical_cores();
-
 //
 // CLI argument parsing
 //
+int32_t get_num_physical_cores();
 
 struct gpt_params {
-    uint32_t seed                 = LLAMA_DEFAULT_SEED; // RNG seed
+    uint32_t seed                 = -1;    // RNG seed
 
     int32_t n_threads             = get_num_physical_cores();
     int32_t n_threads_draft       = -1;
@@ -54,17 +51,17 @@ struct gpt_params {
     int32_t n_threads_batch_draft = -1;
     int32_t n_predict             = -1;    // new tokens to predict
     int32_t n_ctx                 = 512;   // context size
-    int32_t n_batch               = 2048;  // logical batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_ubatch              = 512;   // physical batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_batch               = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft               = 5;     // number of tokens to draft during speculative decoding
+    int32_t n_draft               = 8;     // number of tokens to draft during speculative decoding
     int32_t n_chunks              = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel            = 1;     // number of parallel sequences to decode
     int32_t n_sequences           = 1;     // number of sequences to decode
+    float   p_accept              = 0.5f;  // speculative decoding accept probability
     float   p_split               = 0.1f;  // speculative decoding split probability
     int32_t n_gpu_layers          = -1;    // number of layers to store in VRAM (-1 - use default)
     int32_t n_gpu_layers_draft    = -1;    // number of layers to store in VRAM for the draft model (-1 - use default)
-    llama_split_mode split_mode   = LLAMA_SPLIT_MODE_LAYER; // how to split the model across GPUs
+    llama_split_mode split_mode   = LLAMA_SPLIT_LAYER; // how to split the model across GPUs
     int32_t main_gpu              = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[128]     = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams               = 0;     // if non-zero then use beam search of given width.
@@ -78,32 +75,22 @@ struct gpt_params {
     float   yarn_beta_fast        = 32.0f; // YaRN low correction dim
     float   yarn_beta_slow        = 1.0f;  // YaRN high correction dim
     int32_t yarn_orig_ctx         = 0;     // YaRN original context length
-    float   defrag_thold          = -1.0f; // KV cache defragmentation threshold
-
-    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
+    int32_t rope_scaling_type     = LLAMA_ROPE_SCALING_UNSPECIFIED;
 
     // // sampling parameters
     struct llama_sampling_params sparams;
 
-    std::string model                = "models/7B/ggml-model-f16.gguf"; // model path
-    std::string model_draft          = "";  // draft model for speculative decoding
-    std::string model_alias          = "unknown"; // model alias
-    std::string model_url            = "";  // model url to download
-    std::string hf_repo              = "";  // HF repo
-    std::string hf_file              = "";  // HF file
-    std::string prompt               = "";
-    std::string prompt_file          = "";  // store the external prompt file name
-    std::string path_prompt_cache    = "";  // path to file for saving/loading prompt eval state
-    std::string input_prefix         = "";  // string to prefix user inputs with
-    std::string input_suffix         = "";  // string to suffix user inputs with
+    std::string model             = "models/7B/ggml-model-f16.gguf"; // model path
+    std::string model_draft       = "";                              // draft model for speculative decoding
+    std::string model_alias       = "unknown"; // model alias
+    std::string prompt            = "";
+    std::string prompt_file       = "";  // store the external prompt file name
+    std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
+    std::string input_prefix      = "";  // string to prefix user inputs with
+    std::string input_suffix      = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
-    std::string logdir               = "";  // directory in which to save YAML log files
-    std::string lookup_cache_static  = ""; // path of static ngram cache file for lookup decoding
-    std::string lookup_cache_dynamic = ""; // path of dynamic ngram cache file for lookup decoding
-    std::string logits_file          = "";  // file for saving *all* logits
+    std::string logdir            = "";  // directory in which to save YAML log files
+    std::string logits_file       = "";  // file for saving *all* logits
 
     std::vector<llama_model_kv_override> kv_overrides;
 
@@ -111,11 +98,6 @@ struct gpt_params {
     std::vector<std::tuple<std::string, float>> lora_adapter; // lora adapter path with user defined scale
     std::string lora_base  = "";                              // base model path for the lora adapter
 
-    std::vector<llama_control_vector_load_info> control_vectors; // control vector with user defined scale
-
-    int32_t control_vector_layer_start = -1; // layer range for control vector
-    int32_t control_vector_layer_end   = -1; // layer range for control vector
-
     int  ppl_stride        = 0;     // stride for perplexity calculations. If left at 0, the pre-existing approach will be used.
     int  ppl_output_type   = 0;     // = 0 -> ppl output is as usual, = 1 -> ppl output is num_tokens, ppl, one per line
                                     //                                       (which is more convenient to use for plotting)
@@ -131,6 +113,7 @@ struct gpt_params {
 
     bool   kl_divergence   = false; // compute KL-divergence
 
+    bool mul_mat_q         = true;  // if true, use mul_mat_q kernels instead of cuBLAS
     bool random_prompt     = false; // do not randomize prompt if none provided
     bool use_color         = false; // use color to distinguish generations and inputs
     bool interactive       = false; // interactive mode
@@ -143,7 +126,7 @@ struct gpt_params {
     bool interactive_first = false; // wait for user input immediately
     bool multiline_input   = false; // reverse the usage of `\`
     bool simple_io         = false; // improves compatibility with subprocesses and limited consoles
-    bool cont_batching     = true;  // insert new sequences for decoding on-the-fly
+    bool cont_batching     = false; // insert new sequences for decoding on-the-fly
 
     bool input_prefix_bos  = false; // prefix BOS to user inputs, preceding input_prefix
     bool ignore_eos        = false; // ignore generated EOS tokens
@@ -151,6 +134,7 @@ struct gpt_params {
     bool logits_all        = false; // return logits for all tokens in the batch
     bool use_mmap          = true;  // use mmap for faster loads
     bool use_mlock         = false; // use mlock to keep model in memory
+    bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool display_prompt    = true;  // print prompt before generation
     bool infill            = false; // use infill mode
@@ -171,8 +155,6 @@ 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);
@@ -183,7 +165,7 @@ void process_escapes(std::string& input);
 // String utils
 //
 
-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_names(const std::vector<std::string> & 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 sampler_type_to_name_string(llama_sampler_type sampler_type);
@@ -198,9 +180,6 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
 struct llama_model_params   llama_model_params_from_gpt_params  (const gpt_params & params);
 struct llama_context_params llama_context_params_from_gpt_params(const gpt_params & params);
 
-struct llama_model * llama_load_model_from_url(const char * model_url, const char * path_model, const struct llama_model_params & params);
-struct llama_model * llama_load_model_from_hf(const char * repo, const char * file, const char * path_model, const struct llama_model_params & params);
-
 // Batch utils
 
 void llama_batch_clear(struct llama_batch & batch);
@@ -279,39 +258,3 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-
-//
-// Embedding utils
-//
-
-void llama_embd_normalize(const float * inp, float * out, int n);
-
-float llama_embd_similarity_cos(const float * embd1, const float * embd2, int n);
-
-//
-// Control vector utils
-//
-
-struct llama_control_vector_data {
-    int n_embd;
-
-    // stores data for layers [1, n_layer] where n_layer = data.size() / n_embd
-    std::vector<float> data;
-};
-
-struct llama_control_vector_load_info {
-    float strength;
-
-    std::string fname;
-};
-
-// Load control vectors, scale each by strength, and add them together.
-// On error, returns {-1, empty}
-llama_control_vector_data llama_control_vector_load(const std::vector<llama_control_vector_load_info> & load_infos);
-
-//
-// Split utils
-//
-static const char * const LLM_KV_SPLIT_NO            = "split.no";
-static const char * const LLM_KV_SPLIT_COUNT         = "split.count";
-static const char * const LLM_KV_SPLIT_TENSORS_COUNT = "split.tensors.count";

common/grammar-parser.cpp

@@ -278,22 +278,6 @@ namespace grammar_parser {
             while (*pos) {
                 pos = parse_rule(state, pos);
             }
-            // Validate the state to ensure that all rules are defined
-            for (const auto & rule : state.rules) {
-                for (const auto & elem : rule) {
-                    if (elem.type == LLAMA_GRETYPE_RULE_REF) {
-                        // Ensure that the rule at that location exists
-                        if (elem.value >= state.rules.size() || state.rules[elem.value].empty()) {
-                            // Get the name of the rule that is missing
-                            for (const auto & kv : state.symbol_ids) {
-                                if (kv.second == elem.value) {
-                                    throw std::runtime_error("Undefined rule identifier '" + kv.first + "'");
-                                }
-                            }
-                        }
-                    }
-                }
-            }
             return state;
         } catch (const std::exception & err) {
             fprintf(stderr, "%s: error parsing grammar: %s\n", __func__, err.what());

common/json-schema-to-grammar.cpp

@@ -1,721 +0,0 @@
-#include "json-schema-to-grammar.h"
-#include <algorithm>
-#include <fstream>
-#include <map>
-#include <regex>
-#include <sstream>
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-using json = nlohmann::ordered_json;
-
-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"}
-};
-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"}
-};
-
-static bool is_reserved_name(const std::string & name) {
-    static std::unordered_set<std::string> RESERVED_NAMES;
-    if (RESERVED_NAMES.empty()) {
-        RESERVED_NAMES.insert("root");
-        for (const auto &p : PRIMITIVE_RULES) RESERVED_NAMES.insert(p.first);
-        for (const auto &p : DATE_RULES) RESERVED_NAMES.insert(p.first);
-    }
-    return RESERVED_NAMES.find(name) != RESERVED_NAMES.end();
-}
-
-std::regex INVALID_RULE_CHARS_RE("[^a-zA-Z0-9-]+");
-std::regex GRAMMAR_LITERAL_ESCAPE_RE("[\r\n\"]");
-std::regex GRAMMAR_RANGE_LITERAL_ESCAPE_RE("[\r\n\"\\]\\-\\\\]");
-std::unordered_map<char, std::string> GRAMMAR_LITERAL_ESCAPES = {
-    {'\r', "\\r"}, {'\n', "\\n"}, {'"', "\\\""}, {'-', "\\-"}, {']', "\\]"}
-};
-
-std::unordered_set<char> NON_LITERAL_SET = {'|', '.', '(', ')', '[', ']', '{', '}', '*', '+', '?'};
-std::unordered_set<char> ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS = {'[', ']', '(', ')', '|', '{', '}', '*', '+', '?'};
-
-template <typename Iterator>
-std::string join(Iterator begin, Iterator end, const std::string & separator) {
-    std::ostringstream result;
-    if (begin != end) {
-        result << *begin;
-        for (Iterator it = begin + 1; it != end; ++it) {
-            result << separator << *it;
-        }
-    }
-    return result.str();
-}
-
-static std::vector<std::string> split(const std::string & str, const std::string & delimiter) {
-    std::vector<std::string> tokens;
-    size_t start = 0;
-    size_t end = str.find(delimiter);
-
-    while (end != std::string::npos) {
-        tokens.push_back(str.substr(start, end - start));
-        start = end + delimiter.length();
-        end = str.find(delimiter, start);
-    }
-
-    tokens.push_back(str.substr(start));
-
-    return tokens;
-}
-
-static std::string repeat(const std::string & str, size_t n) {
-    if (n == 0) {
-        return "";
-    }
-
-    std::string result;
-    result.reserve(str.length() * n);
-
-    for (size_t i = 0; i < n; ++i) {
-        result += str;
-    }
-
-    return result;
-}
-
-static std::string replacePattern(const std::string & input, const std::regex & regex, const std::function<std::string(const std::smatch  &)> & replacement) {
-    std::smatch match;
-    std::string result;
-
-    std::string::const_iterator searchStart(input.cbegin());
-    std::string::const_iterator searchEnd(input.cend());
-
-    while (std::regex_search(searchStart, searchEnd, match, regex)) {
-        result.append(searchStart, searchStart + match.position());
-        result.append(replacement(match));
-        searchStart = match.suffix().first;
-    }
-
-    result.append(searchStart, searchEnd);
-
-    return result;
-}
-
-static std::string format_literal(const std::string & literal) {
-    std::string escaped = replacePattern(literal, GRAMMAR_LITERAL_ESCAPE_RE, [&](const std::smatch & match) {
-        char c = match.str()[0];
-        return GRAMMAR_LITERAL_ESCAPES.at(c);
-    });
-    return "\"" + escaped + "\"";
-}
-
-
-class SchemaConverter {
-private:
-    std::function<json(const std::string &)> _fetch_json;
-    bool _dotall;
-    std::map<std::string, std::string> _rules;
-    std::unordered_map<std::string, json> _refs;
-    std::unordered_set<std::string> _refs_being_resolved;
-    std::vector<std::string> _errors;
-    std::vector<std::string> _warnings;
-
-    std::string _add_rule(const std::string & name, const std::string & rule) {
-        std::string esc_name = regex_replace(name, INVALID_RULE_CHARS_RE, "-");
-        if (_rules.find(esc_name) == _rules.end() || _rules[esc_name] == rule) {
-            _rules[esc_name] = rule;
-            return esc_name;
-        } else {
-            int i = 0;
-            while (_rules.find(esc_name + std::to_string(i)) != _rules.end() && _rules[esc_name + std::to_string(i)] != rule) {
-                i++;
-            }
-            std::string key = esc_name + std::to_string(i);
-            _rules[key] = rule;
-            return key;
-        }
-    }
-
-    std::string _generate_union_rule(const std::string & name, const std::vector<json> & alt_schemas) {
-        std::vector<std::string> rules;
-        for (size_t i = 0; i < alt_schemas.size(); i++) {
-            rules.push_back(visit(alt_schemas[i], name + (name.empty() ? "alternative-" : "-") + std::to_string(i)));
-        }
-        return join(rules.begin(), rules.end(), " | ");
-    }
-
-    std::string _visit_pattern(const std::string & pattern, const std::string & name) {
-        if (!(pattern.front() == '^' && pattern.back() == '$')) {
-            _errors.push_back("Pattern must start with '^' and end with '$'");
-            return "";
-        }
-        std::string sub_pattern = pattern.substr(1, pattern.length() - 2);
-        std::unordered_map<std::string, std::string> sub_rule_ids;
-
-        size_t i = 0;
-        size_t length = sub_pattern.length();
-
-        using literal_or_rule = std::pair<std::string, bool>;
-        auto to_rule = [&](const literal_or_rule & ls) {
-            auto is_literal = ls.second;
-            auto s = ls.first;
-            return is_literal ? "\"" + s + "\"" : s;
-        };
-        std::function<literal_or_rule()> transform = [&]() -> literal_or_rule {
-            size_t start = i;
-            std::vector<literal_or_rule> seq;
-
-            auto get_dot = [&]() {
-                std::string rule;
-                if (_dotall) {
-                    rule = "[\\U00000000-\\U0010FFFF]";
-                } else {
-                    rule = "[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]";
-                }
-                return _add_rule("dot", rule);
-            };
-
-            // Joins the sequence, merging consecutive literals together.
-            auto join_seq = [&]() {
-                std::vector<literal_or_rule> ret;
-
-                std::string literal;
-                auto flush_literal = [&]() {
-                    if (literal.empty()) {
-                        return false;
-                    }
-                    ret.push_back(std::make_pair(literal, true));
-                    literal.clear();
-                    return true;
-                };
-
-                for (const auto & item : seq) {
-                    auto is_literal = item.second;
-                    if (is_literal) {
-                        literal += item.first;
-                    } else {
-                        flush_literal();
-                        ret.push_back(item);
-                    }
-                }
-                flush_literal();
-
-                std::vector<std::string> results;
-                for (const auto & item : ret) {
-                    results.push_back(to_rule(item));
-                }
-                return std::make_pair(join(results.begin(), results.end(), " "), false);
-            };
-
-            while (i < length) {
-                char c = sub_pattern[i];
-                if (c == '.') {
-                    seq.push_back(std::make_pair(get_dot(), false));
-                    i++;
-                } else if (c == '(') {
-                    i++;
-                    if (i < length) {
-                        if (sub_pattern[i] == '?') {
-                            _warnings.push_back("Unsupported pattern syntax");
-                        }
-                    }
-                    seq.push_back(std::make_pair("(" + to_rule(transform()) + ")", false));
-                } else if (c == ')') {
-                    i++;
-                    if (start > 0 && sub_pattern[start - 1] != '(') {
-                        _errors.push_back("Unbalanced parentheses");
-                    }
-                    return join_seq();
-                } else if (c == '[') {
-                    std::string square_brackets = std::string(1, c);
-                    i++;
-                    while (i < length && sub_pattern[i] != ']') {
-                        if (sub_pattern[i] == '\\') {
-                            square_brackets += sub_pattern.substr(i, 2);
-                            i += 2;
-                        } else {
-                            square_brackets += sub_pattern[i];
-                            i++;
-                        }
-                    }
-                    if (i >= length) {
-                        _errors.push_back("Unbalanced square brackets");
-                    }
-                    square_brackets += ']';
-                    i++;
-                    seq.push_back(std::make_pair(square_brackets, false));
-                } else if (c == '|') {
-                    seq.push_back(std::make_pair("|", false));
-                    i++;
-                } else if (c == '*' || c == '+' || c == '?') {
-                    seq.back() = std::make_pair(to_rule(seq.back()) + c, false);
-                    i++;
-                } else if (c == '{') {
-                    std::string curly_brackets = std::string(1, c);
-                    i++;
-                    while (i < length && sub_pattern[i] != '}') {
-                        curly_brackets += sub_pattern[i];
-                        i++;
-                    }
-                    if (i >= length) {
-                        _errors.push_back("Unbalanced curly brackets");
-                    }
-                    curly_brackets += '}';
-                    i++;
-                    auto nums = split(curly_brackets.substr(1, curly_brackets.length() - 2), ",");
-                    int min_times = 0;
-                    int max_times = std::numeric_limits<int>::max();
-                    try {
-                        if (nums.size() == 1) {
-                            min_times = max_times = std::stoi(nums[0]);
-                        } else if (nums.size() != 2) {
-                            _errors.push_back("Wrong number of values in curly brackets");
-                        } else {
-                            if (!nums[0].empty()) {
-                                min_times = std::stoi(nums[0]);
-                            }
-                            if (!nums[1].empty()) {
-                                max_times = std::stoi(nums[1]);
-                            }
-                        }
-                    } catch (const std::invalid_argument & e) {
-                        _errors.push_back("Invalid number in curly brackets");
-                        return std::make_pair("", false);
-                    }
-                    auto &last = seq.back();
-                    auto &sub = last.first;
-                    auto sub_is_literal = last.second;
-
-                    if (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;
-                        }
-                        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;
-                    }
-                } else {
-                    std::string literal;
-                    auto is_non_literal = [&](char c) {
-                        return NON_LITERAL_SET.find(c) != NON_LITERAL_SET.end();
-                    };
-                    while (i < length) {
-                        if (sub_pattern[i] == '\\' && i < length - 1) {
-                            char next = sub_pattern[i + 1];
-                            if (ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.find(next) != ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS.end()) {
-                                i++;
-                                literal += sub_pattern[i];
-                                i++;
-                            } else {
-                                literal += sub_pattern.substr(i, 2);
-                                i += 2;
-                            }
-                        } else if (sub_pattern[i] == '"') {
-                            literal += "\\\"";
-                            i++;
-                        } else if (!is_non_literal(sub_pattern[i]) &&
-                                (i == length - 1 || literal.empty() || sub_pattern[i + 1] == '.' || !is_non_literal(sub_pattern[i + 1]))) {
-                            literal += sub_pattern[i];
-                            i++;
-                        } else {
-                            break;
-                        }
-                    }
-                    if (!literal.empty()) {
-                        seq.push_back(std::make_pair(literal, true));
-                    }
-                }
-            }
-            return join_seq();
-        };
-        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
-    }
-
-    std::string _resolve_ref(const std::string & ref) {
-        std::string ref_name = ref.substr(ref.find_last_of('/') + 1);
-        if (_rules.find(ref_name) == _rules.end() && _refs_being_resolved.find(ref) == _refs_being_resolved.end()) {
-            _refs_being_resolved.insert(ref);
-            json resolved = _refs[ref];
-            ref_name = visit(resolved, ref_name);
-            _refs_being_resolved.erase(ref);
-        }
-        return ref_name;
-    }
-
-    std::string _build_object_rule(
-        const std::vector<std::pair<std::string, json>> & properties,
-        const std::unordered_set<std::string> & required,
-        const std::string & name,
-        const json & additional_properties)
-    {
-        std::vector<std::string> required_props;
-        std::vector<std::string> optional_props;
-        std::unordered_map<std::string, std::string> prop_kv_rule_names;
-        for (const auto & kv : properties) {
-            const auto &prop_name = kv.first;
-            const auto &prop_schema = kv.second;
-
-            std::string prop_rule_name = visit(prop_schema, name + (name.empty() ? "" : "-") + prop_name);
-            prop_kv_rule_names[prop_name] = _add_rule(
-                name + (name.empty() ? "" : "-") + prop_name + "-kv",
-                format_literal(json(prop_name).dump()) + " space \":\" space " + prop_rule_name
-            );
-            if (required.find(prop_name) != required.end()) {
-                required_props.push_back(prop_name);
-            } else {
-                optional_props.push_back(prop_name);
-            }
-        }
-        if (additional_properties.is_object() || (additional_properties.is_boolean() && additional_properties.get<bool>())) {
-            std::string sub_name = name + (name.empty() ? "" : "-") + "additional";
-            std::string value_rule = visit(additional_properties.is_object() ? additional_properties : json::object(), sub_name + "-value");
-            std::string kv_rule = _add_rule(sub_name + "-kv", _add_rule("string", PRIMITIVE_RULES.at("string")) + " \":\" space " + value_rule);
-            prop_kv_rule_names["*"] = kv_rule;
-            optional_props.push_back("*");
-        }
-
-        std::string rule = "\"{\" space ";
-        for (size_t i = 0; i < required_props.size(); i++) {
-            if (i > 0) {
-                rule += " \",\" space ";
-            }
-            rule += prop_kv_rule_names[required_props[i]];
-        }
-
-        if (!optional_props.empty()) {
-            rule += " (";
-            if (!required_props.empty()) {
-                rule += " \",\" space ( ";
-            }
-
-            std::function<std::string(const std::vector<std::string> &, bool)> get_recursive_refs = [&](const std::vector<std::string> & ks, bool first_is_optional) {
-                std::string res;
-                if (ks.empty()) {
-                    return res;
-                }
-                std::string k = ks[0];
-                std::string kv_rule_name = prop_kv_rule_names[k];
-                if (k == "*") {
-                    res = _add_rule(
-                        name + (name.empty() ? "" : "-") + "additional-kvs",
-                        kv_rule_name + " ( \",\" space " + kv_rule_name + " )*"
-                    );
-                } else if (first_is_optional) {
-                    res = "( \",\" space " + kv_rule_name + " )?";
-                } else {
-                    res = kv_rule_name;
-                }
-                if (ks.size() > 1) {
-                    res += " " + _add_rule(
-                        name + (name.empty() ? "" : "-") + k + "-rest",
-                        get_recursive_refs(std::vector<std::string>(ks.begin() + 1, ks.end()), true)
-                    );
-                }
-                return res;
-            };
-
-            for (size_t i = 0; i < optional_props.size(); i++) {
-                if (i > 0) {
-                    rule += " | ";
-                }
-                rule += get_recursive_refs(std::vector<std::string>(optional_props.begin() + i, optional_props.end()), false);
-            }
-            if (!required_props.empty()) {
-                rule += " )";
-            }
-            rule += " )?";
-        }
-
-        rule += " \"}\" space";
-
-        return rule;
-    }
-
-public:
-    SchemaConverter(
-        const std::function<json(const std::string &)> & fetch_json,
-        bool dotall)
-          : _fetch_json(fetch_json), _dotall(dotall)
-    {
-        _rules["space"] = SPACE_RULE;
-    }
-
-    void resolve_refs(json & schema, const std::string & url) {
-        /*
-        * Resolves all $ref fields in the given schema, fetching any remote schemas,
-        * replacing each $ref with absolute reference URL and populates _refs with the
-        * respective referenced (sub)schema dictionaries.
-        */
-        std::function<void(json &)> visit_refs = [&](json & n) {
-            if (n.is_array()) {
-                for (auto & x : n) {
-                    visit_refs(x);
-                }
-            } else if (n.is_object()) {
-                if (n.contains("$ref")) {
-                    std::string ref = n["$ref"];
-                    if (_refs.find(ref) == _refs.end()) {
-                        json target;
-                        if (ref.find("https://") == 0) {
-                            std::string base_url = ref.substr(0, ref.find('#'));
-                            auto it = _refs.find(base_url);
-                            if (it != _refs.end()) {
-                                target = it->second;
-                            } else {
-                                // Fetch the referenced schema and resolve its refs
-                                auto referenced = _fetch_json(ref);
-                                resolve_refs(referenced, base_url);
-                                _refs[base_url] = referenced;
-                            }
-                            if (ref.find('#') == std::string::npos || ref.substr(ref.find('#') + 1).empty()) {
-                                return;
-                            }
-                        } else if (ref.find("#/") == 0) {
-                            target = schema;
-                            n["$ref"] = url + ref;
-                            ref = url + ref;
-                        } else {
-                            _errors.push_back("Unsupported ref: " + ref);
-                            return;
-                        }
-                        std::string pointer = ref.substr(ref.find('#') + 1);
-                        std::vector<std::string> tokens = split(pointer, "/");
-                        for (size_t i = 1; i < tokens.size(); ++i) {
-                            std::string sel = tokens[i];
-                            if (target.is_null() || !target.contains(sel)) {
-                                _errors.push_back("Error resolving ref " + ref + ": " + sel + " not in " + target.dump());
-                                return;
-                            }
-                            target = target[sel];
-                        }
-                        _refs[ref] = target;
-                    }
-                } else {
-                    for (auto & kv : n.items()) {
-                        visit_refs(kv.value());
-                    }
-                }
-            }
-        };
-
-        visit_refs(schema);
-    }
-
-    std::string _generate_constant_rule(const json & value) {
-        return format_literal(value.dump());
-    }
-
-    std::string visit(const json & schema, const std::string & name) {
-        json schema_type = schema.contains("type") ? schema["type"] : json();
-        std::string schema_format = schema.contains("format") ? schema["format"].get<std::string>() : "";
-        std::string rule_name = is_reserved_name(name) ? name + "-" : name.empty() ? "root" : name;
-
-        if (schema.contains("$ref")) {
-            return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
-        } else if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
-            return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
-        } else if (schema_type.is_array()) {
-            std::vector<json> schema_types;
-            for (const auto & t : schema_type) {
-                schema_types.push_back({{"type", t}});
-            }
-            return _add_rule(rule_name, _generate_union_rule(name, schema_types));
-        } else if (schema.contains("const")) {
-            return _add_rule(rule_name, _generate_constant_rule(schema["const"]));
-        } else if (schema.contains("enum")) {
-            std::vector<std::string> enum_values;
-            for (const auto & v : schema["enum"]) {
-                enum_values.push_back(_generate_constant_rule(v));
-            }
-            return _add_rule(rule_name, join(enum_values.begin(), enum_values.end(), " | "));
-        } else if ((schema_type.is_null() || schema_type == "object")
-                && (schema.contains("properties") ||
-                    (schema.contains("additionalProperties") && schema["additionalProperties"] != true))) {
-            std::unordered_set<std::string> required;
-            if (schema.contains("required") && schema["required"].is_array()) {
-                for (const auto & item : schema["required"]) {
-                    if (item.is_string()) {
-                        required.insert(item.get<std::string>());
-                    }
-                }
-            }
-            std::vector<std::pair<std::string, json>> properties;
-            if (schema.contains("properties")) {
-                for (const auto & prop : schema["properties"].items()) {
-                    properties.emplace_back(prop.key(), prop.value());
-                }
-            }
-            return _add_rule(rule_name,
-                _build_object_rule(
-                    properties, required, name,
-                    schema.contains("additionalProperties") ? schema["additionalProperties"] : json()));
-        } else if ((schema_type.is_null() || schema_type == "object") && schema.contains("allOf")) {
-            std::unordered_set<std::string> required;
-            std::vector<std::pair<std::string, json>> properties;
-            std::string hybrid_name = name;
-            std::function<void(const json &, bool)> add_component = [&](const json & comp_schema, bool is_required) {
-                if (comp_schema.contains("$ref")) {
-                    add_component(_refs[comp_schema["$ref"]], is_required);
-                } else if (comp_schema.contains("properties")) {
-                    for (const auto & prop : comp_schema["properties"].items()) {
-                        properties.emplace_back(prop.key(), prop.value());
-                        if (is_required) {
-                            required.insert(prop.key());
-                        }
-                    }
-                } else {
-                  // todo warning
-                }
-            };
-            for (auto & t : schema["allOf"]) {
-                if (t.contains("anyOf")) {
-                    for (auto & tt : t["anyOf"]) {
-                        add_component(tt, false);
-                    }
-                } else {
-                    add_component(t, true);
-                }
-            }
-            return _add_rule(rule_name, _build_object_rule(properties, required, hybrid_name, json()));
-        } else if ((schema_type.is_null() || schema_type == "array") && (schema.contains("items") || schema.contains("prefixItems"))) {
-            json items = schema.contains("items") ? schema["items"] : schema["prefixItems"];
-            if (items.is_array()) {
-                std::string rule = "\"[\" space ";
-                for (size_t i = 0; i < items.size(); i++) {
-                    if (i > 0) {
-                        rule += " \",\" space ";
-                    }
-                    rule += visit(items[i], name + (name.empty() ? "" : "-") + "tuple-" + std::to_string(i));
-                }
-                rule += " \"]\" space";
-                return _add_rule(rule_name, rule);
-            } else {
-                std::string item_rule_name = visit(items, name + (name.empty() ? "" : "-") + "item");
-                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);
-            }
-        } 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";
-        } 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");
-        } 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>()));
-        }
-    }
-
-    void check_errors() {
-        if (!_errors.empty()) {
-            throw std::runtime_error("JSON schema conversion failed:\n" + join(_errors.begin(), _errors.end(), "\n"));
-        }
-        if (!_warnings.empty()) {
-            fprintf(stderr, "WARNING: JSON schema conversion was incomplete: %s\n", join(_warnings.begin(), _warnings.end(), "; ").c_str());
-        }
-    }
-
-    std::string format_grammar() {
-        std::stringstream ss;
-        for (const auto & kv : _rules) {
-            ss << kv.first << " ::= " << kv.second << std::endl;
-        }
-        return ss.str();
-    }
-};
-
-std::string json_schema_to_grammar(const json & schema) {
-    SchemaConverter converter([](const std::string &) { return json::object(); }, /* dotall= */ false);
-    auto copy = schema;
-    converter.resolve_refs(copy, "input");
-    converter.visit(copy, "");
-    converter.check_errors();
-    return converter.format_grammar();
-}

common/json-schema-to-grammar.h

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

common/log.h

@@ -234,7 +234,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 // INTERNAL, DO NOT USE
 //  USE LOG() INSTEAD
 //
-#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
+#ifndef _MSC_VER
     #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
 //
-#if !defined(_MSC_VER) or defined(__INTEL_LLVM_COMPILER)
+#ifndef _MSC_VER
     #define LOG_TEE_IMPL(str, ...)                                                                                                      \
     do {                                                                                                                                \
         if (LOG_TARGET != nullptr)                                                                                                      \
@@ -297,7 +297,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG(...) LOG_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // Main TEE macro.
@@ -311,7 +311,7 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
 #ifndef _MSC_VER
     #define LOG_TEE(...) LOG_TEE_IMPL(__VA_ARGS__, "")
 #else
-    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "")
+    #define LOG_TEE(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "")
 #endif
 
 // LOG macro variants with auto endline.
@@ -319,8 +319,8 @@ inline std::string log_filename_generator_impl(LogTriState multilog, const std::
     #define LOGLN(...) LOG_IMPL(__VA_ARGS__, "\n")
     #define LOG_TEELN(...) LOG_TEE_IMPL(__VA_ARGS__, "\n")
 #else
-    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
-    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", ##__VA_ARGS__, "\n")
+    #define LOGLN(str, ...) LOG_IMPL("%s" str, "", __VA_ARGS__, "\n")
+    #define LOG_TEELN(str, ...) LOG_TEE_IMPL("%s" str, "", __VA_ARGS__, "\n")
 #endif
 
 // INTERNAL, DO NOT USE
@@ -566,7 +566,6 @@ 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

@@ -1,282 +0,0 @@
-#include "ngram-cache.h"
-#include "common.h"
-#include "log.h"
-
-#include <cstdint>
-#include <fstream>
-
-void llama_ngram_cache_update(llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max,
-                              std::vector<llama_token> & inp, int nnew, bool print_progress) {
-    const int64_t t_start_ms = ggml_time_ms();
-    const int64_t inp_size = inp.size();
-
-    const int64_t n_todo = inp_size * (ngram_max - ngram_min + 1);
-    int64_t n_done = 0;
-
-    for (int64_t ngram_size = ngram_min; ngram_size <= ngram_max; ++ngram_size) {
-        const int64_t i_start = std::max(inp_size - nnew, ngram_size);
-        for (int64_t i = i_start; i < inp_size; ++i) {
-            const int64_t ngram_start = i - ngram_size;
-            llama_ngram ngram(&inp[ngram_start], ngram_size);
-            const llama_token token = inp[i];
-
-            llama_ngram_cache::iterator part_it = ngram_cache.find(ngram);
-            if (part_it == ngram_cache.end()) {
-                llama_ngram_cache_part part;
-                part.emplace(token, 1);
-                ngram_cache.emplace(ngram, part);
-            } else {
-                llama_ngram_cache_part::iterator token_count_it = part_it->second.find(token);
-                if (token_count_it == part_it->second.end()) {
-                    part_it->second.emplace(token, 1);
-                } else {
-                    token_count_it->second++;
-                }
-            }
-            ++n_done;
-
-            if (print_progress && n_done % 10000000 == 0) {
-                const int64_t t_now_ms = ggml_time_ms();
-                const int64_t eta_ms   = (inp_size*(ngram_max-ngram_min+1) - n_done) * (t_now_ms - t_start_ms) / n_done;
-                const int64_t eta_min  = eta_ms / (60*1000);
-                const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
-
-                fprintf(stderr, "%s: %" PRId64 "/%" PRId64 " done, ETA: %02" PRId64 ":%02" PRId64 "\n", __func__, n_done, n_todo, eta_min, eta_s);
-            }
-        }
-    }
-}
-
-// Helper function to get a token from the combined, speculative sequence of inp and draft.
-static llama_token get_token(const std::vector<llama_token> & inp, const std::vector<llama_token> & draft, const size_t i) {
-    return i < inp.size() ? inp[i] : draft[1 + i - inp.size()];
-}
-
-// If sample size or percentage are below these thresholds the draft is aborted early:
-constexpr int    draft_min_sample_size_lax[LLAMA_NGRAM_MAX] = { 2,  2,  1,  1};
-constexpr int        draft_min_percent_lax[LLAMA_NGRAM_MAX] = {66, 50, 50, 50};
-constexpr int draft_min_sample_size_strict[LLAMA_NGRAM_MAX] = { 4,  3,  2,  2};
-constexpr int     draft_min_percent_strict[LLAMA_NGRAM_MAX] = {75, 66, 66, 66};
-
-// Helper function that tries to draft a token from only the static ngram cache:
-static llama_token try_draft(llama_ngram_cache & nc_static, const llama_ngram ngram_static) {
-    llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
-    if (part_static_it == nc_static.end()) {
-        return -1;
-    }
-    const llama_ngram_cache_part part_static = part_static_it->second;
-
-    int max_count_static  = 0;
-    int sum_count_static  = 0;
-    llama_token max_token = -1;
-
-    for (std::pair<llama_token, int> token_count_static : part_static) {
-        const llama_token token = token_count_static.first;
-        const int32_t count_static  = token_count_static.second;
-
-        if (count_static > max_count_static) {
-            max_token        = token;
-            max_count_static = count_static;
-        }
-        sum_count_static += count_static;
-    }
-
-    if (sum_count_static < draft_min_sample_size_lax[LLAMA_NGRAM_STATIC-1]) {
-        return -1;
-    }
-    if (100*max_count_static < draft_min_percent_lax[LLAMA_NGRAM_STATIC-1]*sum_count_static) {
-        return -1;
-    }
-    return max_token;
-}
-
-// Try to draft a token from primary cache (context/dynamic), validate with static cache:
-static llama_token try_draft(
-    llama_ngram_cache & nc_primary, const std::vector<llama_ngram> & ngrams_primary, llama_ngram_cache_part & part_static,
-    const int * min_sample_size, const int * min_percent) {
-
-    llama_token drafted_token = -1;
-
-    for (int i = ngrams_primary.size()-1; i >= 0 && drafted_token == -1; --i) {
-        const llama_ngram ngram_primary = ngrams_primary[i];
-
-        llama_ngram_cache::iterator part_primary_it = nc_primary.find(ngram_primary);
-        if (part_primary_it == nc_primary.end()) {
-            continue;
-        }
-        const llama_ngram_cache_part part_primary = part_primary_it->second;
-
-        int max_count_primary = 0;
-        int max_count_static  = 0;
-        int sum_count_primary = 0;
-        llama_token max_token = -1;
-
-        for (std::pair<llama_token, int> token_count_primary : part_primary) {
-            const llama_token token = token_count_primary.first;
-
-            llama_ngram_cache_part::iterator token_count_static_it = part_static.find(token);
-
-            const int32_t count_primary = token_count_primary.second;
-            const int32_t count_static  = token_count_static_it != part_static.end() ? 100*token_count_static_it->second : 1;
-
-            if (count_primary*count_static > max_count_primary*max_count_static) {
-                max_token         = token;
-                max_count_primary = count_primary;
-                max_count_static  = count_static;
-            }
-            sum_count_primary += count_primary;
-        }
-
-        if (sum_count_primary < min_sample_size[i]) {
-            continue;
-        }
-        if (100*max_count_primary < min_percent[i]*sum_count_primary) {
-            continue;;
-        }
-        drafted_token = max_token;
-    }
-
-    return drafted_token;
-}
-
-void llama_ngram_cache_draft(
-    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static
-) {
-    GGML_ASSERT(draft.size() == 1);
-    const int inp_size = inp.size();
-
-    if (inp_size < LLAMA_NGRAM_STATIC) {
-        return;
-    }
-
-    while ((int) draft.size()-1 < n_draft) {
-        llama_token drafted_token = -1;
-
-        const int ngram_start_static = inp_size-LLAMA_NGRAM_STATIC + draft.size()-1;
-        llama_ngram ngram_static;
-        for (int j = ngram_start_static; j < ngram_start_static + LLAMA_NGRAM_STATIC; ++j) {
-            ngram_static.tokens[j-ngram_start_static] = get_token(inp, draft, j);
-        }
-        llama_ngram_cache::iterator part_static_it = nc_static.find(ngram_static);
-        llama_ngram_cache_part part_static;
-        if (part_static_it != nc_static.end()) {
-            part_static = part_static_it->second;
-        }
-
-        // cd = context + dynamic
-        std::vector<llama_ngram> ngrams_cd;
-        for (int ngram_size_cd = ngram_min; ngram_size_cd <= ngram_max; ++ngram_size_cd) {
-            const int ngram_start_cd = inp_size-ngram_size_cd + draft.size()-1;
-            llama_ngram ngram_cd;
-            for (int j = ngram_start_cd; j < ngram_start_cd + ngram_size_cd; ++j) {
-                ngram_cd.tokens[j-ngram_start_cd] = get_token(inp, draft, j);
-            }
-            ngrams_cd.push_back(ngram_cd);
-        }
-        if (drafted_token == -1) {
-            drafted_token = try_draft(nc_context, ngrams_cd, part_static, draft_min_sample_size_lax, draft_min_percent_lax);
-        }
-        if (drafted_token == -1) {
-            drafted_token = try_draft(nc_dynamic, ngrams_cd, part_static, draft_min_sample_size_strict, draft_min_percent_strict);
-        }
-        if (drafted_token == -1) {
-            drafted_token = try_draft(nc_static, ngram_static);
-        }
-
-        if (drafted_token == -1) {
-            break;
-        }
-
-        LOG(" - draft candidate: token=%d\n", drafted_token);
-        draft.push_back(drafted_token);
-    }
-}
-
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename) {
-    std::ofstream file_out(filename, std::ios::binary);
-    for (std::pair<llama_ngram, llama_ngram_cache_part> item : ngram_cache) {
-        const llama_ngram      ngram        = item.first;
-        llama_ngram_cache_part token_counts = item.second;
-        GGML_ASSERT(!token_counts.empty());
-        const int32_t ntokens = token_counts.size();
-        GGML_ASSERT(ntokens > 0);
-
-        file_out.write(reinterpret_cast<const char *>(&ngram),   sizeof(llama_ngram));
-        file_out.write(reinterpret_cast<const char *>(&ntokens), sizeof(int32_t));
-        for (std::pair<llama_token, int32_t> item2 : token_counts) {
-            const llama_token token = item2.first;
-            const int32_t     count = item2.second;
-            GGML_ASSERT(count > 0);
-
-            file_out.write(reinterpret_cast<const char *>(&token), sizeof(llama_token));
-            file_out.write(reinterpret_cast<const char *>(&count), sizeof(int32_t));
-        }
-    }
-
-}
-
-llama_ngram_cache llama_ngram_cache_load(std::string & filename) {
-    std::ifstream hashmap_file(filename, std::ios::binary);
-    if (!hashmap_file) {
-        throw std::ifstream::failure("Unable to open file " + filename);
-    }
-    llama_ngram_cache ngram_cache;
-
-    llama_ngram ngram;
-    int32_t     ntokens;
-    llama_token token;
-    int32_t     count;
-
-    char * ngramc   = reinterpret_cast<char*>(&ngram);
-    char * ntokensc = reinterpret_cast<char*>(&ntokens);
-    char * tokenc   = reinterpret_cast<char*>(&token);
-    char * countc   = reinterpret_cast<char*>(&count);
-    while(hashmap_file.read(ngramc, sizeof(llama_ngram))) {
-        GGML_ASSERT(!hashmap_file.eof());
-        GGML_ASSERT(hashmap_file.read(ntokensc, sizeof(int32_t)));
-        GGML_ASSERT(ntokens > 0);
-        llama_ngram_cache_part token_counts;
-
-        for (int i = 0; i < ntokens; ++i) {
-            GGML_ASSERT(!hashmap_file.eof());
-            GGML_ASSERT(hashmap_file.read(tokenc, sizeof(llama_token)));
-            GGML_ASSERT(!hashmap_file.eof());
-            GGML_ASSERT(hashmap_file.read(countc, sizeof(int32_t)));
-            GGML_ASSERT(count > 0);
-            token_counts.emplace(token, count);
-        }
-
-        ngram_cache.emplace(ngram, token_counts);
-    }
-    GGML_ASSERT(hashmap_file.eof());
-
-    return ngram_cache;
-}
-
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add) {
-    for (std::pair<llama_ngram, llama_ngram_cache_part> ngram_part : ngram_cache_add) {
-        const llama_ngram      ngram = ngram_part.first;
-        llama_ngram_cache_part  part = ngram_part.second;
-
-        llama_ngram_cache::iterator part_merged_it = ngram_cache_target.find(ngram);
-        if (part_merged_it == ngram_cache_target.end()) {
-            ngram_cache_target.emplace(ngram, part);
-            continue;
-        }
-
-        for (std::pair<llama_token, int32_t> token_count : part) {
-            const llama_token token = token_count.first;
-            const int32_t     count = token_count.second;
-            GGML_ASSERT(count > 0);
-
-            llama_ngram_cache_part::iterator token_count_merged_it = part_merged_it->second.find(token);
-            if (token_count_merged_it == part_merged_it->second.end()) {
-                part_merged_it->second.emplace(token, count);
-                continue;
-            }
-
-            token_count_merged_it->second += count;
-        }
-    }
-}

common/ngram-cache.h

@@ -1,94 +0,0 @@
-#pragma once
-
-#include "llama.h"
-
-#include <unordered_map>
-#include <string>
-#include <vector>
-
-#define LLAMA_NGRAM_MIN    1
-#define LLAMA_NGRAM_MAX    4
-#define LLAMA_NGRAM_STATIC 2
-
-// Data structures to map n-grams to empirical token probabilities:
-
-struct llama_ngram {
-    llama_token tokens[LLAMA_NGRAM_MAX];
-
-    llama_ngram() {
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            tokens[i] = -1;
-        }
-    }
-
-    llama_ngram(const llama_token * input, const int ngram_size) {
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            tokens[i] = i < ngram_size ? input[i] : -1;
-        }
-    }
-
-    bool operator==(const llama_ngram & other) const {
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            if (tokens[i] != other.tokens[i]) {
-                return false;
-            }
-        }
-        return true;
-    }
-};
-
-struct llama_ngram_hash_function {
-    size_t operator()(const llama_ngram & ngram) const {
-        size_t hash = 0;
-        for (int i = 0; i < LLAMA_NGRAM_MAX; ++i) {
-            hash ^= std::hash<llama_token>{}(ngram.tokens[i]);
-        }
-        return hash;
-    }
-};
-
-// token -> number of times token has been seen
-typedef std::unordered_map<llama_token, int32_t> llama_ngram_cache_part;
-
-// n-gram -> empirical distribution of following tokens
-typedef std::unordered_map<llama_ngram, llama_ngram_cache_part, llama_ngram_hash_function> llama_ngram_cache;
-
-
-// Update an ngram cache with tokens.
-// ngram_cache:         the cache to modify.
-// ngram_min/ngram_max: the min/max size of the ngrams to extract from inp_data.
-// inp_data:            the token sequence with which to update ngram_cache.
-// nnew:                how many new tokens have been appended to inp_data since the last call to this function.
-// print_progress:      whether to print progress to stderr.
-//
-// In order to get correct results inp_data can ONLY BE APPENDED TO.
-// Changes in the middle need a complete rebuild.
-void llama_ngram_cache_update(
-    llama_ngram_cache & ngram_cache, int ngram_min, int ngram_max, std::vector<llama_token> & inp_data, int nnew, bool print_progress);
-
-// Try to draft tokens from ngram caches.
-// inp:                the tokens generated so far.
-// draft:              the token sequence to draft. Expected to initially contain the previously sampled token.
-// n_draft:            maximum number of tokens to add to draft.
-// ngram_min/gram_max: the min/max size of the ngrams in nc_context and nc_dynamic.
-// nc_context:         ngram cache based on current context.
-// nc_dynamic:         ngram cache based on previous user generations.
-// nc_static:          ngram cache generated from a large text corpus, used for validation.
-void llama_ngram_cache_draft(
-    std::vector<llama_token> & inp, std::vector<llama_token> & draft, int n_draft, int ngram_min, int ngram_max,
-    llama_ngram_cache & nc_context, llama_ngram_cache & nc_dynamic, llama_ngram_cache & nc_static);
-
-// Save an ngram cache to a file.
-// ngram_cache: the ngram cache to save.
-// filename:    the path under which to save the ngram cache.
-void llama_ngram_cache_save(llama_ngram_cache & ngram_cache, std::string & filename);
-
-// Load an ngram cache saved with llama_ngram_cache_save.
-// filename: the path from which to load the ngram cache.
-// returns:  an ngram cache containing the information saved to filename.
-llama_ngram_cache llama_ngram_cache_load(std::string & filename);
-
-// Merge two ngram caches.
-// ngram_cache_target: the ngram cache to which to add the information from ngram_cache_add.
-// ngram_cache_add:    the ngram cache to add to ngram_cache_target.
-void llama_ngram_cache_merge(llama_ngram_cache & ngram_cache_target, llama_ngram_cache & ngram_cache_add);

common/sampling.cpp

@@ -17,13 +17,6 @@ struct llama_sampling_context * llama_sampling_init(const struct llama_sampling_
             return nullptr;
         }
 
-        // Ensure that there is a "root" node.
-        if (result->parsed_grammar.symbol_ids.find("root") == result->parsed_grammar.symbol_ids.end()) {
-            fprintf(stderr, "%s: grammar does not contain a 'root' symbol\n", __func__);
-            delete result;
-            return nullptr;
-        }
-
         std::vector<const llama_grammar_element *> grammar_rules(result->parsed_grammar.c_rules());
 
         result->grammar = llama_grammar_init(
@@ -128,7 +121,7 @@ static void sampler_queue(
                    struct llama_context * ctx_main,
             const llama_sampling_params & params,
                  llama_token_data_array & cur_p,
-                                 size_t   min_keep) {
+                                 size_t & min_keep) {
     const float         temp              = params.temp;
     const float         dynatemp_range    = params.dynatemp_range;
     const float         dynatemp_exponent = params.dynatemp_exponent;
@@ -146,7 +139,7 @@ static void sampler_queue(
             case llama_sampler_type::TYPICAL_P: llama_sample_typical  (ctx_main, &cur_p, typical_p, min_keep); break;
             case llama_sampler_type::TOP_P    : llama_sample_top_p    (ctx_main, &cur_p, top_p,     min_keep); break;
             case llama_sampler_type::MIN_P    : llama_sample_min_p    (ctx_main, &cur_p, min_p,     min_keep); break;
-            case llama_sampler_type::TEMPERATURE:
+            case llama_sampler_type::TEMP:
                 if (dynatemp_range > 0) {
                     float dynatemp_min = std::max(0.0f, temp - dynatemp_range);
                     float dynatemp_max = std::max(0.0f, temp + dynatemp_range);
@@ -168,110 +161,32 @@ 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 float   temp            = params.temp;
-    const int     mirostat        = params.mirostat;
-    const float   mirostat_tau    = params.mirostat_tau;
-    const float   mirostat_eta    = params.mirostat_eta;
-
-    std::vector<float> original_logits;
-    auto cur_p = llama_sampling_prepare(ctx_sampling, ctx_main, ctx_cfg, idx, !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);
-
-    if (temp < 0.0) {
-        // greedy sampling, with probs
-        llama_sample_softmax(ctx_main, &cur_p);
-        id = cur_p.data[0].id;
-    } else if (temp == 0.0) {
-        // greedy sampling, no probs
-        id = llama_sample_token_greedy(ctx_main, &cur_p);
-    } else {
-        if (mirostat == 1) {
-            const int mirostat_m = 100;
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
-        } else if (mirostat == 2) {
-            llama_sample_temp(ctx_main, &cur_p, temp);
-            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
-        } else {
-            // temperature sampling
-            size_t min_keep = std::max(1, params.min_keep);
-
-            sampler_queue(ctx_main, params, cur_p, min_keep);
-
-            id = llama_sample_token(ctx_main, &cur_p);
-
-            //{
-            //    const int n_top = 10;
-            //    LOG("top %d candidates:\n", n_top);
-
-            //    for (int i = 0; i < n_top; i++) {
-            //        const llama_token id = cur_p.data[i].id;
-            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
-            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
-            //    }
-            //}
-
-            //LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
-        }
-    }
-
-    if (ctx_sampling->grammar != NULL && !is_resampling) {
-        // Create an array with a single token data element for the sampled id
-        llama_token_data single_token_data = {id, logits[id], 0.0f};
-        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
-
-        // Apply grammar constraints to the single token
-        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
-
-        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
-        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
-
-        // If the token is not valid according to the grammar, perform resampling
-        if (!is_valid) {
-            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
-
-            // Restore logits from the copy
-            std::copy(original_logits.begin(), original_logits.end(), logits);
-
-            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
-        }
-    }
-
-    return id;
-}
-
-static llama_token_data_array llama_sampling_prepare_impl(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool apply_grammar,
-                  std::vector<float> * original_logits) {
-    const llama_sampling_params & params = ctx_sampling->params;
-
     const int n_vocab = llama_n_vocab(llama_get_model(ctx_main));
 
+    const 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;
 
+    llama_token id = 0;
+
     // Get a pointer to the logits
     float * logits = llama_get_logits_ith(ctx_main, idx);
 
-    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))};
+    // 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
@@ -312,12 +227,72 @@ static llama_token_data_array llama_sampling_prepare_impl(
         }
     }
 
-    // apply grammar checks before sampling logic
-    if (apply_grammar && ctx_sampling->grammar != NULL) {
+    // 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);
     }
 
-    return cur_p;
+    if (temp < 0.0) {
+        // greedy sampling, with probs
+        llama_sample_softmax(ctx_main, &cur_p);
+        id = cur_p.data[0].id;
+    } else if (temp == 0.0) {
+        // greedy sampling, no probs
+        id = llama_sample_token_greedy(ctx_main, &cur_p);
+    } else {
+        if (mirostat == 1) {
+            const int mirostat_m = 100;
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat(ctx_main, &cur_p, mirostat_tau, mirostat_eta, mirostat_m, &ctx_sampling->mirostat_mu);
+        } else if (mirostat == 2) {
+            llama_sample_temp(ctx_main, &cur_p, temp);
+            id = llama_sample_token_mirostat_v2(ctx_main, &cur_p, mirostat_tau, mirostat_eta, &ctx_sampling->mirostat_mu);
+        } else {
+            // temperature sampling
+            size_t min_keep = std::max(1, params.n_probs);
+
+            sampler_queue(ctx_main, params, cur_p, min_keep);
+
+            id = llama_sample_token(ctx_main, &cur_p);
+
+            //{
+            //    const int n_top = 10;
+            //    LOG("top %d candidates:\n", n_top);
+
+            //    for (int i = 0; i < n_top; i++) {
+            //        const llama_token id = cur_p.data[i].id;
+            //        (void)id; // To avoid a warning that id is unused when logging is disabled.
+            //        LOG(" - %5d: '%12s' (%.3f)\n", id, llama_token_to_piece(ctx_main, id).c_str(), cur_p.data[i].p);
+            //    }
+            //}
+
+            LOG("sampled token: %5d: '%s'\n", id, llama_token_to_piece(ctx_main, id).c_str());
+        }
+    }
+
+    if (ctx_sampling->grammar != NULL && !is_resampling) {
+        // Create an array with a single token data element for the sampled id
+        llama_token_data single_token_data = {id, logits[id], 0.0f};
+        llama_token_data_array single_token_data_array = { &single_token_data, 1, false };
+
+        // Apply grammar constraints to the single token
+        llama_sample_grammar(ctx_main, &single_token_data_array, ctx_sampling->grammar);
+
+        // Check if the token is valid according to the grammar by seeing if its logit has been set to -INFINITY
+        bool is_valid = single_token_data_array.data[0].logit != -INFINITY;
+
+        // If the token is not valid according to the grammar, perform resampling
+        if (!is_valid) {
+            LOG("Resampling because token %d: '%s' does not meet grammar rules\n", id, llama_token_to_piece(ctx_main, id).c_str());
+
+            // Restore logits from the copy
+            std::copy(original_logits.begin(), original_logits.end(), logits);
+
+            return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, true);  // Pass true for is_resampling
+        }
+    }
+
+    return id;
 }
 
 llama_token llama_sampling_sample(
@@ -329,16 +304,6 @@ llama_token llama_sampling_sample(
     return llama_sampling_sample_impl(ctx_sampling, ctx_main, ctx_cfg, idx, false);
 }
 
-llama_token_data_array llama_sampling_prepare(
-                  struct llama_sampling_context * ctx_sampling,
-                  struct llama_context * ctx_main,
-                  struct llama_context * ctx_cfg,
-                  const int idx,
-                  bool apply_grammar,
-                  std::vector<float> * original_logits) {
-    return llama_sampling_prepare_impl(ctx_sampling,ctx_main, ctx_cfg, idx, apply_grammar, original_logits);
-}
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/sampling.h

@@ -10,19 +10,18 @@
 
 // sampler types
 enum class llama_sampler_type : char {
-    TOP_K       = 'k',
-    TOP_P       = 'p',
-    MIN_P       = 'm',
-    TFS_Z       = 'f',
-    TYPICAL_P   = 'y',
-    TEMPERATURE = 't'
+    TOP_K     = 'k',
+    TOP_P     = 'p',
+    MIN_P     = 'm',
+    TFS_Z     = 'f',
+    TYPICAL_P = 'y',
+    TEMP      = 't'
 };
 
 // sampling parameters
 typedef struct llama_sampling_params {
     int32_t     n_prev                = 64;       // number of previous tokens to remember
     int32_t     n_probs               = 0;        // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t     min_keep              = 0;        // 0 = disabled, otherwise samplers should return at least min_keep tokens
     int32_t     top_k                 = 40;       // <= 0 to use vocab size
     float       top_p                 = 0.95f;    // 1.0 = disabled
     float       min_p                 = 0.05f;    // 0.0 = disabled
@@ -32,13 +31,13 @@ typedef struct llama_sampling_params {
     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_repeat        = 1.10f;    // 1.0 = disabled
     float       penalty_freq          = 0.00f;    // 0.0 = disabled
     float       penalty_present       = 0.00f;    // 0.0 = disabled
     int32_t     mirostat              = 0;        // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
     float       mirostat_tau          = 5.00f;    // target entropy
     float       mirostat_eta          = 0.10f;    // learning rate
-    bool        penalize_nl           = false;     // consider newlines as a repeatable token
+    bool        penalize_nl           = true;     // consider newlines as a repeatable token
 
     std::vector<llama_sampler_type> samplers_sequence = {
         llama_sampler_type::TOP_K,
@@ -46,7 +45,7 @@ typedef struct llama_sampling_params {
         llama_sampler_type::TYPICAL_P,
         llama_sampler_type::TOP_P,
         llama_sampler_type::MIN_P,
-        llama_sampler_type::TEMPERATURE
+        llama_sampler_type::TEMP
     };
 
     std::string grammar;  // optional BNF-like grammar to constrain sampling
@@ -131,15 +130,6 @@ llama_token llama_sampling_sample(
         struct llama_context * ctx_cfg,
         int idx = 0);
 
-// Prepares and adjusts the set of token candidates for sampling based on penalties, biases, and sampling parameters.
-llama_token_data_array llama_sampling_prepare(
-        struct llama_sampling_context * ctx_sampling,
-        struct llama_context * ctx_main,
-        struct llama_context * ctx_cfg,
-        int idx = 0,
-        bool apply_grammar = true,
-        std::vector<float> * original_logits = nullptr);
-
 void llama_sampling_accept(
         struct llama_sampling_context * ctx_sampling,
         struct llama_context * ctx_main,

common/train.cpp

@@ -31,7 +31,7 @@ struct train_state  * init_train_state() {
 
     state->opt = new struct ggml_opt_context;
     state->opt->ctx = NULL;
-    state->opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    state->opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     state->opt->params.graph_size = LLAMA_TRAIN_MAX_NODES;
     state->opt->loss_after = 0.0f;
 
@@ -556,7 +556,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
     std::string opt_type;
     GGUF_GET_KEY(fctx, opt_type, gguf_get_val_str, GGUF_TYPE_STRING, true, LLM_KV_OPTIMIZER_TYPE);
     if (opt_type == LLM_KV_OPTIMIZER_TYPE_ADAM) {
-        opt->params.type = GGML_OPT_TYPE_ADAM;
+        opt->params.type = GGML_OPT_ADAM;
 
         GGUF_GET_KEY(fctx, opt->adam.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS);
         GGUF_GET_KEY(fctx, opt->adam.fx_prev,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_ADAM_PREVIOUS_LOSS);
@@ -568,7 +568,7 @@ void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_g
         copy_tensor_by_name(opt->adam.v,  f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_SECOND_MOMENTS);
         copy_tensor_by_name(opt->adam.pf, f_ggml_ctx, LLM_TENSOR_OPTIMIZER_ADAM_PAST_LOSS_VALUES);
     } else if (opt_type == LLM_KV_OPTIMIZER_TYPE_LBFGS) {
-        opt->params.type = GGML_OPT_TYPE_LBFGS;
+        opt->params.type = GGML_OPT_LBFGS;
 
         GGUF_GET_KEY(fctx, opt->params.lbfgs.m,         gguf_get_val_u32, GGUF_TYPE_UINT32,  true, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT);
         GGUF_GET_KEY(fctx, opt->lbfgs.fx_best,          gguf_get_val_f32, GGUF_TYPE_FLOAT32, true, LLM_KV_OPTIMIZER_LBFGS_BEST_LOSS);
@@ -603,7 +603,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
     gguf_set_val_bool(fctx, LLM_KV_OPTIMIZER_JUST_INITIALIZED, opt->just_initialized);
 
     switch (opt->params.type) {
-        case GGML_OPT_TYPE_ADAM:
+        case GGML_OPT_ADAM:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_ADAM);
                 gguf_set_val_f32(fctx, LLM_KV_OPTIMIZER_ADAM_BEST_LOSS,            opt->adam.fx_best);
@@ -622,7 +622,7 @@ void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context *
                     gguf_add_tensor(fctx, opt->adam.pf);
                 }
             } break;
-        case GGML_OPT_TYPE_LBFGS:
+        case GGML_OPT_LBFGS:
             {
                 gguf_set_val_str(fctx, LLM_KV_OPTIMIZER_TYPE, LLM_KV_OPTIMIZER_TYPE_LBFGS);
                 gguf_set_val_u32(fctx, LLM_KV_OPTIMIZER_LBFGS_APPROX_HESSIAN_COUNT, opt->params.lbfgs.m);

convert-hf-to-gguf.py

@@ -8,10 +8,9 @@ import json
 import os
 import re
 import sys
-from abc import ABC, abstractmethod
 from enum import IntEnum
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterator, Sequence, TypeVar, cast
+from typing import TYPE_CHECKING, Any, ContextManager, Iterator, cast
 
 import numpy as np
 import torch
@@ -23,7 +22,16 @@ if 'NO_LOCAL_GGUF' not in os.environ:
     sys.path.insert(1, str(Path(__file__).parent / 'gguf-py'))
 import gguf
 
-from convert import LlamaHfVocab, permute
+from convert import HfVocab
+
+
+# check for any of the given keys in the dictionary and return the value of the first key found
+def get_key_opts(d, keys):
+    for k in keys:
+        if k in d:
+            return d[k]
+    print(f"Could not find any of {keys}")
+    sys.exit()
 
 
 ###### MODEL DEFINITIONS ######
@@ -37,12 +45,7 @@ class SentencePieceTokenTypes(IntEnum):
     BYTE = 6
 
 
-AnyModel = TypeVar("AnyModel", bound="type[Model]")
-
-
-class Model(ABC):
-    _model_classes: dict[str, type[Model]] = {}
-
+class Model:
     def __init__(self, dir_model: Path, ftype: int, fname_out: Path, is_big_endian: bool):
         self.dir_model = dir_model
         self.ftype = ftype
@@ -53,21 +56,8 @@ class Model(ABC):
         self.num_parts = Model.count_model_parts(self.dir_model, ".safetensors" if self.is_safetensors else ".bin")
         self.part_names = self._get_part_names()
         self.hparams = Model.load_hparams(self.dir_model)
+        self.model_arch = self._get_model_architecture()
         self.gguf_writer = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[self.model_arch], endianess=self.endianess, use_temp_file=False)
-        self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer"])
-
-    @property
-    @abstractmethod
-    def model_arch(self) -> gguf.MODEL_ARCH:
-        pass
-
-    def find_hparam(self, keys: Sequence[str], optional: bool = False) -> Any:
-        key = next((k for k in keys if k in self.hparams), None)
-        if key is not None:
-            return self.hparams[key]
-        if optional:
-            return None
-        raise KeyError(f"could not find any of: {keys}")
 
     def set_vocab(self):
         self._set_vocab_gpt2()
@@ -89,46 +79,28 @@ class Model(ABC):
 
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_block_count(self.block_count)
-
-        if (n_ctx := self.find_hparam(["max_position_embeddings", "n_ctx"], optional=True)) is not None:
+        self.gguf_writer.add_block_count(self.hparams.get(
+            "n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")),
+        ))
+        if (n_ctx := self.hparams.get("max_position_embeddings")) is not None:
             self.gguf_writer.add_context_length(n_ctx)
-            print(f"gguf: context length = {n_ctx}")
-
-        n_embd = self.find_hparam(["hidden_size", "n_embd"])
-        self.gguf_writer.add_embedding_length(n_embd)
-        print(f"gguf: embedding length = {n_embd}")
-
-        if (n_ff := self.find_hparam(["intermediate_size", "n_inner"], optional=True)) is not None:
+        if (n_embd := self.hparams.get("hidden_size")) is not None:
+            self.gguf_writer.add_embedding_length(n_embd)
+        if (n_ff := self.hparams.get("intermediate_size")) is not None:
             self.gguf_writer.add_feed_forward_length(n_ff)
-            print(f"gguf: feed forward length = {n_ff}")
-
-        n_head = self.find_hparam(["num_attention_heads", "n_head"])
-        self.gguf_writer.add_head_count(n_head)
-        print(f"gguf: head count = {n_head}")
-
+        if (n_head := self.hparams.get("num_attention_heads")) is not None:
+            self.gguf_writer.add_head_count(n_head)
         if (n_head_kv := self.hparams.get("num_key_value_heads")) is not None:
             self.gguf_writer.add_head_count_kv(n_head_kv)
-            print(f"gguf: key-value head count = {n_head_kv}")
 
-        if (rope_theta := self.hparams.get("rope_theta")) is not None:
-            self.gguf_writer.add_rope_freq_base(rope_theta)
-            print(f"gguf: rope theta = {rope_theta}")
-        if (f_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
-            self.gguf_writer.add_layer_norm_rms_eps(f_rms_eps)
-            print(f"gguf: rms norm epsilon = {f_rms_eps}")
-        if (f_norm_eps := self.find_hparam(["layer_norm_eps", "layer_norm_epsilon", "norm_epsilon"], optional=True)) is not None:
-            self.gguf_writer.add_layer_norm_eps(f_norm_eps)
-            print(f"gguf: layer norm epsilon = {f_norm_eps}")
+        if (n_rms_eps := self.hparams.get("rms_norm_eps")) is not None:
+            self.gguf_writer.add_layer_norm_rms_eps(n_rms_eps)
         if (n_experts := self.hparams.get("num_local_experts")) is not None:
             self.gguf_writer.add_expert_count(n_experts)
-            print(f"gguf: expert count = {n_experts}")
         if (n_experts_used := self.hparams.get("num_experts_per_tok")) is not None:
             self.gguf_writer.add_expert_used_count(n_experts_used)
-            print(f"gguf: experts used count = {n_experts_used}")
 
-        self.gguf_writer.add_file_type(self.ftype)
-        print(f"gguf: file type = {self.ftype}")
+        self.gguf_writer.add_parallel_residual(self.hparams.get("use_parallel_residual", True))
 
     def write_tensors(self):
         block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
@@ -197,22 +169,49 @@ class Model(ABC):
         with open(dir_model / "config.json", "r", encoding="utf-8") as f:
             return json.load(f)
 
-    @classmethod
-    def register(cls, *names: str) -> Callable[[AnyModel], AnyModel]:
-        assert names
-
-        def func(modelcls: type[Model]):
-            for name in names:
-                cls._model_classes[name] = modelcls
-            return modelcls
-        return func
-
-    @classmethod
-    def from_model_architecture(cls, arch):
-        try:
-            return cls._model_classes[arch]
-        except KeyError:
-            raise NotImplementedError(f'Architecture {arch!r} not supported!') from None
+    @staticmethod
+    def from_model_architecture(model_architecture):
+        if model_architecture == "GPTNeoXForCausalLM":
+            return GPTNeoXModel
+        if model_architecture == "BloomForCausalLM":
+            return BloomModel
+        if model_architecture == "MPTForCausalLM":
+            return MPTModel
+        if model_architecture in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return BaichuanModel
+        if model_architecture in ("FalconForCausalLM", "RWForCausalLM"):
+            return FalconModel
+        if model_architecture == "GPTBigCodeForCausalLM":
+            return StarCoderModel
+        if model_architecture == "GPTRefactForCausalLM":
+            return RefactModel
+        if model_architecture == "PersimmonForCausalLM":
+            return PersimmonModel
+        if model_architecture in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return StableLMModel
+        if model_architecture == "QWenLMHeadModel":
+            return QwenModel
+        if model_architecture == "Qwen2ForCausalLM":
+            return Model
+        if model_architecture == "MixtralForCausalLM":
+            return MixtralModel
+        if model_architecture == "GPT2LMHeadModel":
+            return GPT2Model
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
+        if model_architecture == "PlamoForCausalLM":
+            return PlamoModel
+        if model_architecture == "CodeShellForCausalLM":
+            return CodeShellModel
+        if model_architecture == "OrionForCausalLM":
+            return OrionModel
+        if model_architecture == "InternLM2ForCausalLM":
+            return InternLM2Model
+        if model_architecture == "MiniCPMForCausalLM":
+            return MiniCPMModel
+        if model_architecture == "BertModel":
+            return BertModel
+        return Model
 
     def _is_model_safetensors(self) -> bool:
         return Model.count_model_parts(self.dir_model, ".safetensors") > 0
@@ -227,10 +226,55 @@ class Model(ABC):
             return ("pytorch_model.bin",)
         return (f"pytorch_model-{n:05}-of-{self.num_parts:05}.bin" for n in range(1, self.num_parts + 1))
 
+    def _get_model_architecture(self) -> gguf.MODEL_ARCH:
+        arch = self.hparams["architectures"][0]
+        if arch == "GPTNeoXForCausalLM":
+            return gguf.MODEL_ARCH.GPTNEOX
+        if arch == "BloomForCausalLM":
+            return gguf.MODEL_ARCH.BLOOM
+        if arch == "MPTForCausalLM":
+            return gguf.MODEL_ARCH.MPT
+        if arch in ("BaichuanForCausalLM", "BaiChuanForCausalLM"):
+            return gguf.MODEL_ARCH.BAICHUAN
+        if arch in ("FalconForCausalLM", "RWForCausalLM"):
+            return gguf.MODEL_ARCH.FALCON
+        if arch == "GPTBigCodeForCausalLM":
+            return gguf.MODEL_ARCH.STARCODER
+        if arch == "GPTRefactForCausalLM":
+            return gguf.MODEL_ARCH.REFACT
+        if arch == "PersimmonForCausalLM":
+            return gguf.MODEL_ARCH.PERSIMMON
+        if arch in ("StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM"):
+            return gguf.MODEL_ARCH.STABLELM
+        if arch == "QWenLMHeadModel":
+            return gguf.MODEL_ARCH.QWEN
+        if arch == "Qwen2ForCausalLM":
+            return gguf.MODEL_ARCH.QWEN2
+        if arch == "MixtralForCausalLM":
+            return gguf.MODEL_ARCH.LLAMA
+        if arch == "GPT2LMHeadModel":
+            return gguf.MODEL_ARCH.GPT2
+        if arch == "PhiForCausalLM":
+            return gguf.MODEL_ARCH.PHI2
+        if arch == "PlamoForCausalLM":
+            return gguf.MODEL_ARCH.PLAMO
+        if arch == "CodeShellForCausalLM":
+            return gguf.MODEL_ARCH.CODESHELL
+        if arch == "OrionForCausalLM":
+            return gguf.MODEL_ARCH.ORION
+        if arch == "InternLM2ForCausalLM":
+            return gguf.MODEL_ARCH.INTERNLM2
+        if arch == "MiniCPMForCausalLM":
+            return gguf.MODEL_ARCH.MINICPM
+        if arch == "BertModel":
+            return gguf.MODEL_ARCH.BERT
+
+        raise NotImplementedError(f'Architecture "{arch}" not supported!')
+
     def _set_vocab_gpt2(self):
         dir_model = self.dir_model
         hparams = self.hparams
-        tokens: list[str] = []
+        tokens: list[bytearray] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
@@ -243,7 +287,8 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
+                pad_token = f"[PAD{i}]".encode('utf-8')
+                tokens.append(bytearray(pad_token))
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -265,7 +310,7 @@ class Model(ABC):
     def _set_vocab_qwen(self):
         dir_model = self.dir_model
         hparams = self.hparams
-        tokens: list[str] = []
+        tokens: list[bytearray] = []
         toktypes: list[int] = []
 
         from transformers import AutoTokenizer
@@ -290,7 +335,8 @@ class Model(ABC):
 
         for i in range(vocab_size):
             if i not in reverse_vocab:
-                tokens.append(f"[PAD{i}]")
+                pad_token = f"[PAD{i}]".encode("utf-8")
+                tokens.append(bytearray(pad_token))
                 toktypes.append(gguf.TokenType.USER_DEFINED)
             elif reverse_vocab[i] in added_vocab:
                 tokens.append(reverse_vocab[i])
@@ -329,7 +375,7 @@ class Model(ABC):
         tokenizer = SentencePieceProcessor(str(tokenizer_path))
         vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())
 
-        for token_id in range(tokenizer.vocab_size()):
+        for token_id in range(vocab_size):
             piece = tokenizer.id_to_piece(token_id)
             text = piece.encode("utf-8")
             score = tokenizer.get_score(token_id)
@@ -354,13 +400,9 @@ class Model(ABC):
                 added_tokens_json = json.load(f)
 
                 for key in added_tokens_json:
-                    key = key.encode("utf-8")
-                    if key not in tokens:
-                        tokens.append(key)
-                        scores.append(-1000.0)
-                        toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
-
-        assert len(tokens) == vocab_size
+                    tokens.append(key.encode("utf-8"))
+                    scores.append(-1000.0)
+                    toktypes.append(SentencePieceTokenTypes.USER_DEFINED)
 
         self.gguf_writer.add_tokenizer_model("llama")
         self.gguf_writer.add_token_list(tokens)
@@ -370,8 +412,12 @@ class Model(ABC):
         special_vocab = gguf.SpecialVocab(self.dir_model, n_vocab=len(tokens))
         special_vocab.add_to_gguf(self.gguf_writer)
 
-    def _set_vocab_llama_hf(self):
-        vocab = LlamaHfVocab(self.dir_model)
+    def _set_vocab_hf(self):
+        path = self.dir_model
+        added_tokens_path = self.dir_model
+        vocab = HfVocab(
+            path, added_tokens_path if added_tokens_path.exists() else None
+        )
         tokens = []
         scores = []
         toktypes = []
@@ -392,10 +438,7 @@ class Model(ABC):
         special_vocab.add_to_gguf(self.gguf_writer)
 
 
-@Model.register("GPTNeoXForCausalLM")
 class GPTNeoXModel(Model):
-    model_arch = gguf.MODEL_ARCH.GPTNEOX
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
 
@@ -412,10 +455,7 @@ class GPTNeoXModel(Model):
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
 
 
-@Model.register("BloomForCausalLM")
 class BloomModel(Model):
-    model_arch = gguf.MODEL_ARCH.BLOOM
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name("Bloom")
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
@@ -507,10 +547,7 @@ class BloomModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("MPTForCausalLM")
 class MPTModel(Model):
-    model_arch = gguf.MODEL_ARCH.MPT
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
         self.gguf_writer.add_name(self.dir_model.name)
@@ -572,11 +609,13 @@ class MPTModel(Model):
 
             self.gguf_writer.add_tensor(new_name, data)
 
+            # note: MPT output is tied to (same as) wte in original model;
+            # for easier implementation in llama.cpp it's duplicated in GGUF, though :/
+            if new_name == "token_embd.weight":
+                self.gguf_writer.add_tensor("output.weight", data)
+
 
-@Model.register("OrionForCausalLM")
 class OrionModel(Model):
-    model_arch = gguf.MODEL_ARCH.ORION
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -607,8 +646,6 @@ class OrionModel(Model):
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
-        # note: config provides rms norm but it is actually layer norm
-        # ref:  https://huggingface.co/OrionStarAI/Orion-14B-Chat/blob/276a17221ce42beb45f66fac657a41540e71f4f5/modeling_orion.py#L570-L571
         self.gguf_writer.add_layer_norm_eps(self.hparams["rms_norm_eps"])
 
     def write_tensors(self):
@@ -655,10 +692,7 @@ class OrionModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("BaichuanForCausalLM", "BaiChuanForCausalLM")
 class BaichuanModel(Model):
-    model_arch = gguf.MODEL_ARCH.BAICHUAN
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -773,10 +807,7 @@ class BaichuanModel(Model):
         return weights[r * n_part:r * n_part + r, ...]
 
 
-@Model.register("FalconForCausalLM", "RWForCausalLM")
 class FalconModel(Model):
-    model_arch = gguf.MODEL_ARCH.FALCON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_hidden_layers")
         if block_count is None:
@@ -869,10 +900,7 @@ class FalconModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("GPTBigCodeForCausalLM")
 class StarCoderModel(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -887,10 +915,7 @@ class StarCoderModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
 
-@Model.register("GPTRefactForCausalLM")
 class RefactModel(Model):
-    model_arch = gguf.MODEL_ARCH.REFACT
-
     def set_gguf_parameters(self):
         hidden_dim = self.hparams["n_embd"]
         inner_dim = 4 * hidden_dim
@@ -974,10 +999,7 @@ class RefactModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("PersimmonForCausalLM")
 class PersimmonModel(Model):
-    model_arch = gguf.MODEL_ARCH.PERSIMMON
-
     def set_gguf_parameters(self):
         block_count = self.hparams.get("num_layers", self.hparams.get("num_hidden_layers"))
         head_count = self.hparams["num_attention_heads"]
@@ -1000,6 +1022,7 @@ class PersimmonModel(Model):
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])
         self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
+        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
 
     def set_vocab(self):
         self._set_vocab_sentencepiece()
@@ -1025,10 +1048,7 @@ class PersimmonModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("StableLmForCausalLM", "StableLMEpochForCausalLM", "LlavaStableLMEpochForCausalLM")
 class StableLMModel(Model):
-    model_arch = gguf.MODEL_ARCH.STABLELM
-
     def set_vocab(self):
         if (self.dir_model / "tokenizer.json").is_file():
             self._set_vocab_gpt2()
@@ -1045,100 +1065,18 @@ class StableLMModel(Model):
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        rotary_factor = self.find_hparam(["partial_rotary_factor", "rope_pct"])
-        self.gguf_writer.add_rope_dimension_count(int(rotary_factor * (hparams["hidden_size"] // hparams["num_attention_heads"])))
+        self.gguf_writer.add_rope_dimension_count(int(hparams["rope_pct"] * (hparams["hidden_size"] // hparams["num_attention_heads"])))
         self.gguf_writer.add_head_count(hparams["num_attention_heads"])
         self.gguf_writer.add_parallel_residual(hparams["use_parallel_residual"] if "use_parallel_residual" in hparams else True)
-        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_eps", "norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(1e-5)
 
 
-@Model.register("LlamaForCausalLM", "MistralForCausalLM", "MixtralForCausalLM")
-class LlamaModel(Model):
-    model_arch = gguf.MODEL_ARCH.LLAMA
-
-    def set_vocab(self):
-        try:
-            self. _set_vocab_sentencepiece()
-        except FileNotFoundError:
-            self._set_vocab_llama_hf()
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        hparams = self.hparams
-        self.gguf_writer.add_vocab_size(hparams["vocab_size"])
-        self.gguf_writer.add_rope_dimension_count(hparams["hidden_size"] // hparams["num_attention_heads"])
-
-    # Same as super class, but permuting q_proj, k_proj
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-        n_head = self.hparams.get("num_attention_heads")
-        n_kv_head = self.hparams.get("num_key_value_heads")
-        for name, data_torch in self.get_tensors():
-            # we don't need these
-            if name.endswith((".attention.masked_bias", ".attention.bias", ".attention.rotary_emb.inv_freq")):
-                continue
-
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            data = data_torch.numpy()
-
-            if name.endswith("q_proj.weight"):
-                data = permute(data, n_head, n_head)
-            if name.endswith("k_proj.weight"):
-                data = permute(data, n_head, n_kv_head)
-
-            data = data.squeeze()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("GrokForCausalLM")
-class GrokModel(Model):
-    model_arch = gguf.MODEL_ARCH.GROK
-
+class MixtralModel(Model):
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
 
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_name("Grok")
-
-
-@Model.register("MiniCPMForCausalLM")
 class MiniCPMModel(Model):
-    model_arch = gguf.MODEL_ARCH.MINICPM
-
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
         self.gguf_writer.add_name("MiniCPM")
@@ -1153,7 +1091,7 @@ class MiniCPMModel(Model):
         self.gguf_writer.add_file_type(self.ftype)
 
     def set_vocab(self):
-        self._set_vocab_llama_hf()
+        self._set_vocab_hf()
 
     def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:
         if n_kv_head is not None and n_head != n_kv_head:
@@ -1215,10 +1153,7 @@ class MiniCPMModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("QWenLMHeadModel")
 class QwenModel(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN
-
     @staticmethod
     def token_bytes_to_string(b):
         from transformers.models.gpt2.tokenization_gpt2 import bytes_to_unicode
@@ -1298,15 +1233,7 @@ class QwenModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("Qwen2ForCausalLM")
-class Qwen2Model(Model):
-    model_arch = gguf.MODEL_ARCH.QWEN2
-
-
-@Model.register("GPT2LMHeadModel")
 class GPT2Model(Model):
-    model_arch = gguf.MODEL_ARCH.GPT2
-
     def set_gguf_parameters(self):
         self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_block_count(self.hparams["n_layer"])
@@ -1368,35 +1295,29 @@ class GPT2Model(Model):
                 self.gguf_writer.add_tensor("output.weight", data)
 
 
-@Model.register("PhiForCausalLM")
 class Phi2Model(Model):
-    model_arch = gguf.MODEL_ARCH.PHI2
-
     def set_gguf_parameters(self):
-        block_count = self.find_hparam(["num_hidden_layers", "n_layer"])
+        block_count = get_key_opts(self.hparams, ["num_hidden_layers", "n_layer"])
 
-        rot_pct = self.find_hparam(["partial_rotary_factor"])
-        n_embd = self.find_hparam(["hidden_size", "n_embd"])
-        n_head = self.find_hparam(["num_attention_heads", "n_head"])
+        rot_pct = get_key_opts(self.hparams, ["partial_rotary_factor"])
+        n_embd = get_key_opts(self.hparams, ["hidden_size", "n_embd"])
+        n_head = get_key_opts(self.hparams, ["num_attention_heads", "n_head"])
 
         self.gguf_writer.add_name("Phi2")
-        self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
+        self.gguf_writer.add_context_length(get_key_opts(self.hparams, ["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
         self.gguf_writer.add_feed_forward_length(4 * n_embd)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_head_count(n_head)
         self.gguf_writer.add_head_count_kv(n_head)
-        self.gguf_writer.add_layer_norm_eps(self.find_hparam(["layer_norm_epsilon", "layer_norm_eps"]))
+        self.gguf_writer.add_layer_norm_eps(get_key_opts(self.hparams, ["layer_norm_epsilon", "layer_norm_eps"]))
         self.gguf_writer.add_rope_dimension_count(int(rot_pct * n_embd) // n_head)
         self.gguf_writer.add_file_type(self.ftype)
         self.gguf_writer.add_add_bos_token(False)
 
 
-@Model.register("PlamoForCausalLM")
 class PlamoModel(Model):
-    model_arch = gguf.MODEL_ARCH.PLAMO
-
     def set_vocab(self):
         self._set_vocab_sentencepiece()
 
@@ -1475,10 +1396,7 @@ class PlamoModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("CodeShellForCausalLM")
 class CodeShellModel(Model):
-    model_arch = gguf.MODEL_ARCH.CODESHELL
-
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
@@ -1543,10 +1461,7 @@ class CodeShellModel(Model):
                 print(name, f"=> output.weight, shape = {data.shape}, {old_dtype} --> {data.dtype}")
 
 
-@Model.register("InternLM2ForCausalLM")
 class InternLM2Model(Model):
-    model_arch = gguf.MODEL_ARCH.INTERNLM2
-
     def set_vocab(self):
         # (TODO): Is there a better way?
         # Copy from _set_vocab_sentencepiece, The only difference is that we will treat the character
@@ -1718,47 +1633,31 @@ in chat mode so that the conversation can end normally.")
                 self.post_write_tensors(tensor_map, name, data_torch)
 
 
-@Model.register("BertModel", "CamembertModel")
 class BertModel(Model):
-    model_arch = gguf.MODEL_ARCH.BERT
-
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self.vocab_size = None
+        self.block_count = self.hparams["num_hidden_layers"]
 
     def set_gguf_parameters(self):
-        super().set_gguf_parameters()
+        # TODO(cebtenzzre): merge with parent class
+        self.gguf_writer.add_name(self.dir_model.name)
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
+        self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
+        self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
+        self.gguf_writer.add_block_count(self.block_count)
+        self.gguf_writer.add_head_count(self.hparams["num_attention_heads"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_eps"])
         self.gguf_writer.add_causal_attention(False)
-
-        # get pooling path
-        pooling_path = None
-        module_path = self.dir_model / "modules.json"
-        if module_path.is_file():
-            with open(module_path, encoding="utf-8") as f:
-                modules = json.load(f)
-            for mod in modules:
-                if mod["type"] == "sentence_transformers.models.Pooling":
-                    pooling_path = mod["path"]
-                    break
-
-        # get pooling type
-        if pooling_path is not None:
-            with open(self.dir_model / pooling_path / "config.json", encoding="utf-8") as f:
-                pooling = json.load(f)
-            if pooling["pooling_mode_mean_tokens"]:
-                pooling_type = gguf.PoolingType.MEAN
-            elif pooling["pooling_mode_cls_token"]:
-                pooling_type = gguf.PoolingType.CLS
-            else:
-                raise NotImplementedError("Only MEAN and CLS pooling types supported")
-            self.gguf_writer.add_pooling_type(pooling_type)
+        self.gguf_writer.add_file_type(self.ftype)
 
     def set_vocab(self):
+        path = self.dir_model
+        added_tokens_path = self.dir_model if self.dir_model.exists() else None
+
         # use huggingface vocab to get all tokens
-        vocab = LlamaHfVocab(self.dir_model, ignore_nonllama=True)
+        vocab = HfVocab(path, added_tokens_path)
         tokens, scores, toktypes = zip(*vocab.all_tokens())
         assert len(tokens) == vocab.vocab_size
-        self.vocab_size = vocab.vocab_size
 
         # we need this to validate the size of the token_type embeddings
         # though currently we are passing all zeros to the token_type embeddings
@@ -1772,7 +1671,7 @@ class BertModel(Model):
             if tok.startswith(b"##"):
                 return tok[2:]
             return b"\xe2\x96\x81" + tok
-        tokens = tuple(phantom(t, y) for t, y in zip(tokens, toktypes))
+        tokens = [phantom(t, y) for t, y in zip(tokens, toktypes)]
 
         # set up bos and eos tokens (cls and sep)
         self.gguf_writer.add_bos_token_id(vocab.tokenizer.cls_token_id)
@@ -1824,245 +1723,6 @@ class BertModel(Model):
             self.gguf_writer.add_tensor(new_name, data)
 
 
-@Model.register("NomicBertModel")
-class NomicBertModel(BertModel):
-    model_arch = gguf.MODEL_ARCH.NOMIC_BERT
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # the HF config claims n_ctx=8192, but it uses RoPE scaling
-        self.hparams["n_ctx"] = 2048
-
-        # SwigLU activation
-        assert self.hparams["activation_function"] == "swiglu"
-        # this doesn't do anything in the HF version
-        assert self.hparams["causal"] is False
-        # no bias tensors
-        assert self.hparams["qkv_proj_bias"] is False
-        assert self.hparams["mlp_fc1_bias"] is False
-        assert self.hparams["mlp_fc2_bias"] is False
-        # norm at end of layer
-        assert self.hparams["prenorm"] is False
-        # standard RoPE
-        assert self.hparams["rotary_emb_fraction"] == 1.0
-        assert self.hparams["rotary_emb_interleaved"] is False
-        assert self.hparams["rotary_emb_scale_base"] is None
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_rope_freq_base(self.hparams["rotary_emb_base"])
-
-    def get_tensors(self):
-        assert self.vocab_size is not None
-        for name, data in super().get_tensors():
-            # Nomic Embed's token embeddings tensor is padded, but llama.cpp wants tensor sizes to match exactly.
-            if name == 'embeddings.word_embeddings.weight' and data.shape[1] != self.vocab_size:
-                rounded_vocab_size = (self.vocab_size + 63) // 64 * 64
-                assert data.shape == (rounded_vocab_size, self.hparams["n_embd"])
-                data = data[:self.vocab_size, :]
-            yield name, data
-
-
-@Model.register("GemmaForCausalLM")
-class GemmaModel(Model):
-    model_arch = gguf.MODEL_ARCH.GEMMA
-
-    def set_vocab(self):
-        self._set_vocab_sentencepiece()
-
-    def set_gguf_parameters(self):
-        hparams = self.hparams
-        block_count = hparams["num_hidden_layers"]
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
-        self.gguf_writer.add_embedding_length(hparams["hidden_size"])
-        self.gguf_writer.add_block_count(block_count)
-        self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
-        self.gguf_writer.add_head_count(hparams["num_attention_heads"])
-        self.gguf_writer.add_head_count_kv(self.hparams["num_key_value_heads"] if "num_key_value_heads" in hparams else hparams["num_attention_heads"])
-        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"])
-        self.gguf_writer.add_key_length(hparams["head_dim"])
-        self.gguf_writer.add_value_length(hparams["head_dim"])
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams.get("n_layers", self.hparams.get("num_hidden_layers", self.hparams.get("n_layer")))
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # ref: https://github.com/huggingface/transformers/blob/fc37f38915372c15992b540dfcbbe00a916d4fc6/src/transformers/models/gemma/modeling_gemma.py#L89
-            if name.endswith("norm.weight"):
-                data_torch = data_torch + 1
-            data = data_torch.squeeze().numpy()
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            data = data.astype(np.float32)
-
-            # if f16 desired, convert any float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("Starcoder2ForCausalLM")
-class StarCoder2Model(Model):
-    model_arch = gguf.MODEL_ARCH.STARCODER2
-
-
-@Model.register("MambaForCausalLM", "MambaLMHeadModel")
-class MambaModel(Model):
-    model_arch = gguf.MODEL_ARCH.MAMBA
-
-    def set_vocab(self):
-        vocab_size = self.hparams["vocab_size"]
-        # Round vocab size to next multiple of 8
-        pad_vocab = self.hparams.get("pad_vocab_size_multiple", 8)
-        # pad using ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        vocab_size = -(vocab_size // -pad_vocab) * pad_vocab
-        self.hparams["vocab_size"] = vocab_size
-
-        if (self.dir_model / "tokenizer.json").is_file():
-            self._set_vocab_gpt2()
-        else:
-            # Use the GPT-NeoX tokenizer when no tokenizer files are present
-            tokenizer_path = Path(sys.path[0]) / "models" / "ggml-vocab-gpt-neox.gguf"
-            print(f"Using tokenizer from '{os.path.relpath(tokenizer_path, os.getcwd())}'")
-            neox_reader = gguf.GGUFReader(tokenizer_path, "r")
-
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MODEL)
-            self.gguf_writer.add_tokenizer_model(bytes(field.parts[-1]))
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.LIST)
-            self.gguf_writer.add_token_list([bytes(field.parts[i]) for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.TOKEN_TYPE)
-            self.gguf_writer.add_token_types([field.parts[i].tolist()[0] for i in field.data][:vocab_size])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.MERGES)
-            self.gguf_writer.add_token_merges([bytes(field.parts[i]) for i in field.data])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.BOS_ID)
-            self.gguf_writer.add_bos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.EOS_ID)
-            self.gguf_writer.add_eos_token_id(field.parts[-1].tolist()[0])
-            field = neox_reader.get_field(gguf.Keys.Tokenizer.UNK_ID)
-            self.gguf_writer.add_unk_token_id(field.parts[-1].tolist()[0])
-
-    def set_gguf_parameters(self):
-        d_model = self.find_hparam(["hidden_size", "d_model"])
-        d_conv  = self.find_hparam(["conv_kernel", "d_conv"], optional=True) or 4
-        d_inner = self.find_hparam(["intermediate_size", "d_inner"], optional=True) or 2 * d_model
-        d_state = self.find_hparam(["state_size", "d_state"], optional=True) or 16
-        # ceiling division
-        # ref: https://stackoverflow.com/a/17511341/22827863
-        # ref: https://github.com/state-spaces/mamba/blob/ce59daea3a090d011d6476c6e5b97f6d58ddad8b/mamba_ssm/modules/mamba_simple.py#L58
-        dt_rank = self.find_hparam(["time_step_rank", "dt_rank"], optional=True) or -(d_model // -16)
-        rms_norm_eps = self.find_hparam(["layer_norm_epsilon", "rms_norm_eps"], optional=True) or 1e-5
-
-        # Fail early for models which don't have a block expansion factor of 2
-        assert d_inner == 2 * d_model
-
-        self.gguf_writer.add_name(self.dir_model.name)
-        self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
-        self.gguf_writer.add_embedding_length(d_model)
-        self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_head_count(0) # unused, but seemingly required when loading
-        self.gguf_writer.add_block_count(self.hparams["n_layer"])
-        self.gguf_writer.add_ssm_conv_kernel(d_conv)
-        self.gguf_writer.add_ssm_inner_size(d_inner)
-        self.gguf_writer.add_ssm_state_size(d_state)
-        self.gguf_writer.add_ssm_time_step_rank(dt_rank)
-        self.gguf_writer.add_layer_norm_rms_eps(rms_norm_eps)
-        self.gguf_writer.add_file_type(self.ftype)
-
-    def write_tensors(self):
-        block_count = self.hparams["n_layer"]
-        tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count)
-
-        tok_embd = None
-        tok_embd_name = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.TOKEN_EMBD] + ".weight"
-        output_name   = gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.OUTPUT]     + ".weight"
-
-        for name, data_torch in self.get_tensors():
-            old_dtype = data_torch.dtype
-
-            # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
-                data_torch = data_torch.to(torch.float32)
-
-            # map tensor names
-            new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias"))
-            if new_name is None:
-                print(f"Can not map tensor {name!r}")
-                sys.exit()
-
-            if name.endswith(".A_log"):
-                print("A_log --> A ==> " + new_name)
-                data_torch = -torch.exp(data_torch)
-
-            # assuming token_embd.weight is seen before output.weight
-            if tok_embd is not None and new_name == output_name:
-                if torch.equal(tok_embd, data_torch):
-                    print(f"{output_name} is equivalent to {tok_embd_name}, omitting")
-                    continue
-            if new_name == tok_embd_name:
-                tok_embd = data_torch
-
-            data = data_torch.squeeze().numpy()
-
-            n_dims = len(data.shape)
-            data_dtype = data.dtype
-
-            # if f32 desired, convert any float16 to float32
-            if self.ftype == 0 and data_dtype == np.float16:
-                data = data.astype(np.float32)
-
-            # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32
-            if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1:
-                data = data.astype(np.float32)
-
-            # if f16 desired, convert big float32 2-dim weight tensors to float16
-            if self.ftype == 1 and data_dtype == np.float32 and new_name.removesuffix(".weight").endswith((".ssm_in", ".ssm_out", "token_embd", "output")) and n_dims == 2:
-                data = data.astype(np.float16)
-
-            print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
-
-            self.gguf_writer.add_tensor(new_name, data)
-
-
-@Model.register("CohereForCausalLM")
-class CommandR2Model(Model):
-    model_arch = gguf.MODEL_ARCH.COMMAND_R
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # max_position_embeddings = 8192 in config.json but model was actually
-        # trained on 128k context length
-        self.hparams["max_position_embeddings"] = self.hparams["model_max_length"]
-
-    def set_gguf_parameters(self):
-        super().set_gguf_parameters()
-        self.gguf_writer.add_logit_scale(self.hparams["logit_scale"])
-        self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.NONE)
-
-
 ###### CONVERSION LOGIC ######
 
 

convert-llama-ggml-to-gguf.py

@@ -373,7 +373,7 @@ def handle_metadata(cfg, hp):
         raise ValueError('Unable to load metadata')
     vocab_path = Path(cfg.vocab_dir if cfg.vocab_dir is not None else cfg.model_metadata_dir)
     vocab_factory = convert.VocabFactory(vocab_path)
-    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype.split(","), cfg.model_metadata_dir)
+    vocab, special_vocab = vocab_factory.load_vocab(cfg.vocabtype, cfg.model_metadata_dir)
     convert.check_vocab_size(params, vocab)
     return params, vocab, special_vocab
 
@@ -398,8 +398,8 @@ def handle_args():
                         help ='Load HuggingFace/.pth vocab and metadata from the specified directory')
     parser.add_argument("--vocab-dir", type=Path,
                         help="directory containing tokenizer.model, if separate from model file - only meaningful with --model-metadata-dir")
-    parser.add_argument("--vocabtype", default="spm,hfft",
-                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm,hfft)")
+    parser.add_argument("--vocabtype", choices=["spm", "bpe"], default="spm",
+                        help="vocab format - only meaningful with --model-metadata-dir and/or --vocab-dir (default: spm)")
     return parser.parse_args()
 
 

convert-persimmon-to-gguf.py

@@ -106,12 +106,12 @@ def main():
     tensor_map = gguf.get_tensor_name_map(arch, block_count)
     print(tensor_map)
     for name in tensors.keys():
-        data_torch = tensors[name]
+        data = tensors[name]
         if name.endswith(".self_attention.rotary_emb.inv_freq"):
             continue
-        old_dtype = data_torch.dtype
+        old_dtype = data.dtype
         # TODO: FP16 conversion produces garbage outputs. (Q8_0 does not, so..?)
-        data = data_torch.to(torch.float32).squeeze().numpy()
+        data = data.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

@@ -16,14 +16,13 @@ import re
 import signal
 import struct
 import sys
-import textwrap
 import time
 import zipfile
-from abc import ABC, abstractmethod
+from abc import ABCMeta, abstractmethod
 from concurrent.futures import ProcessPoolExecutor, ThreadPoolExecutor
 from dataclasses import dataclass
 from pathlib import Path
-from typing import TYPE_CHECKING, Any, Callable, ClassVar, IO, Iterable, Literal, Protocol, TypeVar, runtime_checkable
+from typing import IO, TYPE_CHECKING, Any, Callable, Iterable, Literal, TypeVar
 
 import numpy as np
 from sentencepiece import SentencePieceProcessor
@@ -44,9 +43,6 @@ ARCH = gguf.MODEL_ARCH.LLAMA
 
 DEFAULT_CONCURRENCY = 8
 
-ADDED_TOKENS_FILE = 'added_tokens.json'
-FAST_TOKENIZER_FILE = 'tokenizer.json'
-
 #
 # data types
 #
@@ -192,10 +188,8 @@ 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:
-            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))
+            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.")
 
         n_head = n_embd // 128 # guessed
         n_mult = 256           # guessed
@@ -217,8 +211,7 @@ class Params:
 
     @staticmethod
     def loadHFTransformerJson(model: LazyModel, config_path: Path) -> Params:
-        with open(config_path) as f:
-            config = json.load(f)
+        config = json.load(open(config_path))
 
         rope_scaling_type = f_rope_scale = n_orig_ctx = rope_finetuned = None
         rope_scaling = config.get("rope_scaling")
@@ -240,10 +233,8 @@ class Params:
         elif "max_position_embeddings" in config:
             n_ctx = config["max_position_embeddings"]
         else:
-            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))
+            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.")
 
         n_experts      = None
         n_experts_used = None
@@ -274,8 +265,7 @@ 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:
-        with open(config_path) as f:
-            config = json.load(f)
+        config = json.load(open(config_path))
 
         n_experts      = None
         n_experts_used = None
@@ -341,86 +331,44 @@ class Params:
 # vocab
 #
 
-@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, base_path: Path):
-        added_tokens: dict[str, int] = {}
-
-        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
+class BpeVocab:
+    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:
-            # "fast" tokenizer
-            fname_tokenizer = base_path / FAST_TOKENIZER_FILE
+            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)
 
-            # 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())
+        vocab_size: int = 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 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}")
+            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}")
 
         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      = vocab_size
-        self.vocab_size           = self.vocab_size_base + len(self.added_tokens_list)
+        self.vocab_size_base: int = vocab_size
+        self.vocab_size: int      = 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()}
@@ -441,25 +389,16 @@ class BpeVocab(Vocab):
         return f"<BpeVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class SentencePieceVocab(Vocab):
-    tokenizer_model = "llama"
-    name = "spm"
-
-    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')
-
+class SentencePieceVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None) -> None:
         self.sentencepiece_tokenizer = SentencePieceProcessor(str(fname_tokenizer))
-        vocab_size = self.sentencepiece_tokenizer.vocab_size()
+        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 = {}
+
+        vocab_size: int = 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)))
@@ -469,17 +408,18 @@ class SentencePieceVocab(Vocab):
             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         = piece.encode("utf-8")
+            text: bytes = piece.encode("utf-8")
             score: float = tokenizer.get_score(i)
 
             toktype = gguf.TokenType.NORMAL
@@ -512,42 +452,24 @@ class SentencePieceVocab(Vocab):
         return f"<SentencePieceVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
 
 
-class LlamaHfVocab(Vocab):
-    tokenizer_model = "llama"
-    name = "hfft"
-
-    def __init__(self, base_path: Path, ignore_nonllama: bool = False):
-        fname_tokenizer = base_path / FAST_TOKENIZER_FILE
-        # if this fails, FileNotFoundError propagates to caller
-        with open(fname_tokenizer, encoding='utf-8') as f:
-            tokenizer_json = json.load(f)
-
-        # pre-check so we know if we need transformers
-        tokenizer_model: dict[str, Any] = tokenizer_json['model']
-        if ignore_nonllama:
-            pass  # workaround incorrect use of this class for WordPiece
-        elif (
-            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')
-
+class HfVocab:
+    def __init__(self, fname_tokenizer: Path, fname_added_tokens: Path | None = None) -> None:
         try:
             from transformers import AutoTokenizer
         except ImportError as e:
             raise ImportError(
-                "To use LlamaHfVocab, please install the `transformers` package. "
+                "To use HfVocab, 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(
-            base_path,
-            cache_dir=base_path,
+            fname_tokenizer,
+            cache_dir=fname_tokenizer,
             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 = []
@@ -575,7 +497,8 @@ class LlamaHfVocab(Vocab):
         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_tokenizer    = fname_tokenizer
+        self.fname_added_tokens = fname_added_tokens
 
     def hf_tokens(self) -> Iterable[tuple[bytes, float, gguf.TokenType]]:
         reverse_vocab = {
@@ -627,7 +550,10 @@ class LlamaHfVocab(Vocab):
         yield from self.added_tokens()
 
     def __repr__(self) -> str:
-        return f"<LlamaHfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+        return f"<HfVocab with {self.vocab_size_base} base tokens and {len(self.added_tokens_list)} added tokens>"
+
+
+Vocab: TypeAlias = "BpeVocab | SentencePieceVocab | HfVocab"
 
 
 #
@@ -645,7 +571,7 @@ def permute(weights: NDArray, n_head: int, n_head_kv: int) -> NDArray:
             .reshape(weights.shape))
 
 
-class Tensor(ABC):
+class Tensor(metaclass=ABCMeta):
     data_type: DataType
 
     @abstractmethod
@@ -667,7 +593,7 @@ def bf16_to_fp32(bf16_arr: np.ndarray[Any, np.dtype[np.uint16]]) -> NDArray:
 
 
 class UnquantizedTensor(Tensor):
-    def __init__(self, ndarray: NDArray):
+    def __init__(self, ndarray: NDArray) -> None:
         assert isinstance(ndarray, np.ndarray)
         self.ndarray = ndarray
         self.data_type = NUMPY_TYPE_TO_DATA_TYPE[ndarray.dtype]
@@ -746,7 +672,7 @@ class ModelPlus:
     model: LazyModel
     paths: list[Path]  # Where this was read from.
     format: Literal['ggml', 'torch', 'safetensors', 'none']
-    vocab: BaseVocab | None  # For GGML models (which have vocab built in), the vocab.
+    vocab: Vocab | None  # For GGML models (which have vocab built in), the vocab.
 
 
 def merge_sharded(models: list[LazyModel]) -> LazyModel:
@@ -755,7 +681,7 @@ 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 = [model[name] for model in models]
+        lazy_tensors: list[LazyTensor] = [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
@@ -776,7 +702,7 @@ def merge_sharded(models: list[LazyModel]) -> LazyModel:
 
         def load() -> UnquantizedTensor:
             ndarrays = [load_unquantized(tensor) for tensor in lazy_tensors]
-            concatenated = np.concatenate(ndarrays, axis=axis)
+            concatenated: NDArray = 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)
@@ -864,10 +790,10 @@ class LazyUnpickler(pickle.Unpickler):
 
         def load(offset: int, elm_count: int) -> NDArray:
             dtype = data_type.dtype
-            with self.zip_file.open(info) as fp:
-                fp.seek(offset * dtype.itemsize)
-                size = elm_count * dtype.itemsize
-                data = fp.read(size)
+            fp = self.zip_file.open(info)
+            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}'
@@ -888,7 +814,7 @@ class LazyUnpickler(pickle.Unpickler):
     def rebuild_from_type_v2(func, new_type, args, state):
         return func(*args)
 
-    CLASSES = {
+    CLASSES: dict[tuple[str, str], Any] = {
         # 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__'),
@@ -947,7 +873,7 @@ 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 EOFError("unexpectedly reached end of file")
+        raise Exception("unexpectedly reached end of file")
     return ret
 
 
@@ -1005,15 +931,12 @@ def bounded_parallel_map(func: Callable[[In], Out], iterable: Iterable[In], conc
             yield result
 
 
-def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False) -> None:
+def check_vocab_size(params: Params, vocab: Vocab, pad_vocab: bool = False) -> None:
     # Handle special case where the model's vocab size is not set
     if params.n_vocab == -1:
         raise ValueError(
-            "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 ""),
+            f"The model's vocab size is set to -1 in params.json. Please update it manually. Maybe {vocab.vocab_size}?"
         )
-    if not isinstance(vocab, Vocab):
-        return  # model has no vocab
 
     # Check for a vocab size mismatch
     if params.n_vocab == vocab.vocab_size:
@@ -1037,11 +960,11 @@ def check_vocab_size(params: Params, vocab: BaseVocab, pad_vocab: bool = False)
     if vocab.vocab_size < params.n_vocab:
         msg += " Add the --pad-vocab option and try again."
 
-    raise ValueError(msg)
+    raise Exception(msg)
 
 
 class OutputFile:
-    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE):
+    def __init__(self, fname_out: Path, endianess:gguf.GGUFEndian = gguf.GGUFEndian.LITTLE) -> None:
         self.gguf = gguf.GGUFWriter(fname_out, gguf.MODEL_ARCH_NAMES[ARCH], endianess=endianess)
 
     def add_meta_arch(self, params: Params) -> None:
@@ -1054,7 +977,6 @@ class OutputFile:
             name = str(params.path_model.parent).split('/')[-1]
 
         self.gguf.add_name                (name)
-        self.gguf.add_vocab_size          (params.n_vocab)
         self.gguf.add_context_length      (params.n_ctx)
         self.gguf.add_embedding_length    (params.n_embd)
         self.gguf.add_block_count         (params.n_layer)
@@ -1091,6 +1013,20 @@ class OutputFile:
         if params.ftype is not None:
             self.gguf.add_file_type(params.ftype)
 
+    def handle_tokenizer_model(self, vocab: Vocab) -> str:
+        # Map the vocab types to the supported tokenizer models
+        tokenizer_model = {
+            SentencePieceVocab: "llama",
+            HfVocab: "llama",
+            BpeVocab: "gpt2",
+        }.get(type(vocab))
+
+        # Block if vocab type is not predefined
+        if tokenizer_model is None:
+            raise ValueError("Unknown vocab type: Not supported")
+
+        return tokenizer_model
+
     def extract_vocabulary_from_model(self, vocab: Vocab) -> tuple[list[bytes], list[float], list[gguf.TokenType]]:
         tokens = []
         scores = []
@@ -1107,8 +1043,11 @@ class OutputFile:
         return tokens, scores, toktypes
 
     def add_meta_vocab(self, vocab: Vocab) -> None:
+        # Handle the tokenizer model
+        tokenizer_model = self.handle_tokenizer_model(vocab)
+
         # Ensure that tokenizer_model is added to the GGUF model
-        self.gguf.add_tokenizer_model(vocab.tokenizer_model)
+        self.gguf.add_tokenizer_model(tokenizer_model)
 
         # Extract model vocabulary for model conversion
         tokens, scores, toktypes = self.extract_vocabulary_from_model(vocab)
@@ -1135,26 +1074,6 @@ class OutputFile:
     def write_tensor_info(self) -> None:
         self.gguf.write_ti_data_to_file()
 
-    def write_tensor_data(self, ftype: GGMLFileType, model: LazyModel, concurrency: int) -> None:
-        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency=concurrency)
-        if ftype == GGMLFileType.MostlyQ8_0:
-            ndarrays = bounded_parallel_map(
-                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
-                use_processpool_executor=True,
-            )
-        else:
-            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
-
-        start = time.time()
-        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
-            elapsed = time.time() - start
-            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
-            padi = len(str(len(model)))
-            print(
-                f"[{i + 1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
-            )
-            self.gguf.write_tensor_data(ndarray)
-
     def close(self) -> None:
         self.gguf.close()
 
@@ -1163,7 +1082,7 @@ class OutputFile:
         fname_out: Path, params: Params, vocab: Vocab, svocab: gguf.SpecialVocab,
         endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE, pad_vocab: bool = False,
     ) -> None:
-        check_vocab_size(params, vocab, pad_vocab=pad_vocab)
+        check_vocab_size(params, vocab, pad_vocab = pad_vocab)
 
         of = OutputFile(fname_out, endianess=endianess)
 
@@ -1191,7 +1110,7 @@ class OutputFile:
 
     @staticmethod
     def write_all(
-        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: BaseVocab, svocab: gguf.SpecialVocab,
+        fname_out: Path, ftype: GGMLFileType, params: Params, model: LazyModel, vocab: Vocab, svocab: gguf.SpecialVocab,
         concurrency: int = DEFAULT_CONCURRENCY, endianess: gguf.GGUFEndian = gguf.GGUFEndian.LITTLE,
         pad_vocab: bool = False,
     ) -> None:
@@ -1201,11 +1120,8 @@ class OutputFile:
 
         # meta data
         of.add_meta_arch(params)
-        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)
+        of.add_meta_vocab(vocab)
+        of.add_meta_special_vocab(svocab)
 
         # tensor info
         for name, lazy_tensor in model.items():
@@ -1215,7 +1131,24 @@ class OutputFile:
         of.write_tensor_info()
 
         # tensor data
-        of.write_tensor_data(ftype, model, concurrency)
+        ndarrays_inner = bounded_parallel_map(OutputFile.do_item, model.items(), concurrency = concurrency)
+        if ftype == GGMLFileType.MostlyQ8_0:
+            ndarrays = bounded_parallel_map(
+                OutputFile.maybe_do_quantize, ndarrays_inner, concurrency=concurrency, max_workers=concurrency,
+                use_processpool_executor=True,
+            )
+        else:
+            ndarrays = map(OutputFile.maybe_do_quantize, ndarrays_inner)
+
+        start = time.time()
+        for i, ((name, lazy_tensor), ndarray) in enumerate(zip(model.items(), ndarrays)):
+            elapsed = time.time() - start
+            size = ' x '.join(f"{dim:6d}" for dim in lazy_tensor.shape)
+            padi = len(str(len(model)))
+            print(
+                f"[{i+1:{padi}d}/{len(model)}] Writing tensor {name:38s} | size {size:16} | type {lazy_tensor.data_type.name:4} | T+{int(elapsed):4}"
+            )
+            of.gguf.write_tensor_data(ndarray)
 
         of.close()
 
@@ -1223,16 +1156,16 @@ class OutputFile:
 def pick_output_type(model: LazyModel, output_type_str: str | None) -> GGMLFileType:
     wq_type = model[gguf.TENSOR_NAMES[gguf.MODEL_TENSOR.ATTN_Q].format(bid=0) + ".weight"].data_type
 
-    if output_type_str == "f32" or (output_type_str is None and wq_type in (DT_F32, DT_BF16)):
+    if output_type_str == "f32" or (output_type_str is None and wq_type == DT_F32):
         return GGMLFileType.AllF32
-    if output_type_str == "f16" or (output_type_str is None and wq_type == DT_F16):
+    if output_type_str == "f16" or (output_type_str is None and wq_type in (DT_F16, DT_BF16)):
         return GGMLFileType.MostlyF16
     if output_type_str == "q8_0":
         return GGMLFileType.MostlyQ8_0
 
     name_to_type = {name: lazy_tensor.data_type for (name, lazy_tensor) in model.items()}
 
-    raise ValueError(f"Unexpected combination of types: {name_to_type}")
+    raise Exception(f"Unexpected combination of types: {name_to_type}")
 
 
 def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyModel:
@@ -1240,9 +1173,9 @@ def convert_to_output_type(model: LazyModel, output_type: GGMLFileType) -> LazyM
             for (name, tensor) in model.items()}
 
 
-def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) -> LazyModel:
+def convert_model_names(model: LazyModel, params: Params) -> LazyModel:
     tmap = gguf.TensorNameMap(ARCH, params.n_layer)
-    should_skip = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
+    should_skip: set[gguf.MODEL_TENSOR] = set(gguf.MODEL_TENSOR_SKIP.get(ARCH, []))
 
     tmp = model
 
@@ -1266,10 +1199,7 @@ def convert_model_names(model: LazyModel, params: Params, skip_unknown: bool) ->
     for name, lazy_tensor in model.items():
         tensor_type, name_new = tmap.get_type_and_name(name, try_suffixes = (".weight", ".bias")) or (None, None)
         if name_new is None:
-            if skip_unknown:
-                print(f"Unexpected tensor name: {name} - skipping")
-                continue
-            raise ValueError(f"Unexpected tensor name: {name}. Use --skip-unknown to ignore it (e.g. LLaVA)")
+            raise Exception(f"Unexpected tensor name: {name}")
 
         if tensor_type in should_skip:
             print(f"skipping tensor {name_new}")
@@ -1286,7 +1216,7 @@ def nth_multifile_path(path: Path, n: int) -> Path | None:
     the nth path in the model.
     '''
     # Support the following patterns:
-    patterns = [
+    patterns: list[tuple[str, str]] = [
         # - 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.
@@ -1332,9 +1262,9 @@ def load_some_model(path: Path) -> ModelPlus:
             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 FileNotFoundError(f"Can't find model in directory {path}")
+            raise Exception(f"Can't find model in directory {path}")
         if len(files) > 1:
-            raise ValueError(f"Found multiple models in {path}, not sure which to pick: {files}")
+            raise Exception(f"Found multiple models in {path}, not sure which to pick: {files}")
         path = files[0]
 
     paths = find_multifile_paths(path)
@@ -1348,14 +1278,39 @@ def load_some_model(path: Path) -> ModelPlus:
 
 
 class VocabFactory:
-    _VOCAB_CLASSES: list[type[Vocab]] = [SentencePieceVocab, BpeVocab, LlamaHfVocab]
-
     def __init__(self, path: Path):
         self.path = path
+        self.files: dict[str, Path | None] = {
+            "tokenizer.model": None,
+            "vocab.json": None,
+            "tokenizer.json": None,
+        }
+        self._detect_files()
 
-    def _create_special_vocab(self, vocab: BaseVocab, model_parent_path: Path) -> gguf.SpecialVocab:
-        load_merges = vocab.name == "bpe"
-        n_vocab = vocab.vocab_size if isinstance(vocab, Vocab) else None
+    def _detect_files(self):
+        for file in self.files.keys():
+            file_path = self.path / file
+            parent_file_path = self.path.parent / file
+            if file_path.exists():
+                self.files[file] = file_path
+            elif parent_file_path.exists():
+                self.files[file] = parent_file_path
+        print(f"Found vocab files: {self.files}")
+
+    def _select_file(self, vocabtype: str | None) -> Path:
+        if vocabtype in ["spm", "bpe"]:
+            for file_key in self.files.keys():
+                if (file := self.files[file_key]) is not None:
+                    return file
+            raise FileNotFoundError(f"{vocabtype} vocab not found.")
+        if vocabtype == "hfft":
+            # For Hugging Face Fast Tokenizer, return the directory path instead of a specific file
+            return self.path
+        raise ValueError(f"Unsupported vocabulary type {vocabtype}")
+
+    def _create_special_vocab(self, vocab: Vocab, vocabtype: str, model_parent_path: Path) -> gguf.SpecialVocab:
+        load_merges = vocabtype == "bpe"
+        n_vocab = vocab.vocab_size if hasattr(vocab, "vocab_size") else None
         return gguf.SpecialVocab(
             model_parent_path,
             load_merges=load_merges,
@@ -1363,36 +1318,30 @@ class VocabFactory:
             n_vocab=n_vocab,
         )
 
-    def _create_vocab_by_path(self, vocab_types: list[str]) -> Vocab:
-        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
+    def load_vocab(self, vocabtype: str, model_parent_path: Path) -> tuple[Vocab, gguf.SpecialVocab]:
+        path = self._select_file(vocabtype)
+        print(f"Loading vocab file '{path}', type '{vocabtype}'")
 
-        for vtype, cls in selected_vocabs.items():
-            try:
-                vocab = cls(self.path)
-                break
-            except FileNotFoundError:
-                pass  # ignore unavailable tokenizers
+        added_tokens_path = path.parent / "added_tokens.json"
+        vocab: Vocab
+        if vocabtype == "bpe":
+            vocab = BpeVocab(
+                path, added_tokens_path if added_tokens_path.exists() else None
+            )
+        elif vocabtype == "spm":
+            vocab = SentencePieceVocab(
+                path, added_tokens_path if added_tokens_path.exists() else None
+            )
+        elif vocabtype == "hfft":
+            vocab = HfVocab(
+                path, added_tokens_path if added_tokens_path.exists() else None
+            )
         else:
-            raise FileNotFoundError(f"Could not find a tokenizer matching any of {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)
+            raise ValueError(f"Unsupported vocabulary type {vocabtype}")
         # FIXME: Respect --vocab-dir?
         special_vocab = self._create_special_vocab(
             vocab,
+            vocabtype,
             model_parent_path,
         )
         return vocab, special_vocab
@@ -1426,25 +1375,35 @@ def main(args_in: list[str] | None = None) -> None:
     if np.uint32(1) == np.uint32(1).newbyteorder("<"):
         # We currently only support Q8_0 output on little endian systems.
         output_choices.append("q8_0")
-    parser = argparse.ArgumentParser(description="Convert a LLaMA model to a GGML compatible file")
-    parser.add_argument("--dump",         action="store_true",    help="don't convert, just show what's in the model")
-    parser.add_argument("--dump-single",  action="store_true",    help="don't convert, just show what's in a single model file")
-    parser.add_argument("--vocab-only",   action="store_true",    help="extract only the vocab")
-    parser.add_argument("--no-vocab",     action="store_true",    help="store model without the vocab")
-    parser.add_argument("--outtype",      choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
-    parser.add_argument("--vocab-dir",    type=Path,              help="directory containing tokenizer.model, if separate from model file")
-    parser.add_argument("--vocab-type",                           help="vocab types to try in order, choose from 'spm', 'bpe', 'hfft' (default: spm,hfft)", default="spm,hfft")
-    parser.add_argument("--outfile",      type=Path,              help="path to write to; default: based on input")
-    parser.add_argument("model",          type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
-    parser.add_argument("--ctx",          type=int,               help="model training context (default: based on input)")
-    parser.add_argument("--concurrency",  type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
-    parser.add_argument("--big-endian",   action="store_true",    help="model is executed on big endian machine")
-    parser.add_argument("--pad-vocab",    action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
-    parser.add_argument("--skip-unknown", action="store_true",    help="skip unknown tensor names instead of failing")
+    vocab_types = ["spm", "bpe", "hfft"]
+    parser = argparse.ArgumentParser(description="Convert a LLaMa model to a GGML compatible file")
+    parser.add_argument("--awq-path",    type=Path,              help="Path to scale awq cache file", default=None)
+    parser.add_argument("--dump",        action="store_true",    help="don't convert, just show what's in the model")
+    parser.add_argument("--dump-single", action="store_true",    help="don't convert, just show what's in a single model file")
+    parser.add_argument("--vocab-only",  action="store_true",    help="extract only the vocab")
+    parser.add_argument("--outtype",     choices=output_choices, help="output format - note: q8_0 may be very slow (default: f16 or f32 based on input)")
+    parser.add_argument("--vocab-dir",   type=Path,              help="directory containing tokenizer.model, if separate from model file")
+    parser.add_argument("--vocab-type",  choices=vocab_types,    help="The vocabulary format used to define the tokenizer model (default: spm)", default="spm")
+    parser.add_argument("--outfile",     type=Path,              help="path to write to; default: based on input")
+    parser.add_argument("model",         type=Path,              help="directory containing model file, or model file itself (*.pth, *.pt, *.bin)")
+    parser.add_argument("--ctx",         type=int,               help="model training context (default: based on input)")
+    parser.add_argument("--concurrency", type=int,               help=f"concurrency used for conversion (default: {DEFAULT_CONCURRENCY})", default=DEFAULT_CONCURRENCY)
+    parser.add_argument("--big-endian",  action="store_true",    help="model is executed on big endian machine")
+    parser.add_argument("--pad-vocab",   action="store_true",    help="add pad tokens when model vocab expects more than tokenizer metadata provides")
 
     args = parser.parse_args(args_in)
-    if args.no_vocab and args.vocab_only:
-        raise ValueError("--vocab-only does not make sense with --no-vocab")
+    if args.awq_path:
+        sys.path.insert(1, str(Path(__file__).parent / 'awq-py'))
+        from awq.apply_awq import add_scale_weights  # type: ignore[import-not-found]
+        tmp_model_path = args.model / "weighted_model"
+        if tmp_model_path.is_dir():
+            print(f"{tmp_model_path} exists as a weighted model.")
+        else:
+            tmp_model_path.mkdir(parents=True, exist_ok=True)
+            print("Saving new weighted model ...")
+            add_scale_weights(str(args.model), str(args.awq_path), str(tmp_model_path))
+            print(f"Saved weighted model at {tmp_model_path}.")
+        args.model = tmp_model_path
 
     if args.dump_single:
         model_plus = lazy_load_file(args.model)
@@ -1466,12 +1425,10 @@ def main(args_in: list[str] | None = None) -> None:
     params = Params.load(model_plus)
     if params.n_ctx == -1:
         if args.ctx is None:
-            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))
+            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")
         params.n_ctx = args.ctx
 
     if args.outtype:
@@ -1486,11 +1443,9 @@ 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_types = None if args.no_vocab else args.vocab_type.split(",")
-    vocab, special_vocab = vocab_factory.load_vocab(vocab_types, model_parent_path)
+    vocab, special_vocab = vocab_factory.load_vocab(args.vocab_type, 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
@@ -1499,14 +1454,14 @@ def main(args_in: list[str] | None = None) -> None:
         print(f"Wrote {outfile}")
         return
 
-    if model_plus.vocab is not None and args.vocab_dir is None and not args.no_vocab:
+    if model_plus.vocab is not None and args.vocab_dir is None:
         vocab = model_plus.vocab
 
     print(f"Vocab info: {vocab}")
     print(f"Special vocab info: {special_vocab}")
 
     model   = model_plus.model
-    model   = convert_model_names(model, params, args.skip_unknown)
+    model   = convert_model_names(model, params)
     ftype   = pick_output_type(model, args.outtype)
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype)

docs/token_generation_performance_tips.md

@@ -1,7 +1,7 @@
 # Token generation performance troubleshooting
 
-## Verifying that the model is running on the GPU with CUDA
-Make sure you compiled llama with the correct env variables according to [this guide](../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:
+## 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:
 ```shell
 ./main -m "path/to/model.gguf" -ngl 200000 -p "Please sir, may I have some "
 ```

examples/CMakeLists.txt

@@ -20,8 +20,6 @@ else()
     add_subdirectory(convert-llama2c-to-ggml)
     add_subdirectory(embedding)
     add_subdirectory(finetune)
-    add_subdirectory(gritlm)
-    add_subdirectory(gguf-split)
     add_subdirectory(infill)
     add_subdirectory(llama-bench)
     add_subdirectory(llava)
@@ -34,14 +32,12 @@ 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)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
-    add_subdirectory(gguf)
     add_subdirectory(train-text-from-scratch)
     add_subdirectory(imatrix)
     if (LLAMA_BUILD_SERVER)

examples/baby-llama/baby-llama.cpp

@@ -1533,28 +1533,27 @@ int main(int argc, char ** argv) {
 
         int n_past = 0;
 
-        struct ggml_cgraph * gf = NULL;
-        gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+        ggml_cgraph gf = {};
 
         get_example_targets_batch(ctx0, 64*ex+0,  tokens_input, targets);
 
-        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, gf, tokens_input, n_tokens, n_past, n_batch);
+        struct ggml_tensor * logits = forward_batch(&model, &kv_self, ctx0, &gf, tokens_input, n_tokens, n_past, n_batch);
         // struct ggml_tensor * e = cross_entropy_loss(ctx0, targets, logits);
         struct ggml_tensor * e = square_error_loss(ctx0, targets, logits);
 
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_before_opt = ggml_get_f32_1d(e, 0);
 
-        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_TYPE_LBFGS);
+        struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
         opt_params_lbfgs.print_forward_graph = false;
         opt_params_lbfgs.print_backward_graph = false;
         opt_params_lbfgs.lbfgs.n_iter = 16;
         ggml_opt(ctx0, opt_params_lbfgs, e);
         //
-        ggml_build_forward_expand(gf, e);
-        ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+        ggml_build_forward_expand(&gf, e);
+        ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
         float error_after_opt = ggml_get_f32_1d(e, 0);
 
@@ -1601,14 +1600,13 @@ int main(int argc, char ** argv) {
             };
             struct ggml_context * ctx0 = ggml_init(params);
 
-            struct ggml_cgraph * gf = NULL;
-            gf = ggml_new_graph_custom(ctx0, LLAMA_TRAIN_MAX_NODES, true);
+            ggml_cgraph gf = {};
 
             int n_past = 0;
-            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, gf, tokens_input, sample_ctx, n_past);
+            struct ggml_tensor * logits = forward(&model, &kv_self, ctx0, &gf, tokens_input, sample_ctx, n_past);
 
-            ggml_build_forward_expand(gf, logits);
-            ggml_graph_compute_helper(work_buffer, gf, /*n_threads*/ 1);
+            ggml_build_forward_expand(&gf, logits);
+            ggml_graph_compute_helper(work_buffer, &gf, /*n_threads*/ 1);
 
             struct ggml_tensor * best_samples = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, sample_ctx);
             struct ggml_tensor * probs        = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_vocab, sample_ctx);

examples/batched-bench/batched-bench.cpp

@@ -32,15 +32,16 @@ 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] [IS_PP_SHARED] [NGL] [MMQ] <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 0 999 0 128,256,512 128,256 1,2,4,8,16,32\n\n", argv[0]);
         return 1 ;
     }
 
     int n_kv_max     = 2048;
     int is_pp_shared = 0;
     int n_gpu_layers = 0;
+    int mmq          = 0;
 
     std::vector<int> n_pp = { 128, 256, 512, 1024, 2048, 3584, 7680, };
     std::vector<int> n_tg = { 128, 256, };
@@ -64,21 +65,24 @@ int main(int argc, char ** argv) {
     }
 
     if (argc >= 6) {
-        n_pp = parse_list(argv[5]);
+        mmq = std::atoi(argv[5]);
     }
 
     if (argc >= 7) {
-        n_tg = parse_list(argv[6]);
+        n_pp = parse_list(argv[6]);
     }
 
     if (argc >= 8) {
-        n_pl = parse_list(argv[7]);
+        n_tg = parse_list(argv[7]);
+    }
+
+    if (argc >= 9) {
+        n_pl = parse_list(argv[8]);
     }
 
     // init LLM
 
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     // initialize the model
 
@@ -101,13 +105,11 @@ int main(int argc, char ** argv) {
     ctx_params.seed      = 1234;
     ctx_params.n_ctx     = n_kv_max;
     ctx_params.n_batch   = 512;
+    ctx_params.mul_mat_q = mmq;
 
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
-    // ensure enough sequences are available
-    ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
-
     llama_context * ctx = llama_new_context_with_model(model, ctx_params);
 
     if (ctx == NULL) {
@@ -138,8 +140,6 @@ int main(int argc, char ** argv) {
                 LOG_TEE("failed to decode the batch, n_batch = %d, ret = %d\n", n_batch, ret);
                 return false;
             }
-
-            llama_synchronize(ctx);
         }
 
         return true;
@@ -158,7 +158,7 @@ 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, is_pp_shared = %d, n_gpu_layers = %d, mmq = %d, n_threads = %d, n_threads_batch = %d\n", __func__, n_kv_max, is_pp_shared, n_gpu_layers, mmq, 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");
@@ -179,10 +179,10 @@ int main(int argc, char ** argv) {
 
                 llama_batch_clear(batch);
 
-                for (int i = 0; i < pp; ++i) {
-                    for (int j = 0; j < (is_pp_shared ? 1 : pl); ++j) {
-                        llama_batch_add(batch, 0, i, { j }, false);
-                    }
+                const int n_tokens = is_pp_shared ? pp : pl*pp;
+
+                for (int i = 0; i < n_tokens; ++i) {
+                    llama_batch_add(batch, 0, i, { 0 }, false);
                 }
                 batch.logits[batch.n_tokens - 1] = true;
 
@@ -197,7 +197,7 @@ int main(int argc, char ** argv) {
 
                 if (is_pp_shared) {
                     for (int32_t i = 1; i < pl; ++i) {
-                        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+                        llama_kv_cache_seq_cp(ctx, 0, i, 0, pp);
                     }
                 }
 

examples/batched.swift/Sources/main.swift

@@ -17,7 +17,7 @@ let n_parallel: Int = arguments.count > 3 && Int(arguments[3]) != nil ? Int(argu
 let n_len: Int = 32
 
 // init LLM
-llama_backend_init()
+llama_backend_init(false)
 defer {
     llama_backend_free()
 }

examples/batched/batched.cpp

@@ -48,12 +48,9 @@ int main(int argc, char ** argv) {
         params.prompt = "Hello my name is";
     }
 
-    process_escapes(params.prompt);
-
     // init LLM
 
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     // initialize the model
 
@@ -80,9 +77,8 @@ int main(int argc, char ** argv) {
     llama_context_params ctx_params = llama_context_default_params();
 
     ctx_params.seed  = 1234;
-    ctx_params.n_ctx   = n_kv_req;
+    ctx_params.n_ctx = n_kv_req;
     ctx_params.n_batch = std::max(n_len, n_parallel);
-    ctx_params.n_seq_max       = n_parallel;
     ctx_params.n_threads       = params.n_threads;
     ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
 
@@ -95,7 +91,7 @@ int main(int argc, char ** argv) {
 
     const int n_ctx    = llama_n_ctx(ctx);
 
-    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %u, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_batch = %d, n_parallel = %d, n_kv_req = %d\n", __func__, n_len, n_ctx, ctx_params.n_batch, n_parallel, n_kv_req);
 
     // make sure the KV cache is big enough to hold all the prompt and generated tokens
     if (n_kv_req > n_ctx) {
@@ -135,7 +131,7 @@ int main(int argc, char ** argv) {
     // assign the system KV cache to all parallel sequences
     // this way, the parallel sequences will "reuse" the prompt tokens without having to copy them
     for (int32_t i = 1; i < n_parallel; ++i) {
-        llama_kv_cache_seq_cp(ctx, 0, i, -1, -1);
+        llama_kv_cache_seq_cp(ctx, 0, i, 0, batch.n_tokens);
     }
 
     if (n_parallel > 1) {

examples/beam-search/beam-search.cpp

@@ -119,8 +119,7 @@ int main(int argc, char ** argv)
     // Init LLM :
     //---------------------------------
 
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;

examples/benchmark/benchmark-matmult.cpp

@@ -189,10 +189,12 @@ int main(int argc, char ** argv)  {
 
     int32_t nelements = sizex*sizey;
 
+    std::vector<int64_t> hist_cur(1 << 4, 0);
+
     // Set up a the benchmark matrices
     // printf("Creating new tensor q11 & Running quantize\n");
     struct ggml_tensor * q11 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m11->data, q11->data, 0, nelements/m11->ne[0], m11->ne[0], hist_cur.data(), nullptr);
 
     // Set up a the compute graph
     // printf("Creating new tensor q31\n");
@@ -205,7 +207,7 @@ int main(int argc, char ** argv)  {
     // Set up a second graph computation to make sure we override the CPU cache lines
     // printf("Creating new tensor q12 & Running quantize\n");
     struct ggml_tensor * q12 = ggml_new_tensor_2d(ctx, qtype, sizex, sizey);
-    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], nullptr);
+    ggml_quantize_chunk(qtype, (const float *) m12->data, q12->data, 0, nelements/m12->ne[0], m12->ne[0], hist_cur.data(), nullptr);
 
     // printf("Creating new tensor q32\n");
     struct ggml_tensor * q32 = ggml_mul_mat(ctx, q12, m2);

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

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

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

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

examples/embedding/embedding.cpp

@@ -7,52 +7,6 @@
 #pragma warning(disable: 4244 4267) // possible loss of data
 #endif
 
-static std::vector<std::string> split_lines(const std::string & s) {
-    std::string line;
-    std::vector<std::string> lines;
-    std::stringstream ss(s);
-    while (std::getline(ss, line)) {
-        lines.push_back(line);
-    }
-    return lines;
-}
-
-static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & tokens, int seq_id) {
-    for (size_t i = 0; i < tokens.size(); i++) {
-        llama_batch_add(batch, tokens[i], i, { seq_id }, i == tokens.size() - 1);
-    }
-}
-
-static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd) {
-    // clear previous kv_cache values (irrelevant for embeddings)
-    llama_kv_cache_clear(ctx);
-
-    // run model
-    fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
-    if (llama_decode(ctx, batch) < 0) {
-        fprintf(stderr, "%s : failed to decode\n", __func__);
-    }
-
-    for (int i = 0; i < batch.n_tokens; i++) {
-        if (!batch.logits[i]) {
-            continue;
-        }
-
-        // try to get sequence embeddings - supported only when pooling_type is not NONE
-        const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
-        if (embd == NULL) {
-            embd = llama_get_embeddings_ith(ctx, i);
-            if (embd == NULL) {
-                fprintf(stderr, "%s: failed to get embeddings for token %d\n", __func__, i);
-                continue;
-            }
-        }
-
-        float * out = output + batch.seq_id[i][0] * n_embd;
-        llama_embd_normalize(embd, out, n_embd);
-    }
-}
-
 int main(int argc, char ** argv) {
     gpt_params params;
 
@@ -61,8 +15,6 @@ 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();
 
@@ -77,8 +29,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -104,107 +55,59 @@ int main(int argc, char ** argv) {
         fprintf(stderr, "%s\n", get_system_info(params).c_str());
     }
 
-    // split the prompt into lines
-    std::vector<std::string> prompts = split_lines(params.prompt);
+    int n_past = 0;
 
-    // max batch size
-    const uint64_t n_batch = params.n_batch;
-    GGML_ASSERT(params.n_batch >= params.n_ctx);
+    // tokenize the prompt
+    auto embd_inp = ::llama_tokenize(ctx, params.prompt, true);
 
-    // tokenize the prompts and trim
-    std::vector<std::vector<int32_t>> inputs;
-    for (const auto & prompt : prompts) {
-        auto inp = ::llama_tokenize(ctx, prompt, true, false);
-        if (inp.size() > n_batch) {
-            fprintf(stderr, "%s: error: number of tokens in input line (%lld) exceeds batch size (%lld), increase batch size and re-run\n",
-                    __func__, (long long int) inp.size(), (long long int) n_batch);
+    if (params.verbose_prompt) {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+        fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+        for (int i = 0; i < (int) embd_inp.size(); i++) {
+            fprintf(stderr, "%6d -> '%s'\n", embd_inp[i], llama_token_to_piece(ctx, embd_inp[i]).c_str());
+        }
+        fprintf(stderr, "\n");
+    }
+
+    if (embd_inp.size() > (size_t)n_ctx) {
+        fprintf(stderr, "%s: error: prompt is longer than the context window (%zu tokens, n_ctx = %d)\n",
+                __func__, embd_inp.size(), n_ctx);
+        return 1;
+    }
+
+    while (!embd_inp.empty()) {
+        int n_tokens = std::min(params.n_batch, (int) embd_inp.size());
+        if (llama_decode(ctx, llama_batch_get_one(embd_inp.data(), n_tokens, n_past, 0))) {
+            fprintf(stderr, "%s : failed to eval\n", __func__);
             return 1;
         }
-        inputs.push_back(inp);
+        n_past += n_tokens;
+        embd_inp.erase(embd_inp.begin(), embd_inp.begin() + n_tokens);
     }
 
-    // add eos if not present
-    for (auto & inp : inputs) {
-        if (inp.empty() || inp.back() != llama_token_eos(model)) {
-            inp.push_back(llama_token_eos(model));
-        }
-    }
-
-    // tokenization stats
-    if (params.verbose_prompt) {
-        for (int i = 0; i < (int) inputs.size(); i++) {
-            fprintf(stderr, "%s: prompt %d: '%s'\n", __func__, i, prompts[i].c_str());
-            fprintf(stderr, "%s: number of tokens in prompt = %zu\n", __func__, inputs[i].size());
-            for (int j = 0; j < (int) inputs[i].size(); j++) {
-                fprintf(stderr, "%6d -> '%s'\n", inputs[i][j], llama_token_to_piece(ctx, inputs[i][j]).c_str());
-            }
-            fprintf(stderr, "\n\n");
-        }
-    }
-
-    // initialize batch
-    const int n_prompts = prompts.size();
-    struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
-
-    // allocate output
     const int n_embd = llama_n_embd(model);
-    std::vector<float> embeddings(n_prompts * n_embd, 0);
-    float * emb = embeddings.data();
+    auto * embeddings = llama_get_embeddings(ctx);
 
-    // break into batches
-    int p = 0; // number of prompts processed already
-    int s = 0; // number of prompts in current batch
-    for (int k = 0; k < n_prompts; k++) {
-        // clamp to n_batch tokens
-        auto & inp = inputs[k];
-
-        const uint64_t n_toks = inp.size();
-
-        // encode if at capacity
-        if (batch.n_tokens + n_toks > n_batch) {
-            float * out = emb + p * n_embd;
-            batch_decode(ctx, batch, out, s, n_embd);
-            llama_batch_clear(batch);
-            p += s;
-            s = 0;
-        }
-
-        // add to batch
-        batch_add_seq(batch, inp, s);
-        s += 1;
+    // l2-normalize embeddings
+    float norm = 0;
+    for (int i = 0; i < n_embd; i++) {
+        norm += embeddings[i] * embeddings[i];
+    }
+    norm = sqrt(norm);
+    for (int i = 0; i < n_embd; i++) {
+        embeddings[i] /= norm;
     }
 
-    // final batch
-    float * out = emb + p * n_embd;
-    batch_decode(ctx, batch, out, s, n_embd);
-
-    // print the first part of the embeddings or for a single prompt, the full embedding
-    fprintf(stdout, "\n");
-    for (int j = 0; j < n_prompts; j++) {
-        fprintf(stdout, "embedding %d: ", j);
-        for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
-            fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
-        }
-        fprintf(stdout, "\n");
+    for (int i = 0; i < n_embd; i++) {
+        printf("%f ", embeddings[i]);
     }
+    printf("\n");
 
-    // print cosine similarity matrix
-    if (n_prompts > 1) {
-        fprintf(stdout, "\n");
-        printf("cosine similarity matrix:\n\n");
-        for (int i = 0; i < n_prompts; i++) {
-            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
     llama_print_timings(ctx);
     llama_free(ctx);
     llama_free_model(model);
+
     llama_backend_free();
 
     return 0;

examples/export-lora/export-lora.cpp

@@ -7,6 +7,8 @@
 #include <string>
 #include <thread>
 
+static const size_t tensor_alignment = 32;
+
 struct lora_info {
     std::string filename;
     float scale;

examples/finetune/README.md

@@ -80,9 +80,9 @@ The LORA rank can be configured for each model tensor type separately with these
   --rank-wk N                LORA rank for wk tensor (default 4)
   --rank-wv N                LORA rank for wv tensor (default 4)
   --rank-wo N                LORA rank for wo tensor (default 4)
-  --rank-ffn_gate N          LORA rank for ffn_gate tensor (default 4)
-  --rank-ffn_down N          LORA rank for ffn_down tensor (default 4)
-  --rank-ffn_up N            LORA rank for ffn_up tensor (default 4)
+  --rank-w1 N                LORA rank for w1 tensor (default 4)
+  --rank-w2 N                LORA rank for w2 tensor (default 4)
+  --rank-w3 N                LORA rank for w3 tensor (default 4)
 ```
 
 The LORA rank of 'norm' tensors should always be 1.

examples/finetune/finetune.cpp

@@ -60,9 +60,9 @@ struct my_llama_layer {
     struct ggml_tensor * ffn_norm;
 
     // ff
-    struct ggml_tensor * ffn_gate; // w1
-    struct ggml_tensor * ffn_down; // w2
-    struct ggml_tensor * ffn_up;   // w3
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
 };
 
 struct my_llama_model {
@@ -85,9 +85,9 @@ struct my_llama_lora_hparams {
     uint32_t n_rank_wv = 4;
     uint32_t n_rank_wo = 4;
     uint32_t n_rank_ffn_norm = 1;
-    uint32_t n_rank_ffn_gate = 4;
-    uint32_t n_rank_ffn_down = 4;
-    uint32_t n_rank_ffn_up = 4;
+    uint32_t n_rank_w1 = 4;
+    uint32_t n_rank_w2 = 4;
+    uint32_t n_rank_w3 = 4;
     uint32_t n_rank_tok_embeddings = 4;
     uint32_t n_rank_norm = 1;
     uint32_t n_rank_output = 4;
@@ -117,12 +117,12 @@ struct my_llama_lora_layer {
     struct ggml_tensor * ffn_norm_b;
 
     // ff
-    struct ggml_tensor * ffn_gate_a;
-    struct ggml_tensor * ffn_gate_b;
-    struct ggml_tensor * ffn_down_a;
-    struct ggml_tensor * ffn_down_b;
-    struct ggml_tensor * ffn_up_a;
-    struct ggml_tensor * ffn_up_b;
+    struct ggml_tensor * w1_a;
+    struct ggml_tensor * w1_b;
+    struct ggml_tensor * w2_a;
+    struct ggml_tensor * w2_b;
+    struct ggml_tensor * w3_a;
+    struct ggml_tensor * w3_b;
 };
 
 struct my_llama_lora {
@@ -208,9 +208,9 @@ static void print_lora_params(struct my_llama_lora_hparams * params) {
     printf("%s: n_rank_wv             : %u\n", __func__, params->n_rank_wv);
     printf("%s: n_rank_wo             : %u\n", __func__, params->n_rank_wo);
     printf("%s: n_rank_ffn_norm       : %u\n", __func__, params->n_rank_ffn_norm);
-    printf("%s: n_rank_ffn_gate       : %u\n", __func__, params->n_rank_ffn_gate);
-    printf("%s: n_rank_ffn_down       : %u\n", __func__, params->n_rank_ffn_down);
-    printf("%s: n_rank_ffn_up         : %u\n", __func__, params->n_rank_ffn_up);
+    printf("%s: n_rank_w1             : %u\n", __func__, params->n_rank_w1);
+    printf("%s: n_rank_w2             : %u\n", __func__, params->n_rank_w2);
+    printf("%s: n_rank_w3             : %u\n", __func__, params->n_rank_w3);
     printf("%s: n_rank_tok_embeddings : %u\n", __func__, params->n_rank_tok_embeddings);
     printf("%s: n_rank_norm           : %u\n", __func__, params->n_rank_norm);
     printf("%s: n_rank_output         : %u\n", __func__, params->n_rank_output);
@@ -319,9 +319,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         layer.wv             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_V, i));
         layer.wo             = llama_get_model_tensor(input, tni(LLM_TENSOR_ATTN_OUT, i));
         layer.ffn_norm       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_NORM, i));
-        layer.ffn_gate       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
-        layer.ffn_down       = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
-        layer.ffn_up         = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
+        layer.w1             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_GATE, i));
+        layer.w2             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_DOWN, i));
+        layer.w3             = llama_get_model_tensor(input, tni(LLM_TENSOR_FFN_UP, i));
 
         assert_shape_1d(layer.attention_norm, hparams.n_embd);
         assert_shape_2d(layer.wq,             hparams.n_embd, hparams.n_embd);
@@ -329,9 +329,9 @@ static void init_model(struct llama_model * input, struct my_llama_model * model
         assert_shape_2d(layer.wv,             hparams.n_embd, hparams.n_embd_gqa());
         assert_shape_2d(layer.wo,             hparams.n_embd, hparams.n_embd);
         assert_shape_1d(layer.ffn_norm,       hparams.n_embd);
-        assert_shape_2d(layer.ffn_gate,       hparams.n_embd, hparams.n_ff);
-        assert_shape_2d(layer.ffn_down,       hparams.n_ff,   hparams.n_embd);
-        assert_shape_2d(layer.ffn_up,         hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.w1,             hparams.n_embd, hparams.n_ff);
+        assert_shape_2d(layer.w2,             hparams.n_ff,   hparams.n_embd);
+        assert_shape_2d(layer.w3,             hparams.n_embd, hparams.n_ff);
     }
 }
 
@@ -362,12 +362,12 @@ static void set_param_lora(struct my_llama_lora * lora) {
         ggml_set_param(ctx, layer.wo_b);
         ggml_set_param(ctx, layer.ffn_norm_a);
         ggml_set_param(ctx, layer.ffn_norm_b);
-        ggml_set_param(ctx, layer.ffn_gate_a);
-        ggml_set_param(ctx, layer.ffn_gate_b);
-        ggml_set_param(ctx, layer.ffn_down_a);
-        ggml_set_param(ctx, layer.ffn_down_b);
-        ggml_set_param(ctx, layer.ffn_up_a);
-        ggml_set_param(ctx, layer.ffn_up_b);
+        ggml_set_param(ctx, layer.w1_a);
+        ggml_set_param(ctx, layer.w1_b);
+        ggml_set_param(ctx, layer.w2_a);
+        ggml_set_param(ctx, layer.w2_b);
+        ggml_set_param(ctx, layer.w3_a);
+        ggml_set_param(ctx, layer.w3_b);
     }
 }
 
@@ -435,12 +435,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         layer.ffn_norm_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, n_embd);
         layer.ffn_norm_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_norm, 1);
 
-        layer.ffn_gate_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_embd);
-        layer.ffn_gate_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_gate, n_ff);
-        layer.ffn_down_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_ff);
-        layer.ffn_down_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_down, n_embd);
-        layer.ffn_up_a   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_embd);
-        layer.ffn_up_b   = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_ffn_up,   n_ff);
+        layer.w1_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_embd);
+        layer.w1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w1, n_ff);
+        layer.w2_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_ff);
+        layer.w2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w2, n_embd);
+        layer.w3_a = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_embd);
+        layer.w3_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, lparams.n_rank_w3, n_ff);
 
         ggml_set_name(layer.attention_norm_a, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_a", i));
         ggml_set_name(layer.attention_norm_b, tni(LLM_TENSOR_ATTN_NORM, ".weight.lora_b", i));
@@ -454,12 +454,12 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
         ggml_set_name(layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  ".weight.lora_b", i));
         ggml_set_name(layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_a", i));
         ggml_set_name(layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  ".weight.lora_b", i));
-        ggml_set_name(layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
-        ggml_set_name(layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
-        ggml_set_name(layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
-        ggml_set_name(layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
-        ggml_set_name(layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
-        ggml_set_name(layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
+        ggml_set_name(layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_a", i));
+        ggml_set_name(layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  ".weight.lora_b", i));
+        ggml_set_name(layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_a", i));
+        ggml_set_name(layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  ".weight.lora_b", i));
+        ggml_set_name(layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_a", i));
+        ggml_set_name(layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    ".weight.lora_b", i));
     }
 
     set_param_lora(lora);
@@ -497,12 +497,12 @@ static void randomize_lora(struct my_llama_lora * lora, int seed, float mean, fl
         randomize_tensor_normal(layer.ffn_norm_a, rnd);
         ggml_set_zero(layer.ffn_norm_b);
 
-        randomize_tensor_normal(layer.ffn_gate_a, rnd);
-        ggml_set_zero(layer.ffn_gate_b);
-        randomize_tensor_normal(layer.ffn_down_a, rnd);
-        ggml_set_zero(layer.ffn_down_b);
-        randomize_tensor_normal(layer.ffn_up_a, rnd);
-        ggml_set_zero(layer.ffn_up_b);
+        randomize_tensor_normal(layer.w1_a, rnd);
+        ggml_set_zero(layer.w1_b);
+        randomize_tensor_normal(layer.w2_a, rnd);
+        ggml_set_zero(layer.w2_b);
+        randomize_tensor_normal(layer.w3_a, rnd);
+        ggml_set_zero(layer.w3_b);
     }
 
     free_random_normal_distribution(rnd);
@@ -610,13 +610,13 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
 
         struct ggml_tensor * attention_norm = add_to_f32(ctx, layer.attention_norm, ggml_mul_mat(ctx, llayer.attention_norm_a, llayer.attention_norm_b));
         struct ggml_tensor * ffn_norm = add_to_f32(ctx, layer.ffn_norm, ggml_mul_mat(ctx, llayer.ffn_norm_a, llayer.ffn_norm_b));
-        struct ggml_tensor * wq       = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
-        struct ggml_tensor * wk       = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
-        struct ggml_tensor * wv       = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
-        struct ggml_tensor * wo       = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
-        struct ggml_tensor * ffn_gate = add_to_f32(ctx, layer.ffn_gate, ggml_mul_mat(ctx, llayer.ffn_gate_a, llayer.ffn_gate_b));
-        struct ggml_tensor * ffn_down = add_to_f32(ctx, layer.ffn_down, ggml_mul_mat(ctx, llayer.ffn_down_a, llayer.ffn_down_b));
-        struct ggml_tensor * ffn_up   = add_to_f32(ctx, layer.ffn_up, ggml_mul_mat(ctx, llayer.ffn_up_a, llayer.ffn_up_b));
+        struct ggml_tensor * wq = add_to_f32(ctx, layer.wq, ggml_mul_mat(ctx, llayer.wq_a, llayer.wq_b));
+        struct ggml_tensor * wk = add_to_f32(ctx, layer.wk, ggml_mul_mat(ctx, llayer.wk_a, llayer.wk_b));
+        struct ggml_tensor * wv = add_to_f32(ctx, layer.wv, ggml_mul_mat(ctx, llayer.wv_a, llayer.wv_b));
+        struct ggml_tensor * wo = add_to_f32(ctx, layer.wo, ggml_mul_mat(ctx, llayer.wo_a, llayer.wo_b));
+        struct ggml_tensor * w1 = add_to_f32(ctx, layer.w1, ggml_mul_mat(ctx, llayer.w1_a, llayer.w1_b));
+        struct ggml_tensor * w2 = add_to_f32(ctx, layer.w2, ggml_mul_mat(ctx, llayer.w2_a, llayer.w2_b));
+        struct ggml_tensor * w3 = add_to_f32(ctx, layer.w3, ggml_mul_mat(ctx, llayer.w3_a, llayer.w3_b));
 
         struct ggml_tensor * t02 = ggml_rms_norm     (ctx, cur, rms_norm_eps);                       set_name(t02, "t02");     assert_shape_2d(t02, n_embd, N*n_batch);
         struct ggml_tensor * t03 = ggml_repeat       (ctx, attention_norm, t02);                     set_name(t03, "t03");     assert_shape_2d(t03, n_embd, N*n_batch);
@@ -659,11 +659,11 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         struct ggml_tensor * t22 = ggml_rms_norm     (ctx, t21, rms_norm_eps);                       set_name(t22, "t22");     assert_shape_2d(t22, n_embd, N*n_batch);
         struct ggml_tensor * t23 = ggml_repeat       (ctx, ffn_norm, t22);                           set_name(t23, "t23");     assert_shape_2d(t23, n_embd, N*n_batch);
         struct ggml_tensor * t24 = ggml_mul          (ctx, t23, t22);                                set_name(t24, "t24");     assert_shape_2d(t24, n_embd, N*n_batch);
-        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, ffn_up, t24);                             set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
-        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, ffn_gate, t24);                           set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
+        struct ggml_tensor * t25 = ggml_mul_mat      (ctx, w3, t24);                                 set_name(t25, "t25");     assert_shape_2d(t25, n_ff, N*n_batch);
+        struct ggml_tensor * t26 = ggml_mul_mat      (ctx, w1, t24);                                 set_name(t26, "t26");     assert_shape_2d(t26, n_ff, N*n_batch);
         struct ggml_tensor * t27 = ggml_silu         (ctx, t26);                                     set_name(t27, "t27");     assert_shape_2d(t27, n_ff, N*n_batch);
         struct ggml_tensor * t28 = ggml_mul          (ctx, t27, t25);                                set_name(t28, "t28");     assert_shape_2d(t28, n_ff, N*n_batch);
-        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, ffn_down, t28);                           set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
+        struct ggml_tensor * t29 = ggml_mul_mat      (ctx, w2, t28);                                 set_name(t29, "t29");     assert_shape_2d(t29, n_embd, N*n_batch);
         struct ggml_tensor * t30 = ggml_add          (ctx, t29, t21);                                set_name(t30, "t30");     assert_shape_2d(t30, n_embd, N*n_batch);
         cur = t30;
         if (enable_checkpointing) {
@@ -723,9 +723,9 @@ static struct ggml_tensor * llama_build_lora_finetune_graphs(
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wk, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wv, 1.0f));
         ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.wo, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_gate, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_down, 1.0f));
-        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.ffn_up, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w1, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w2, 1.0f));
+        ggml_build_forward_expand(gb, ggml_scale_inplace(ctx, layer.w3, 1.0f));
     }
 
     // allocating checkpoints in one block to reduce memory fragmentation
@@ -798,9 +798,9 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wv,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_V);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_wo,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT);
     GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_norm,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_NORM);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_gate,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_down,       gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
-    GGUF_GET_KEY(fctx, lora->hparams.n_rank_ffn_up,         gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w1,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_GATE);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w2,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN);
+    GGUF_GET_KEY(fctx, lora->hparams.n_rank_w3,             gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_LORA_RANK_FFN_UP);
 
     init_lora(model, lora);
 
@@ -825,12 +825,12 @@ static void load_llama_lora_gguf(struct gguf_context * fctx, struct ggml_context
         copy_tensor_by_name(layer.wo_b,             f_ggml_ctx, ggml_get_name(layer.wo_b));
         copy_tensor_by_name(layer.ffn_norm_a,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_a));
         copy_tensor_by_name(layer.ffn_norm_b,       f_ggml_ctx, ggml_get_name(layer.ffn_norm_b));
-        copy_tensor_by_name(layer.ffn_gate_a,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_a));
-        copy_tensor_by_name(layer.ffn_gate_b,       f_ggml_ctx, ggml_get_name(layer.ffn_gate_b));
-        copy_tensor_by_name(layer.ffn_down_a,       f_ggml_ctx, ggml_get_name(layer.ffn_down_a));
-        copy_tensor_by_name(layer.ffn_down_b,       f_ggml_ctx, ggml_get_name(layer.ffn_down_b));
-        copy_tensor_by_name(layer.ffn_up_a,         f_ggml_ctx, ggml_get_name(layer.ffn_up_a));
-        copy_tensor_by_name(layer.ffn_up_b,         f_ggml_ctx, ggml_get_name(layer.ffn_up_b));
+        copy_tensor_by_name(layer.w1_a,             f_ggml_ctx, ggml_get_name(layer.w1_a));
+        copy_tensor_by_name(layer.w1_b,             f_ggml_ctx, ggml_get_name(layer.w1_b));
+        copy_tensor_by_name(layer.w2_a,             f_ggml_ctx, ggml_get_name(layer.w2_a));
+        copy_tensor_by_name(layer.w2_b,             f_ggml_ctx, ggml_get_name(layer.w2_b));
+        copy_tensor_by_name(layer.w3_a,             f_ggml_ctx, ggml_get_name(layer.w3_a));
+        copy_tensor_by_name(layer.w3_b,             f_ggml_ctx, ggml_get_name(layer.w3_b));
     }
 }
 
@@ -868,9 +868,9 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_V,       lora->hparams.n_rank_wv);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_ATTN_OUT,     lora->hparams.n_rank_wo);
     gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_NORM,     lora->hparams.n_rank_ffn_norm);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_ffn_gate);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_ffn_down);
-    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_ffn_up);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_GATE,     lora->hparams.n_rank_w1);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_DOWN,     lora->hparams.n_rank_w2);
+    gguf_set_val_u32(fctx, LLM_KV_TRAINING_LORA_RANK_FFN_UP,       lora->hparams.n_rank_w3);
 
     gguf_add_tensor(fctx, lora->tok_embeddings_a);
     gguf_add_tensor(fctx, lora->tok_embeddings_b);
@@ -894,12 +894,12 @@ static void save_llama_lora_gguf(struct gguf_context * fctx, struct my_llama_mod
         gguf_add_tensor(fctx, layer.wo_b);
         gguf_add_tensor(fctx, layer.ffn_norm_a);
         gguf_add_tensor(fctx, layer.ffn_norm_b);
-        gguf_add_tensor(fctx, layer.ffn_gate_a);
-        gguf_add_tensor(fctx, layer.ffn_gate_b);
-        gguf_add_tensor(fctx, layer.ffn_down_a);
-        gguf_add_tensor(fctx, layer.ffn_down_b);
-        gguf_add_tensor(fctx, layer.ffn_up_a);
-        gguf_add_tensor(fctx, layer.ffn_up_b);
+        gguf_add_tensor(fctx, layer.w1_a);
+        gguf_add_tensor(fctx, layer.w1_b);
+        gguf_add_tensor(fctx, layer.w2_a);
+        gguf_add_tensor(fctx, layer.w2_b);
+        gguf_add_tensor(fctx, layer.w3_a);
+        gguf_add_tensor(fctx, layer.w3_b);
     }
 }
 
@@ -1104,12 +1104,12 @@ static void save_as_llama_lora(const char * filename, struct my_llama_lora * lor
         write_tensor(&file, layer.wo_b,             tni(LLM_TENSOR_ATTN_OUT,  i, ".weight.loraB"));
         write_tensor(&file, layer.ffn_norm_a,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraA"));
         write_tensor(&file, layer.ffn_norm_b,       tni(LLM_TENSOR_FFN_NORM,  i, ".weight.loraB"));
-        write_tensor(&file, layer.ffn_gate_a,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
-        write_tensor(&file, layer.ffn_gate_b,       tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
-        write_tensor(&file, layer.ffn_down_a,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
-        write_tensor(&file, layer.ffn_down_b,       tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
-        write_tensor(&file, layer.ffn_up_a,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
-        write_tensor(&file, layer.ffn_up_b,         tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
+        write_tensor(&file, layer.w1_a,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w1_b,             tni(LLM_TENSOR_FFN_GATE,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w2_a,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraA"));
+        write_tensor(&file, layer.w2_b,             tni(LLM_TENSOR_FFN_DOWN,  i, ".weight.loraB"));
+        write_tensor(&file, layer.w3_a,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraA"));
+        write_tensor(&file, layer.w3_b,             tni(LLM_TENSOR_FFN_UP,    i, ".weight.loraB"));
     }
 }
 
@@ -1139,9 +1139,9 @@ struct train_params {
     uint32_t n_rank_wv;
     uint32_t n_rank_wo;
     uint32_t n_rank_ffn_norm;
-    uint32_t n_rank_ffn_gate;
-    uint32_t n_rank_ffn_down;
-    uint32_t n_rank_ffn_up;
+    uint32_t n_rank_w1;
+    uint32_t n_rank_w2;
+    uint32_t n_rank_w3;
     uint32_t n_rank_tok_embeddings;
     uint32_t n_rank_norm;
     uint32_t n_rank_output;
@@ -1152,9 +1152,9 @@ struct train_params {
     bool custom_n_rank_wv;
     bool custom_n_rank_wo;
     bool custom_n_rank_ffn_norm;
-    bool custom_n_rank_ffn_gate;
-    bool custom_n_rank_ffn_down;
-    bool custom_n_rank_ffn_up;
+    bool custom_n_rank_w1;
+    bool custom_n_rank_w2;
+    bool custom_n_rank_w3;
     bool custom_n_rank_tok_embeddings;
     bool custom_n_rank_norm;
     bool custom_n_rank_output;
@@ -1186,9 +1186,9 @@ static struct train_params get_default_train_params() {
     params.n_rank_wv             = 4;
     params.n_rank_wo             = 4;
     params.n_rank_ffn_norm       = 1;
-    params.n_rank_ffn_gate       = 4;
-    params.n_rank_ffn_down       = 4;
-    params.n_rank_ffn_up         = 4;
+    params.n_rank_w1             = 4;
+    params.n_rank_w2             = 4;
+    params.n_rank_w3             = 4;
     params.n_rank_tok_embeddings = 4;
     params.n_rank_norm           = 1;
     params.n_rank_output         = 4;
@@ -1199,9 +1199,9 @@ static struct train_params get_default_train_params() {
     params.custom_n_rank_wv             = false;
     params.custom_n_rank_wo             = false;
     params.custom_n_rank_ffn_norm       = false;
-    params.custom_n_rank_ffn_gate       = false;
-    params.custom_n_rank_ffn_down       = false;
-    params.custom_n_rank_ffn_up         = false;
+    params.custom_n_rank_w1             = false;
+    params.custom_n_rank_w2             = false;
+    params.custom_n_rank_w3             = false;
     params.custom_n_rank_tok_embeddings = false;
     params.custom_n_rank_norm           = false;
     params.custom_n_rank_output         = false;
@@ -1232,9 +1232,9 @@ static void train_print_usage(int argc, char ** argv, const struct train_params
     fprintf(stderr, "  --rank-wk N                LORA rank for wk tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wv N                LORA rank for wv tensor, overrides default rank.\n");
     fprintf(stderr, "  --rank-wo N                LORA rank for wo tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-ffn_gate N          LORA rank for ffn_gate tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-ffn_down N          LORA rank for ffn_down tensor, overrides default rank.\n");
-    fprintf(stderr, "  --rank-ffn_up N            LORA rank for ffn_up tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w1 N                LORA rank for w1 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w2 N                LORA rank for w2 tensor, overrides default rank.\n");
+    fprintf(stderr, "  --rank-w3 N                LORA rank for w3 tensor, overrides default rank.\n");
 
     print_common_train_usage(argc, argv, &params->common);
 }
@@ -1369,27 +1369,27 @@ static bool train_params_parse(int argc, char ** argv, struct train_params * par
             }
             params->n_rank_wo = std::stoi(argv[i]);
             params->custom_n_rank_wo = true;
-        } else if (arg == "--rank-ffn_gate") {
+        } else if (arg == "--rank-w1") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_ffn_gate = std::stoi(argv[i]);
-            params->custom_n_rank_ffn_gate = true;
-        } else if (arg == "--rank-ffn_down") {
+            params->n_rank_w1 = std::stoi(argv[i]);
+            params->custom_n_rank_w1 = true;
+        } else if (arg == "--rank-w2") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_ffn_down = std::stoi(argv[i]);
-            params->custom_n_rank_ffn_down = true;
-        } else if (arg == "--rank-ffn_up") {
+            params->n_rank_w2 = std::stoi(argv[i]);
+            params->custom_n_rank_w2 = true;
+        } else if (arg == "--rank-w3") {
             if (++i >= argc) {
                 invalid_param = true;
                 break;
             }
-            params->n_rank_ffn_up = std::stoi(argv[i]);
-            params->custom_n_rank_ffn_up = true;
+            params->n_rank_w3 = std::stoi(argv[i]);
+            params->custom_n_rank_w3 = true;
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             train_print_usage(argc, argv, &default_params);
@@ -1452,12 +1452,12 @@ static int64_t get_parameter_count(struct my_llama_lora* lora) {
         nx += ggml_nelements(layer.wo_b);
         nx += ggml_nelements(layer.ffn_norm_a);
         nx += ggml_nelements(layer.ffn_norm_b);
-        nx += ggml_nelements(layer.ffn_gate_a);
-        nx += ggml_nelements(layer.ffn_gate_b);
-        nx += ggml_nelements(layer.ffn_down_a);
-        nx += ggml_nelements(layer.ffn_down_b);
-        nx += ggml_nelements(layer.ffn_up_a);
-        nx += ggml_nelements(layer.ffn_up_b);
+        nx += ggml_nelements(layer.w1_a);
+        nx += ggml_nelements(layer.w1_b);
+        nx += ggml_nelements(layer.w2_a);
+        nx += ggml_nelements(layer.w2_b);
+        nx += ggml_nelements(layer.w3_a);
+        nx += ggml_nelements(layer.w3_b);
     }
     return nx;
 }
@@ -1511,9 +1511,9 @@ int main(int argc, char ** argv) {
     uint32_t n_rank_wv                 = params.custom_n_rank_wv             ? params.n_rank_wv             : params.lora_r;
     uint32_t n_rank_wo                 = params.custom_n_rank_wo             ? params.n_rank_wo             : params.lora_r;
     uint32_t n_rank_ffn_norm           = params.custom_n_rank_ffn_norm       ? params.n_rank_ffn_norm       : 1;
-    uint32_t n_rank_ffn_gate           = params.custom_n_rank_ffn_gate       ? params.n_rank_ffn_gate       : params.lora_r;
-    uint32_t n_rank_ffn_down           = params.custom_n_rank_ffn_down       ? params.n_rank_ffn_down       : params.lora_r;
-    uint32_t n_rank_ffn_up             = params.custom_n_rank_ffn_up         ? params.n_rank_ffn_up         : params.lora_r;
+    uint32_t n_rank_w1                 = params.custom_n_rank_w1             ? params.n_rank_w1             : params.lora_r;
+    uint32_t n_rank_w2                 = params.custom_n_rank_w2             ? params.n_rank_w2             : params.lora_r;
+    uint32_t n_rank_w3                 = params.custom_n_rank_w3             ? params.n_rank_w3             : params.lora_r;
     uint32_t n_rank_tok_embeddings     = params.custom_n_rank_tok_embeddings ? params.n_rank_tok_embeddings : params.lora_r;
     uint32_t n_rank_norm               = params.custom_n_rank_norm           ? params.n_rank_norm           : 1;
     uint32_t n_rank_output             = params.custom_n_rank_output         ? params.n_rank_output         : params.lora_r;
@@ -1523,15 +1523,15 @@ int main(int argc, char ** argv) {
     lora.hparams.n_rank_wv             = n_rank_wv;
     lora.hparams.n_rank_wo             = n_rank_wo;
     lora.hparams.n_rank_ffn_norm       = n_rank_ffn_norm;
-    lora.hparams.n_rank_ffn_gate       = n_rank_ffn_gate;
-    lora.hparams.n_rank_ffn_down       = n_rank_ffn_down;
-    lora.hparams.n_rank_ffn_up         = n_rank_ffn_up;
+    lora.hparams.n_rank_w1             = n_rank_w1;
+    lora.hparams.n_rank_w2             = n_rank_w2;
+    lora.hparams.n_rank_w3             = n_rank_w3;
     lora.hparams.n_rank_tok_embeddings = n_rank_tok_embeddings;
     lora.hparams.n_rank_norm           = n_rank_norm;
     lora.hparams.n_rank_output         = n_rank_output;
 
     // set opt params from command line
-    opt->params = ggml_opt_default_params(GGML_OPT_TYPE_ADAM);
+    opt->params = ggml_opt_default_params(GGML_OPT_ADAM);
     opt->params.print_forward_graph     = false;
     opt->params.print_backward_graph    = false;
     opt->params.graph_size              = LLAMA_TRAIN_MAX_NODES;
@@ -1566,9 +1566,9 @@ int main(int argc, char ** argv) {
         || (lora.hparams.n_rank_wv             != n_rank_wv)
         || (lora.hparams.n_rank_wo             != n_rank_wo)
         || (lora.hparams.n_rank_ffn_norm       != n_rank_ffn_norm)
-        || (lora.hparams.n_rank_ffn_gate       != n_rank_ffn_gate)
-        || (lora.hparams.n_rank_ffn_down       != n_rank_ffn_down)
-        || (lora.hparams.n_rank_ffn_up         != n_rank_ffn_up)
+        || (lora.hparams.n_rank_w1             != n_rank_w1)
+        || (lora.hparams.n_rank_w2             != n_rank_w2)
+        || (lora.hparams.n_rank_w3             != n_rank_w3)
         || (lora.hparams.n_rank_tok_embeddings != n_rank_tok_embeddings)
         || (lora.hparams.n_rank_norm           != n_rank_norm)
         || (lora.hparams.n_rank_output         != n_rank_output)

examples/gguf-split/CMakeLists.txt

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

examples/gguf-split/README.md

@@ -1,9 +0,0 @@
-## GGUF split Example
-
-CLI to split / merge GGUF files.
-
-**Command line options:**
-
-- `--split`: split GGUF to multiple GGUF, default operation.
-- `--split-max-tensors`: maximum tensors in each split: default(128)
-- `--merge`: merge multiple GGUF to a single GGUF.

examples/gguf-split/gguf-split.cpp

@@ -1,468 +0,0 @@
-#include "llama.h"
-#include "common.h"
-
-#include <algorithm>
-#include <cmath>
-#include <cstdlib>
-#include <fstream>
-#include <string>
-#include <vector>
-
-#include <stdio.h>
-#include <string.h>
-#include <climits>
-#include <stdexcept>
-
-#if defined(_WIN32)
-    #include <windows.h>
-    #ifndef PATH_MAX
-        #define PATH_MAX MAX_PATH
-    #endif
-    #include <io.h>
-#endif
-
-enum split_operation : uint8_t {
-    SPLIT_OP_SPLIT,
-    SPLIT_OP_MERGE,
-};
-
-struct split_params {
-    split_operation operation = SPLIT_OP_SPLIT;
-    int n_split_tensors = 128;
-    std::string input;
-    std::string output;
-};
-
-static void split_print_usage(const char * executable) {
-    const split_params default_params;
-    printf("\n");
-    printf("usage: %s [options] GGUF_IN GGUF_OUT\n", executable);
-    printf("\n");
-    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("\n");
-}
-
-static bool split_params_parse_ex(int argc, const char ** argv, split_params & params) {
-    std::string arg;
-    const std::string arg_prefix = "--";
-    bool invalid_param = false;
-
-    int arg_idx = 1;
-    for (; arg_idx < argc && strncmp(argv[arg_idx], "--", 2) == 0; arg_idx++) {
-        arg = argv[arg_idx];
-        if (arg.compare(0, arg_prefix.size(), arg_prefix) == 0) {
-            std::replace(arg.begin(), arg.end(), '_', '-');
-        }
-
-        bool arg_found = false;
-        if (arg == "-h" || arg == "--help") {
-            split_print_usage(argv[0]);
-            exit(0);
-        }
-        if (arg == "--version") {
-            fprintf(stderr, "version: %d (%s)\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
-            fprintf(stderr, "built with %s for %s\n", LLAMA_COMPILER, LLAMA_BUILD_TARGET);
-            exit(0);
-        }
-
-        if (arg == "--merge") {
-            arg_found = true;
-            params.operation = SPLIT_OP_MERGE;
-        }
-        if (arg == "--split") {
-            arg_found = true;
-            params.operation = SPLIT_OP_SPLIT;
-        }
-        if (arg == "--split-max-tensors") {
-            if (++arg_idx >= argc) {
-                invalid_param = true;
-                break;
-            }
-            arg_found = true;
-            params.n_split_tensors = atoi(argv[arg_idx]);
-        }
-
-        if (!arg_found) {
-            throw std::invalid_argument("error: unknown argument: " + arg);
-        }
-    }
-
-    if (invalid_param) {
-        throw std::invalid_argument("error: invalid parameter for argument: " + arg);
-    }
-
-    if (argc - arg_idx < 2) {
-        printf("%s: bad arguments\n", argv[0]);
-        split_print_usage(argv[0]);
-        return false;
-    }
-
-    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);
-        }
-    }
-    catch (const std::invalid_argument & ex) {
-        fprintf(stderr, "%s\n", ex.what());
-        split_print_usage(argv[0]);
-        exit(EXIT_FAILURE);
-    }
-    return result;
-}
-
-static void zeros(std::ofstream & file, size_t n) {
-    char zero = 0;
-    for (size_t i = 0; i < n; ++i) {
-        file.write(&zero, 1);
-    }
-}
-
-struct split_strategy {
-    const split_params params;
-    std::ifstream & f_input;
-    struct gguf_context * ctx_gguf;
-    struct ggml_context * ctx_meta = NULL;
-    const int n_tensors;
-
-    const int n_split;
-    int i_split = 0;
-
-    int i_tensor = 0;
-
-    std::vector<uint8_t> read_data;
-
-    struct gguf_context * ctx_out;
-    std::ofstream fout;
-
-    split_strategy(const split_params & params,
-            std::ifstream & f_input,
-            struct gguf_context * ctx_gguf,
-            struct ggml_context * ctx_meta) :
-        params(params),
-        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)) {
-        }
-
-    bool should_split() const {
-        return i_tensor < n_tensors && i_tensor % params.n_split_tensors == 0;
-    }
-
-    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);
-        }
-        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);
-        }
-
-        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());
-
-        fout.close();
-        gguf_free(ctx_out);
-
-        fprintf(stderr, "\033[3Ddone\n");
-    }
-};
-
-static void gguf_split(const split_params & split_params) {
-    struct ggml_context * ctx_meta = NULL;
-
-    struct gguf_init_params params = {
-        /*.no_alloc = */ true,
-        /*.ctx      = */ &ctx_meta,
-    };
-
-    std::ifstream f_input(split_params.input.c_str(), std::ios::binary);
-    if (!f_input.is_open()) {
-        fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_params.input.c_str());
-        exit(EXIT_FAILURE);
-    }
-
-    auto * ctx_gguf = gguf_init_from_file(split_params.input.c_str(), params);
-    if (!ctx_gguf) {
-        fprintf(stderr, "%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
-        exit(EXIT_FAILURE);
-    }
-
-    split_strategy strategy(split_params, f_input, ctx_gguf, ctx_meta);
-
-    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();
-        }
-    }
-    strategy.split_end();
-
-    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);
-}
-
-static void gguf_merge(const split_params & split_params) {
-    fprintf(stderr, "%s: %s -> %s\n",
-            __func__, split_params.input.c_str(),
-            split_params.output.c_str());
-    int n_split = 1;
-    int total_tensors = 0;
-
-    auto * ctx_out = gguf_init_empty();
-    std::ofstream fout(split_params.output.c_str(), std::ios::binary);
-    fout.exceptions(std::ofstream::failbit); // fail fast on write errors
-
-    std::vector<uint8_t> read_data;
-    std::vector<ggml_context *> ctx_metas;
-    std::vector<gguf_context *> ctx_ggufs;
-
-    char split_path[PATH_MAX] = {0};
-    strncpy(split_path, split_params.input.c_str(), sizeof(split_path) - 1);
-    char split_prefix[PATH_MAX] = {0};
-
-    // First pass to find KV and tensors metadata
-    for (int i_split = 0; i_split < n_split; i_split++) {
-        struct ggml_context * ctx_meta = NULL;
-
-        struct gguf_init_params params = {
-            /*.no_alloc = */ true,
-            /*.ctx      = */ &ctx_meta,
-        };
-
-        if (i_split > 0) {
-            llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
-        }
-        fprintf(stderr, "%s: reading metadata %s ...", __func__, split_path);
-
-        auto * ctx_gguf = gguf_init_from_file(split_path, params);
-        if (!ctx_gguf) {
-            fprintf(stderr, "\n%s:  failed to load input GGUF from %s\n", __func__, split_params.input.c_str());
-            exit(EXIT_FAILURE);
-        }
-        ctx_ggufs.push_back(ctx_gguf);
-        ctx_metas.push_back(ctx_meta);
-
-        if (i_split == 0) {
-            auto key_n_split = gguf_find_key(ctx_gguf, LLM_KV_SPLIT_COUNT);
-            if (key_n_split < 0) {
-                fprintf(stderr,
-                        "\n%s: input file does not contain %s metadata\n",
-                        __func__,
-                        LLM_KV_SPLIT_COUNT);
-                gguf_free(ctx_gguf);
-                ggml_free(ctx_meta);
-                gguf_free(ctx_out);
-                fout.close();
-                exit(EXIT_FAILURE);
-            }
-
-            n_split = gguf_get_val_u16(ctx_gguf, key_n_split);
-            if (n_split < 1) {
-                fprintf(stderr,
-                        "\n%s: input file does not contain a valid split count %d\n",
-                        __func__,
-                        n_split);
-                gguf_free(ctx_gguf);
-                ggml_free(ctx_meta);
-                gguf_free(ctx_out);
-                fout.close();
-                exit(EXIT_FAILURE);
-            }
-
-            // Verify the file naming and extract split_prefix
-            if (!llama_split_prefix(split_prefix, sizeof (split_prefix), split_path, i_split, n_split)) {
-                fprintf(stderr, "\n%s: unexpected input file name: %s"
-                                " i_split=%d"
-                                " n_split=%d\n", __func__,
-                        split_path, i_split, n_split);
-                gguf_free(ctx_gguf);
-                ggml_free(ctx_meta);
-                gguf_free(ctx_out);
-                fout.close();
-                exit(EXIT_FAILURE);
-            }
-
-            // Do not trigger merge if we try to merge again the output
-            gguf_set_val_u16(ctx_gguf, LLM_KV_SPLIT_COUNT, 0);
-
-            // Set metadata from the first split
-            gguf_set_kv(ctx_out, ctx_gguf);
-        }
-
-        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
-        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
-            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
-            struct ggml_tensor * t = ggml_get_tensor(ctx_meta, t_name);
-            gguf_add_tensor(ctx_out, t);
-        }
-        total_tensors += n_tensors;
-
-        fprintf(stderr, "\033[3Ddone\n");
-    }
-
-    // placeholder for the meta data
-    {
-        auto meta_size = gguf_get_meta_size(ctx_out);
-        ::zeros(fout, meta_size);
-    }
-
-    // Write tensors data
-    for (int i_split = 0; i_split < n_split; i_split++) {
-        llama_split_path(split_path, sizeof(split_path), split_prefix, i_split, n_split);
-        std::ifstream f_input(split_path, std::ios::binary);
-        if (!f_input.is_open()) {
-            fprintf(stderr, "%s:  failed to open input GGUF from %s\n", __func__, split_path);
-            for (uint32_t i = 0; i < ctx_ggufs.size(); i++) {
-                gguf_free(ctx_ggufs[i]);
-                ggml_free(ctx_metas[i]);
-            }
-            gguf_free(ctx_out);
-            fout.close();
-            exit(EXIT_FAILURE);
-        }
-        fprintf(stderr, "%s: writing tensors %s ...", __func__, split_path);
-
-        auto * ctx_gguf = ctx_ggufs[i_split];
-        auto * ctx_meta = ctx_metas[i_split];
-
-        auto n_tensors = gguf_get_n_tensors(ctx_gguf);
-        for (int i_tensor = 0; i_tensor < n_tensors; i_tensor++) {
-            const char * t_name = gguf_get_tensor_name(ctx_gguf, i_tensor);
-            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);
-            }
-
-            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);
-
-            // write tensor data + padding
-            fout.write((const char *)read_data.data(), n_bytes);
-            zeros(fout, GGML_PAD(n_bytes, GGUF_DEFAULT_ALIGNMENT) - n_bytes);
-        }
-
-        gguf_free(ctx_gguf);
-        ggml_free(ctx_meta);
-        f_input.close();
-        fprintf(stderr, "\033[3Ddone\n");
-    }
-
-    {
-        // 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());
-
-        fout.close();
-        gguf_free(ctx_out);
-    }
-
-    fprintf(stderr, "%s: %s merged from %d split with %d tensors.\n",
-            __func__, split_params.output.c_str(), n_split, total_tensors);
-}
-
-int main(int argc, const char ** argv) {
-    if (argc < 3) {
-        split_print_usage(argv[0]);
-    }
-
-    split_params params;
-    split_params_parse(argc, argv, params);
-
-    switch (params.operation) {
-        case SPLIT_OP_SPLIT: gguf_split(params);
-            break;
-        case SPLIT_OP_MERGE: gguf_merge(params);
-            break;
-        default: split_print_usage(argv[0]);
-            exit(EXIT_FAILURE);
-    }
-
-    return 0;
-}

examples/gguf/gguf.cpp

@@ -211,7 +211,6 @@ static bool gguf_ex_read_1(const std::string & fname) {
                 for (int j = 0; j < ggml_nelements(cur); ++j) {
                     if (data[j] != 100 + i) {
                         fprintf(stderr, "%s: tensor[%d]: data[%d] = %f\n", __func__, i, j, data[j]);
-                        gguf_free(ctx);
                         return false;
                     }
                 }

examples/gritlm/CMakeLists.txt

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

examples/gritlm/README.md

@@ -1,62 +0,0 @@
-## Generative Representational Instruction Tuning (GRIT) Example
-[gritlm] a model which can generate embeddings as well as "normal" text
-generation depending on the instructions in the prompt.
-
-* Paper: https://arxiv.org/pdf/2402.09906.pdf
-
-### Retrieval-Augmented Generation (RAG) use case
-One use case for `gritlm` is to use it with RAG. If we recall how RAG works is
-that we take documents that we want to use as context, to ground the large
-language model (LLM), and we create token embeddings for them. We then store
-these token embeddings in a vector database.
-
-When we perform a query, prompt the LLM, we will first create token embeddings
-for the query and then search the vector database to retrieve the most
-similar vectors, and return those documents so they can be passed to the LLM as
-context. Then the query and the context will be passed to the LLM which will
-have to _again_ create token embeddings for the query. But because gritlm is used
-the first query can be cached and the second query tokenization generation does
-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
-```
-
-Run the example using the downloaded model:
-```console
-$ ./gritlm -m 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
-Cosine similarity between "Generative Representational Instruction Tuning" and "A purely peer-to-peer version of electronic cash w" is: 0.112
-Cosine similarity between "Generative Representational Instruction Tuning" and "All text-based language problems can be reduced to" is: 0.547
-
-Oh, brave adventurer, who dared to climb
-The lofty peak of Mt. Fuji in the night,
-When shadows lurk and ghosts do roam,
-And darkness reigns, a fearsome sight.
-
-Thou didst set out, with heart aglow,
-To conquer this mountain, so high,
-And reach the summit, where the stars do glow,
-And the moon shines bright, up in the sky.
-
-Through the mist and fog, thou didst press on,
-With steadfast courage, and a steadfast will,
-Through the darkness, thou didst not be gone,
-But didst climb on, with a steadfast skill.
-
-At last, thou didst reach the summit's crest,
-And gazed upon the world below,
-And saw the beauty of the night's best,
-And felt the peace, that only nature knows.
-
-Oh, brave adventurer, who dared to climb
-The lofty peak of Mt. Fuji in the night,
-Thou art a hero, in the eyes of all,
-For thou didst conquer this mountain, so bright.
-```
-
-[gritlm]: https://github.com/ContextualAI/gritlm

examples/gritlm/gritlm.cpp

@@ -1,215 +0,0 @@
-#include "common.h"
-#include "llama.h"
-
-#include <string>
-#include <vector>
-
-// #define GRIT_DEBUG
-
-static std::vector<std::vector<float>> encode(llama_context * ctx, const std::vector<std::string> & sentences, const std::string & instruction) {
-    std::vector<std::vector<float>> result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-
-    llama_batch batch = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    for (uint64_t i = 0; i < sentences.size(); i++) {
-        llama_batch_clear(batch);
-
-        const std::string input_string = instruction + sentences[i];
-
-        std::vector<llama_token> inputs = llama_tokenize(mdl, input_string, true, false);
-
-        const int32_t n_toks = inputs.size();
-
-        // GritLM seems to have EOS = ""
-        // https://github.com/ContextualAI/gritlm/blob/92025b16534712b31b3c4aaaf069350e222bd5f8/gritlm/gritlm.py#L18
-        // inputs.push_back(llama_token_eos(mdl));
-
-        // we want to ignore instruction tokens for mean pooling
-        const int32_t n_inst = llama_tokenize(mdl, instruction, true, false).size();
-
-#ifdef GRIT_DEBUG
-        // debug tokens - should be matching as referenced in the GritLM sample
-        std::for_each(inputs.begin(), inputs.end(), [&ctx](llama_token t) {
-            std::printf("[%u:%s]", t, llama_token_to_piece(ctx, t).c_str());
-        });
-        std::printf("\n");
-#endif
-
-        // add input to batch (this increments n_tokens)
-        for (int32_t j = 0; j < n_toks; j++) {
-            llama_batch_add(batch, inputs[j], j, { 0 }, j >= n_inst);
-        }
-
-        // clear previous kv_cache values (irrelevant for embeddings)
-        llama_kv_cache_clear(ctx);
-        llama_set_causal_attn(ctx, false);
-
-        // run model
-        llama_decode(ctx, batch);
-
-        // get embedding dimensions
-        uint64_t n_embd = llama_n_embd(mdl);
-
-        // allocate embedding output
-        std::vector<float> emb_unorm(n_embd, 0.0f);
-
-        // sum up all token embeddings
-        for (int32_t k = n_inst; k < n_toks; k++) {
-            float * emb = llama_get_embeddings_ith(ctx, k);
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] += emb[j];
-            }
-        }
-
-        // divide by number of tokens (mean pooling)
-        {
-            const uint64_t n_sent = n_toks - n_inst;
-
-            for (uint64_t j = 0; j < n_embd; j++) {
-                emb_unorm[j] /= n_sent;
-            }
-        }
-
-        std::vector<float> emb_norm(emb_unorm.size());
-        llama_embd_normalize(emb_unorm.data(), emb_norm.data(), n_embd);
-        result.push_back(emb_norm);
-
-#ifdef GRIT_DEBUG
-        // print out emb_norm
-        std::printf("embedding %ld: ", i);
-        for (uint64_t j = 0; j < n_embd; j++) {
-            std::printf("%.5f ", emb_norm[j]);
-        }
-        std::printf("\n\n");
-#endif
-    }
-
-    llama_batch_free(batch);
-
-    return result;
-}
-
-static std::string generate(llama_context * ctx, const std::string & prompt, bool stream) {
-    std::string result;
-
-    const llama_model * mdl = llama_get_model(ctx);
-    llama_token eos_token = llama_token_eos(mdl);
-
-    llama_kv_cache_clear(ctx);
-    llama_set_causal_attn(ctx, true);
-    llama_batch bat = llama_batch_init(llama_n_batch(ctx), 0, 1);
-
-    std::vector<llama_token> inputs = llama_tokenize(mdl, prompt, false, true);
-    int32_t i_current_token = 0;
-
-    while (true) {
-        llama_batch_clear(bat);
-        auto n_inputs = (int32_t)inputs.size();
-        for (int32_t i = 0; i < n_inputs; i++) {
-            llama_batch_add(bat, inputs[i], i_current_token++, { 0 }, i == n_inputs - 1);
-        }
-        inputs.clear();
-
-        llama_decode(ctx, bat);
-        auto logits = llama_get_logits_ith(ctx, bat.n_tokens - 1);
-
-        auto candidates = std::vector<llama_token_data>(llama_n_vocab(mdl));
-        auto n_candidates = (int32_t)candidates.size();
-        for (int32_t token = 0; token < n_candidates; token++) {
-            candidates[token] = llama_token_data{ token, logits[token], 0.0f };
-        }
-        auto candidates_p = llama_token_data_array{ candidates.data(), candidates.size(), false };
-
-        llama_token token = llama_sample_token_greedy(ctx, &candidates_p);
-        if (token == eos_token) {
-            break;
-        }
-
-        std::string piece = llama_token_to_piece(ctx, token);
-        if (stream) {
-            std::printf("%s", piece.c_str());
-            std::fflush(stdout);
-        }
-
-        inputs.push_back(token);
-
-        result += piece;
-    }
-
-    if (stream) {
-        std::printf("\n");
-    }
-
-    llama_batch_free(bat);
-
-    return result;
-}
-
-static std::string gritlm_instruction(const std::string & instruction) {
-    return !instruction.empty() ? "<|user|>\n" + instruction + "\n<|embed|>\n" : "<|embed|>\n";
-}
-
-int main(int argc, char * argv[]) {
-    gpt_params params;
-    if (!gpt_params_parse(argc, argv, params)) {
-        return 1;
-    }
-
-    llama_model_params mparams = llama_model_params_from_gpt_params(params);
-    llama_context_params cparams = llama_context_params_from_gpt_params(params);
-
-    llama_backend_init();
-
-    llama_model * mdl = llama_load_model_from_file(params.model.c_str(), mparams);
-
-    // create new context - set to embedding mode
-    cparams.embeddings = true;
-    llama_context * ctx = llama_new_context_with_model(mdl, cparams);
-
-    // ### Embedding/Representation ###
-    // samples taken from: https://github.com/ContextualAI/gritlm#basic
-    {
-        const std::string instruction = "Given a scientific paper title, retrieve the paper's abstract";
-
-        const std::vector<std::string> queries = {
-            "Bitcoin: A Peer-to-Peer Electronic Cash System",
-            "Generative Representational Instruction Tuning",
-        };
-
-        const std::vector<std::string> documents = {
-            "A purely peer-to-peer version of electronic cash would allow online payments to be sent directly from one party to another without going through a financial institution. Digital signatures provide part of the solution, but the main benefits are lost if a trusted third party is still required to prevent double-spending. We propose a solution to the double-spending problem using a peer-to-peer network. The network timestamps transactions by hashing them into an ongoing chain of hash-based proof-of-work, forming a record that cannot be changed without redoing the proof-of-work. The longest chain not only serves as proof of the sequence of events witnessed, but proof that it came from the largest pool of CPU power. As long as a majority of CPU power is controlled by nodes that are not cooperating to attack the network, they'll generate the longest chain and outpace attackers. The network itself requires minimal structure. Messages are broadcast on a best effort basis, and nodes can leave and rejoin the network at will, accepting the longest proof-of-work chain as proof of what happened while they were gone.",
-            "All text-based language problems can be reduced to either generation or embedding. Current models only perform well at one or the other. We introduce generative representational instruction tuning (GRIT) whereby a large language model is trained to handle both generative and embedding tasks by distinguishing between them through instructions. Compared to other open models, our resulting GritLM 7B sets a new state of the art on the Massive Text Embedding Benchmark (MTEB) and outperforms all models up to its size on a range of generative tasks. By scaling up further, GritLM 8X7B outperforms all open generative language models that we tried while still being among the best embedding models. Notably, we find that GRIT matches training on only generative or embedding data, thus we can unify both at no performance loss. Among other benefits, the unification via GRIT speeds up Retrieval-Augmented Generation (RAG) by > 60% for long documents, by no longer requiring separate retrieval and generation models. Models, code, etc. are freely available at https://github.com/ContextualAI/gritlm.",
-        };
-
-        // No need to add instruction for retrieval documents
-        const std::vector<std::vector<float>> d_rep = encode(ctx, documents, gritlm_instruction(""));
-        const std::vector<std::vector<float>> q_rep = encode(ctx, queries,   gritlm_instruction(instruction));
-
-        const int n_embd = llama_n_embd(mdl);
-
-        const float cosine_sim_q0_d0 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q0_d1 = llama_embd_similarity_cos(q_rep[0].data(), d_rep[1].data(), n_embd);
-        const float cosine_sim_q1_d0 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[0].data(), n_embd);
-        const float cosine_sim_q1_d1 = llama_embd_similarity_cos(q_rep[1].data(), d_rep[1].data(), n_embd);
-
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[0].c_str(), cosine_sim_q0_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[0].c_str(), documents[1].c_str(), cosine_sim_q0_d1);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[0].c_str(), cosine_sim_q1_d0);
-        std::printf("Cosine similarity between \"%.50s\" and \"%.50s\" is: %.3f\n", queries[1].c_str(), documents[1].c_str(), cosine_sim_q1_d1);
-    }
-
-    // ### Generation ###
-    // GritLM models are not finetuned with system prompts, as you can just include system-like instructions together with your user instruction
-    {
-        const std::string prompt = "<|user|>\nPlease write me a poem about my recent hike of Mt. Fuji at midnight in the style of Shakespeare.\n<|assistant|>\n";
-        std::string response = generate(ctx, prompt, true);
-    }
-
-    llama_free(ctx);
-    llama_free_model(mdl);
-    llama_backend_free();
-
-    return 0;
-}

examples/imatrix/README.md

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

examples/imatrix/imatrix.cpp

@@ -50,31 +50,11 @@ private:
     void keep_imatrix(int ncall) const;
 };
 
-// remove any prefix and suffixes from the name
-// CUDA0#blk.0.attn_k.weight#0 => blk.0.attn_k.weight
-static std::string filter_tensor_name(const char * name) {
-    std::string wname;
-    const char * p = strchr(name, '#');
-    if (p != NULL) {
-        p = p + 1;
-        const char * q = strchr(p, '#');
-        if (q != NULL) {
-            wname = std::string(p, q - p);
-        } else {
-            wname = p;
-        }
-    } else {
-        wname = name;
-    }
-    return wname;
-}
-
 bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void * user_data) {
     GGML_UNUSED(user_data);
 
     const struct ggml_tensor * src0 = t->src[0];
     const struct ggml_tensor * src1 = t->src[1];
-    std::string wname = 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
@@ -82,7 +62,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         if (t->op == GGML_OP_MUL_MAT_ID) return true; // collect all indirect matrix multiplications
         if (t->op != GGML_OP_MUL_MAT) return false;
         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;
+        if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return false;
         return true;
     }
 
@@ -114,13 +94,12 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
         // 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];
-            wname = filter_tensor_name(src0->name);
-            auto& e = m_stats[wname];
+            auto& e = m_stats[src0->name];
             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]);
+                fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (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
@@ -128,7 +107,7 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             //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);
+                printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, 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];
@@ -150,17 +129,17 @@ bool IMatrixCollector::collect_imatrix(struct ggml_tensor * t, bool ask, void *
             }
         }
     } else {
-        auto& e = m_stats[wname];
+        auto& e = m_stats[src0->name];
         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]);
+            fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
             exit(1); //GGML_ASSERT(false);
         }
         ++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);
+            printf("%s[%d]: %32s, %s, %5d x %5d, %d\n", __func__, m_last_call, src0->name, ggml_op_name(t->op), (int)src1->ne[0], (int)src1->ne[1], (int)src1->type);
         }
         for (int row = 0; row < (int)src1->ne[1]; ++row) {
             const float * x = data + row * src1->ne[0];
@@ -424,7 +403,6 @@ 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;
@@ -590,8 +568,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model_params mparams = llama_model_params_from_gpt_params(params);
 

examples/infill/infill.cpp

@@ -202,8 +202,7 @@ int main(int argc, char ** argv) {
     std::mt19937 rng(params.seed);
 
     LOG("%s: llama backend init\n", __func__);
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model;
     llama_context * ctx;
@@ -378,10 +377,10 @@ int main(int argc, char ** argv) {
     if (params.interactive) {
         const char *control_message;
         if (params.multiline_input) {
-            control_message = " - To return control to LLaMA, end your input with '\\'.\n"
+            control_message = " - To return control to LLaMa, end your input with '\\'.\n"
                               " - To return control without starting a new line, end your input with '/'.\n";
         } else {
-            control_message = " - Press Return to return control to LLaMA.\n"
+            control_message = " - Press Return to return control to LLaMa.\n"
                               " - To return control without starting a new line, end your input with '/'.\n"
                               " - If you want to submit another line, end your input with '\\'.\n";
         }
@@ -447,8 +446,8 @@ int main(int argc, char ** argv) {
                 LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
                     n_past, n_left, n_ctx, params.n_keep, n_discard);
 
-                llama_kv_cache_seq_rm (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_add(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
 
                 n_past -= n_discard;
 

examples/json-schema-pydantic-example.py

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

examples/json-schema-to-grammar.py

@@ -1,10 +1,8 @@
 #!/usr/bin/env python3
 import argparse
-import itertools
 import json
 import re
 import sys
-from typing import Any, Dict, List, Set, Tuple, Union
 
 # whitespace is constrained to a single space char to prevent model "running away" in
 # whitespace. Also maybe improves generation quality?
@@ -14,54 +12,26 @@ 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''',
+      )* "\"" space ''',
     'null': '"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',
-}
-
-RESERVED_NAMES = set(["root", *PRIMITIVE_RULES.keys(), *DATE_RULES.keys()])
 
 INVALID_RULE_CHARS_RE = re.compile(r'[^a-zA-Z0-9-]+')
 GRAMMAR_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"]')
-GRAMMAR_RANGE_LITERAL_ESCAPE_RE = re.compile(r'[\r\n"\]\-\\]')
-GRAMMAR_LITERAL_ESCAPES = {'\r': '\\r', '\n': '\\n', '"': '\\"', '-': '\\-', ']': '\\]'}
+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):
+    def __init__(self, prop_order):
         self._prop_order = prop_order
-        self._allow_fetch = allow_fetch
-        self._dotall = dotall
-        self._raw_pattern = raw_pattern
         self._rules = {'space': SPACE_RULE}
-        self._refs = {}
-        self._refs_being_resolved = set()
 
     def _format_literal(self, literal):
         escaped = GRAMMAR_LITERAL_ESCAPE_RE.sub(
-            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), literal
+            lambda m: GRAMMAR_LITERAL_ESCAPES.get(m.group(0)), json.dumps(literal)
         )
         return f'"{escaped}"'
 
@@ -71,420 +41,64 @@ class SchemaConverter:
             key = esc_name
         else:
             i = 0
-            while f'{esc_name}{i}' in self._rules and self._rules[f'{esc_name}{i}'] != rule:
+            while f'{esc_name}{i}' in self._rules:
                 i += 1
             key = f'{esc_name}{i}'
         self._rules[key] = rule
         return key
 
-    def resolve_refs(self, schema: dict, url: str):
-        '''
-            Resolves all $ref fields in the given schema, fetching any remote schemas,
-            replacing $ref with absolute reference URL and populating self._refs with the
-            respective referenced (sub)schema dictionaries.
-        '''
-        def visit(n: dict):
-            if isinstance(n, list):
-                return [visit(x) for x in n]
-            elif isinstance(n, dict):
-                ref = n.get('$ref')
-                if ref is not None and ref not in self._refs:
-                    if ref.startswith('https://'):
-                        assert self._allow_fetch, 'Fetching remote schemas is not allowed (use --allow-fetch for force)'
-                        import requests
-
-                        frag_split = ref.split('#')
-                        base_url = frag_split[0]
-
-                        target = self._refs.get(base_url)
-                        if target is None:
-                            target = self.resolve_refs(requests.get(ref).json(), base_url)
-                            self._refs[base_url] = target
-
-                        if len(frag_split) == 1 or frag_split[-1] == '':
-                            return target
-                    elif ref.startswith('#/'):
-                        target = schema
-                        ref = f'{url}{ref}'
-                        n['$ref'] = ref
-                    else:
-                        raise ValueError(f'Unsupported ref {ref}')
-
-                    for sel in ref.split('#')[-1].split('/')[1:]:
-                        assert target is not None and sel in target, f'Error resolving ref {ref}: {sel} not in {target}'
-                        target = target[sel]
-
-                    self._refs[ref] = target
-                else:
-                    for v in n.values():
-                        visit(v)
-
-            return n
-        return visit(schema)
-
-    def _generate_union_rule(self, name, alt_schemas):
-        return ' | '.join((
-            self.visit(alt_schema, f'{name}{"-" if name else "alternative-"}{i}')
-            for i, alt_schema in enumerate(alt_schemas)
-        ))
-
-    def _visit_pattern(self, pattern, name):
-        '''
-            Transforms a regular expression pattern into a GBNF rule.
-
-            Input: https://json-schema.org/understanding-json-schema/reference/regular_expressions
-            Output: https://github.com/ggerganov/llama.cpp/blob/master/grammars/README.md
-
-            Unsupported features: negative/positive lookaheads, greedy/non-greedy modifiers.
-
-            Mostly a 1:1 translation, except for {x} / {x,} / {x,y} quantifiers for which
-            we define sub-rules to keep the output lean.
-        '''
-
-        assert pattern.startswith('^') and pattern.endswith('$'), 'Pattern must start with "^" and end with "$"'
-        pattern = pattern[1:-1]
-        sub_rule_ids = {}
-
-        i = 0
-        length = len(pattern)
-
-        def to_rule(s: Tuple[str, bool]) -> str:
-            (txt, is_literal) = s
-            return "\"" + txt + "\"" if is_literal else txt
-
-        def transform() -> Tuple[str, bool]:
-            '''
-                Parse a unit at index i (advancing it), and return its string representation + whether it's a literal.
-            '''
-            nonlocal i
-            nonlocal pattern
-            nonlocal sub_rule_ids
-
-            start = i
-            # For each component of this sequence, store its string representation and whether it's a literal.
-            # We only need a flat structure here to apply repetition operators to the last item, and
-            # to merge literals at the and (we're parsing grouped ( sequences ) recursively and don't treat '|' specially
-            # (GBNF's syntax is luckily very close to regular expressions!)
-            seq: list[Tuple[str, bool]] = []
-
-            def get_dot():
-                if self._dotall:
-                    rule = '[\\U00000000-\\U0010FFFF]'
-                else:
-                    # Accept any character... except \n and \r line break chars (\x0A and \xOD)
-                    rule = '[\\U00000000-\\x09\\x0B\\x0C\\x0E-\\U0010FFFF]'
-                return self._add_rule(f'dot', rule)
-
-            def join_seq():
-                nonlocal seq
-                ret = []
-                for is_literal, g in itertools.groupby(seq, lambda x: x[1]):
-                    if is_literal:
-                        ret.append((''.join(x[0] for x in g), True))
-                    else:
-                        ret.extend(g)
-                if len(ret) == 1:
-                    return ret[0]
-                return (' '.join(to_rule(x) for x in seq), False)
-
-            while i < length:
-                c = pattern[i]
-                if c == '.':
-                    seq.append((get_dot(), False))
-                    i += 1
-                elif c == '(':
-                    i += 1
-                    if i < length:
-                        assert pattern[i] != '?', f'Unsupported pattern syntax "{pattern[i]}" at index {i} of /{pattern}/'
-                    seq.append((f'({to_rule(transform())})', False))
-                elif c == ')':
-                    i += 1
-                    assert start > 0 and pattern[start-1] == '(', f'Unbalanced parentheses; start = {start}, i = {i}, pattern = {pattern}'
-                    return join_seq()
-                elif c == '[':
-                    square_brackets = c
-                    i += 1
-                    while i < length and pattern[i] != ']':
-                        if pattern[i] == '\\':
-                            square_brackets += pattern[i:i+2]
-                            i += 2
-                        else:
-                            square_brackets += pattern[i]
-                            i += 1
-                    assert i < length, f'Unbalanced square brackets; start = {start}, i = {i}, pattern = {pattern}'
-                    square_brackets += ']'
-                    i += 1
-                    seq.append((square_brackets, False))
-                elif c == '|':
-                    seq.append(('|', False))
-                    i += 1
-                elif c in ('*', '+', '?'):
-                    seq[-1] = (to_rule(seq[-1]) + c, False)
-                    i += 1
-                elif c == '{':
-                    curly_brackets = c
-                    i += 1
-                    while i < length and pattern[i] != '}':
-                        curly_brackets += pattern[i]
-                        i += 1
-                    assert i < length, f'Unbalanced curly brackets; start = {start}, i = {i}, pattern = {pattern}'
-                    curly_brackets += '}'
-                    i += 1
-                    nums = [s.strip() for s in curly_brackets[1:-1].split(',')]
-                    min_times = 0
-                    max_times = None
-                    try:
-                        if len(nums) == 1:
-                            min_times = int(nums[0])
-                            max_times = min_times
-                        else:
-                            assert len(nums) == 2
-                            min_times = int(nums[0]) if nums[0] else 0
-                            max_times = int(nums[1]) if nums[1] else None
-                    except ValueError:
-                        raise ValueError(f'Invalid quantifier {curly_brackets} in /{pattern}/')
-
-                    (sub, sub_is_literal) = seq[-1]
-
-                    if 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
-
-                        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
-                        )
-                else:
-                    literal = ''
-                    while i < length:
-                        if pattern[i] == '\\' and i < length - 1:
-                            next = pattern[i + 1]
-                            if next in ESCAPED_IN_REGEXPS_BUT_NOT_IN_LITERALS:
-                                i += 1
-                                literal += pattern[i]
-                                i += 1
-                            else:
-                                literal += pattern[i:i+2]
-                                i += 2
-                        elif pattern[i] == '"' and not self._raw_pattern:
-                            literal += '\\"'
-                            i += 1
-                        elif pattern[i] not in NON_LITERAL_SET and \
-                                (i == length - 1 or literal == '' or pattern[i+1] == '.' or pattern[i+1] not in NON_LITERAL_SET):
-                            literal += pattern[i]
-                            i += 1
-                        else:
-                            break
-                    if literal:
-                        seq.append((literal, True))
-
-            return join_seq()
-
-        return self._add_rule(
-            name,
-            to_rule(transform()) if self._raw_pattern \
-                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
-
-
-    def _resolve_ref(self, ref):
-        ref_name = ref.split('/')[-1]
-        if ref_name not in self._rules and ref not in self._refs_being_resolved:
-            self._refs_being_resolved.add(ref)
-            resolved = self._refs[ref]
-            ref_name = self.visit(resolved, ref_name)
-            self._refs_being_resolved.remove(ref)
-        return ref_name
-
-    def _generate_constant_rule(self, value):
-        return self._format_literal(json.dumps(value))
-
     def visit(self, schema, name):
         schema_type = schema.get('type')
-        schema_format = schema.get('format')
-        rule_name = name + '-' if name in RESERVED_NAMES else name or 'root'
+        rule_name = name or 'root'
 
-        if (ref := schema.get('$ref')) is not None:
-            return self._add_rule(rule_name, self._resolve_ref(ref))
-
-        elif 'oneOf' in schema or 'anyOf' in schema:
-            return self._add_rule(rule_name, self._generate_union_rule(name, schema.get('oneOf') or schema['anyOf']))
-
-        elif isinstance(schema_type, list):
-            return self._add_rule(rule_name, self._generate_union_rule(name, [{'type': t} for t in schema_type]))
-
-        elif 'const' in schema:
-            return self._add_rule(rule_name, self._generate_constant_rule(schema['const']))
-
-        elif 'enum' in schema:
-            rule = ' | '.join((self._generate_constant_rule(v) for v in schema['enum']))
+        if 'oneOf' in schema or 'anyOf' in schema:
+            rule = ' | '.join((
+                self.visit(alt_schema, f'{name}{"-" if name else ""}{i}')
+                for i, alt_schema in enumerate(schema.get('oneOf') or schema['anyOf'])
+            ))
             return self._add_rule(rule_name, rule)
 
-        elif schema_type in (None, 'object') and \
-             ('properties' in schema or \
-              ('additionalProperties' in schema and schema['additionalProperties'] is not True)):
-            required = set(schema.get('required', []))
-            properties = list(schema.get('properties', {}).items())
-            return self._add_rule(rule_name, self._build_object_rule(properties, required, name, schema.get('additionalProperties')))
+        elif 'const' in schema:
+            return self._add_rule(rule_name, self._format_literal(schema['const']))
 
-        elif schema_type in (None, 'object') and 'allOf' in schema:
-            required = set()
-            properties = []
-            hybrid_name = name
-            def add_component(comp_schema, is_required):
-                if (ref := comp_schema.get('$ref')) is not None:
-                    comp_schema = self._refs[ref]
+        elif 'enum' in schema:
+            rule = ' | '.join((self._format_literal(v) for v in schema['enum']))
+            return self._add_rule(rule_name, rule)
 
-                if 'properties' in comp_schema:
-                    for prop_name, prop_schema in comp_schema['properties'].items():
-                        properties.append((prop_name, prop_schema))
-                        if is_required:
-                            required.add(prop_name)
-
-            for t in schema['allOf']:
-                if 'anyOf' in t:
-                    for tt in t['anyOf']:
-                        add_component(tt, is_required=False)
-                else:
-                    add_component(t, is_required=True)
-
-            return self._add_rule(rule_name, self._build_object_rule(properties, required, hybrid_name, additional_properties=[]))
-
-        elif schema_type in (None, 'array') and ('items' in schema or 'prefixItems' in schema):
-            items = schema.get('items') or schema['prefixItems']
-            if isinstance(items, list):
-                return self._add_rule(
-                    rule_name,
-                    '"[" space ' +
-                    ' "," space '.join(
-                        self.visit(item, f'{name}{"-" if name else ""}tuple-{i}')
-                        for i, item in enumerate(items)) +
-                    ' "]" space')
-            else:
-                item_rule_name = self.visit(items, f'{name}{"-" if name else ""}item')
-                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)
-
-        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(
-                'root' if rule_name == 'root' else schema_format,
-                PRIMITIVE_RULES['uuid']
+        elif schema_type == 'object' and 'properties' in schema:
+            # TODO: `required` keyword
+            prop_order = self._prop_order
+            prop_pairs = sorted(
+                schema['properties'].items(),
+                # sort by position in prop_order (if specified) then by key
+                key=lambda kv: (prop_order.get(kv[0], len(prop_order)), kv[0]),
             )
 
-        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'
+            rule = '"{" space'
+            for i, (prop_name, prop_schema) in enumerate(prop_pairs):
+                prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
+                if i > 0:
+                    rule += ' "," space'
+                rule += fr' {self._format_literal(prop_name)} space ":" space {prop_rule_name}'
+            rule += ' "}" space'
 
-        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, rule)
+
+        elif schema_type == 'array' and 'items' in schema:
+            # TODO `prefixItems` keyword
+            item_rule_name = self.visit(schema['items'], f'{name}{"-" if name else ""}item')
+            rule = f'"[" space ({item_rule_name} ("," space {item_rule_name})*)? "]" space'
+            return self._add_rule(rule_name, rule)
 
         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]
             )
 
-    def _build_object_rule(self, properties: List[Tuple[str, Any]], required: Set[str], name: str, additional_properties: Union[bool, Any]):
-        prop_order = self._prop_order
-        # sort by position in prop_order (if specified) then by original order
-        sorted_props = [kv[0] for _, kv in sorted(enumerate(properties), key=lambda ikv: (prop_order.get(ikv[1][0], len(prop_order)), ikv[0]))]
-
-        prop_kv_rule_names = {}
-        for prop_name, prop_schema in properties:
-            prop_rule_name = self.visit(prop_schema, f'{name}{"-" if name else ""}{prop_name}')
-            prop_kv_rule_names[prop_name] = self._add_rule(
-                f'{name}{"-" if name else ""}{prop_name}-kv',
-                fr'{self._format_literal(json.dumps(prop_name))} space ":" space {prop_rule_name}'
-            )
-        required_props = [k for k in sorted_props if k in required]
-        optional_props = [k for k in sorted_props if k not in required]
-
-        if additional_properties == True or isinstance(additional_properties, dict):
-            sub_name = f'{name}{"-" if name else ""}additional'
-            value_rule = self.visit({} if additional_properties == True else additional_properties, f'{sub_name}-value')
-            prop_kv_rule_names["*"] = self._add_rule(
-                f'{sub_name}-kv',
-                self._add_rule('string', PRIMITIVE_RULES['string']) + f' ":" space {value_rule}'
-            )
-            optional_props.append("*")
-
-        rule = '"{" space '
-        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)
-
-        if optional_props:
-            rule += ' ('
-            if required_props:
-                rule += ' "," space ( '
-
-            def get_recursive_refs(ks, first_is_optional):
-                [k, *rest] = ks
-                kv_rule_name = prop_kv_rule_names[k]
-                if k == '*':
-                    res = self._add_rule(
-                        f'{name}{"-" if name else ""}additional-kvs',
-                        f'{kv_rule_name} ( "," space ' + kv_rule_name + ' )*'
-                    )
-                elif first_is_optional:
-                    res = f'( "," space {kv_rule_name} )?'
-                else:
-                    res = kv_rule_name
-                if len(rest) > 0:
-                    res += ' ' + self._add_rule(
-                        f'{name}{"-" if name else ""}{k}-rest',
-                        get_recursive_refs(rest, first_is_optional=True)
-                    )
-                return res
-
-            rule += ' | '.join(
-                get_recursive_refs(optional_props[i:], first_is_optional=False)
-                for i in range(len(optional_props))
-            )
-            if required_props:
-                rule += ' )'
-            rule += ' )?'
-
-        rule += ' "}" space'
-
-        return rule
-
     def format_grammar(self):
-        return '\n'.join(
-            f'{name} ::= {rule}'
-            for name, rule in sorted(self._rules.items(), key=lambda kv: kv[0])
-        )
+        return '\n'.join((f'{name} ::= {rule}' for name, rule in self._rules.items()))
 
 
 def main(args_in = None):
@@ -501,47 +115,16 @@ def main(args_in = None):
         type=lambda s: s.split(','),
         help='''
             comma-separated property names defining the order of precedence for object properties;
-            properties not specified here are given lower precedence than those that are, and
-            are kept in their original order from the schema. Required properties are always
-            given precedence over optional properties.
+            properties not specified here are given lower precedence than those that are, and are
+            sorted alphabetically
         '''
     )
-    parser.add_argument(
-        '--allow-fetch',
-        action='store_true',
-        default=False,
-        help='Whether to allow fetching referenced schemas over HTTPS')
-    parser.add_argument(
-        '--dotall',
-        action='store_true',
-        default=False,
-        help='Whether to treat dot (".") as matching all chars including line breaks in regular expression patterns')
-    parser.add_argument(
-        '--raw-pattern',
-        action='store_true',
-        default=False,
-        help='Treats string patterns as raw patterns w/o quotes (or quote escapes)')
-
     parser.add_argument('schema', help='file containing JSON schema ("-" for stdin)')
     args = parser.parse_args(args_in)
 
-    if args.schema.startswith('https://'):
-        url = args.schema
-        import requests
-        schema = requests.get(url).json()
-    elif args.schema == '-':
-        url = 'stdin'
-        schema = json.load(sys.stdin)
-    else:
-        url = f'file://{args.schema}'
-        with open(args.schema) as f:
-            schema = json.load(f)
-    converter = SchemaConverter(
-        prop_order={name: idx for idx, name in enumerate(args.prop_order)},
-        allow_fetch=args.allow_fetch,
-        dotall=args.dotall,
-        raw_pattern=args.raw_pattern)
-    schema = converter.resolve_refs(schema, url)
+    schema = json.load(sys.stdin if args.schema == '-' else open(args.schema))
+    prop_order = {name: idx for idx, name in enumerate(args.prop_order)}
+    converter = SchemaConverter(prop_order)
     converter.visit(schema, '')
     print(converter.format_grammar())
 

examples/llama-bench/README.md

@@ -35,6 +35,7 @@ options:
   -mg, --main-gpu <i>                 (default: 0)
   -nkvo, --no-kv-offload <0|1>        (default: 0)
   -mmp, --mmap <0|1>                  (default: 1)
+  -mmq, --mul-mat-q <0|1>             (default: 1)
   -ts, --tensor_split <ts0/ts1/..>    (default: 0)
   -r, --repetitions <n>               (default: 5)
   -o, --output <csv|json|md|sql>      (default: md)

examples/llama-bench/llama-bench.cpp

@@ -8,7 +8,6 @@
 #include <cstdio>
 #include <cstring>
 #include <ctime>
-#include <cstdlib>
 #include <iterator>
 #include <map>
 #include <numeric>
@@ -104,7 +103,6 @@ static std::string get_cpu_info() {
                 }
             }
         }
-        fclose(f);
     }
 #endif
     // TODO: other platforms
@@ -113,11 +111,11 @@ static std::string get_cpu_info() {
 
 static std::string get_gpu_info() {
     std::string id;
-#ifdef GGML_USE_CUDA
-    int count = ggml_backend_cuda_get_device_count();
+#ifdef GGML_USE_CUBLAS
+    int count = ggml_cuda_get_device_count();
     for (int i = 0; i < count; i++) {
         char buf[128];
-        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
+        ggml_cuda_get_device_description(i, buf, sizeof(buf));
         id += buf;
         if (i < count - 1) {
             id += "/";
@@ -125,15 +123,20 @@ static std::string get_gpu_info() {
     }
 #endif
 #ifdef GGML_USE_SYCL
-    int count = ggml_backend_sycl_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_sycl_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
+    int device_list[GGML_SYCL_MAX_DEVICES];
+    ggml_sycl_get_gpu_list(device_list, GGML_SYCL_MAX_DEVICES);
+
+    for (int i = 0; i < GGML_SYCL_MAX_DEVICES; i++) {
+        if (device_list[i] >0 ){
+            char buf[128];
+            ggml_sycl_get_device_description(i, buf, sizeof(buf));
+            id += buf;
             id += "/";
         }
     }
+    if (id.length() >2 ) {
+        id.pop_back();
+    }
 #endif
     // TODO: other backends
     return id;
@@ -154,9 +157,9 @@ static const char * output_format_str(output_formats format) {
 
 static const char * split_mode_str(llama_split_mode mode) {
     switch (mode) {
-        case LLAMA_SPLIT_MODE_NONE:  return "none";
-        case LLAMA_SPLIT_MODE_LAYER: return "layer";
-        case LLAMA_SPLIT_MODE_ROW:   return "row";
+        case LLAMA_SPLIT_NONE:  return "none";
+        case LLAMA_SPLIT_LAYER: return "layer";
+        case LLAMA_SPLIT_ROW:   return "row";
         default: GGML_ASSERT(!"invalid split mode");
     }
 }
@@ -166,7 +169,6 @@ struct cmd_params {
     std::vector<int> n_prompt;
     std::vector<int> n_gen;
     std::vector<int> n_batch;
-    std::vector<int> n_ubatch;
     std::vector<ggml_type> type_k;
     std::vector<ggml_type> type_v;
     std::vector<int> n_threads;
@@ -174,9 +176,9 @@ struct cmd_params {
     std::vector<llama_split_mode> split_mode;
     std::vector<int> main_gpu;
     std::vector<bool> no_kv_offload;
+    std::vector<bool> mul_mat_q;
     std::vector<std::vector<float>> tensor_split;
     std::vector<bool> use_mmap;
-    std::vector<bool> embeddings;
     int reps;
     bool verbose;
     output_formats output_format;
@@ -186,18 +188,17 @@ static const cmd_params cmd_params_defaults = {
     /* model         */ {"models/7B/ggml-model-q4_0.gguf"},
     /* n_prompt      */ {512},
     /* n_gen         */ {128},
-    /* n_batch       */ {2048},
-    /* n_ubatch      */ {512},
+    /* n_batch       */ {512},
     /* type_k        */ {GGML_TYPE_F16},
     /* type_v        */ {GGML_TYPE_F16},
     /* n_threads     */ {get_num_physical_cores()},
     /* n_gpu_layers  */ {99},
-    /* split_mode    */ {LLAMA_SPLIT_MODE_LAYER},
+    /* split_mode    */ {LLAMA_SPLIT_LAYER},
     /* main_gpu      */ {0},
     /* no_kv_offload */ {false},
+    /* mul_mat_q     */ {true},
     /* tensor_split  */ {std::vector<float>(llama_max_devices(), 0.0f)},
     /* use_mmap      */ {true},
-    /* embeddings    */ {false},
     /* reps          */ 5,
     /* verbose       */ false,
     /* output_format */ MARKDOWN
@@ -212,7 +213,6 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -p, --n-prompt <n>                  (default: %s)\n", join(cmd_params_defaults.n_prompt, ",").c_str());
     printf("  -n, --n-gen <n>                     (default: %s)\n", join(cmd_params_defaults.n_gen, ",").c_str());
     printf("  -b, --batch-size <n>                (default: %s)\n", join(cmd_params_defaults.n_batch, ",").c_str());
-    printf("  -ub N, --ubatch-size <n>            (default: %s)\n", join(cmd_params_defaults.n_ubatch, ",").c_str());
     printf("  -ctk <t>, --cache-type-k <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_k, ggml_type_name), ",").c_str());
     printf("  -ctv <t>, --cache-type-v <t>        (default: %s)\n", join(transform_to_str(cmd_params_defaults.type_v, ggml_type_name), ",").c_str());
     printf("  -t, --threads <n>                   (default: %s)\n", join(cmd_params_defaults.n_threads, ",").c_str());
@@ -221,8 +221,8 @@ static void print_usage(int /* argc */, char ** argv) {
     printf("  -mg, --main-gpu <i>                 (default: %s)\n", join(cmd_params_defaults.main_gpu, ",").c_str());
     printf("  -nkvo, --no-kv-offload <0|1>        (default: %s)\n", join(cmd_params_defaults.no_kv_offload, ",").c_str());
     printf("  -mmp, --mmap <0|1>                  (default: %s)\n", join(cmd_params_defaults.use_mmap, ",").c_str());
-    printf("  -embd, --embeddings <0|1>           (default: %s)\n", join(cmd_params_defaults.embeddings, ",").c_str());
-    printf("  -ts, --tensor-split <ts0/ts1/..>    (default: 0)\n");
+    printf("  -mmq, --mul-mat-q <0|1>             (default: %s)\n", join(cmd_params_defaults.mul_mat_q, ",").c_str());
+    printf("  -ts, --tensor_split <ts0/ts1/..>    (default: 0)\n");
     printf("  -r, --repetitions <n>               (default: %d)\n", cmd_params_defaults.reps);
     printf("  -o, --output <csv|json|md|sql>      (default: %s)\n", output_format_str(cmd_params_defaults.output_format));
     printf("  -v, --verbose                       (default: %s)\n", cmd_params_defaults.verbose ? "1" : "0");
@@ -249,9 +249,6 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "q5_1") {
         return GGML_TYPE_Q5_1;
     }
-    if (s == "iq4_nl") {
-        return GGML_TYPE_IQ4_NL;
-    }
 
     return GGML_TYPE_COUNT;
 }
@@ -305,13 +302,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<int>(argv[i], split_delim);
             params.n_batch.insert(params.n_batch.end(), p.begin(), p.end());
-        } else if (arg == "-ub" || arg == "--ubatch-size") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<int>(argv[i], split_delim);
-            params.n_ubatch.insert(params.n_ubatch.end(), p.begin(), p.end());
         } else if (arg == "-ctk" || arg == "--cache-type-k") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -368,11 +358,11 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             for (const auto & m : p) {
                 llama_split_mode mode;
                 if (m == "none") {
-                    mode = LLAMA_SPLIT_MODE_NONE;
+                    mode = LLAMA_SPLIT_NONE;
                 } else if (m == "layer") {
-                    mode = LLAMA_SPLIT_MODE_LAYER;
+                    mode = LLAMA_SPLIT_LAYER;
                 } else if (m == "row") {
-                    mode = LLAMA_SPLIT_MODE_ROW;
+                    mode = LLAMA_SPLIT_ROW;
                 } else {
                     invalid_param = true;
                     break;
@@ -393,6 +383,13 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.no_kv_offload.insert(params.no_kv_offload.end(), p.begin(), p.end());
+        } else if (arg == "-mmq" || arg == "--mul-mat-q") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            auto p = split<bool>(argv[i], split_delim);
+            params.mul_mat_q.insert(params.mul_mat_q.end(), p.begin(), p.end());
         } else if (arg == "-mmp" || arg == "--mmap") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -400,13 +397,6 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
             }
             auto p = split<bool>(argv[i], split_delim);
             params.use_mmap.insert(params.use_mmap.end(), p.begin(), p.end());
-        } else if (arg == "-embd" || arg == "--embeddings") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            auto p = split<bool>(argv[i], split_delim);
-            params.embeddings.insert(params.embeddings.end(), p.begin(), p.end());
         } else if (arg == "-ts" || arg == "--tensor-split") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -470,16 +460,15 @@ static cmd_params parse_cmd_params(int argc, char ** argv) {
     if (params.n_prompt.empty())     { params.n_prompt = cmd_params_defaults.n_prompt; }
     if (params.n_gen.empty())        { params.n_gen = cmd_params_defaults.n_gen; }
     if (params.n_batch.empty())      { params.n_batch = cmd_params_defaults.n_batch; }
-    if (params.n_ubatch.empty())     { params.n_ubatch = cmd_params_defaults.n_ubatch; }
     if (params.type_k.empty())       { params.type_k = cmd_params_defaults.type_k; }
     if (params.type_v.empty())       { params.type_v = cmd_params_defaults.type_v; }
     if (params.n_gpu_layers.empty()) { params.n_gpu_layers = cmd_params_defaults.n_gpu_layers; }
     if (params.split_mode.empty())   { params.split_mode = cmd_params_defaults.split_mode; }
     if (params.main_gpu.empty())     { params.main_gpu = cmd_params_defaults.main_gpu; }
     if (params.no_kv_offload.empty()){ params.no_kv_offload = cmd_params_defaults.no_kv_offload; }
+    if (params.mul_mat_q.empty())    { params.mul_mat_q = cmd_params_defaults.mul_mat_q; }
     if (params.tensor_split.empty()) { params.tensor_split = cmd_params_defaults.tensor_split; }
     if (params.use_mmap.empty())     { params.use_mmap = cmd_params_defaults.use_mmap; }
-    if (params.embeddings.empty())   { params.embeddings = cmd_params_defaults.embeddings; }
     if (params.n_threads.empty())    { params.n_threads = cmd_params_defaults.n_threads; }
 
     return params;
@@ -490,7 +479,6 @@ struct cmd_params_instance {
     int n_prompt;
     int n_gen;
     int n_batch;
-    int n_ubatch;
     ggml_type type_k;
     ggml_type type_v;
     int n_threads;
@@ -498,9 +486,9 @@ struct cmd_params_instance {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
 
     llama_model_params to_llama_mparams() const {
         llama_model_params mparams = llama_model_default_params();
@@ -528,11 +516,10 @@ struct cmd_params_instance {
 
         cparams.n_ctx = n_prompt + n_gen;
         cparams.n_batch = n_batch;
-        cparams.n_ubatch = n_ubatch;
         cparams.type_k = type_k;
         cparams.type_v = type_v;
+        cparams.mul_mat_q = mul_mat_q;
         cparams.offload_kqv = !no_kv_offload;
-        cparams.embeddings = embeddings;
 
         return cparams;
     }
@@ -548,11 +535,10 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
     for (const auto & mg : params.main_gpu)
     for (const auto & ts : params.tensor_split)
     for (const auto & mmp : params.use_mmap)
-    for (const auto & embd : params.embeddings)
     for (const auto & nb : params.n_batch)
-    for (const auto & nub : params.n_ubatch)
     for (const auto & tk : params.type_k)
     for (const auto & tv : params.type_v)
+    for (const auto & mmq : params.mul_mat_q)
     for (const auto & nkvo : params.no_kv_offload)
     for (const auto & nt : params.n_threads) {
         for (const auto & n_prompt : params.n_prompt) {
@@ -564,7 +550,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ n_prompt,
                 /* .n_gen        = */ 0,
                 /* .n_batch      = */ nb,
-                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -572,9 +557,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -588,7 +573,6 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .n_prompt     = */ 0,
                 /* .n_gen        = */ n_gen,
                 /* .n_batch      = */ nb,
-                /* .n_ubatch     = */ nub,
                 /* .type_k       = */ tk,
                 /* .type_v       = */ tv,
                 /* .n_threads    = */ nt,
@@ -596,9 +580,9 @@ static std::vector<cmd_params_instance> get_cmd_params_instances(const cmd_param
                 /* .split_mode   = */ sm,
                 /* .main_gpu     = */ mg,
                 /* .no_kv_offload= */ nkvo,
+                /* .mul_mat_q    = */ mmq,
                 /* .tensor_split = */ ts,
                 /* .use_mmap     = */ mmp,
-                /* .embeddings   = */ embd,
             };
             instances.push_back(instance);
         }
@@ -625,7 +609,6 @@ struct test {
     uint64_t model_size;
     uint64_t model_n_params;
     int n_batch;
-    int n_ubatch;
     int n_threads;
     ggml_type type_k;
     ggml_type type_v;
@@ -633,9 +616,9 @@ struct test {
     llama_split_mode split_mode;
     int main_gpu;
     bool no_kv_offload;
+    bool mul_mat_q;
     std::vector<float> tensor_split;
     bool use_mmap;
-    bool embeddings;
     int n_prompt;
     int n_gen;
     std::string test_time;
@@ -649,7 +632,6 @@ struct test {
         model_size = llama_model_size(lmodel);
         model_n_params = llama_model_n_params(lmodel);
         n_batch = inst.n_batch;
-        n_ubatch = inst.n_ubatch;
         n_threads = inst.n_threads;
         type_k = inst.type_k;
         type_v = inst.type_v;
@@ -657,9 +639,9 @@ struct test {
         split_mode = inst.split_mode;
         main_gpu = inst.main_gpu;
         no_kv_offload = inst.no_kv_offload;
+        mul_mat_q = inst.mul_mat_q;
         tensor_split = inst.tensor_split;
         use_mmap = inst.use_mmap;
-        embeddings = inst.embeddings;
         n_prompt = inst.n_prompt;
         n_gen = inst.n_gen;
         // RFC 3339 date-time format
@@ -728,11 +710,10 @@ struct test {
             "cuda", "opencl", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas",
             "cpu_info", "gpu_info",
             "model_filename", "model_type", "model_size", "model_n_params",
-            "n_batch", "n_ubatch",
-            "n_threads", "type_k", "type_v",
+            "n_batch", "n_threads", "type_k", "type_v",
             "n_gpu_layers", "split_mode",
             "main_gpu", "no_kv_offload",
-            "tensor_split", "use_mmap", "embeddings",
+            "mul_mat_q", "tensor_split", "use_mmap",
             "n_prompt", "n_gen", "test_time",
             "avg_ns", "stddev_ns",
             "avg_ts", "stddev_ts"
@@ -743,8 +724,7 @@ struct test {
     enum field_type {STRING, BOOL, INT, FLOAT};
 
     static field_type get_field_type(const std::string & field) {
-        if (field == "build_number" || field == "n_batch" || field == "n_ubatch" ||
-            field == "n_threads" ||
+        if (field == "build_number" || field == "n_batch" || field == "n_threads" ||
             field == "model_size" || field == "model_n_params" ||
             field == "n_gpu_layers" || field == "main_gpu" ||
             field == "n_prompt" || field == "n_gen" ||
@@ -753,7 +733,7 @@ struct test {
         }
         if (field == "cuda" || field == "opencl"  || field == "vulkan" || field == "kompute" || field == "metal" ||
             field == "gpu_blas" || field == "blas" || field == "sycl" ||field == "f16_kv" || field == "no_kv_offload" ||
-            field == "use_mmap" || field == "embeddings") {
+            field == "mul_mat_q" || field == "use_mmap") {
             return BOOL;
         }
         if (field == "avg_ts" || field == "stddev_ts") {
@@ -784,11 +764,10 @@ struct test {
             std::to_string(metal), std::to_string(sycl), std::to_string(gpu_blas), std::to_string(blas),
             cpu_info, gpu_info,
             model_filename, model_type, std::to_string(model_size), std::to_string(model_n_params),
-            std::to_string(n_batch), std::to_string(n_ubatch),
-            std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
+            std::to_string(n_batch), std::to_string(n_threads), ggml_type_name(type_k), ggml_type_name(type_v),
             std::to_string(n_gpu_layers), split_mode_str(split_mode),
             std::to_string(main_gpu), std::to_string(no_kv_offload),
-            tensor_split_str, std::to_string(use_mmap), std::to_string(embeddings),
+            std::to_string(mul_mat_q), tensor_split_str, std::to_string(use_mmap),
             std::to_string(n_prompt), std::to_string(n_gen), test_time,
             std::to_string(avg_ns()), std::to_string(stdev_ns()),
             std::to_string(avg_ts()), std::to_string(stdev_ts())
@@ -808,7 +787,7 @@ 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_cuda();
+const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::vulkan       = !!ggml_cpu_has_vulkan();
 const bool        test::kompute      = !!ggml_cpu_has_kompute();
@@ -952,15 +931,15 @@ struct markdown_printer : public printer {
         if (field == "n_threads") {
             return "threads";
         }
+        if (field == "mul_mat_q") {
+            return "mmq";
+        }
         if (field == "no_kv_offload") {
             return "nkvo";
         }
         if (field == "use_mmap") {
             return "mmap";
         }
-        if (field == "embeddings") {
-            return "embd";
-        }
         if (field == "tensor_split") {
             return "ts";
         }
@@ -983,9 +962,6 @@ struct markdown_printer : public printer {
         if (params.n_batch.size() > 1 || params.n_batch != cmd_params_defaults.n_batch) {
             fields.emplace_back("n_batch");
         }
-        if (params.n_ubatch.size() > 1 || params.n_ubatch != cmd_params_defaults.n_ubatch) {
-            fields.emplace_back("n_ubatch");
-        }
         if (params.type_k.size() > 1 || params.type_k != cmd_params_defaults.type_k) {
             fields.emplace_back("type_k");
         }
@@ -998,6 +974,9 @@ struct markdown_printer : public printer {
         if (params.split_mode.size() > 1 || params.split_mode != cmd_params_defaults.split_mode) {
             fields.emplace_back("split_mode");
         }
+        if (params.mul_mat_q.size() > 1 || params.mul_mat_q != cmd_params_defaults.mul_mat_q) {
+            fields.emplace_back("mul_mat_q");
+        }
         if (params.no_kv_offload.size() > 1 || params.no_kv_offload != cmd_params_defaults.no_kv_offload) {
             fields.emplace_back("no_kv_offload");
         }
@@ -1007,9 +986,6 @@ struct markdown_printer : public printer {
         if (params.use_mmap.size() > 1 || params.use_mmap != cmd_params_defaults.use_mmap) {
             fields.emplace_back("use_mmap");
         }
-        if (params.embeddings.size() > 1 || params.embeddings != cmd_params_defaults.embeddings) {
-            fields.emplace_back("embeddings");
-        }
         fields.emplace_back("test");
         fields.emplace_back("t/s");
 
@@ -1125,40 +1101,25 @@ struct sql_printer : public printer {
 };
 
 static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
-    llama_set_n_threads(ctx, n_threads, n_threads);
-
-    const llama_model * model = llama_get_model(ctx);
-    const int32_t n_vocab = llama_n_vocab(model);
-
-    std::vector<llama_token> tokens(n_batch);
-
+    std::vector<llama_token> tokens(n_batch, llama_token_bos(llama_get_model(ctx)));
     int n_processed = 0;
 
+    llama_set_n_threads(ctx, n_threads, n_threads);
+
     while (n_processed < n_prompt) {
         int n_tokens = std::min(n_prompt - n_processed, n_batch);
-        tokens[0] = n_processed == 0 && llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
-        for (int i = 1; i < n_tokens; i++) {
-            tokens[i] = std::rand() % n_vocab;
-        }
         llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
         n_processed += n_tokens;
     }
-
-    llama_synchronize(ctx);
 }
 
 static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
+    llama_token token = llama_token_bos(llama_get_model(ctx));
+
     llama_set_n_threads(ctx, n_threads, n_threads);
 
-    const llama_model * model = llama_get_model(ctx);
-    const int32_t n_vocab = llama_n_vocab(model);
-
-    llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
-
     for (int i = 0; i < n_gen; i++) {
         llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
-        llama_synchronize(ctx);
-        token = std::rand() % n_vocab;
     }
 }
 
@@ -1190,7 +1151,8 @@ int main(int argc, char ** argv) {
     if (!params.verbose) {
         llama_log_set(llama_null_log_callback, NULL);
     }
-    llama_backend_init();
+    bool numa = false;
+    llama_backend_init(numa);
 
     // initialize printer
     std::unique_ptr<printer> p;
@@ -1247,8 +1209,7 @@ int main(int argc, char ** argv) {
 
         // warmup run
         if (t.n_prompt > 0) {
-            //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, std::min(2, t.n_batch), 0, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             test_gen(ctx, 1, 0, t.n_threads);
@@ -1264,7 +1225,6 @@ int main(int argc, char ** argv) {
             if (t.n_gen > 0) {
                 test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
             }
-
             uint64_t t_ns = get_time_ns() - t_start;
             t.samples_ns.push_back(t_ns);
         }

examples/llama.android/app/build.gradle.kts

@@ -21,8 +21,12 @@ android {
             useSupportLibrary = true
         }
         ndk {
-            // Add NDK properties if wanted, e.g.
-            // abiFilters += listOf("arm64-v8a")
+            // Workaround for https://github.com/llvm/llvm-project/issues/65820
+            // affecting armeabi-v7a. Skip armeabi-v7a when invoked with
+            // -Pskip-armeabi-v7a (e.g., ./gradlew build -Pskip-armeabi-v7a).
+            if (project.hasProperty("skip-armeabi-v7a")) {
+                abiFilters += listOf("arm64-v8a", "x86_64", "x86")
+            }
         }
         externalNativeBuild {
             cmake {

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

@@ -33,45 +33,6 @@ jclass la_int_var;
 jmethodID la_int_var_value;
 jmethodID la_int_var_inc;
 
-std::string cached_token_chars;
-
-bool is_valid_utf8(const char * string) {
-    if (!string) {
-        return true;
-    }
-
-    const unsigned char * bytes = (const unsigned char *)string;
-    int num;
-
-    while (*bytes != 0x00) {
-        if ((*bytes & 0x80) == 0x00) {
-            // U+0000 to U+007F
-            num = 1;
-        } else if ((*bytes & 0xE0) == 0xC0) {
-            // U+0080 to U+07FF
-            num = 2;
-        } else if ((*bytes & 0xF0) == 0xE0) {
-            // U+0800 to U+FFFF
-            num = 3;
-        } else if ((*bytes & 0xF8) == 0xF0) {
-            // U+10000 to U+10FFFF
-            num = 4;
-        } else {
-            return false;
-        }
-
-        bytes += 1;
-        for (int i = 1; i < num; ++i) {
-            if ((*bytes & 0xC0) != 0x80) {
-                return false;
-            }
-            bytes += 1;
-        }
-    }
-
-    return true;
-}
-
 static void log_callback(ggml_log_level level, const char * fmt, void * data) {
     if (level == GGML_LOG_LEVEL_ERROR)     __android_log_print(ANDROID_LOG_ERROR, TAG, fmt, data);
     else if (level == GGML_LOG_LEVEL_INFO) __android_log_print(ANDROID_LOG_INFO, TAG, fmt, data);
@@ -313,8 +274,8 @@ Java_com_example_llama_Llm_new_1batch(JNIEnv *, jobject, jint n_tokens, jint emb
 
 extern "C"
 JNIEXPORT void JNICALL
-Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject) {
-    llama_backend_init();
+Java_com_example_llama_Llm_backend_1init(JNIEnv *, jobject, jboolean numa) {
+    llama_backend_init(numa);
 }
 
 extern "C"
@@ -334,8 +295,6 @@ Java_com_example_llama_Llm_completion_1init(
         jint n_len
     ) {
 
-    cached_token_chars.clear();
-
     const auto text = env->GetStringUTFChars(jtext, 0);
     const auto context = reinterpret_cast<llama_context *>(context_pointer);
     const auto batch = reinterpret_cast<llama_batch *>(batch_pointer);
@@ -413,16 +372,8 @@ Java_com_example_llama_Llm_completion_1loop(
     }
 
     auto new_token_chars = llama_token_to_piece(context, new_token_id);
-    cached_token_chars += new_token_chars;
-
-    jstring new_token = nullptr;
-    if (is_valid_utf8(cached_token_chars.c_str())) {
-        new_token = env->NewStringUTF(cached_token_chars.c_str());
-        LOGi("cached: %s, new_token_chars: `%s`, id: %d", cached_token_chars.c_str(), new_token_chars.c_str(), new_token_id);
-        cached_token_chars.clear();
-    } else {
-        new_token = env->NewStringUTF("");
-    }
+    LOGi("new_token_chars: `%s`", new_token_chars.c_str());
+    auto new_token = env->NewStringUTF(new_token_chars.c_str());
 
     llama_batch_clear(*batch);
     llama_batch_add(*batch, new_token_id, n_cur, { 0 }, true);

examples/llama.android/app/src/main/java/com/example/llama/Llm.kt

@@ -71,7 +71,7 @@ class Llm {
         batch: Long,
         nLen: Int,
         ncur: IntVar
-    ): String?
+    ): String
 
     private external fun kv_cache_clear(context: Long)
 
@@ -115,7 +115,7 @@ class Llm {
                 val ncur = IntVar(completion_init(state.context, state.batch, message, nlen))
                 while (ncur.value <= nlen) {
                     val str = completion_loop(state.context, state.batch, nlen, ncur)
-                    if (str == null) {
+                    if (str.isEmpty()) {
                         break
                     }
                     emit(str)

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

@@ -51,7 +51,7 @@ actor LlamaContext {
     }
 
     static func create_context(path: String) throws -> LlamaContext {
-        llama_backend_init()
+        llama_backend_init(false)
         var model_params = llama_model_default_params()
 
 #if targetEnvironment(simulator)
@@ -221,7 +221,6 @@ actor LlamaContext {
             if llama_decode(context, batch) != 0 {
                 print("llama_decode() failed during prompt")
             }
-            llama_synchronize(context)
 
             let t_pp_end = ggml_time_us()
 
@@ -241,7 +240,6 @@ actor LlamaContext {
                 if llama_decode(context, batch) != 0 {
                     print("llama_decode() failed during text generation")
                 }
-                llama_synchronize(context)
             }
 
             let t_tg_end = ggml_time_us()

examples/llava/MobileVLM-README.md

@@ -1,13 +1,11 @@
 # MobileVLM
 
-Currently this implementation supports [MobileVLM-1.7B](https://huggingface.co/mtgv/MobileVLM-1.7B) / [MobileVLM_V2-1.7B](https://huggingface.co/mtgv/MobileVLM_V2-1.7B) variants.
+Currently this implementation supports [MobileVLM-v1.7](https://huggingface.co/mtgv/MobileVLM-1.7B) variants.
 
 for more information, please go to [Meituan-AutoML/MobileVLM](https://github.com/Meituan-AutoML/MobileVLM)
 
 The implementation is based on llava, and is compatible with llava and mobileVLM. The usage is basically same as llava.
 
-Notice: The overall process of model inference for both **MobileVLM** and **MobileVLM_V2** models is the same, but the process of model conversion is a little different. Therefore, using **MobileVLM-1.7B** as an example, the different conversion step will be shown.
-
 ## Usage
 Build with cmake or run `make llava-cli` to build it.
 
@@ -36,7 +34,7 @@ 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** please use `--projector-type ldpv2`) to convert the LLaVA image encoder to GGUF:
+3. Use `convert-image-encoder-to-gguf.py` with `--projector-type ldp` to convert the LLaVA image encoder to GGUF:
 
 ```sh
 python ./examples/llava/convert-image-encoder-to-gguf \
@@ -46,14 +44,6 @@ python ./examples/llava/convert-image-encoder-to-gguf \
     --projector-type ldp
 ```
 
-```sh
-python ./examples/llava/convert-image-encoder-to-gguf \
-    -m path/to/clip-vit-large-patch14-336 \
-    --llava-projector path/to/MobileVLM-1.7B_V2/llava.projector \
-    --output-dir path/to/MobileVLM-1.7B_V2 \
-    --projector-type ldpv2
-```
-
 4. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
@@ -78,7 +68,7 @@ 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
@@ -109,6 +99,7 @@ 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)
@@ -120,82 +111,12 @@ 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_CUDA=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
+make LLAMA_CUBLAS=1 CUDA_DOCKER_ARCH=sm_87 LLAMA_CUDA_F16=1 -j 32
 ```
+
 ### run on Orin
 ### case 1
 **input**
@@ -244,121 +165,8 @@ llama_print_timings:        eval time =     166.65 ms /    11 runs   (   15.15 m
 llama_print_timings:       total time =    1365.47 ms /   243 tokens
 ```
 
-## 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
-```
+## 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.
 
 ## TODO
 
@@ -373,5 +181,5 @@ llama_print_timings:       total time =   14371.19 ms /   446 tokens
 
 ## contributor
 ```sh
-zhangjidong05, yangyang260, huyiming03, chenxiaotao03, ZiangWu-77
+zhangjidong05, yangyang260, huyiming03, chenxiaotao03
 ```

examples/llava/README.md

@@ -1,12 +1,10 @@
 # LLaVA
 
-Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants,
-as well as llava-1.6 [llava-v1.6](https://huggingface.co/collections/liuhaotian/llava-16-65b9e40155f60fd046a5ccf2) variants.
+Currently this implementation supports [llava-v1.5](https://huggingface.co/liuhaotian/llava-v1.5-7b) variants.
 
 The pre-converted [7b](https://huggingface.co/mys/ggml_llava-v1.5-7b)
 and [13b](https://huggingface.co/mys/ggml_llava-v1.5-13b)
 models are available.
-For llava-1.6 a variety of prepared gguf models are available as well [7b-34b](https://huggingface.co/cmp-nct/llava-1.6-gguf)
 
 After API is confirmed, more models will be supported / uploaded.
 
@@ -20,11 +18,10 @@ After building, run: `./llava-cli` to see the usage. For example:
 ```
 
 **note**: A lower temperature like 0.1 is recommended for better quality. add `--temp 0.1` to the command to do so.
-**note**: For GPU offloading ensure to use the `-ngl` flag just like usual
 
-## LLaVA 1.5
+## Model conversion
 
-- Clone a LLaVA and a CLIP model ([available options](https://github.com/haotian-liu/LLaVA/blob/main/docs/MODEL_ZOO.md)). For example:
+- Clone `llava-v15-7b` and `clip-vit-large-patch14-336` locally:
 
 ```sh
 git clone https://huggingface.co/liuhaotian/llava-v1.5-7b
@@ -53,87 +50,13 @@ python ./examples/llava/convert-image-encoder-to-gguf.py -m ../clip-vit-large-pa
 5. Use `convert.py` to convert the LLaMA part of LLaVA to GGUF:
 
 ```sh
-python ./convert.py ../llava-v1.5-7b --skip-unknown
+python ./convert.py ../llava-v1.5-7b
 ```
 
 Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
 
-## LLaVA 1.6 gguf conversion
-1) First clone a LLaVA 1.6 model:
-```console
-git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
-```
-
-2) Install the required Python packages:
-
-```sh
-pip install -r examples/llava/requirements.txt
-```
-
-3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
-```console
-python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
-```
-- you will find a llava.projector and a llava.clip file in your model directory
-
-4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
-```console
-mkdir vit
-cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
-cp ../llava-v1.6-vicuna-7b/llava.projector vit/
-curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
-```
-
-5) Create the visual gguf model:
-```console
-python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
-```
-- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
-
-6) Then convert the model to gguf format:
-```console
-python ./convert.py ../llava-v1.6-vicuna-7b/ --skip-unknown
-```
-
-7) And finally we can run the llava-cli using the 1.6 model version:
-```console
-./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
-```
-
-**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
-**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
-
-## llava-cli templating and llava-1.6 prompting
-
-llava-1.5 models all use the same vicuna prompt, here you can just add your image question like `-p "Provide a full description."`
-For llava-1.5 models which are not vicuna (mistral and Yi) you need to adapt system prompt as well as user prompt, for this purpose llava-cli has a basic templating system:
-
-**For Mistral and using llava-cli binary:**
-Add this: `-p "<image>\nUSER:\nProvide a full description.\nASSISTANT:\n"`
-The mistral template for llava-1.6 seems to be no system print and a USER/ASSISTANT role
-
-**For the 34B this should work:**
-Add this: `-e -p <|im_start|>system\nAnswer the questions.<|im_end|><|im_start|>user\n<image>\nProvide a full description.<|im_end|><|im_start|>assistant\n`
-
-
-## How to know if you are running in llava-1.5 or llava-1.6 mode
-
-When running llava-cli you will see a visual information right before the prompt is being processed:
-
-**Llava-1.5:**
-`encode_image_with_clip: image embedding created: 576 tokens`
-
-**Llava-1.6 (anything above 576):**
-`encode_image_with_clip: image embedding created: 2880 tokens`
-
-
-Alternatively just pay notice to how many "tokens" have been used for your prompt, it will also show 1000+ tokens for llava-1.6
-
-
-
-
 ## TODO
 
-- [x] Support non-CPU backend for the image encoding part.
+- [ ] Support non-CPU backend for the image encoding part.
 - [ ] Support different sampling methods.
 - [ ] Support more model variants.

examples/llava/clip.h

@@ -24,7 +24,25 @@ struct clip_ctx;
 extern "C" {
 #endif
 
-struct clip_ctx;
+struct clip_vision_hparams {
+    int32_t image_size;
+    int32_t patch_size;
+    int32_t hidden_size;
+    int32_t n_intermediate;
+    int32_t projection_dim;
+    int32_t n_head;
+    int32_t n_layer;
+    float eps;
+};
+
+CLIP_API struct clip_ctx * clip_model_load(const char * fname, int verbosity);
+
+CLIP_API void clip_free(struct clip_ctx * ctx);
+
+CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
+
+CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
+CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
 
 struct clip_image_u8_batch {
     struct clip_image_u8 * data;
@@ -36,43 +54,18 @@ struct clip_image_f32_batch {
     size_t size;
 };
 
-CLIP_API struct clip_ctx * clip_model_load    (const char * fname, int verbosity);
-CLIP_API struct clip_ctx * clip_model_load_cpu(const char * fname, int verbosity);
-
-CLIP_API void clip_free(struct clip_ctx * ctx);
-
-CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
-
-CLIP_API int32_t clip_image_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_patch_size (const struct clip_ctx * ctx);
-CLIP_API int32_t clip_hidden_size(const struct clip_ctx * ctx);
-
-// TODO: should be enum, not string
-CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
-
-CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
-
-CLIP_API int clip_n_patches    (const struct clip_ctx * ctx);
-CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
-
 CLIP_API struct clip_image_u8  * clip_image_u8_init ();
 CLIP_API struct clip_image_f32 * clip_image_f32_init();
 
-CLIP_API void clip_image_u8_free (struct clip_image_u8  * img);
+CLIP_API void clip_image_u8_free (struct clip_image_u8 * img);
 CLIP_API void clip_image_f32_free(struct clip_image_f32 * img);
-CLIP_API void clip_image_u8_batch_free (struct clip_image_u8_batch  * batch);
-CLIP_API void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
 
 CLIP_API bool clip_image_load_from_file(const char * fname, struct clip_image_u8 * img);
 
 /** interpret bytes as an image file with length bytes_length, and use the result to populate img */
 CLIP_API bool clip_image_load_from_bytes(const unsigned char * bytes, size_t bytes_length, struct clip_image_u8 * img);
 
-/** preprocess img and store the result in res_imgs, pad_to_square may be overriden to false depending on model configuration */
-CLIP_API bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
-
-CLIP_API struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
-
+CLIP_API bool clip_image_preprocess  (struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32 * res, bool pad2square);
 CLIP_API bool clip_image_encode      (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
 CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
 

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

@@ -1,7 +1,6 @@
 import argparse
 import os
 import json
-import re
 
 import torch
 import numpy as np
@@ -39,11 +38,9 @@ def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_llava: b
 def get_tensor_name(name: str) -> str:
     if "projection" in name:
         return name
+
     if "mm_projector" in name:
-        name = name.replace("model.mm_projector", "mm")
-        name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
-        name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
-        return name
+        return name.replace("model.mm_projector", "mm")
 
     return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
 
@@ -81,19 +78,18 @@ ap.add_argument("--text-only", action="store_true", required=False,
                 help="Save a text-only model. It can't be used to encode images")
 ap.add_argument("--vision-only", action="store_true", required=False,
                 help="Save a vision-only model. It can't be used to encode texts")
-ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
+ap.add_argument("--clip_model_is_vision", action="store_true", required=False,
                 help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
-ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
-                help="The clip model is from openclip (for ViT-SO400M type))")
 ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
-ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
+ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp", choices=["mlp", "ldp"], default="mlp")
+ap.add_argument("--image-mean", nargs=3, type=float, required=False, help="Override image mean values")
+ap.add_argument("--image-std", nargs=3, type=float, required=False, help="Override image std values")
 ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
 # Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
-# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
 default_image_mean = [0.48145466, 0.4578275, 0.40821073]
 default_image_std = [0.26862954, 0.26130258, 0.27577711]
-ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
-ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
+ap.add_argument('--image_mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
+ap.add_argument('--image_std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
 
 # with proper
 args = ap.parse_args()
@@ -109,7 +105,7 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
-if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
+if args.clip_model_is_vision:
     vocab = None
     tokens = None
 else:
@@ -137,7 +133,7 @@ ftype = 1
 if args.use_f32:
     ftype = 0
 
-if args.clip_model_is_vision or args.clip_model_is_openclip:
+if args.clip_model_is_vision:
     model = CLIPVisionModel.from_pretrained(dir_model)
     processor = None
 else:
@@ -206,57 +202,6 @@ if has_vision_encoder:
     fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
     block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
     fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
-                            #     /**
-                            #      "image_grid_pinpoints": [
-                            #         [
-                            #         336,
-                            #         672
-                            #         ],
-                            #         [
-                            #         672,
-                            #         336
-                            #         ],
-                            #         [
-                            #         672,
-                            #         672
-                            #         ],
-                            #         [
-                            #         1008,
-                            #         336
-                            #         ],
-                            #         [
-                            #         336,
-                            #         1008
-                            #         ]
-                            #     ],
-                            #     Flattened:
-                            #     [
-                            #         336, 672,
-                            #         672, 336,
-                            #         672, 672,
-                            #         1008, 336,
-                            #         336, 1008
-                            #     ]
-                            #  *
-                            #  */
-    if "image_grid_pinpoints" in v_hparams:
-        # flatten it
-        image_grid_pinpoints = []
-        for pinpoint in v_hparams["image_grid_pinpoints"]:
-            for p in pinpoint:
-                image_grid_pinpoints.append(p)
-        fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
-    if "image_crop_resolution" in v_hparams:
-        fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
-    if "image_aspect_ratio" in v_hparams:
-        fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
-    if "image_split_resolution" in v_hparams:
-        fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
-    if "mm_patch_merge_type" in v_hparams:
-        fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
-    if "mm_projector_type" in v_hparams:
-        fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
-
 
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean

examples/llava/llava-cli.cpp

@@ -155,29 +155,11 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_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());
-        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());
-            }
-        }
         printf("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());
-            }
-        }
     } else {
         // llava-1.5 native mode
         system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
         user_prompt = prompt + "\nASSISTANT:";
-        if (params->verbose_prompt) {
-            auto tmp = ::llama_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
-            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());
-            }
-        }
     }
 
     eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, add_bos);
@@ -189,17 +171,13 @@ static void process_prompt(struct llava_context * ctx_llava, struct llava_image_
     fprintf(stderr, "\n");
 
     struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
-    std::string response = "";
+
     for (int i = 0; i < max_tgt_len; i++) {
         const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
-        response += tmp;
         if (strcmp(tmp, "</s>") == 0) break;
         if (strstr(tmp, "###")) break; // Yi-VL behavior
-        printf("%s", tmp);
-        if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
-        if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
-        if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
 
+        printf("%s", tmp);
         fflush(stdout);
     }
 
@@ -218,8 +196,7 @@ static struct llava_context * llava_init(gpt_params * params) {
 
     auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
 
-    llama_backend_init();
-    llama_numa_init(params->numa);
+    llama_backend_init(params->numa);
 
     llama_model_params model_params = llama_model_params_from_gpt_params(*params);
 

examples/llava/llava-surgery-v2.py

@@ -1,155 +0,0 @@
-import argparse
-import glob
-import os
-import torch
-from safetensors.torch import load as safe_load, save as safe_save, safe_open, save_file
-
-# Function to determine if file is a SafeTensor file
-def is_safetensor_file(file_path):
-    return file_path.endswith('.safetensors')
-
-
-# Unified loading function
-def load_model(file_path):
-    if is_safetensor_file(file_path):
-        tensors = {}
-        with safe_open(file_path, framework="pt", device="cpu") as f:
-            for key in f.keys():
-                tensors[key] = f.get_tensor(key).clone()
-                # output shape
-                print(f"{key} : {tensors[key].shape}")
-        return tensors, 'safetensor'
-    else:
-        return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
-
-
-# Unified saving function
-def save_model(model, file_path, file_type):
-    if file_type == 'safetensor':
-        # safe_save(model, file_path)
-        save_file(model, file_path)
-    else:
-        torch.save(model, file_path)
-
-
-# Adapted function to clean vision tower from checkpoint
-def clean_vision_tower_from_checkpoint(checkpoint_path):
-    checkpoint, file_type = load_model(checkpoint_path)
-    # file_type = 'pytorch'
-    model_path = os.path.dirname(checkpoint_path)
-    print(f"Searching for vision tower tensors in {checkpoint_path}")
-    clip_tensors = [k for k, v in checkpoint.items() if (k.startswith("model.vision_tower") or k.startswith("vit."))]
-
-    if len(clip_tensors) > 0:
-        print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
-        # Adapted for file type
-        clip_path = os.path.join(model_path, "llava.clip")
-
-        if os.path.exists(clip_path):
-            print(f"Loading existing llava.clip from {clip_path}")
-            existing_clip, _ = load_model(clip_path)
-        else:
-            print(f"Creating new llava.clip at {clip_path}")
-            existing_clip = {}
-        # Update existing_clip with new tensors, avoid duplicates
-        for name in clip_tensors:
-            simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
-            print(f"Adding {simple_name} to llava.clip")
-            if simple_name not in existing_clip:
-                existing_clip[simple_name] = checkpoint[name]
-
-        # Save the updated clip tensors back to llava.clip
-        save_model(existing_clip, clip_path, 'pytorch')
-
-        # Remove the tensors from the original checkpoint
-        for name in clip_tensors:
-            del checkpoint[name]
-
-        checkpoint_path = checkpoint_path
-        return True
-    return False
-
-def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
-    newline_checkpoint_path = None
-    projector_checkpoint_path = None
-
-    for path in checkpoint_paths:
-        checkpoint, _ = load_model(path)
-        if newline_criteria(checkpoint) and newline_checkpoint_path is None:
-            newline_checkpoint_path = path
-        if projector(checkpoint):
-            projector_checkpoint_path = path
-
-    return newline_checkpoint_path, projector_checkpoint_path
-
-def newline_criteria(checkpoint):
-    return any(k.startswith("model.image_newline") for k in checkpoint.keys())
-
-def proj_criteria(checkpoint):
-    return any(k.startswith("model.mm_projector") or k.startswith("vision_proj.") for k in checkpoint.keys())
-
-
-# Command-line interface setup
-ap = argparse.ArgumentParser()
-ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
-ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
-args = ap.parse_args()
-
-if args.clean_vision_tower:
-    # Generalized to handle both PyTorch and SafeTensors models
-    model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
-    # checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
-    checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
-    for projector_checkpoint_path in checkpoint_paths:
-        print(f"Cleaning {projector_checkpoint_path}")
-        if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
-            print(f"No vision tower found in {projector_checkpoint_path}")
-            # we break once none is found, so far all models append them at the end
-            # break
-    print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
-
-# Now we look for the projector in the last checkpoint
-model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
-checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
-# last_checkpoint_path = checkpoint_paths[0]
-# first_checkpoint_path = checkpoint_paths[-1]
-newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
-
-print(f"Taking projector from {projector_checkpoint_path}")
-first_mm_tensors = []
-first_checkpoint = None
-if newline_checkpoint_path is not None:
-    print(f"Taking newline from {newline_checkpoint_path}")
-    first_checkpoint, file_type = load_model(newline_checkpoint_path)
-    first_mm_tensors = [k for k, v in first_checkpoint.items() if k.startswith("model.image_newline")]
-
-# Load the checkpoint
-mm_tensors = []
-last_checkpoint = None
-if projector_checkpoint_path is not None:
-    last_checkpoint, file_type = load_model(projector_checkpoint_path)
-    mm_tensors = [k for k, v in last_checkpoint.items() if k.startswith("model.mm_projector") or k.startswith("vision_proj.")]
-
-if len(mm_tensors) == 0:
-    if last_checkpoint is not None:
-        for k, v in last_checkpoint.items():
-            print(k)
-    print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint)} tensors.")
-    print("No tensors found. Is this a LLaVA model?")
-    exit()
-
-print(f"Found {len(mm_tensors)} tensors to extract.")
-print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
-# projector = {name: checkpoint.[name].float() for name in mm_tensors}
-projector = {}
-for name in mm_tensors:
-    projector[name] = last_checkpoint[name].float()
-for name in first_mm_tensors:
-    projector[name] = first_checkpoint[name].float()
-
-if len(projector) > 0:
-    save_model(projector, f"{args.model}/llava.projector", 'pytorch')
-
-print("Done!")
-print(f"Now you can convert {args.model} to a a regular LLaMA GGUF file.")
-print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

examples/llava/llava-surgery.py

@@ -19,12 +19,19 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_project
 projector = {name: checkpoint[name].float() for name in mm_tensors}
 torch.save(projector, f"{args.model}/llava.projector")
 
+# remove these tensors from the checkpoint and save it again
+for name in mm_tensors:
+    del checkpoint[name]
+
 # BakLLaVA models contain CLIP tensors in it
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
 if len(clip_tensors) > 0:
     clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
     torch.save(clip, f"{args.model}/llava.clip")
 
+    # remove these tensors
+    for name in clip_tensors:
+        del checkpoint[name]
 
     # added tokens should be removed to be able to convert Mistral models
     if os.path.exists(f"{args.model}/added_tokens.json"):
@@ -32,6 +39,7 @@ if len(clip_tensors) > 0:
             f.write("{}\n")
 
 
+torch.save(checkpoint, path)
 
 print("Done!")
 print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")

examples/llava/llava.cpp

@@ -2,296 +2,32 @@
 #include "common.h"
 #include "llama.h"
 #include "llava.h"
-#include "base64.hpp"
 
 #include <cstdio>
 #include <cstdlib>
 #include <vector>
-#include <numeric>
-
-// RGB uint8 image
-struct clip_image_u8 {
-    int nx;
-    int ny;
-
-    std::vector<uint8_t> buf;
-};
-
-// RGB float32 image (NHWC)
-// Memory layout: RGBRGBRGB...
-struct clip_image_f32 {
-    int nx;
-    int ny;
-
-    std::vector<float> buf;
-};
-
-struct clip_image_grid_shape {
-    int first;
-    int second;
-};
-
-/**
- * Selects the best resolution from a list of possible resolutions based on the original size.
- *
- * @param original_size The original size of the image in the format (width, height).
- * @param possible_resolutions A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
- * @return The best fit resolution in the format (width, height).
- */
-static std::pair<int, int> select_best_resolution(const std::pair<int, int>& original_size, const std::vector<std::pair<int, int>>& possible_resolutions) {
-    int original_width  = original_size.first;
-    int original_height = original_size.second;
-
-    std::pair<int, int> best_fit;
-    int max_effective_resolution = 0;
-    int min_wasted_resolution = std::numeric_limits<int>::max();
-
-    for (const auto& resolution : possible_resolutions) {
-        int width = resolution.first;
-        int height = resolution.second;
-        float scale = std::min(static_cast<float>(width) / original_width, static_cast<float>(height) / original_height);
-        int downscaled_width  = static_cast<int>(original_width * scale);
-        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);
-        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;
-            best_fit = resolution;
-        }
-    }
-
-    return best_fit;
-}
-
-/**
- * @brief Get the anyres image grid shape object
- *
- * @param image_size
- * @param grid_pinpoints
- * @param image_patch_size
- * @return <int, int>
- */
-static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<int, int> & image_size, const std::vector<std::pair<int, int>> & grid_pinpoints, int image_patch_size) {
-    /**
-        Conversion from gguf flat array to vector:
-        std::vector<std::pair<int, int>> possible_resolutions;
-        for (int i = 0; i < 32 && params.image_grid_pinpoints[i] != 0; i+=2) {
-            possible_resolutions.push_back({params.image_grid_pinpoints[i], params.image_grid_pinpoints[i+1]});
-        }
-     */
-    auto best_resolution = select_best_resolution(image_size, grid_pinpoints);
-    return {best_resolution.first / image_patch_size, best_resolution.second / image_patch_size};
-}
-
-// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
-static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
-    struct {
-        struct ggml_tensor * newline;
-        struct ggml_context * ctx;
-    } model;
-
-    const int32_t image_size = clip_image_size(ctx_clip);
-    const int32_t patch_size = clip_patch_size(ctx_clip);
-
-    int32_t num_patches_per_side = image_size / patch_size; // 336 / 14 = 24 - used for embedding-patching boxes (24*24 = 576 patches)
-
-    int num_patches_width  = grid_shape.first;  // grid 1-4
-    int num_patches_height = grid_shape.second; // grid 1-4
-
-    const size_t num_images = num_patches_width * num_patches_height + 1;
-
-    // TODO: size calculation is not calculated - it's only tens of MB
-    size_t ctx_size = 0;
-
-    {
-        ctx_size += clip_embd_nbytes(ctx_clip) * num_images * 8; // image_features
-        ctx_size += 1024*1024 * ggml_type_size(GGML_TYPE_F32);
-    }
-
-    struct ggml_init_params params {
-        /*.mem_size   =*/ ctx_size,
-        /*.mem_buffer =*/ NULL,
-        /*.no_alloc   =*/ false, // NOTE: this should be false when using the legacy API
-    };
-
-    // Python reference code for full unpad:
-    /*
-        base_image_feature = image_feature[0]
-        image_feature = image_feature[1:]
-        image_feature = image_feature.permute(4, 0, 2, 1, 3).contiguous()
-        image_feature = image_feature.flatten(1, 2).flatten(2, 3)
-        image_feature = unpad_image(image_feature, image_sizes[image_idx])
-        image_feature = torch.cat((
-            image_feature,
-            self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1)
-        ), dim=-1)
-        image_feature = image_feature.flatten(1, 2).transpose(0, 1)
-        image_feature = torch.cat((base_image_feature, image_feature), dim=0)
-    */
-    // We now have two options: unpad or no unpad. Unpad removes tokens for faster llm eval.
-    // In terms of result quality it appears to make no difference, so we'll start with the easier approach given 5D tensors are not supported in ggml yet.
-    // Without unpad we have to split the sub-image embeddings into patches of 24 features each and permute them.
-    // Once all images are processed to prepended the base_image_features without any changes.
-
-    // Pytorch reference simplified, modified for ggml compatibility - confirmed identical output in python (for a 2x2 grid image (676x676 scaling))
-    /*
-        image_feature = image_feature.view(2, 2, 24, 24, 4096)
-        image_feature = image_feature.permute(0, 2, 1, 3, 4).contiguous()
-        image_feature = image_feature.view(2, 24, 2, 24, 4096)
-        image_feature = image_feature.flatten(0, 3)
-
-        // Reshape to 4D tensor by merging the last two dimensions
-        image_feature = image_feature.view(2, 2, 24, 24*4096)
-        image_feature = image_feature.permute(0, 2, 1, 3).contiguous()
-        image_feature = image_feature.view(-1, 4096)
-    */
-
-    model.ctx = ggml_init(params);
-
-    ggml_tensor * newline_tmp = clip_get_newline_tensor(ctx_clip);
-    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");
-        }
-        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");
-        }
-    }
-
-    struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
-    // ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
-    // fill it with the image embeddings, ignoring the base
-    for (size_t i = 1; i < num_images; i++) {
-        size_t offset = (i-1) * clip_embd_nbytes(ctx_clip);
-        memcpy((uint8_t *)(image_features->data) + offset, image_embd_v[i], clip_embd_nbytes(ctx_clip));
-    }
-
-    struct ggml_cgraph  * gf = ggml_new_graph(model.ctx);
-    size_t size_ele = ggml_type_size(GGML_TYPE_F32);
-
-    struct ggml_tensor *image_features_patchview = ggml_view_4d(model.ctx, image_features,
-                                                                num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
-                                                                num_patches_per_side,
-                                                                num_patches_width,
-                                                                num_patches_height,
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip),
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side,
-                                                                size_ele * num_patches_per_side * clip_n_mmproj_embd(ctx_clip) * num_patches_per_side * num_patches_width, 0);
-    // ggml_tensor_printf(image_features_patchview,"image_features_patchview",__LINE__,false,false);
-    struct ggml_tensor *permuted_cont = ggml_cont(model.ctx, ggml_permute(model.ctx, image_features_patchview, 0, 2, 1, 3));
-    /**
-     At the end of each row we have to add the row_end embeddings, which are the same as the newline embeddings
-         image_feature = torch.cat((
-        image_feature,
-        self.model.image_newline[:, None, None].expand(*image_feature.shape[:-1], 1).to(image_feature.device)
-    ), dim=-1)
-     *
-     */
-
-    // ggml_tensor_printf(permuted_cont,"permuted_cont",__LINE__,false,false);
-    struct ggml_tensor *flatten = ggml_view_2d(model.ctx, permuted_cont, clip_n_mmproj_embd(ctx_clip), num_patches_height * num_patches_width * num_patches_per_side * num_patches_per_side,  size_ele * clip_n_mmproj_embd(ctx_clip), 0);
-    // ggml_tensor_printf(flatten,"flatten",__LINE__,false,false);
-    ggml_build_forward_expand(gf, flatten);
-    ggml_graph_compute_with_ctx(model.ctx, gf, 1);
-    struct ggml_tensor* result = gf->nodes[gf->n_nodes - 1];
-
-    memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
-    // append without newline tokens (default behavior in llava_arch when not using unpad ):
-    memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
-    *n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
-
-    // Debug: Test single segments
-    // Current findings: sending base image, sending a segment embedding all works similar to python
-    // However, permuted embeddings do not work yet (stride issue?)
-    // memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as context
-    // memcpy(image_embd_out, (float*)prepared_cont->data, clip_embd_nbytes(ctx_clip)); // main image as context
-    // *n_img_pos_out=576;
-
-    ggml_free(model.ctx);
-    return true;
-}
 
+#include "base64.hpp"
 
 static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
-    // std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
-    clip_image_f32_batch img_res_v;
-    img_res_v.size = 0;
-    img_res_v.data = nullptr;
-    if (!clip_image_preprocess(ctx_clip, img, &img_res_v)) {
+    clip_image_f32 * img_res = clip_image_f32_init();
+    if (!clip_image_preprocess(ctx_clip, img, img_res, /*pad2square =*/ true)) {
         fprintf(stderr, "%s: unable to preprocess image\n", __func__);
-        delete[] img_res_v.data;
+        clip_image_f32_free(img_res);
         return false;
     }
 
+    *n_img_pos = clip_n_patches(ctx_clip);
+
     const int64_t t_img_enc_start_us = ggml_time_us();
+    bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd);
+    clip_image_f32_free(img_res);
+    if (!encoded) {
+        fprintf(stderr, "Unable to encode image\n");
 
-    const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
-
-    if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
-        // flat / default llava-1.5 type embedding
-        *n_img_pos = clip_n_patches(ctx_clip);
-        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");
-
-            return false;
-        }
-    } else {
-        // spatial_unpad llava-1.6 type embedding
-        // TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
-        std::vector<float *> image_embd_v;
-        image_embd_v.resize(img_res_v.size);
-        for (size_t i = 0; i < img_res_v.size; i++) {
-            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);
-                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);
-
-        const int32_t * image_grid = clip_image_grid(ctx_clip);
-
-        std::vector<std::pair<int, int>> grid_pinpoints;
-        for (int i = 0; i < 32 && image_grid[i] != 0; i += 2) {
-            grid_pinpoints.push_back({image_grid[i], image_grid[i+1]});
-        }
-
-        // free all img_res_v - not needed anymore
-        delete[] img_res_v.data;
-        img_res_v.size = 0;
-        img_res_v.data = nullptr;
-
-        const int32_t image_size = clip_image_size(ctx_clip);
-
-        struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
-
-        int n_img_pos_out;
-        clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
-        *n_img_pos = n_img_pos_out;
-
-        for (size_t i = 0; i < image_embd_v.size(); i++) {
-            free(image_embd_v[i]);
-        }
-        image_embd_v.clear();
-
-        // debug image/segment/normalization content:
-        // clip_image_u8 * tmp = clip_image_u8_init();
-        // clip_image_convert_f32_to_u8(*image_feature, *tmp);
-        // clip_image_save_to_bmp(*tmp, "image_feature.bmp");
+        return false;
     }
 
-    printf("%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;
 
@@ -311,10 +47,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
     return true;
 }
 
-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
+static 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));
     if (!image_embd) {
         fprintf(stderr, "Unable to allocate memory for image embeddings\n");
+        free(image_embd);
         return false;
     }
 
@@ -348,7 +85,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
     return true;
 }
 
-struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length) {
     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);
@@ -405,7 +142,7 @@ static bool load_file_to_bytes(const char* path, unsigned char** bytesOut, long
     return true;
 }
 
-struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
+LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path) {
     unsigned char* image_bytes;
     long image_bytes_length;
     auto loaded = load_file_to_bytes(image_path, &image_bytes, &image_bytes_length);
@@ -414,13 +151,13 @@ struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx
         return NULL;
     }
 
-    llava_image_embed *embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
+    auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, image_bytes, image_bytes_length);
     free(image_bytes);
 
     return embed;
 }
 
-void llava_image_embed_free(struct llava_image_embed * embed) {
+LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed) {
     free(embed->embed);
     free(embed);
 }

examples/llava/llava.h

@@ -3,6 +3,7 @@
 
 #include "ggml.h"
 
+
 #ifdef LLAMA_SHARED
 #    if defined(_WIN32) && !defined(__MINGW32__)
 #        ifdef LLAMA_BUILD
@@ -29,9 +30,7 @@ struct llava_image_embed {
 };
 
 /** sanity check for clip <-> llava embed size match */
-LLAVA_API bool llava_validate_embed_size(const struct llama_context * ctx_llama, const struct clip_ctx * ctx_clip);
-
-LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip, int n_threads, const struct clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
+LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
 
 /** build an image embed from image file bytes */
 LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
@@ -43,6 +42,7 @@ LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
 /** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
 LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);
 
+
 #ifdef __cplusplus
 }
 #endif

examples/lookahead/lookahead.cpp

@@ -54,8 +54,7 @@ int main(int argc, char ** argv) {
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;

examples/lookup/CMakeLists.txt

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

examples/lookup/lookup-create.cpp

@@ -1,43 +0,0 @@
-#include "ggml.h"
-#include "llama.h"
-#include "common.h"
-#include "ngram-cache.h"
-
-#include <cstdint>
-#include <fstream>
-#include <iostream>
-#include <string>
-#include <unordered_map>
-#include <vector>
-
-int main(int argc, char ** argv){
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        return 1;
-    }
-    // init llama.cpp
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
-    // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    GGML_ASSERT(model != nullptr);
-
-    // tokenize the prompt
-    const bool add_bos = llama_should_add_bos_token(model);
-
-    std::vector<llama_token> inp;
-    inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
-    fprintf(stderr, "%s: tokenization done\n", __func__);
-
-
-    llama_ngram_cache ngram_cache;
-    llama_ngram_cache_update(ngram_cache, LLAMA_NGRAM_STATIC, LLAMA_NGRAM_STATIC, inp, inp.size(), true);
-    fprintf(stderr, "%s: hashing done, writing file to %s\n", __func__, params.lookup_cache_static.c_str());
-
-    llama_ngram_cache_save(ngram_cache, params.lookup_cache_static);
-}

examples/lookup/lookup-merge.cpp

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

examples/lookup/lookup-stats.cpp

@@ -1,163 +0,0 @@
-#include "ggml.h"
-#include "common.h"
-#include "llama.h"
-#include "log.h"
-#include "ngram-cache.h"
-
-#include <cmath>
-#include <cstdint>
-#include <cstdio>
-#include <fstream>
-#include <string>
-#include <vector>
-#include <unordered_map>
-
-int main(int argc, char ** argv){
-    gpt_params params;
-
-    if (!gpt_params_parse(argc, argv, params)) {
-        return 1;
-    }
-
-    const int n_draft = params.n_draft;
-
-    // init llama.cpp
-    llama_backend_init();
-    llama_numa_init(params.numa);
-
-    llama_model * model = NULL;
-    llama_context * ctx = NULL;
-
-    // load the model
-    std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    llama_set_rng_seed(ctx, params.seed);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
-
-    // tokenize the prompt
-    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);
-
-    llama_ngram_cache ngram_cache_context;
-    llama_ngram_cache ngram_cache_dynamic;
-    llama_ngram_cache ngram_cache_static;
-    int64_t t_draft_flat_us = 0;
-    int64_t t_draft_us = 0;
-
-    {
-        const int64_t t_start_draft_us = ggml_time_us();
-
-        if (!params.lookup_cache_static.empty()) {
-            try {
-                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
-            } catch (std::ifstream::failure const &) {
-                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
-                exit(1);
-            }
-        }
-
-        if (!params.lookup_cache_dynamic.empty()) {
-            try {
-                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
-            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
-        }
-
-        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
-    }
-
-    const int n_input = inp.size();
-    const int n_ctx = params.n_ctx;
-
-    int n_drafted = 0;
-    int n_accept  = 0;
-
-    const int64_t t_start_ms = ggml_time_ms();
-
-    // Iterate over input tokens in chunks of size n_ctx.
-    // Each chunk is treated as if a sequential generation but with pre-determined tokens to ensure reproducibility.
-    for (int i_start = 0; i_start + n_ctx < n_input; i_start += n_ctx) {
-        const std::vector<llama_token> inp_slice(inp.begin() + i_start, inp.begin() + i_start + n_ctx);
-        std::vector<llama_token> pseudo_output;
-        pseudo_output.push_back(inp_slice[0]);
-
-        while ((int) pseudo_output.size() < n_ctx) {
-            // Simulate drafting and decoding from draft:
-            std::vector<llama_token> draft;
-            draft.push_back(pseudo_output.back());
-
-            {
-                const int64_t t_start_draft_us = ggml_time_us();
-                llama_ngram_cache_draft(pseudo_output, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
-                t_draft_us += ggml_time_us() - t_start_draft_us;
-            }
-
-            n_drafted += draft.size() - 1;
-
-            for (size_t j = 1; j < draft.size() && (int) pseudo_output.size() < n_ctx; ++j) {
-                const llama_token ground_truth = inp_slice[pseudo_output.size()];
-                const llama_token drafted = draft[j];
-
-                if (ground_truth != drafted) {
-                    break;
-                }
-
-                ++n_accept;
-                pseudo_output.push_back(ground_truth);
-
-                {
-                    const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
-                    t_draft_us += ggml_time_us() - t_start_draft_us;
-                }
-            }
-
-            // After each simulated batch decoding simulate the sampling of a single token:
-            if ((int) pseudo_output.size() < n_ctx) {
-                pseudo_output.push_back(inp_slice[pseudo_output.size()]);
-                {
-                    const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, pseudo_output, 1, false);
-                    t_draft_us += ggml_time_us() - t_start_draft_us;
-                }
-            }
-
-            draft.erase(draft.begin());
-
-        }
-        if (i_start > 0 && i_start / 100000 != (i_start - n_ctx) / 100000) {
-            const int64_t t_now_ms = ggml_time_ms();
-            const int64_t eta_ms   = (n_input - i_start) * (t_now_ms - t_start_ms) / i_start;
-            const int64_t eta_min  = eta_ms / (60*1000);
-            const int64_t eta_s    = (eta_ms - 60*1000*eta_min) / 1000;
-
-            LOG_TEE("lookup-stats: %d/%d done, ETA: %02" PRId64 ":%02" PRId64 "\n", i_start, n_input, eta_min, eta_s);
-        }
-
-        // After each chunk, update the dynamic ngram cache with the context ngram cache:
-        llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-        ngram_cache_context.clear();
-    }
-
-    LOG_TEE("\n");
-
-    LOG_TEE("\n");
-    LOG_TEE("n_draft      = %d\n", n_draft);
-    LOG_TEE("n_predict    = %d\n", n_input - n_input % n_ctx);
-    LOG_TEE("n_drafted    = %d\n", n_drafted);
-    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
-    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
-            t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
-    LOG_TEE("n_accept     = %d\n", n_accept);
-    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
-
-    llama_free(ctx);
-    llama_free_model(model);
-
-    llama_backend_free();
-
-    fprintf(stderr, "\n\n");
-
-    return 0;
-}

examples/lookup/lookup.cpp

@@ -1,15 +1,12 @@
+#include "common.h"
 #include "ggml.h"
 #include "llama.h"
-#include "common.h"
-#include "ngram-cache.h"
 
 #include <cmath>
 #include <cstdint>
 #include <cstdio>
-#include <fstream>
 #include <string>
 #include <vector>
-#include <unordered_map>
 
 int main(int argc, char ** argv){
     gpt_params params;
@@ -18,7 +15,11 @@ int main(int argc, char ** argv){
         return 1;
     }
 
-    // max. number of additional tokens to draft if match is found
+    // max/min n-grams size to search for in prompt
+    const int ngram_max = 4;
+    const int ngram_min = 1;
+
+    // length of the candidate / draft sequence, if match is found
     const int n_draft = params.n_draft;
 
     const bool dump_kv_cache = params.dump_kv_cache;
@@ -30,16 +31,13 @@ int main(int argc, char ** argv){
 #endif // LOG_DISABLE_LOGS
 
     // init llama.cpp
-    llama_backend_init();
-    llama_numa_init(params.numa);
+    llama_backend_init(params.numa);
 
     llama_model * model = NULL;
     llama_context * ctx = NULL;
 
     // load the model
     std::tie(model, ctx) = llama_init_from_gpt_params(params);
-    llama_set_rng_seed(ctx, params.seed);
-    GGML_ASSERT(llama_n_vocab(model) < (1 << 16));
 
     // tokenize the prompt
     const bool add_bos = llama_should_add_bos_token(model);
@@ -48,35 +46,6 @@ int main(int argc, char ** argv){
     std::vector<llama_token> inp;
     inp = ::llama_tokenize(ctx, params.prompt, add_bos, true);
 
-    llama_ngram_cache ngram_cache_context;
-    llama_ngram_cache ngram_cache_dynamic;
-    llama_ngram_cache ngram_cache_static;
-    int64_t t_draft_flat_us = 0;
-    int64_t t_draft_us = 0;
-
-    {
-        // Fill up context ngram cache with tokens from user input:
-        const int64_t t_start_draft_us = ggml_time_us();
-        llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, inp.size(), false);
-
-        if (!params.lookup_cache_static.empty()) {
-            try {
-                ngram_cache_static = llama_ngram_cache_load(params.lookup_cache_static);
-            } catch (std::ifstream::failure const &) {
-                fprintf(stderr, "error: failed to open static lookup cache: %s", params.lookup_cache_static.c_str());
-                exit(1);
-            }
-        }
-
-        if (!params.lookup_cache_dynamic.empty()) {
-            try {
-                ngram_cache_dynamic = llama_ngram_cache_load(params.lookup_cache_dynamic);
-            } catch (std::ifstream::failure const &) {} // if the file does not exist it will simply be created at the end of the program
-        }
-
-        t_draft_flat_us += ggml_time_us() - t_start_draft_us;
-    }
-
     const int max_context_size     = llama_n_ctx(ctx);
     const int max_tokens_list_size = max_context_size - 4;
 
@@ -106,6 +75,8 @@ int main(int argc, char ** argv){
     int n_drafted = 0;
     int n_accept  = 0;
 
+    int64_t t_draft_us = 0;
+
     int n_past = inp.size();
 
     bool has_eos = false;
@@ -157,12 +128,6 @@ int main(int argc, char ** argv){
                 ++n_past;
                 ++i_dft;
                 inp.push_back(id);
-                {
-                    // Update context ngram cache with the newly accepted token:
-                    const int64_t t_start_draft_us = ggml_time_us();
-                    llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
-                    t_draft_us += ggml_time_us() - t_start_draft_us;
-                }
 
                 if (params.use_color) {
                     // color accepted draft token
@@ -183,12 +148,6 @@ int main(int argc, char ** argv){
             draft.clear();
             draft.push_back(id);
             inp.push_back(id);
-            {
-                // Update context ngram cache with the newly accepted token:
-                const int64_t t_start_draft_us = ggml_time_us();
-                llama_ngram_cache_update(ngram_cache_context, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, inp, 1, false);
-                t_draft_us += ggml_time_us() - t_start_draft_us;
-            }
             break;
         }
 
@@ -203,19 +162,44 @@ int main(int argc, char ** argv){
         llama_batch_clear(batch_tgt);
         llama_batch_add(batch_tgt, draft[0], n_past, { 0 }, true);
 
-        // Draft already contains a single token sampled from the model:
-        GGML_ASSERT(draft.size() == 1);
-        GGML_ASSERT(draft[0] == inp.back());
+        // generate n_pred tokens through prompt lookup
+        auto prompt_lookup = [&]() -> void {
+            const int inp_size = inp.size();
+            for (int ngram_size = ngram_max ; ngram_size > ngram_min; --ngram_size){
+                const llama_token * ngram = &inp[inp_size - ngram_size];
+
+                for (int i = 0; i <= (int) inp_size - (ngram_size * 2); ++i) {
+                    bool match = true;
+                    for (int j = 0; j < ngram_size; ++j) {
+                        if (inp[i + j] != ngram[j]) {
+                            match = false;
+                            break;
+                        }
+                    }
+
+                    if (match) {
+                        const int startIdx = i + ngram_size;
+                        const int endIdx = startIdx + n_draft;
+                        if (endIdx < inp_size) {
+                            for (int j = startIdx; j < endIdx; ++j) {
+                                LOG(" - draft candidate %d: %d\n", j, inp[j]);
+                                draft.push_back(inp[j]);
+                                llama_batch_add(batch_tgt, inp[j], n_past + (j - startIdx) + 1, { 0 }, true);
+                                ++n_drafted;
+                            }
+                            return;
+                        }
+                    }
+                }
+            }
+            return;
+        };
+
         const int64_t t_start_draft_us = ggml_time_us();
 
-        llama_ngram_cache_draft(inp, draft, n_draft, LLAMA_NGRAM_MIN, LLAMA_NGRAM_MAX, ngram_cache_context, ngram_cache_dynamic, ngram_cache_static);
-
-        for (size_t i = 1; i < draft.size(); ++i) {
-            llama_batch_add(batch_tgt, draft[i], n_past + i, { 0 }, true);
-        }
+        prompt_lookup();
 
         t_draft_us += ggml_time_us() - t_start_draft_us;
-        n_drafted += draft.size() - 1;
 
         llama_decode(ctx, batch_tgt);
         ++n_past;
@@ -225,24 +209,19 @@ int main(int argc, char ** argv){
 
     auto t_dec_end = ggml_time_us();
 
-    // Update dynamic ngram cache with context ngram cache and save it to disk:
-    llama_ngram_cache_merge(ngram_cache_dynamic, ngram_cache_context);
-    llama_ngram_cache_save(ngram_cache_dynamic, params.lookup_cache_dynamic);
-
     LOG_TEE("\n\n");
 
     LOG_TEE("encoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_input,   (t_enc_end - t_enc_start) / 1e6f, inp.size() / ((t_enc_end - t_enc_start) / 1e6f));
     LOG_TEE("decoded %4d tokens in %8.3f seconds, speed: %8.3f t/s\n", n_predict, (t_dec_end - t_dec_start) / 1e6f, n_predict  / ((t_dec_end - t_dec_start) / 1e6f));
 
     LOG_TEE("\n");
-    LOG_TEE("n_draft      = %d\n", n_draft);
-    LOG_TEE("n_predict    = %d\n", n_predict);
-    LOG_TEE("n_drafted    = %d\n", n_drafted);
-    LOG_TEE("t_draft_flat = %.2f ms\n", t_draft_flat_us*1e-3);
-    LOG_TEE("t_draft      = %.2f ms, %.2f us per token, %.2f tokens per second\n",
+    LOG_TEE("n_draft   = %d\n", n_draft);
+    LOG_TEE("n_predict = %d\n", n_predict);
+    LOG_TEE("n_drafted = %d\n", n_drafted);
+    LOG_TEE("t_draft   = %.2f ms, %.2f us per token, %.2f tokens per second\n",
             t_draft_us*1e-3, 1.0f*t_draft_us/n_drafted, n_drafted/(1e-6*t_draft_us));
-    LOG_TEE("n_accept     = %d\n", n_accept);
-    LOG_TEE("accept       = %.3f%%\n", 100.0f * n_accept / n_drafted);
+    LOG_TEE("n_accept  = %d\n", n_accept);
+    LOG_TEE("accept    = %.3f%%\n", 100.0f * n_accept / n_drafted);
 
     LOG_TEE("\ntarget:\n");
     llama_print_timings(ctx);

examples/main-cmake-pkg/README.md

@@ -8,7 +8,7 @@ Because this example is "outside of the source tree", it is important to first b
 
 ### Considerations
 
-When hardware acceleration libraries are used (e.g. 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_.
+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_.
 
 ### Build llama.cpp and install to C:\LlamaCPP directory